llama.cpp/ggml-metal.metal

2696 lines
96 KiB
Metal
Raw Normal View History

llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
#include <metal_stdlib>
using namespace metal;
#define MAX(x, y) ((x) > (y) ? (x) : (y))
#define QK4_0 32
#define QR4_0 2
typedef struct {
half d; // delta
uint8_t qs[QK4_0 / 2]; // nibbles / quants
} block_q4_0;
#define QK4_1 32
typedef struct {
half d; // delta
half m; // min
uint8_t qs[QK4_1 / 2]; // nibbles / quants
} block_q4_1;
#define QK5_0 32
typedef struct {
half d; // delta
uint8_t qh[4]; // 5-th bit of quants
uint8_t qs[QK5_0 / 2]; // nibbles / quants
} block_q5_0;
#define QK5_1 32
typedef struct {
half d; // delta
half m; // min
uint8_t qh[4]; // 5-th bit of quants
uint8_t qs[QK5_1 / 2]; // nibbles / quants
} block_q5_1;
#define QK8_0 32
typedef struct {
half d; // delta
int8_t qs[QK8_0]; // quants
} block_q8_0;
llama : custom attention mask + parallel decoding + no context swaps (#3228) * tests : verify that RoPE is "additive" * llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) * ggml : ggml_rope now takes a vector with positions instead of n_past * metal : add rope_f16 kernel + optimize cpy kernels * llama : unified KV cache + batch inference API * llama : add new llama_decode() API that works with llama_batch * llama : add cell_max heuristic for more efficient kv_cache * llama : extend llama_kv_cache API * llama : more robust cell_max heuristic + wip shift * metal : disable concurrency optimization * llama : add llama_kv_cache_shift_seq + no more context swaps * llama : apply K-cache roping for Falcon and Baichuan * speculative : fix KV cache management * parallel : example for serving multiple users in parallel * parallel : disable hot-plug to avoid cache fragmentation * fixes : speculative KV cache + llama worst-case graph * llama : extend batch API to select which logits to output * llama : fix worst case graph build * ggml-cuda : update rope implementation for parallel decoding (#3254) * ggml-cuda : update rope implementation for parallel decoding * better solution for p0 computation * fix rope * simpler rope implementation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * make : add parallel to build + fix static functions in llama.cpp * simple : fix token counting * parallel : various improvements * llama : fix cell_max logic + rename functions * parallel : try smaller batches when the KV cache is fragmented * parallel : fix sequence termination criteria * llama : silence errors KV cache errors * parallel : remove new line from prompt * parallel : process system prompt once + configurable paramters + llama API * parallel : remove question with short answers * parallel : count cache misses * parallel : print misses on each request * parallel : minor * llama : fix n_kv to never become 0 * parallel : rename hot-plug to continuous-batching * llama : improve llama_batch API + simplify parallel example * simple : add parallel decoding support * simple : improve comments + free batch * ggml-cuda : add rope f16, restore performance with parallel decoding (#3272) * ggml-cuda : add rope f16, restore performance * offload KQ_mask with all models * fix rope shift --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : disable MPI for now ggml-ci * train : make KQ_pos memory buffer permanent via dummy scale op * ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275) ggml-ci * parallel : fix bug (extra BOS) + smaller token_prev array * parallel : fix cases where the input prompts can overflow the batch * parallel : add disabled experimental batch chunking in powers of two * llama : llama.h formatting + comments * simple : add README.md * llama : fix kv cache heuristic when context is less than 32 * parallel : fix crash when `-n -1` * llama : simplify returns if/else branches * metal : use mm kernels for batch size > 2 * examples : utilize new llama_get_logits_ith() * examples : add example for batched decoding * examples : do not eval prompt 2 times (close #3348) * server : clear the KV cache beyond n_past before llama_decode * server : avoid context swaps by shifting the KV cache --------- Co-authored-by: slaren <slarengh@gmail.com>
2023-09-28 18:04:36 +02:00
// general-purpose kernel for addition of two tensors
// pros: works for non-contiguous tensors, supports broadcast across dims 1, 2 and 3
// cons: not very efficient
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
kernel void kernel_add(
llama : custom attention mask + parallel decoding + no context swaps (#3228) * tests : verify that RoPE is "additive" * llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) * ggml : ggml_rope now takes a vector with positions instead of n_past * metal : add rope_f16 kernel + optimize cpy kernels * llama : unified KV cache + batch inference API * llama : add new llama_decode() API that works with llama_batch * llama : add cell_max heuristic for more efficient kv_cache * llama : extend llama_kv_cache API * llama : more robust cell_max heuristic + wip shift * metal : disable concurrency optimization * llama : add llama_kv_cache_shift_seq + no more context swaps * llama : apply K-cache roping for Falcon and Baichuan * speculative : fix KV cache management * parallel : example for serving multiple users in parallel * parallel : disable hot-plug to avoid cache fragmentation * fixes : speculative KV cache + llama worst-case graph * llama : extend batch API to select which logits to output * llama : fix worst case graph build * ggml-cuda : update rope implementation for parallel decoding (#3254) * ggml-cuda : update rope implementation for parallel decoding * better solution for p0 computation * fix rope * simpler rope implementation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * make : add parallel to build + fix static functions in llama.cpp * simple : fix token counting * parallel : various improvements * llama : fix cell_max logic + rename functions * parallel : try smaller batches when the KV cache is fragmented * parallel : fix sequence termination criteria * llama : silence errors KV cache errors * parallel : remove new line from prompt * parallel : process system prompt once + configurable paramters + llama API * parallel : remove question with short answers * parallel : count cache misses * parallel : print misses on each request * parallel : minor * llama : fix n_kv to never become 0 * parallel : rename hot-plug to continuous-batching * llama : improve llama_batch API + simplify parallel example * simple : add parallel decoding support * simple : improve comments + free batch * ggml-cuda : add rope f16, restore performance with parallel decoding (#3272) * ggml-cuda : add rope f16, restore performance * offload KQ_mask with all models * fix rope shift --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : disable MPI for now ggml-ci * train : make KQ_pos memory buffer permanent via dummy scale op * ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275) ggml-ci * parallel : fix bug (extra BOS) + smaller token_prev array * parallel : fix cases where the input prompts can overflow the batch * parallel : add disabled experimental batch chunking in powers of two * llama : llama.h formatting + comments * simple : add README.md * llama : fix kv cache heuristic when context is less than 32 * parallel : fix crash when `-n -1` * llama : simplify returns if/else branches * metal : use mm kernels for batch size > 2 * examples : utilize new llama_get_logits_ith() * examples : add example for batched decoding * examples : do not eval prompt 2 times (close #3348) * server : clear the KV cache beyond n_past before llama_decode * server : avoid context swaps by shifting the KV cache --------- Co-authored-by: slaren <slarengh@gmail.com>
2023-09-28 18:04:36 +02:00
device const char * src0,
device const char * src1,
device char * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant int64_t & nb00,
constant int64_t & nb01,
constant int64_t & nb02,
constant int64_t & nb03,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant int64_t & ne13,
constant int64_t & nb10,
constant int64_t & nb11,
constant int64_t & nb12,
constant int64_t & nb13,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant int64_t & nb0,
constant int64_t & nb1,
constant int64_t & nb2,
constant int64_t & nb3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig.z;
const int64_t i02 = tgpig.y;
const int64_t i01 = tgpig.x;
const int64_t i13 = i03 % ne13;
const int64_t i12 = i02 % ne12;
const int64_t i11 = i01 % ne11;
device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01 + tpitg.x*nb00;
device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11 + tpitg.x*nb10;
device char * dst_ptr = dst + i03*nb3 + i02*nb2 + i01*nb1 + tpitg.x*nb0;
for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
((device float *)dst_ptr)[0] = ((device float *)src0_ptr)[0] + ((device float *)src1_ptr)[0];
src0_ptr += ntg.x*nb00;
src1_ptr += ntg.x*nb10;
dst_ptr += ntg.x*nb0;
}
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
}
// assumption: src1 is a row
// broadcast src1 into src0
kernel void kernel_add_row(
device const float4 * src0,
device const float4 * src1,
device float4 * dst,
llama : custom attention mask + parallel decoding + no context swaps (#3228) * tests : verify that RoPE is "additive" * llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) * ggml : ggml_rope now takes a vector with positions instead of n_past * metal : add rope_f16 kernel + optimize cpy kernels * llama : unified KV cache + batch inference API * llama : add new llama_decode() API that works with llama_batch * llama : add cell_max heuristic for more efficient kv_cache * llama : extend llama_kv_cache API * llama : more robust cell_max heuristic + wip shift * metal : disable concurrency optimization * llama : add llama_kv_cache_shift_seq + no more context swaps * llama : apply K-cache roping for Falcon and Baichuan * speculative : fix KV cache management * parallel : example for serving multiple users in parallel * parallel : disable hot-plug to avoid cache fragmentation * fixes : speculative KV cache + llama worst-case graph * llama : extend batch API to select which logits to output * llama : fix worst case graph build * ggml-cuda : update rope implementation for parallel decoding (#3254) * ggml-cuda : update rope implementation for parallel decoding * better solution for p0 computation * fix rope * simpler rope implementation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * make : add parallel to build + fix static functions in llama.cpp * simple : fix token counting * parallel : various improvements * llama : fix cell_max logic + rename functions * parallel : try smaller batches when the KV cache is fragmented * parallel : fix sequence termination criteria * llama : silence errors KV cache errors * parallel : remove new line from prompt * parallel : process system prompt once + configurable paramters + llama API * parallel : remove question with short answers * parallel : count cache misses * parallel : print misses on each request * parallel : minor * llama : fix n_kv to never become 0 * parallel : rename hot-plug to continuous-batching * llama : improve llama_batch API + simplify parallel example * simple : add parallel decoding support * simple : improve comments + free batch * ggml-cuda : add rope f16, restore performance with parallel decoding (#3272) * ggml-cuda : add rope f16, restore performance * offload KQ_mask with all models * fix rope shift --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : disable MPI for now ggml-ci * train : make KQ_pos memory buffer permanent via dummy scale op * ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275) ggml-ci * parallel : fix bug (extra BOS) + smaller token_prev array * parallel : fix cases where the input prompts can overflow the batch * parallel : add disabled experimental batch chunking in powers of two * llama : llama.h formatting + comments * simple : add README.md * llama : fix kv cache heuristic when context is less than 32 * parallel : fix crash when `-n -1` * llama : simplify returns if/else branches * metal : use mm kernels for batch size > 2 * examples : utilize new llama_get_logits_ith() * examples : add example for batched decoding * examples : do not eval prompt 2 times (close #3348) * server : clear the KV cache beyond n_past before llama_decode * server : avoid context swaps by shifting the KV cache --------- Co-authored-by: slaren <slarengh@gmail.com>
2023-09-28 18:04:36 +02:00
constant int64_t & nb [[buffer(27)]],
uint tpig[[thread_position_in_grid]]) {
dst[tpig] = src0[tpig] + src1[tpig % nb];
}
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
kernel void kernel_mul(
device const float4 * src0,
device const float4 * src1,
device float4 * dst,
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
uint tpig[[thread_position_in_grid]]) {
dst[tpig] = src0[tpig] * src1[tpig];
}
// assumption: src1 is a row
// broadcast src1 into src0
kernel void kernel_mul_row(
device const float4 * src0,
device const float4 * src1,
device float4 * dst,
constant int64_t & nb,
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
uint tpig[[thread_position_in_grid]]) {
dst[tpig] = src0[tpig] * src1[tpig % nb];
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
}
kernel void kernel_scale(
device const float * src0,
device float * dst,
constant float & scale,
uint tpig[[thread_position_in_grid]]) {
dst[tpig] = src0[tpig] * scale;
}
kernel void kernel_scale_4(
device const float4 * src0,
device float4 * dst,
constant float & scale,
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
uint tpig[[thread_position_in_grid]]) {
dst[tpig] = src0[tpig] * scale;
}
kernel void kernel_silu(
device const float4 * src0,
device float4 * dst,
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
uint tpig[[thread_position_in_grid]]) {
device const float4 & x = src0[tpig];
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
dst[tpig] = x / (1.0f + exp(-x));
}
kernel void kernel_relu(
device const float * src0,
device float * dst,
uint tpig[[thread_position_in_grid]]) {
dst[tpig] = max(0.0f, src0[tpig]);
}
kernel void kernel_sqr(
device const float * src0,
device float * dst,
uint tpig[[thread_position_in_grid]]) {
dst[tpig] = src0[tpig] * src0[tpig];
}
constant float GELU_COEF_A = 0.044715f;
constant float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
kernel void kernel_gelu(
device const float4 * src0,
device float4 * dst,
uint tpig[[thread_position_in_grid]]) {
device const float4 & x = src0[tpig];
llm : add Falcon support (#2717) * llama : refactor GGUF constants into static maps * llama : check if model architecture is known * llama : refactor llama_model_load_internal() * gguf : add KV constant maps * llm : read arch-specific KVs * convert : add dummy scores + types * falcon : load tensor data (CPU only) * llama : fix loading progress bar * llama : add arch member to llama_model * falcon : CPU inference working * falcon : support non-40B models * falcon : minor * llama : minor updates ggml-ci * convert-falcon-hf-to-gguf.py : fix special token mapping * llama.cpp : llama default UNK token = id 0 * llama.cpp : fix bpe tokenizer * llama.cpp : fix the fix of bpe tokenizer * ggml : pass eps to ggml_norm * metal : implement RoPE (mode = 2) + avoid ggml_repeat * ggml : ggml_repeat always creates new tensor * falcon : copy-paste self-attention from LLaMA * metal : print extra compute pipeline info * falcon : minor changes (still chasing the Metal problem) * llama.cpp : fix linefeed token * metal : fix GELU kernel numerical stability by using precise::tanh * metal : temporary workaround for the concurrency optimization bug * falcon : add CUDA offloading (#2739) * llama : better model naming and size reporting * llama : prep new tokenizer support * llama : advanced BPE tokenizer based on ggllm.cpp imlpementation * llama : remove oboslete comment ggml-ci * common : remove obsolete BPE API + disable test-tokenizer-1 * llama : revert BPE special-case in llama_byte_to_token() * cuda : add TODOs for RoPE NeoX implementation * llama : default special tokens based on vocab type * perplexity : add log for start of tokenization --------- Co-authored-by: klosax <131523366+klosax@users.noreply.github.com> Co-authored-by: slaren <slarengh@gmail.com>
2023-08-23 22:08:04 +02:00
// BEWARE !!!
// Simply using "tanh" instead of "precise::tanh" will sometimes results in NaNs!
// This was observed with Falcon 7B and 40B models
//
dst[tpig] = 0.5f*x*(1.0f + precise::tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
}
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
kernel void kernel_soft_max(
device const float * src0,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig[2];
const int64_t i02 = tgpig[1];
const int64_t i01 = tgpig[0];
device const float * psrc0 = src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
device float * pdst = dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
// parallel max
float lmax = tpitg[0] < ne00 ? psrc0[tpitg[0]] : -INFINITY;
for (int i00 = tpitg[0] + ntg[0]; i00 < ne00; i00 += ntg[0]) {
lmax = MAX(lmax, psrc0[i00]);
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
}
const float max = simd_max(lmax);
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
// parallel sum
float lsum = 0.0f;
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) {
const float exp_psrc0 = exp(psrc0[i00] - max);
lsum += exp_psrc0;
// Remember the result of exp here. exp is expensive, so we really do not
// whish to compute it twice.
pdst[i00] = exp_psrc0;
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
}
const float sum = simd_sum(lsum);
for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) {
pdst[i00] /= sum;
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
}
}
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
kernel void kernel_soft_max_4(
device const float * src0,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig[2];
const int64_t i02 = tgpig[1];
const int64_t i01 = tgpig[0];
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
device const float4 * psrc4 = (device const float4 *)(src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
device float4 * pdst4 = (device float4 *)(dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
// parallel max
float4 lmax4 = tpitg[0] < ne00/4 ? psrc4[tpitg[0]] : -INFINITY;
for (int i00 = tpitg[0] + ntg[0]; i00 < ne00/4; i00 += ntg[0]) {
lmax4 = fmax(lmax4, psrc4[i00]);
}
float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3]));
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
const float max = simd_max(lmax);
// parallel sum
float4 lsum4 = 0.0f;
for (int i00 = tpitg[0]; i00 < ne00/4; i00 += ntg[0]) {
const float4 exp_psrc4 = exp(psrc4[i00] - max);
lsum4 += exp_psrc4;
pdst4[i00] = exp_psrc4;
}
float lsum = lsum4[0] + lsum4[1] + lsum4[2] + lsum4[3];
const float sum = simd_sum(lsum);
for (int i00 = tpitg[0]; i00 < ne00/4; i00 += ntg[0]) {
pdst4[i00] /= sum;
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
}
}
kernel void kernel_diag_mask_inf(
device const float * src0,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int & n_past,
uint3 tpig[[thread_position_in_grid]]) {
const int64_t i02 = tpig[2];
const int64_t i01 = tpig[1];
const int64_t i00 = tpig[0];
if (i00 > n_past + i01) {
dst[i02*ne01*ne00 + i01*ne00 + i00] = -INFINITY;
} else {
dst[i02*ne01*ne00 + i01*ne00 + i00] = src0[i02*ne01*ne00 + i01*ne00 + i00];
}
}
kernel void kernel_diag_mask_inf_8(
device const float4 * src0,
device float4 * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int & n_past,
uint3 tpig[[thread_position_in_grid]]) {
const int64_t i = 2*tpig[0];
dst[i+0] = src0[i+0];
dst[i+1] = src0[i+1];
int64_t i4 = 4*i;
const int64_t i02 = i4/(ne00*ne01); i4 -= i02*ne00*ne01;
const int64_t i01 = i4/(ne00); i4 -= i01*ne00;
const int64_t i00 = i4;
for (int k = 3; k >= 0; --k) {
if (i00 + 4 + k <= n_past + i01) {
break;
}
dst[i+1][k] = -INFINITY;
if (i00 + k > n_past + i01) {
dst[i][k] = -INFINITY;
}
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
}
}
kernel void kernel_norm(
device const void * src0,
device float * dst,
constant int64_t & ne00,
constant uint64_t & nb01,
constant float & eps,
threadgroup float * sum [[threadgroup(0)]],
uint tgpig[[threadgroup_position_in_grid]],
uint tpitg[[thread_position_in_threadgroup]],
uint ntg[[threads_per_threadgroup]]) {
device const float * x = (device const float *) ((device const char *) src0 + tgpig*nb01);
// MEAN
// parallel sum
sum[tpitg] = 0.0f;
for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
sum[tpitg] += x[i00];
}
// reduce
threadgroup_barrier(mem_flags::mem_threadgroup);
for (uint i = ntg/2; i > 0; i /= 2) {
if (tpitg < i) {
sum[tpitg] += sum[tpitg + i];
}
threadgroup_barrier(mem_flags::mem_threadgroup);
}
const float mean = sum[0] / ne00;
// recenter and VARIANCE
threadgroup_barrier(mem_flags::mem_threadgroup);
device float * y = dst + tgpig*ne00;
sum[tpitg] = 0.0f;
for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
y[i00] = x[i00] - mean;
sum[tpitg] += y[i00] * y[i00];
}
// reduce
threadgroup_barrier(mem_flags::mem_threadgroup);
for (uint i = ntg/2; i > 0; i /= 2) {
if (tpitg < i) {
sum[tpitg] += sum[tpitg + i];
}
threadgroup_barrier(mem_flags::mem_threadgroup);
}
const float variance = sum[0] / ne00;
const float scale = 1.0f/sqrt(variance + eps);
for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
y[i00] = y[i00] * scale;
}
}
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
kernel void kernel_rms_norm(
device const void * src0,
device float * dst,
constant int64_t & ne00,
constant uint64_t & nb01,
constant float & eps,
threadgroup float * sum [[threadgroup(0)]],
uint tgpig[[threadgroup_position_in_grid]],
uint tpitg[[thread_position_in_threadgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]],
uint tiisg[[thread_index_in_simdgroup]],
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
uint ntg[[threads_per_threadgroup]]) {
device const float4 * x = (device const float4 *) ((device const char *) src0 + tgpig*nb01);
device const float * x_scalar = (device const float *) x;
float4 sumf = 0;
float all_sum = 0;
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
// parallel sum
for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
sumf += x[i00] * x[i00];
}
all_sum = sumf[0] + sumf[1] + sumf[2] + sumf[3];
all_sum = simd_sum(all_sum);
if (tiisg == 0) {
sum[sgitg] = all_sum;
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
}
threadgroup_barrier(mem_flags::mem_threadgroup);
// broadcast, simd group number is ntg / 32
for (uint i = ntg / 32 / 2; i > 0; i /= 2) {
if (tpitg < i) {
sum[tpitg] += sum[tpitg + i];
}
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
}
if (tpitg == 0) {
for (int i = 4 * (ne00 / 4); i < ne00; i++) {
sum[0] += x_scalar[i];
}
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
sum[0] /= ne00;
}
threadgroup_barrier(mem_flags::mem_threadgroup);
const float mean = sum[0];
const float scale = 1.0f/sqrt(mean + eps);
device float4 * y = (device float4 *) (dst + tgpig*ne00);
device float * y_scalar = (device float *) y;
for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
y[i00] = x[i00] * scale;
}
if (tpitg == 0) {
for (int i00 = 4 * (ne00 / 4); i00 < ne00; i00++) {
y_scalar[i00] = x_scalar[i00] * scale;
}
}
}
// function for calculate inner product between half a q4_0 block and 16 floats (yl), sumy is SUM(yl[i])
// il indicates where the q4 quants begin (0 or QK4_0/4)
// we assume that the yl's have been multiplied with the appropriate scale factor
// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
inline float block_q_n_dot_y(device const block_q4_0 * qb_curr, float sumy, thread float * yl, int il) {
float d = qb_curr->d;
float2 acc = 0.f;
device const uint16_t * qs = ((device const uint16_t *)qb_curr + 1 + il/2);
for (int i = 0; i < 8; i+=2) {
acc[0] += yl[i + 0] * (qs[i / 2] & 0x000F)
+ yl[i + 1] * (qs[i / 2] & 0x0F00);
acc[1] += yl[i + 8] * (qs[i / 2] & 0x00F0)
+ yl[i + 9] * (qs[i / 2] & 0xF000);
}
return d * (sumy * -8.f + acc[0] + acc[1]);
}
// function for calculate inner product between half a q4_1 block and 16 floats (yl), sumy is SUM(yl[i])
// il indicates where the q4 quants begin (0 or QK4_0/4)
// we assume that the yl's have been multiplied with the appropriate scale factor
// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
inline float block_q_n_dot_y(device const block_q4_1 * qb_curr, float sumy, thread float * yl, int il) {
float d = qb_curr->d;
float m = qb_curr->m;
float2 acc = 0.f;
device const uint16_t * qs = ((device const uint16_t *)qb_curr + 2 + il/2);
for (int i = 0; i < 8; i+=2) {
acc[0] += yl[i + 0] * (qs[i / 2] & 0x000F)
+ yl[i + 1] * (qs[i / 2] & 0x0F00);
acc[1] += yl[i + 8] * (qs[i / 2] & 0x00F0)
+ yl[i + 9] * (qs[i / 2] & 0xF000);
}
return d * (acc[0] + acc[1]) + sumy * m;
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
}
// function for calculate inner product between half a q5_0 block and 16 floats (yl), sumy is SUM(yl[i])
// il indicates where the q5 quants begin (0 or QK5_0/4)
// we assume that the yl's have been multiplied with the appropriate scale factor
// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
inline float block_q_n_dot_y(device const block_q5_0 * qb_curr, float sumy, thread float * yl, int il) {
float d = qb_curr->d;
float2 acc = 0.f;
device const uint16_t * qs = ((device const uint16_t *)qb_curr + 3 + il/2);
const uint32_t qh = *((device const uint32_t *)qb_curr->qh);
for (int i = 0; i < 8; i+=2) {
acc[0] += yl[i + 0] * ((qs[i / 2] & 0x000F) | ((qh >> (i+0+il ) << 4 ) & 0x00010))
+ yl[i + 1] * ((qs[i / 2] & 0x0F00) | ((qh >> (i+1+il ) << 12) & 0x01000));
acc[1] += yl[i + 8] * ((qs[i / 2] & 0x00F0) | ((qh >> (i+0+il+QK5_0/2) << 8 ) & 0x00100))
+ yl[i + 9] * ((qs[i / 2] & 0xF000) | ((qh >> (i+1+il+QK5_0/2) << 16) & 0x10000));
}
return d * (sumy * -16.f + acc[0] + acc[1]);
}
// function for calculate inner product between half a q5_1 block and 16 floats (yl), sumy is SUM(yl[i])
// il indicates where the q5 quants begin (0 or QK5_1/4)
// we assume that the yl's have been multiplied with the appropriate scale factor
// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
inline float block_q_n_dot_y(device const block_q5_1 * qb_curr, float sumy, thread float * yl, int il) {
float d = qb_curr->d;
float m = qb_curr->m;
float2 acc = 0.f;
device const uint16_t * qs = ((device const uint16_t *)qb_curr + 4 + il/2);
const uint32_t qh = *((device const uint32_t *)qb_curr->qh);
for (int i = 0; i < 8; i+=2) {
acc[0] += yl[i + 0] * ((qs[i / 2] & 0x000F) | ((qh >> (i+0+il ) << 4 ) & 0x00010))
+ yl[i + 1] * ((qs[i / 2] & 0x0F00) | ((qh >> (i+1+il ) << 12) & 0x01000));
acc[1] += yl[i + 8] * ((qs[i / 2] & 0x00F0) | ((qh >> (i+0+il+QK5_0/2) << 8 ) & 0x00100))
+ yl[i + 9] * ((qs[i / 2] & 0xF000) | ((qh >> (i+1+il+QK5_0/2) << 16) & 0x10000));
}
return d * (acc[0] + acc[1]) + sumy * m;
}
// putting them in the kernel cause a significant performance penalty
#define N_DST 4 // each SIMD group works on 4 rows
#define N_SIMDGROUP 2 // number of SIMD groups in a thread group
#define N_SIMDWIDTH 32 // assuming SIMD group size is 32
//Note: This is a template, but strictly speaking it only applies to
// quantizations where the block size is 32. It also does not
// giard against the number of rows not being divisible by
// N_DST, so this is another explicit assumption of the implementation.
template<typename block_q_type, int nr, int nsg, int nw>
void mul_vec_q_n_f32(device const void * src0, device const float * src1, device float * dst,
int64_t ne00, int64_t ne01, int64_t ne02, int64_t ne10, int64_t ne12, int64_t ne0, int64_t ne1, uint gqa,
uint3 tgpig, uint tiisg, uint sgitg) {
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
const int nb = ne00/QK4_0;
const int r0 = tgpig.x;
const int r1 = tgpig.y;
const int im = tgpig.z;
const int first_row = (r0 * nsg + sgitg) * nr;
const uint offset0 = first_row * nb + im/gqa*(nb*ne0);
device const block_q_type * x = (device const block_q_type *) src0 + offset0;
device const float * y = (device const float *) src1 + r1*ne10 + im*ne00*ne1;
float yl[16]; // src1 vector cache
float sumf[nr] = {0.f};
const int ix = (tiisg/2);
const int il = (tiisg%2)*8;
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
device const float * yb = y + ix * QK4_0 + il;
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
// each thread in a SIMD group deals with half a block.
for (int ib = ix; ib < nb; ib += nw/2) {
float sumy = 0;
for (int i = 0; i < 8; i += 2) {
sumy += yb[i] + yb[i+1];
yl[i+0] = yb[i+ 0];
yl[i+1] = yb[i+ 1]/256.f;
sumy += yb[i+16] + yb[i+17];
yl[i+8] = yb[i+16]/16.f;
yl[i+9] = yb[i+17]/4096.f;
}
for (int row = 0; row < nr; row++) {
sumf[row] += block_q_n_dot_y(x+ib+row*nb, sumy, yl, il);
}
yb += QK4_0 * 16;
}
for (int row = 0; row < nr; ++row) {
const float tot = simd_sum(sumf[row]);
if (tiisg == 0 && first_row + row < ne01) {
dst[im*ne0*ne1 + r1*ne0 + first_row + row] = tot;
}
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
}
}
kernel void kernel_mul_mv_q4_0_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01[[buffer(4)]],
constant int64_t & ne02[[buffer(5)]],
constant int64_t & ne10[[buffer(9)]],
constant int64_t & ne12[[buffer(11)]],
constant int64_t & ne0[[buffer(15)]],
constant int64_t & ne1[[buffer(16)]],
constant uint & gqa[[buffer(17)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
mul_vec_q_n_f32<block_q4_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,gqa,tgpig,tiisg,sgitg);
}
kernel void kernel_mul_mv_q4_1_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01[[buffer(4)]],
constant int64_t & ne02[[buffer(5)]],
constant int64_t & ne10[[buffer(9)]],
constant int64_t & ne12[[buffer(11)]],
constant int64_t & ne0[[buffer(15)]],
constant int64_t & ne1[[buffer(16)]],
constant uint & gqa[[buffer(17)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
mul_vec_q_n_f32<block_q4_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,gqa,tgpig,tiisg,sgitg);
}
kernel void kernel_mul_mv_q5_0_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01[[buffer(4)]],
constant int64_t & ne02[[buffer(5)]],
constant int64_t & ne10[[buffer(9)]],
constant int64_t & ne12[[buffer(11)]],
constant int64_t & ne0[[buffer(15)]],
constant int64_t & ne1[[buffer(16)]],
constant uint & gqa[[buffer(17)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
mul_vec_q_n_f32<block_q5_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,gqa,tgpig,tiisg,sgitg);
}
kernel void kernel_mul_mv_q5_1_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01[[buffer(4)]],
constant int64_t & ne02[[buffer(5)]],
constant int64_t & ne10[[buffer(9)]],
constant int64_t & ne12[[buffer(11)]],
constant int64_t & ne0[[buffer(15)]],
constant int64_t & ne1[[buffer(16)]],
constant uint & gqa[[buffer(17)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
mul_vec_q_n_f32<block_q5_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,gqa,tgpig,tiisg,sgitg);
}
#define NB_Q8_0 8
kernel void kernel_mul_mv_q8_0_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01[[buffer(4)]],
constant int64_t & ne02[[buffer(5)]],
constant int64_t & ne10[[buffer(9)]],
constant int64_t & ne12[[buffer(11)]],
constant int64_t & ne0[[buffer(15)]],
constant int64_t & ne1[[buffer(16)]],
constant uint & gqa[[buffer(17)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
const int nr = N_DST;
const int nsg = N_SIMDGROUP;
const int nw = N_SIMDWIDTH;
const int nb = ne00/QK8_0;
const int r0 = tgpig.x;
const int r1 = tgpig.y;
const int im = tgpig.z;
const int first_row = (r0 * nsg + sgitg) * nr;
const uint offset0 = first_row * nb + im/gqa*(nb*ne0);
device const block_q8_0 * x = (device const block_q8_0 *) src0 + offset0;
device const float * y = (device const float *) src1 + r1*ne10 + im*ne00*ne1;
float yl[NB_Q8_0];
float sumf[nr]={0.f};
const int ix = tiisg/4;
const int il = tiisg%4;
device const float * yb = y + ix * QK8_0 + NB_Q8_0*il;
// each thread in a SIMD group deals with NB_Q8_0 quants at a time
for (int ib = ix; ib < nb; ib += nw/4) {
for (int i = 0; i < NB_Q8_0; ++i) {
yl[i] = yb[i];
}
for (int row = 0; row < nr; row++) {
device const int8_t * qs = x[ib+row*nb].qs + NB_Q8_0*il;
float sumq = 0.f;
for (int iq = 0; iq < NB_Q8_0; ++iq) {
sumq += qs[iq] * yl[iq];
}
sumf[row] += sumq*x[ib+row*nb].d;
}
yb += NB_Q8_0 * nw;
}
for (int row = 0; row < nr; ++row) {
const float tot = simd_sum(sumf[row]);
if (tiisg == 0 && first_row + row < ne01) {
dst[r1*ne0 + im*ne0*ne1 + first_row + row] = tot;
}
}
}
#define N_F32_F32 4
kernel void kernel_mul_mv_f32_f32(
device const char * src0,
device const char * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]]) {
const int64_t r0 = tgpig.x;
const int64_t rb = tgpig.y*N_F32_F32;
const int64_t im = tgpig.z;
device const float * x = (device const float *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02);
if (ne00 < 128) {
for (int row = 0; row < N_F32_F32; ++row) {
int r1 = rb + row;
if (r1 >= ne11) {
break;
}
device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
float sumf = 0;
for (int i = tiisg; i < ne00; i += 32) {
sumf += (float) x[i] * (float) y[i];
}
float all_sum = simd_sum(sumf);
if (tiisg == 0) {
dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
}
}
} else {
device const float4 * x4 = (device const float4 *)x;
for (int row = 0; row < N_F32_F32; ++row) {
int r1 = rb + row;
if (r1 >= ne11) {
break;
}
device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
device const float4 * y4 = (device const float4 *) y;
float sumf = 0;
for (int i = tiisg; i < ne00/4; i += 32) {
for (int k = 0; k < 4; ++k) sumf += (float) x4[i][k] * y4[i][k];
}
float all_sum = simd_sum(sumf);
if (tiisg == 0) {
for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (float) x[i] * y[i];
dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
}
}
}
}
kernel void kernel_mul_mv_f16_f32_1row(
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
device const char * src0,
device const char * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]]) {
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
const int64_t r0 = tgpig.x;
const int64_t r1 = tgpig.y;
const int64_t im = tgpig.z;
device const half * x = (device const half *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02);
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
float sumf = 0;
if (ne00 < 128) {
for (int i = tiisg; i < ne00; i += 32) {
sumf += (float) x[i] * (float) y[i];
}
float all_sum = simd_sum(sumf);
if (tiisg == 0) {
dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
}
} else {
device const half4 * x4 = (device const half4 *) x;
device const float4 * y4 = (device const float4 *) y;
for (int i = tiisg; i < ne00/4; i += 32) {
for (int k = 0; k < 4; ++k) sumf += (float)x4[i][k] * y4[i][k];
}
float all_sum = simd_sum(sumf);
if (tiisg == 0) {
for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (float) x[i] * y[i];
dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
}
}
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
}
#define N_F16_F32 4
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
kernel void kernel_mul_mv_f16_f32(
device const char * src0,
device const char * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]]) {
const int64_t r0 = tgpig.x;
const int64_t rb = tgpig.y*N_F16_F32;
const int64_t im = tgpig.z;
device const half * x = (device const half *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02);
if (ne00 < 128) {
for (int row = 0; row < N_F16_F32; ++row) {
int r1 = rb + row;
if (r1 >= ne11) {
break;
}
device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
float sumf = 0;
for (int i = tiisg; i < ne00; i += 32) {
sumf += (float) x[i] * (float) y[i];
}
float all_sum = simd_sum(sumf);
if (tiisg == 0) {
dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
}
}
} else {
device const half4 * x4 = (device const half4 *)x;
for (int row = 0; row < N_F16_F32; ++row) {
int r1 = rb + row;
if (r1 >= ne11) {
break;
}
device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
device const float4 * y4 = (device const float4 *) y;
float sumf = 0;
for (int i = tiisg; i < ne00/4; i += 32) {
for (int k = 0; k < 4; ++k) sumf += (float) x4[i][k] * y4[i][k];
}
float all_sum = simd_sum(sumf);
if (tiisg == 0) {
for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (float) x[i] * y[i];
dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
}
}
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
}
}
// Assumes row size (ne00) is a multiple of 4
kernel void kernel_mul_mv_f16_f32_l4(
device const char * src0,
device const char * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]]) {
const int nrows = ne11;
const int64_t r0 = tgpig.x;
const int64_t im = tgpig.z;
device const half4 * x4 = (device const half4 *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02);
for (int r1 = 0; r1 < nrows; ++r1) {
device const float4 * y4 = (device const float4 *) (src1 + r1*nb11 + im*nb12);
float sumf = 0;
for (int i = tiisg; i < ne00/4; i += 32) {
for (int k = 0; k < 4; ++k) sumf += (float) x4[i][k] * y4[i][k];
}
float all_sum = simd_sum(sumf);
if (tiisg == 0) {
dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
}
}
}
kernel void kernel_alibi_f32(
device const float * src0,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
constant float & m0,
constant float & m1,
constant int & n_heads_log2_floor,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig[2];
const int64_t i02 = tgpig[1];
const int64_t i01 = tgpig[0];
const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
const int64_t i3 = n / (ne2*ne1*ne0);
const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
device float * dst_data = (device float *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
float m_k;
if (i2 < n_heads_log2_floor) {
m_k = pow(m0, i2 + 1);
} else {
m_k = pow(m1, 2 * (i2 - n_heads_log2_floor) + 1);
}
for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
dst_data[i00] = src[0] + m_k * (i00 - ne00 + 1);
}
}
llama : custom attention mask + parallel decoding + no context swaps (#3228) * tests : verify that RoPE is "additive" * llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) * ggml : ggml_rope now takes a vector with positions instead of n_past * metal : add rope_f16 kernel + optimize cpy kernels * llama : unified KV cache + batch inference API * llama : add new llama_decode() API that works with llama_batch * llama : add cell_max heuristic for more efficient kv_cache * llama : extend llama_kv_cache API * llama : more robust cell_max heuristic + wip shift * metal : disable concurrency optimization * llama : add llama_kv_cache_shift_seq + no more context swaps * llama : apply K-cache roping for Falcon and Baichuan * speculative : fix KV cache management * parallel : example for serving multiple users in parallel * parallel : disable hot-plug to avoid cache fragmentation * fixes : speculative KV cache + llama worst-case graph * llama : extend batch API to select which logits to output * llama : fix worst case graph build * ggml-cuda : update rope implementation for parallel decoding (#3254) * ggml-cuda : update rope implementation for parallel decoding * better solution for p0 computation * fix rope * simpler rope implementation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * make : add parallel to build + fix static functions in llama.cpp * simple : fix token counting * parallel : various improvements * llama : fix cell_max logic + rename functions * parallel : try smaller batches when the KV cache is fragmented * parallel : fix sequence termination criteria * llama : silence errors KV cache errors * parallel : remove new line from prompt * parallel : process system prompt once + configurable paramters + llama API * parallel : remove question with short answers * parallel : count cache misses * parallel : print misses on each request * parallel : minor * llama : fix n_kv to never become 0 * parallel : rename hot-plug to continuous-batching * llama : improve llama_batch API + simplify parallel example * simple : add parallel decoding support * simple : improve comments + free batch * ggml-cuda : add rope f16, restore performance with parallel decoding (#3272) * ggml-cuda : add rope f16, restore performance * offload KQ_mask with all models * fix rope shift --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : disable MPI for now ggml-ci * train : make KQ_pos memory buffer permanent via dummy scale op * ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275) ggml-ci * parallel : fix bug (extra BOS) + smaller token_prev array * parallel : fix cases where the input prompts can overflow the batch * parallel : add disabled experimental batch chunking in powers of two * llama : llama.h formatting + comments * simple : add README.md * llama : fix kv cache heuristic when context is less than 32 * parallel : fix crash when `-n -1` * llama : simplify returns if/else branches * metal : use mm kernels for batch size > 2 * examples : utilize new llama_get_logits_ith() * examples : add example for batched decoding * examples : do not eval prompt 2 times (close #3348) * server : clear the KV cache beyond n_past before llama_decode * server : avoid context swaps by shifting the KV cache --------- Co-authored-by: slaren <slarengh@gmail.com>
2023-09-28 18:04:36 +02:00
typedef void (rope_t)(
device const void * src0,
device const int32_t * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
constant int & n_past,
constant int & n_dims,
constant int & mode,
constant float & freq_base,
constant float & freq_scale,
uint tiitg[[thread_index_in_threadgroup]],
uint3 tptg[[threads_per_threadgroup]],
uint3 tgpig[[threadgroup_position_in_grid]]);
template<typename T>
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
kernel void kernel_rope(
llama : custom attention mask + parallel decoding + no context swaps (#3228) * tests : verify that RoPE is "additive" * llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) * ggml : ggml_rope now takes a vector with positions instead of n_past * metal : add rope_f16 kernel + optimize cpy kernels * llama : unified KV cache + batch inference API * llama : add new llama_decode() API that works with llama_batch * llama : add cell_max heuristic for more efficient kv_cache * llama : extend llama_kv_cache API * llama : more robust cell_max heuristic + wip shift * metal : disable concurrency optimization * llama : add llama_kv_cache_shift_seq + no more context swaps * llama : apply K-cache roping for Falcon and Baichuan * speculative : fix KV cache management * parallel : example for serving multiple users in parallel * parallel : disable hot-plug to avoid cache fragmentation * fixes : speculative KV cache + llama worst-case graph * llama : extend batch API to select which logits to output * llama : fix worst case graph build * ggml-cuda : update rope implementation for parallel decoding (#3254) * ggml-cuda : update rope implementation for parallel decoding * better solution for p0 computation * fix rope * simpler rope implementation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * make : add parallel to build + fix static functions in llama.cpp * simple : fix token counting * parallel : various improvements * llama : fix cell_max logic + rename functions * parallel : try smaller batches when the KV cache is fragmented * parallel : fix sequence termination criteria * llama : silence errors KV cache errors * parallel : remove new line from prompt * parallel : process system prompt once + configurable paramters + llama API * parallel : remove question with short answers * parallel : count cache misses * parallel : print misses on each request * parallel : minor * llama : fix n_kv to never become 0 * parallel : rename hot-plug to continuous-batching * llama : improve llama_batch API + simplify parallel example * simple : add parallel decoding support * simple : improve comments + free batch * ggml-cuda : add rope f16, restore performance with parallel decoding (#3272) * ggml-cuda : add rope f16, restore performance * offload KQ_mask with all models * fix rope shift --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : disable MPI for now ggml-ci * train : make KQ_pos memory buffer permanent via dummy scale op * ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275) ggml-ci * parallel : fix bug (extra BOS) + smaller token_prev array * parallel : fix cases where the input prompts can overflow the batch * parallel : add disabled experimental batch chunking in powers of two * llama : llama.h formatting + comments * simple : add README.md * llama : fix kv cache heuristic when context is less than 32 * parallel : fix crash when `-n -1` * llama : simplify returns if/else branches * metal : use mm kernels for batch size > 2 * examples : utilize new llama_get_logits_ith() * examples : add example for batched decoding * examples : do not eval prompt 2 times (close #3348) * server : clear the KV cache beyond n_past before llama_decode * server : avoid context swaps by shifting the KV cache --------- Co-authored-by: slaren <slarengh@gmail.com>
2023-09-28 18:04:36 +02:00
device const void * src0,
device const int32_t * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
constant int & n_past,
constant int & n_dims,
constant int & mode,
constant float & freq_base,
constant float & freq_scale,
uint tiitg[[thread_index_in_threadgroup]],
uint3 tptg[[threads_per_threadgroup]],
uint3 tgpig[[threadgroup_position_in_grid]]) {
const int64_t i3 = tgpig[2];
const int64_t i2 = tgpig[1];
const int64_t i1 = tgpig[0];
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
const bool is_neox = mode & 2;
llama : custom attention mask + parallel decoding + no context swaps (#3228) * tests : verify that RoPE is "additive" * llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) * ggml : ggml_rope now takes a vector with positions instead of n_past * metal : add rope_f16 kernel + optimize cpy kernels * llama : unified KV cache + batch inference API * llama : add new llama_decode() API that works with llama_batch * llama : add cell_max heuristic for more efficient kv_cache * llama : extend llama_kv_cache API * llama : more robust cell_max heuristic + wip shift * metal : disable concurrency optimization * llama : add llama_kv_cache_shift_seq + no more context swaps * llama : apply K-cache roping for Falcon and Baichuan * speculative : fix KV cache management * parallel : example for serving multiple users in parallel * parallel : disable hot-plug to avoid cache fragmentation * fixes : speculative KV cache + llama worst-case graph * llama : extend batch API to select which logits to output * llama : fix worst case graph build * ggml-cuda : update rope implementation for parallel decoding (#3254) * ggml-cuda : update rope implementation for parallel decoding * better solution for p0 computation * fix rope * simpler rope implementation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * make : add parallel to build + fix static functions in llama.cpp * simple : fix token counting * parallel : various improvements * llama : fix cell_max logic + rename functions * parallel : try smaller batches when the KV cache is fragmented * parallel : fix sequence termination criteria * llama : silence errors KV cache errors * parallel : remove new line from prompt * parallel : process system prompt once + configurable paramters + llama API * parallel : remove question with short answers * parallel : count cache misses * parallel : print misses on each request * parallel : minor * llama : fix n_kv to never become 0 * parallel : rename hot-plug to continuous-batching * llama : improve llama_batch API + simplify parallel example * simple : add parallel decoding support * simple : improve comments + free batch * ggml-cuda : add rope f16, restore performance with parallel decoding (#3272) * ggml-cuda : add rope f16, restore performance * offload KQ_mask with all models * fix rope shift --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : disable MPI for now ggml-ci * train : make KQ_pos memory buffer permanent via dummy scale op * ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275) ggml-ci * parallel : fix bug (extra BOS) + smaller token_prev array * parallel : fix cases where the input prompts can overflow the batch * parallel : add disabled experimental batch chunking in powers of two * llama : llama.h formatting + comments * simple : add README.md * llama : fix kv cache heuristic when context is less than 32 * parallel : fix crash when `-n -1` * llama : simplify returns if/else branches * metal : use mm kernels for batch size > 2 * examples : utilize new llama_get_logits_ith() * examples : add example for batched decoding * examples : do not eval prompt 2 times (close #3348) * server : clear the KV cache beyond n_past before llama_decode * server : avoid context swaps by shifting the KV cache --------- Co-authored-by: slaren <slarengh@gmail.com>
2023-09-28 18:04:36 +02:00
device const int32_t * pos = src1;
const int64_t p = pos[i2];
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
const float theta_0 = freq_scale * (float)p;
const float inv_ndims = -1.f/n_dims;
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
if (!is_neox) {
for (int64_t i0 = 2*tiitg; i0 < ne0; i0 += 2*tptg.x) {
const float theta = theta_0 * pow(freq_base, inv_ndims*i0);
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
const float cos_theta = cos(theta);
const float sin_theta = sin(theta);
llama : custom attention mask + parallel decoding + no context swaps (#3228) * tests : verify that RoPE is "additive" * llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) * ggml : ggml_rope now takes a vector with positions instead of n_past * metal : add rope_f16 kernel + optimize cpy kernels * llama : unified KV cache + batch inference API * llama : add new llama_decode() API that works with llama_batch * llama : add cell_max heuristic for more efficient kv_cache * llama : extend llama_kv_cache API * llama : more robust cell_max heuristic + wip shift * metal : disable concurrency optimization * llama : add llama_kv_cache_shift_seq + no more context swaps * llama : apply K-cache roping for Falcon and Baichuan * speculative : fix KV cache management * parallel : example for serving multiple users in parallel * parallel : disable hot-plug to avoid cache fragmentation * fixes : speculative KV cache + llama worst-case graph * llama : extend batch API to select which logits to output * llama : fix worst case graph build * ggml-cuda : update rope implementation for parallel decoding (#3254) * ggml-cuda : update rope implementation for parallel decoding * better solution for p0 computation * fix rope * simpler rope implementation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * make : add parallel to build + fix static functions in llama.cpp * simple : fix token counting * parallel : various improvements * llama : fix cell_max logic + rename functions * parallel : try smaller batches when the KV cache is fragmented * parallel : fix sequence termination criteria * llama : silence errors KV cache errors * parallel : remove new line from prompt * parallel : process system prompt once + configurable paramters + llama API * parallel : remove question with short answers * parallel : count cache misses * parallel : print misses on each request * parallel : minor * llama : fix n_kv to never become 0 * parallel : rename hot-plug to continuous-batching * llama : improve llama_batch API + simplify parallel example * simple : add parallel decoding support * simple : improve comments + free batch * ggml-cuda : add rope f16, restore performance with parallel decoding (#3272) * ggml-cuda : add rope f16, restore performance * offload KQ_mask with all models * fix rope shift --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : disable MPI for now ggml-ci * train : make KQ_pos memory buffer permanent via dummy scale op * ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275) ggml-ci * parallel : fix bug (extra BOS) + smaller token_prev array * parallel : fix cases where the input prompts can overflow the batch * parallel : add disabled experimental batch chunking in powers of two * llama : llama.h formatting + comments * simple : add README.md * llama : fix kv cache heuristic when context is less than 32 * parallel : fix crash when `-n -1` * llama : simplify returns if/else branches * metal : use mm kernels for batch size > 2 * examples : utilize new llama_get_logits_ith() * examples : add example for batched decoding * examples : do not eval prompt 2 times (close #3348) * server : clear the KV cache beyond n_past before llama_decode * server : avoid context swaps by shifting the KV cache --------- Co-authored-by: slaren <slarengh@gmail.com>
2023-09-28 18:04:36 +02:00
device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
device T * dst_data = (device T *)((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
llama : custom attention mask + parallel decoding + no context swaps (#3228) * tests : verify that RoPE is "additive" * llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) * ggml : ggml_rope now takes a vector with positions instead of n_past * metal : add rope_f16 kernel + optimize cpy kernels * llama : unified KV cache + batch inference API * llama : add new llama_decode() API that works with llama_batch * llama : add cell_max heuristic for more efficient kv_cache * llama : extend llama_kv_cache API * llama : more robust cell_max heuristic + wip shift * metal : disable concurrency optimization * llama : add llama_kv_cache_shift_seq + no more context swaps * llama : apply K-cache roping for Falcon and Baichuan * speculative : fix KV cache management * parallel : example for serving multiple users in parallel * parallel : disable hot-plug to avoid cache fragmentation * fixes : speculative KV cache + llama worst-case graph * llama : extend batch API to select which logits to output * llama : fix worst case graph build * ggml-cuda : update rope implementation for parallel decoding (#3254) * ggml-cuda : update rope implementation for parallel decoding * better solution for p0 computation * fix rope * simpler rope implementation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * make : add parallel to build + fix static functions in llama.cpp * simple : fix token counting * parallel : various improvements * llama : fix cell_max logic + rename functions * parallel : try smaller batches when the KV cache is fragmented * parallel : fix sequence termination criteria * llama : silence errors KV cache errors * parallel : remove new line from prompt * parallel : process system prompt once + configurable paramters + llama API * parallel : remove question with short answers * parallel : count cache misses * parallel : print misses on each request * parallel : minor * llama : fix n_kv to never become 0 * parallel : rename hot-plug to continuous-batching * llama : improve llama_batch API + simplify parallel example * simple : add parallel decoding support * simple : improve comments + free batch * ggml-cuda : add rope f16, restore performance with parallel decoding (#3272) * ggml-cuda : add rope f16, restore performance * offload KQ_mask with all models * fix rope shift --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : disable MPI for now ggml-ci * train : make KQ_pos memory buffer permanent via dummy scale op * ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275) ggml-ci * parallel : fix bug (extra BOS) + smaller token_prev array * parallel : fix cases where the input prompts can overflow the batch * parallel : add disabled experimental batch chunking in powers of two * llama : llama.h formatting + comments * simple : add README.md * llama : fix kv cache heuristic when context is less than 32 * parallel : fix crash when `-n -1` * llama : simplify returns if/else branches * metal : use mm kernels for batch size > 2 * examples : utilize new llama_get_logits_ith() * examples : add example for batched decoding * examples : do not eval prompt 2 times (close #3348) * server : clear the KV cache beyond n_past before llama_decode * server : avoid context swaps by shifting the KV cache --------- Co-authored-by: slaren <slarengh@gmail.com>
2023-09-28 18:04:36 +02:00
const T x0 = src[0];
const T x1 = src[1];
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
dst_data[0] = x0*cos_theta - x1*sin_theta;
dst_data[1] = x0*sin_theta + x1*cos_theta;
}
} else {
llm : add Falcon support (#2717) * llama : refactor GGUF constants into static maps * llama : check if model architecture is known * llama : refactor llama_model_load_internal() * gguf : add KV constant maps * llm : read arch-specific KVs * convert : add dummy scores + types * falcon : load tensor data (CPU only) * llama : fix loading progress bar * llama : add arch member to llama_model * falcon : CPU inference working * falcon : support non-40B models * falcon : minor * llama : minor updates ggml-ci * convert-falcon-hf-to-gguf.py : fix special token mapping * llama.cpp : llama default UNK token = id 0 * llama.cpp : fix bpe tokenizer * llama.cpp : fix the fix of bpe tokenizer * ggml : pass eps to ggml_norm * metal : implement RoPE (mode = 2) + avoid ggml_repeat * ggml : ggml_repeat always creates new tensor * falcon : copy-paste self-attention from LLaMA * metal : print extra compute pipeline info * falcon : minor changes (still chasing the Metal problem) * llama.cpp : fix linefeed token * metal : fix GELU kernel numerical stability by using precise::tanh * metal : temporary workaround for the concurrency optimization bug * falcon : add CUDA offloading (#2739) * llama : better model naming and size reporting * llama : prep new tokenizer support * llama : advanced BPE tokenizer based on ggllm.cpp imlpementation * llama : remove oboslete comment ggml-ci * common : remove obsolete BPE API + disable test-tokenizer-1 * llama : revert BPE special-case in llama_byte_to_token() * cuda : add TODOs for RoPE NeoX implementation * llama : default special tokens based on vocab type * perplexity : add log for start of tokenization --------- Co-authored-by: klosax <131523366+klosax@users.noreply.github.com> Co-authored-by: slaren <slarengh@gmail.com>
2023-08-23 22:08:04 +02:00
for (int64_t ib = 0; ib < ne0/n_dims; ++ib) {
for (int64_t ic = 2*tiitg; ic < n_dims; ic += 2*tptg.x) {
const float theta = theta_0 * pow(freq_base, inv_ndims*ic - ib);
llm : add Falcon support (#2717) * llama : refactor GGUF constants into static maps * llama : check if model architecture is known * llama : refactor llama_model_load_internal() * gguf : add KV constant maps * llm : read arch-specific KVs * convert : add dummy scores + types * falcon : load tensor data (CPU only) * llama : fix loading progress bar * llama : add arch member to llama_model * falcon : CPU inference working * falcon : support non-40B models * falcon : minor * llama : minor updates ggml-ci * convert-falcon-hf-to-gguf.py : fix special token mapping * llama.cpp : llama default UNK token = id 0 * llama.cpp : fix bpe tokenizer * llama.cpp : fix the fix of bpe tokenizer * ggml : pass eps to ggml_norm * metal : implement RoPE (mode = 2) + avoid ggml_repeat * ggml : ggml_repeat always creates new tensor * falcon : copy-paste self-attention from LLaMA * metal : print extra compute pipeline info * falcon : minor changes (still chasing the Metal problem) * llama.cpp : fix linefeed token * metal : fix GELU kernel numerical stability by using precise::tanh * metal : temporary workaround for the concurrency optimization bug * falcon : add CUDA offloading (#2739) * llama : better model naming and size reporting * llama : prep new tokenizer support * llama : advanced BPE tokenizer based on ggllm.cpp imlpementation * llama : remove oboslete comment ggml-ci * common : remove obsolete BPE API + disable test-tokenizer-1 * llama : revert BPE special-case in llama_byte_to_token() * cuda : add TODOs for RoPE NeoX implementation * llama : default special tokens based on vocab type * perplexity : add log for start of tokenization --------- Co-authored-by: klosax <131523366+klosax@users.noreply.github.com> Co-authored-by: slaren <slarengh@gmail.com>
2023-08-23 22:08:04 +02:00
const float cos_theta = cos(theta);
const float sin_theta = sin(theta);
const int64_t i0 = ib*n_dims + ic/2;
llama : custom attention mask + parallel decoding + no context swaps (#3228) * tests : verify that RoPE is "additive" * llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) * ggml : ggml_rope now takes a vector with positions instead of n_past * metal : add rope_f16 kernel + optimize cpy kernels * llama : unified KV cache + batch inference API * llama : add new llama_decode() API that works with llama_batch * llama : add cell_max heuristic for more efficient kv_cache * llama : extend llama_kv_cache API * llama : more robust cell_max heuristic + wip shift * metal : disable concurrency optimization * llama : add llama_kv_cache_shift_seq + no more context swaps * llama : apply K-cache roping for Falcon and Baichuan * speculative : fix KV cache management * parallel : example for serving multiple users in parallel * parallel : disable hot-plug to avoid cache fragmentation * fixes : speculative KV cache + llama worst-case graph * llama : extend batch API to select which logits to output * llama : fix worst case graph build * ggml-cuda : update rope implementation for parallel decoding (#3254) * ggml-cuda : update rope implementation for parallel decoding * better solution for p0 computation * fix rope * simpler rope implementation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * make : add parallel to build + fix static functions in llama.cpp * simple : fix token counting * parallel : various improvements * llama : fix cell_max logic + rename functions * parallel : try smaller batches when the KV cache is fragmented * parallel : fix sequence termination criteria * llama : silence errors KV cache errors * parallel : remove new line from prompt * parallel : process system prompt once + configurable paramters + llama API * parallel : remove question with short answers * parallel : count cache misses * parallel : print misses on each request * parallel : minor * llama : fix n_kv to never become 0 * parallel : rename hot-plug to continuous-batching * llama : improve llama_batch API + simplify parallel example * simple : add parallel decoding support * simple : improve comments + free batch * ggml-cuda : add rope f16, restore performance with parallel decoding (#3272) * ggml-cuda : add rope f16, restore performance * offload KQ_mask with all models * fix rope shift --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : disable MPI for now ggml-ci * train : make KQ_pos memory buffer permanent via dummy scale op * ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275) ggml-ci * parallel : fix bug (extra BOS) + smaller token_prev array * parallel : fix cases where the input prompts can overflow the batch * parallel : add disabled experimental batch chunking in powers of two * llama : llama.h formatting + comments * simple : add README.md * llama : fix kv cache heuristic when context is less than 32 * parallel : fix crash when `-n -1` * llama : simplify returns if/else branches * metal : use mm kernels for batch size > 2 * examples : utilize new llama_get_logits_ith() * examples : add example for batched decoding * examples : do not eval prompt 2 times (close #3348) * server : clear the KV cache beyond n_past before llama_decode * server : avoid context swaps by shifting the KV cache --------- Co-authored-by: slaren <slarengh@gmail.com>
2023-09-28 18:04:36 +02:00
device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
device T * dst_data = (device T *)((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
llm : add Falcon support (#2717) * llama : refactor GGUF constants into static maps * llama : check if model architecture is known * llama : refactor llama_model_load_internal() * gguf : add KV constant maps * llm : read arch-specific KVs * convert : add dummy scores + types * falcon : load tensor data (CPU only) * llama : fix loading progress bar * llama : add arch member to llama_model * falcon : CPU inference working * falcon : support non-40B models * falcon : minor * llama : minor updates ggml-ci * convert-falcon-hf-to-gguf.py : fix special token mapping * llama.cpp : llama default UNK token = id 0 * llama.cpp : fix bpe tokenizer * llama.cpp : fix the fix of bpe tokenizer * ggml : pass eps to ggml_norm * metal : implement RoPE (mode = 2) + avoid ggml_repeat * ggml : ggml_repeat always creates new tensor * falcon : copy-paste self-attention from LLaMA * metal : print extra compute pipeline info * falcon : minor changes (still chasing the Metal problem) * llama.cpp : fix linefeed token * metal : fix GELU kernel numerical stability by using precise::tanh * metal : temporary workaround for the concurrency optimization bug * falcon : add CUDA offloading (#2739) * llama : better model naming and size reporting * llama : prep new tokenizer support * llama : advanced BPE tokenizer based on ggllm.cpp imlpementation * llama : remove oboslete comment ggml-ci * common : remove obsolete BPE API + disable test-tokenizer-1 * llama : revert BPE special-case in llama_byte_to_token() * cuda : add TODOs for RoPE NeoX implementation * llama : default special tokens based on vocab type * perplexity : add log for start of tokenization --------- Co-authored-by: klosax <131523366+klosax@users.noreply.github.com> Co-authored-by: slaren <slarengh@gmail.com>
2023-08-23 22:08:04 +02:00
const float x0 = src[0];
const float x1 = src[n_dims/2];
dst_data[0] = x0*cos_theta - x1*sin_theta;
dst_data[n_dims/2] = x0*sin_theta + x1*cos_theta;
}
}
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
}
}
llama : custom attention mask + parallel decoding + no context swaps (#3228) * tests : verify that RoPE is "additive" * llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) * ggml : ggml_rope now takes a vector with positions instead of n_past * metal : add rope_f16 kernel + optimize cpy kernels * llama : unified KV cache + batch inference API * llama : add new llama_decode() API that works with llama_batch * llama : add cell_max heuristic for more efficient kv_cache * llama : extend llama_kv_cache API * llama : more robust cell_max heuristic + wip shift * metal : disable concurrency optimization * llama : add llama_kv_cache_shift_seq + no more context swaps * llama : apply K-cache roping for Falcon and Baichuan * speculative : fix KV cache management * parallel : example for serving multiple users in parallel * parallel : disable hot-plug to avoid cache fragmentation * fixes : speculative KV cache + llama worst-case graph * llama : extend batch API to select which logits to output * llama : fix worst case graph build * ggml-cuda : update rope implementation for parallel decoding (#3254) * ggml-cuda : update rope implementation for parallel decoding * better solution for p0 computation * fix rope * simpler rope implementation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * make : add parallel to build + fix static functions in llama.cpp * simple : fix token counting * parallel : various improvements * llama : fix cell_max logic + rename functions * parallel : try smaller batches when the KV cache is fragmented * parallel : fix sequence termination criteria * llama : silence errors KV cache errors * parallel : remove new line from prompt * parallel : process system prompt once + configurable paramters + llama API * parallel : remove question with short answers * parallel : count cache misses * parallel : print misses on each request * parallel : minor * llama : fix n_kv to never become 0 * parallel : rename hot-plug to continuous-batching * llama : improve llama_batch API + simplify parallel example * simple : add parallel decoding support * simple : improve comments + free batch * ggml-cuda : add rope f16, restore performance with parallel decoding (#3272) * ggml-cuda : add rope f16, restore performance * offload KQ_mask with all models * fix rope shift --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : disable MPI for now ggml-ci * train : make KQ_pos memory buffer permanent via dummy scale op * ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275) ggml-ci * parallel : fix bug (extra BOS) + smaller token_prev array * parallel : fix cases where the input prompts can overflow the batch * parallel : add disabled experimental batch chunking in powers of two * llama : llama.h formatting + comments * simple : add README.md * llama : fix kv cache heuristic when context is less than 32 * parallel : fix crash when `-n -1` * llama : simplify returns if/else branches * metal : use mm kernels for batch size > 2 * examples : utilize new llama_get_logits_ith() * examples : add example for batched decoding * examples : do not eval prompt 2 times (close #3348) * server : clear the KV cache beyond n_past before llama_decode * server : avoid context swaps by shifting the KV cache --------- Co-authored-by: slaren <slarengh@gmail.com>
2023-09-28 18:04:36 +02:00
template [[host_name("kernel_rope_f32")]] kernel rope_t kernel_rope<float>;
template [[host_name("kernel_rope_f16")]] kernel rope_t kernel_rope<half>;
kernel void kernel_cpy_f16_f16(
device const half * src0,
device half * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig[2];
const int64_t i02 = tgpig[1];
const int64_t i01 = tgpig[0];
const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
const int64_t i3 = n / (ne2*ne1*ne0);
const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
device half * dst_data = (device half *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
device const half * src = (device half *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
dst_data[i00] = src[0];
}
}
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
kernel void kernel_cpy_f32_f16(
device const float * src0,
device half * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig[2];
const int64_t i02 = tgpig[1];
const int64_t i01 = tgpig[0];
const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
const int64_t i3 = n / (ne2*ne1*ne0);
const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
device half * dst_data = (device half *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
dst_data[i00] = src[0];
}
}
kernel void kernel_cpy_f32_f32(
device const float * src0,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig[2];
const int64_t i02 = tgpig[1];
const int64_t i01 = tgpig[0];
const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
const int64_t i3 = n / (ne2*ne1*ne0);
const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
device float * dst_data = (device float *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
dst_data[i00] = src[0];
}
}
kernel void kernel_concat(
device const char * src0,
device const char * src1,
device char * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant int64_t & ne13,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant uint64_t & nb13,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig.z;
const int64_t i02 = tgpig.y;
const int64_t i01 = tgpig.x;
const int64_t i13 = i03 % ne13;
const int64_t i12 = i02 % ne12;
const int64_t i11 = i01 % ne11;
device const char * src0_ptr = src0 + i03 * nb03 + i02 * nb02 + i01 * nb01 + tpitg.x*nb00;
device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11 + tpitg.x*nb10;
device char * dst_ptr = dst + i03*nb3 + i02*nb2 + i01*nb1 + tpitg.x*nb0;
for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
if (i02 < ne02) {
((device float *)dst_ptr)[0] = ((device float *)src0_ptr)[0];
src0_ptr += ntg.x*nb00;
} else {
((device float *)dst_ptr)[0] = ((device float *)src1_ptr)[0];
src1_ptr += ntg.x*nb10;
}
dst_ptr += ntg.x*nb0;
}
}
//============================================ k-quants ======================================================
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
#ifndef QK_K
#define QK_K 256
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
#else
static_assert(QK_K == 256 || QK_K == 64, "QK_K must be 256 or 64");
#endif
#if QK_K == 256
#define K_SCALE_SIZE 12
#else
#define K_SCALE_SIZE 4
#endif
typedef struct {
uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
uint8_t qs[QK_K/4]; // quants
half d; // super-block scale for quantized scales
half dmin; // super-block scale for quantized mins
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
} block_q2_K;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
// 84 bytes / block
typedef struct {
uint8_t hmask[QK_K/8]; // quants - high bit
uint8_t qs[QK_K/4]; // quants - low 2 bits
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
#if QK_K == 64
uint8_t scales[2];
#else
uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
#endif
half d; // super-block scale
} block_q3_K;
#if QK_K == 64
typedef struct {
half d[2]; // super-block scales/mins
uint8_t scales[2];
uint8_t qs[QK_K/2]; // 4-bit quants
} block_q4_K;
#else
typedef struct {
half d; // super-block scale for quantized scales
half dmin; // super-block scale for quantized mins
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
uint8_t qs[QK_K/2]; // 4--bit quants
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
} block_q4_K;
#endif
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
#if QK_K == 64
typedef struct {
half d; // super-block scales/mins
int8_t scales[QK_K/16]; // 8-bit block scales
uint8_t qh[QK_K/8]; // quants, high bit
uint8_t qs[QK_K/2]; // quants, low 4 bits
} block_q5_K;
#else
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
typedef struct {
half d; // super-block scale for quantized scales
half dmin; // super-block scale for quantized mins
uint8_t scales[3*QK_K/64]; // scales and mins, quantized with 6 bits
uint8_t qh[QK_K/8]; // quants, high bit
uint8_t qs[QK_K/2]; // quants, low 4 bits
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
} block_q5_K;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
// 176 bytes / block
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
#endif
typedef struct {
uint8_t ql[QK_K/2]; // quants, lower 4 bits
uint8_t qh[QK_K/4]; // quants, upper 2 bits
int8_t scales[QK_K/16]; // scales, quantized with 8 bits
half d; // super-block scale
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
} block_q6_K;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
// 210 bytes / block
static inline uchar4 get_scale_min_k4(int j, device const uint8_t * q) {
uchar4 r;
if (j < 4) {
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
r[0] = q[j+0] & 63;
r[2] = q[j+1] & 63;
r[1] = q[j+4] & 63;
r[3] = q[j+5] & 63;
} else {
r[0] = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4);
r[2] = (q[j+5] & 0xF) | ((q[j-3] >> 6) << 4);
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
r[1] = (q[j+4] >> 4) | ((q[j-0] >> 6) << 4);
r[3] = (q[j+5] >> 4) | ((q[j+1] >> 6) << 4);
}
return r;
}
//====================================== dot products =========================
kernel void kernel_mul_mv_q2_K_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01[[buffer(4)]],
constant int64_t & ne02[[buffer(5)]],
constant int64_t & ne10[[buffer(9)]],
constant int64_t & ne12[[buffer(11)]],
constant int64_t & ne0[[buffer(15)]],
constant int64_t & ne1[[buffer(16)]],
constant uint & gqa[[buffer(17)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
const int nb = ne00/QK_K;
const int r0 = tgpig.x;
const int r1 = tgpig.y;
const int r2 = tgpig.z;
const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
const int ib_row = first_row * nb;
const uint offset0 = r2/gqa*(nb*ne0);
device const block_q2_K * x = (device const block_q2_K *) src0 + ib_row + offset0;
device const float * y = (device const float *) src1 + r1*ne10 + r2*ne00*ne1;
float yl[32];
float sumf[N_DST]={0.f}, all_sum;
const int step = sizeof(block_q2_K) * nb;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
#if QK_K == 256
const int ix = tiisg/8; // 0...3
const int it = tiisg%8; // 0...7
const int im = it/4; // 0 or 1
const int ir = it%4; // 0...3
const int is = (8*ir)/16;// 0 or 1
device const float * y4 = y + ix * QK_K + 128 * im + 8 * ir;
for (int ib = ix; ib < nb; ib += 4) {
float4 sumy = {0.f, 0.f, 0.f, 0.f};
for (int i = 0; i < 8; ++i) {
yl[i+ 0] = y4[i+ 0]; sumy[0] += yl[i+ 0];
yl[i+ 8] = y4[i+32]; sumy[1] += yl[i+ 8];
yl[i+16] = y4[i+64]; sumy[2] += yl[i+16];
yl[i+24] = y4[i+96]; sumy[3] += yl[i+24];
}
device const uint8_t * sc = (device const uint8_t *)x[ib].scales + 8*im + is;
device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 16 * im + 4 * ir;
device const half * dh = &x[ib].d;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
for (int row = 0; row < N_DST; row++) {
float4 acc1 = {0.f, 0.f, 0.f, 0.f};
float4 acc2 = {0.f, 0.f, 0.f, 0.f};
for (int i = 0; i < 8; i += 2) {
acc1[0] += yl[i+ 0] * (qs[i/2] & 0x0003);
acc2[0] += yl[i+ 1] * (qs[i/2] & 0x0300);
acc1[1] += yl[i+ 8] * (qs[i/2] & 0x000c);
acc2[1] += yl[i+ 9] * (qs[i/2] & 0x0c00);
acc1[2] += yl[i+16] * (qs[i/2] & 0x0030);
acc2[2] += yl[i+17] * (qs[i/2] & 0x3000);
acc1[3] += yl[i+24] * (qs[i/2] & 0x00c0);
acc2[3] += yl[i+25] * (qs[i/2] & 0xc000);
}
float dall = dh[0];
float dmin = dh[1] * 1.f/16.f;
sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc2[0]) * (sc[0] & 0xF) * 1.f/ 1.f +
(acc1[1] + 1.f/256.f * acc2[1]) * (sc[2] & 0xF) * 1.f/ 4.f +
(acc1[2] + 1.f/256.f * acc2[2]) * (sc[4] & 0xF) * 1.f/16.f +
(acc1[3] + 1.f/256.f * acc2[3]) * (sc[6] & 0xF) * 1.f/64.f) -
dmin * (sumy[0] * (sc[0] & 0xF0) + sumy[1] * (sc[2] & 0xF0) + sumy[2] * (sc[4] & 0xF0) + sumy[3] * (sc[6] & 0xF0));
qs += step/2;
sc += step;
dh += step/2;
}
y4 += 4 * QK_K;
}
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
#else
const int ix = tiisg/2; // 0...15
const int it = tiisg%2; // 0...1
device const float * y4 = y + ix * QK_K + 8 * it;
for (int ib = ix; ib < nb; ib += 16) {
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
float4 sumy = {0.f, 0.f, 0.f, 0.f};
for (int i = 0; i < 8; ++i) {
yl[i+ 0] = y4[i+ 0]; sumy[0] += yl[i+ 0];
yl[i+ 8] = y4[i+16]; sumy[1] += yl[i+ 8];
yl[i+16] = y4[i+32]; sumy[2] += yl[i+16];
yl[i+24] = y4[i+48]; sumy[3] += yl[i+24];
}
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
device const uint8_t * sc = (device const uint8_t *)x[ib].scales;
device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 4 * it;
device const half * dh = &x[ib].d;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
for (int row = 0; row < N_DST; row++) {
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
float4 acc1 = {0.f, 0.f, 0.f, 0.f};
float4 acc2 = {0.f, 0.f, 0.f, 0.f};
for (int i = 0; i < 8; i += 2) {
acc1[0] += yl[i+ 0] * (qs[i/2] & 0x0003);
acc2[0] += yl[i+ 1] * (qs[i/2] & 0x0300);
acc1[1] += yl[i+ 8] * (qs[i/2] & 0x000c);
acc2[1] += yl[i+ 9] * (qs[i/2] & 0x0c00);
acc1[2] += yl[i+16] * (qs[i/2] & 0x0030);
acc2[2] += yl[i+17] * (qs[i/2] & 0x3000);
acc1[3] += yl[i+24] * (qs[i/2] & 0x00c0);
acc2[3] += yl[i+25] * (qs[i/2] & 0xc000);
}
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
float dall = dh[0];
float dmin = dh[1];
sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc2[0]) * (sc[0] & 0xF) * 1.f/ 1.f +
(acc1[1] + 1.f/256.f * acc2[1]) * (sc[1] & 0xF) * 1.f/ 4.f +
(acc1[2] + 1.f/256.f * acc2[2]) * (sc[2] & 0xF) * 1.f/16.f +
(acc1[3] + 1.f/256.f * acc2[3]) * (sc[3] & 0xF) * 1.f/64.f) -
dmin * (sumy[0] * (sc[0] >> 4) + sumy[1] * (sc[1] >> 4) + sumy[2] * (sc[2] >> 4) + sumy[3] * (sc[3] >> 4));
qs += step/2;
sc += step;
dh += step/2;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
}
y4 += 16 * QK_K;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
}
#endif
for (int row = 0; row < N_DST; ++row) {
all_sum = simd_sum(sumf[row]);
if (tiisg == 0) {
dst[r1*ne0 + r2*ne0*ne1 + first_row + row] = all_sum;
}
}
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
}
#if QK_K == 256
kernel void kernel_mul_mv_q3_K_f32(
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01[[buffer(4)]],
constant int64_t & ne02[[buffer(5)]],
constant int64_t & ne10[[buffer(9)]],
constant int64_t & ne12[[buffer(11)]],
constant int64_t & ne0[[buffer(15)]],
constant int64_t & ne1[[buffer(16)]],
constant uint & gqa[[buffer(17)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
const int nb = ne00/QK_K;
const int64_t r0 = tgpig.x;
const int64_t r1 = tgpig.y;
const int64_t r2 = tgpig.z;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
const int first_row = (r0 * N_SIMDGROUP + sgitg) * 2;
const uint offset0 = r2/gqa*(nb*ne0);
device const block_q3_K * x = (device const block_q3_K *) src0 + first_row*nb + offset0;
device const float * yy = (device const float *) src1 + r1*ne10 + r2*ne00*ne1;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
float yl[32];
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
llama : custom attention mask + parallel decoding + no context swaps (#3228) * tests : verify that RoPE is "additive" * llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) * ggml : ggml_rope now takes a vector with positions instead of n_past * metal : add rope_f16 kernel + optimize cpy kernels * llama : unified KV cache + batch inference API * llama : add new llama_decode() API that works with llama_batch * llama : add cell_max heuristic for more efficient kv_cache * llama : extend llama_kv_cache API * llama : more robust cell_max heuristic + wip shift * metal : disable concurrency optimization * llama : add llama_kv_cache_shift_seq + no more context swaps * llama : apply K-cache roping for Falcon and Baichuan * speculative : fix KV cache management * parallel : example for serving multiple users in parallel * parallel : disable hot-plug to avoid cache fragmentation * fixes : speculative KV cache + llama worst-case graph * llama : extend batch API to select which logits to output * llama : fix worst case graph build * ggml-cuda : update rope implementation for parallel decoding (#3254) * ggml-cuda : update rope implementation for parallel decoding * better solution for p0 computation * fix rope * simpler rope implementation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * make : add parallel to build + fix static functions in llama.cpp * simple : fix token counting * parallel : various improvements * llama : fix cell_max logic + rename functions * parallel : try smaller batches when the KV cache is fragmented * parallel : fix sequence termination criteria * llama : silence errors KV cache errors * parallel : remove new line from prompt * parallel : process system prompt once + configurable paramters + llama API * parallel : remove question with short answers * parallel : count cache misses * parallel : print misses on each request * parallel : minor * llama : fix n_kv to never become 0 * parallel : rename hot-plug to continuous-batching * llama : improve llama_batch API + simplify parallel example * simple : add parallel decoding support * simple : improve comments + free batch * ggml-cuda : add rope f16, restore performance with parallel decoding (#3272) * ggml-cuda : add rope f16, restore performance * offload KQ_mask with all models * fix rope shift --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : disable MPI for now ggml-ci * train : make KQ_pos memory buffer permanent via dummy scale op * ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275) ggml-ci * parallel : fix bug (extra BOS) + smaller token_prev array * parallel : fix cases where the input prompts can overflow the batch * parallel : add disabled experimental batch chunking in powers of two * llama : llama.h formatting + comments * simple : add README.md * llama : fix kv cache heuristic when context is less than 32 * parallel : fix crash when `-n -1` * llama : simplify returns if/else branches * metal : use mm kernels for batch size > 2 * examples : utilize new llama_get_logits_ith() * examples : add example for batched decoding * examples : do not eval prompt 2 times (close #3348) * server : clear the KV cache beyond n_past before llama_decode * server : avoid context swaps by shifting the KV cache --------- Co-authored-by: slaren <slarengh@gmail.com>
2023-09-28 18:04:36 +02:00
//const uint16_t kmask1 = 0x3030;
//const uint16_t kmask2 = 0x0f0f;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
const int tid = tiisg/4;
const int ix = tiisg%4;
const int ip = tid/4; // 0 or 1
const int il = 2*((tid%4)/2); // 0 or 2
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
const int ir = tid%2;
const int n = 8;
const int l0 = n*ir;
// One would think that the Metal compiler would figure out that ip and il can only have
// 4 possible states, and optimize accordingly. Well, no. It needs help, and we do it
// with these two tales.
//
// Possible masks for the high bit
const ushort4 mm[4] = {{0x0001, 0x0100, 0x0002, 0x0200}, // ip = 0, il = 0
{0x0004, 0x0400, 0x0008, 0x0800}, // ip = 0, il = 2
{0x0010, 0x1000, 0x0020, 0x2000}, // ip = 1, il = 0
{0x0040, 0x4000, 0x0080, 0x8000}}; // ip = 1, il = 2
// Possible masks for the low 2 bits
const int4 qm[2] = {{0x0003, 0x0300, 0x000c, 0x0c00}, {0x0030, 0x3000, 0x00c0, 0xc000}};
const ushort4 hm = mm[2*ip + il/2];
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
const int shift = 2*il;
const float v1 = il == 0 ? 4.f : 64.f;
const float v2 = 4.f * v1;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
const uint16_t s_shift1 = 4*ip;
const uint16_t s_shift2 = s_shift1 + il;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
const int q_offset = 32*ip + l0;
const int y_offset = 128*ip + 32*il + l0;
const int step = sizeof(block_q3_K) * nb / 2;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
device const float * y1 = yy + ix*QK_K + y_offset;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
uint32_t scales32, aux32;
thread uint16_t * scales16 = (thread uint16_t *)&scales32;
thread const int8_t * scales = (thread const int8_t *)&scales32;
float sumf1[2] = {0.f};
float sumf2[2] = {0.f};
for (int i = ix; i < nb; i += 4) {
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
for (int l = 0; l < 8; ++l) {
yl[l+ 0] = y1[l+ 0];
yl[l+ 8] = y1[l+16];
yl[l+16] = y1[l+32];
yl[l+24] = y1[l+48];
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
}
device const uint16_t * q = (device const uint16_t *)(x[i].qs + q_offset);
device const uint16_t * h = (device const uint16_t *)(x[i].hmask + l0);
device const uint16_t * a = (device const uint16_t *)(x[i].scales);
device const half * dh = &x[i].d;
for (int row = 0; row < 2; ++row) {
const float d_all = (float)dh[0];
scales16[0] = a[4];
scales16[1] = a[5];
aux32 = ((scales32 >> s_shift2) << 4) & 0x30303030;
scales16[0] = a[il+0];
scales16[1] = a[il+1];
scales32 = ((scales32 >> s_shift1) & 0x0f0f0f0f) | aux32;
float s1 = 0, s2 = 0, s3 = 0, s4 = 0, s5 = 0, s6 = 0;
for (int l = 0; l < n; l += 2) {
const int32_t qs = q[l/2];
s1 += yl[l+0] * (qs & qm[il/2][0]);
s2 += yl[l+1] * (qs & qm[il/2][1]);
s3 += ((h[l/2] & hm[0]) ? 0.f : yl[l+0]) + ((h[l/2] & hm[1]) ? 0.f : yl[l+1]);
s4 += yl[l+16] * (qs & qm[il/2][2]);
s5 += yl[l+17] * (qs & qm[il/2][3]);
s6 += ((h[l/2] & hm[2]) ? 0.f : yl[l+16]) + ((h[l/2] & hm[3]) ? 0.f : yl[l+17]);
}
float d1 = d_all * (s1 + 1.f/256.f * s2 - s3*v1);
float d2 = d_all * (s4 + 1.f/256.f * s5 - s6*v2);
sumf1[row] += d1 * (scales[0] - 32);
sumf2[row] += d2 * (scales[2] - 32);
s1 = s2 = s3 = s4 = s5 = s6 = 0;
for (int l = 0; l < n; l += 2) {
const int32_t qs = q[l/2+8];
s1 += yl[l+8] * (qs & qm[il/2][0]);
s2 += yl[l+9] * (qs & qm[il/2][1]);
s3 += ((h[l/2+8] & hm[0]) ? 0.f : yl[l+8]) + ((h[l/2+8] & hm[1]) ? 0.f : yl[l+9]);
s4 += yl[l+24] * (qs & qm[il/2][2]);
s5 += yl[l+25] * (qs & qm[il/2][3]);
s6 += ((h[l/2+8] & hm[2]) ? 0.f : yl[l+24]) + ((h[l/2+8] & hm[3]) ? 0.f : yl[l+25]);
}
d1 = d_all * (s1 + 1.f/256.f * s2 - s3*v1);
d2 = d_all * (s4 + 1.f/256.f * s5 - s6*v2);
sumf1[row] += d1 * (scales[1] - 32);
sumf2[row] += d2 * (scales[3] - 32);
q += step;
h += step;
a += step;
dh += step;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
}
y1 += 4 * QK_K;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
}
for (int row = 0; row < 2; ++row) {
const float sumf = (sumf1[row] + 0.25f * sumf2[row]) / (1 << shift);
sumf1[row] = simd_sum(sumf);
}
if (tiisg == 0) {
for (int row = 0; row < 2; ++row) {
dst[r1*ne0 + r2*ne0*ne1 + first_row + row] = sumf1[row];
}
}
}
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
#else
kernel void kernel_mul_mv_q3_K_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01[[buffer(4)]],
constant int64_t & ne02[[buffer(5)]],
constant int64_t & ne10[[buffer(9)]],
constant int64_t & ne12[[buffer(11)]],
constant int64_t & ne0[[buffer(15)]],
constant int64_t & ne1[[buffer(16)]],
constant uint & gqa[[buffer(17)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
const int nb = ne00/QK_K;
const int64_t r0 = tgpig.x;
const int64_t r1 = tgpig.y;
const int64_t r2 = tgpig.z;
const int row = 2 * r0 + sgitg;
const uint offset0 = r2/gqa*(nb*ne0);
device const block_q3_K * x = (device const block_q3_K *) src0 + row*nb + offset0;
device const float * yy = (device const float *) src1 + r1*ne10 + r2*ne00*ne1;
const int ix = tiisg/4;
const int il = 4 * (tiisg%4);// 0, 4, 8, 12
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
const int im = il/8; // 0, 0, 1, 1
const int in = il%8; // 0, 4, 0, 4
float2 sum = {0.f, 0.f};
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
for (int i = ix; i < nb; i += 8) {
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
const float d_all = (float)(x[i].d);
device const uint16_t * q = (device const uint16_t *)(x[i].qs + il);
device const uint16_t * h = (device const uint16_t *)(x[i].hmask + in);
device const uint16_t * s = (device const uint16_t *)(x[i].scales);
device const float * y = yy + i * QK_K + il;
const float d1 = d_all * ((int32_t)(s[0] & 0x000F) - 8);
const float d2 = d_all * ((int32_t)(s[0] & 0x00F0) - 128) * 1.f/64.f;
const float d3 = d_all * ((int32_t)(s[0] & 0x0F00) - 2048) * 1.f/4096.f;
const float d4 = d_all * ((int32_t)(s[0] & 0xF000) - 32768) * 1.f/262144.f;
for (int l = 0; l < 4; l += 2) {
const uint16_t hm = h[l/2] >> im;
sum[0] += y[l+ 0] * d1 * ((int32_t)(q[l/2] & 0x0003) - ((hm & 0x0001) ? 0 : 4))
+ y[l+16] * d2 * ((int32_t)(q[l/2] & 0x000c) - ((hm & 0x0004) ? 0 : 16))
+ y[l+32] * d3 * ((int32_t)(q[l/2] & 0x0030) - ((hm & 0x0010) ? 0 : 64))
+ y[l+48] * d4 * ((int32_t)(q[l/2] & 0x00c0) - ((hm & 0x0040) ? 0 : 256));
sum[1] += y[l+ 1] * d1 * ((int32_t)(q[l/2] & 0x0300) - ((hm & 0x0100) ? 0 : 1024))
+ y[l+17] * d2 * ((int32_t)(q[l/2] & 0x0c00) - ((hm & 0x0400) ? 0 : 4096))
+ y[l+33] * d3 * ((int32_t)(q[l/2] & 0x3000) - ((hm & 0x1000) ? 0 : 16384))
+ y[l+49] * d4 * ((int32_t)(q[l/2] & 0xc000) - ((hm & 0x4000) ? 0 : 65536));
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
}
}
const float sumf = sum[0] + sum[1] * 1.f/256.f;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
const float tot = simd_sum(sumf);
if (tiisg == 0) {
dst[r1*ne0 + r2*ne0*ne1 + row] = tot;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
}
}
#endif
#if QK_K == 256
kernel void kernel_mul_mv_q4_K_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01 [[buffer(4)]],
constant int64_t & ne02 [[buffer(5)]],
constant int64_t & ne10 [[buffer(9)]],
constant int64_t & ne12 [[buffer(11)]],
constant int64_t & ne0 [[buffer(15)]],
constant int64_t & ne1 [[buffer(16)]],
constant uint & gqa [[buffer(17)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
const uint16_t kmask1 = 0x3f3f;
const uint16_t kmask2 = 0x0f0f;
const uint16_t kmask3 = 0xc0c0;
const int ix = tiisg/8; // 0...3
const int it = tiisg%8; // 0...7
const int im = it/4; // 0 or 1
const int ir = it%4; // 0...3
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
const int nb = ne00/QK_K;
const int r0 = tgpig.x;
const int r1 = tgpig.y;
const int r2 = tgpig.z;
//const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
const int first_row = r0 * N_DST;
const int ib_row = first_row * nb;
const uint offset0 = r2/gqa*(nb*ne0);
device const block_q4_K * x = (device const block_q4_K *) src0 + ib_row + offset0;
device const float * y = (device const float *) src1 + r1*ne10 + r2*ne00*ne1;
float yl[16];
float yh[16];
float sumf[N_DST]={0.f}, all_sum;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
const int step = sizeof(block_q4_K) * nb / 2;
device const float * y4 = y + ix * QK_K + 64 * im + 8 * ir;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
uint16_t sc16[4];
thread const uint8_t * sc8 = (thread const uint8_t *)sc16;
for (int ib = ix; ib < nb; ib += 4) {
float4 sumy = {0.f, 0.f, 0.f, 0.f};
for (int i = 0; i < 8; ++i) {
yl[i+0] = y4[i+ 0]; sumy[0] += yl[i+0];
yl[i+8] = y4[i+ 32]; sumy[1] += yl[i+8];
yh[i+0] = y4[i+128]; sumy[2] += yh[i+0];
yh[i+8] = y4[i+160]; sumy[3] += yh[i+8];
}
device const uint16_t * sc = (device const uint16_t *)x[ib].scales + im;
device const uint16_t * q1 = (device const uint16_t *)x[ib].qs + 16 * im + 4 * ir;
device const half * dh = &x[ib].d;
for (int row = 0; row < N_DST; row++) {
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
sc16[0] = sc[0] & kmask1;
sc16[1] = sc[2] & kmask1;
sc16[2] = ((sc[4] >> 0) & kmask2) | ((sc[0] & kmask3) >> 2);
sc16[3] = ((sc[4] >> 4) & kmask2) | ((sc[2] & kmask3) >> 2);
device const uint16_t * q2 = q1 + 32;
float4 acc1 = {0.f, 0.f, 0.f, 0.f};
float4 acc2 = {0.f, 0.f, 0.f, 0.f};
for (int i = 0; i < 8; i += 2) {
acc1[0] += yl[i+0] * (q1[i/2] & 0x000F);
acc1[1] += yl[i+1] * (q1[i/2] & 0x0F00);
acc1[2] += yl[i+8] * (q1[i/2] & 0x00F0);
acc1[3] += yl[i+9] * (q1[i/2] & 0xF000);
acc2[0] += yh[i+0] * (q2[i/2] & 0x000F);
acc2[1] += yh[i+1] * (q2[i/2] & 0x0F00);
acc2[2] += yh[i+8] * (q2[i/2] & 0x00F0);
acc2[3] += yh[i+9] * (q2[i/2] & 0xF000);
}
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
float dall = dh[0];
float dmin = dh[1];
sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc1[1]) * sc8[0] +
(acc1[2] + 1.f/256.f * acc1[3]) * sc8[1] * 1.f/16.f +
(acc2[0] + 1.f/256.f * acc2[1]) * sc8[4] +
(acc2[2] + 1.f/256.f * acc2[3]) * sc8[5] * 1.f/16.f) -
dmin * (sumy[0] * sc8[2] + sumy[1] * sc8[3] + sumy[2] * sc8[6] + sumy[3] * sc8[7]);
q1 += step;
sc += step;
dh += step;
}
y4 += 4 * QK_K;
}
for (int row = 0; row < N_DST; ++row) {
all_sum = simd_sum(sumf[row]);
if (tiisg == 0) {
dst[r1*ne0 + r2*ne0*ne1 + first_row + row] = all_sum;
}
}
}
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
#else
kernel void kernel_mul_mv_q4_K_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01[[buffer(4)]],
constant int64_t & ne02[[buffer(5)]],
constant int64_t & ne10[[buffer(9)]],
constant int64_t & ne12[[buffer(11)]],
constant int64_t & ne0[[buffer(15)]],
constant int64_t & ne1[[buffer(16)]],
constant uint & gqa[[buffer(17)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
const int ix = tiisg/4; // 0...7
const int it = tiisg%4; // 0...3
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
const int nb = ne00/QK_K;
const int r0 = tgpig.x;
const int r1 = tgpig.y;
const int r2 = tgpig.z;
const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
const int ib_row = first_row * nb;
const uint offset0 = r2/gqa*(nb*ne0);
device const block_q4_K * x = (device const block_q4_K *) src0 + ib_row + offset0;
device const float * y = (device const float *) src1 + r1*ne10 + r2*ne00*ne1;
float yl[8];
float yh[8];
float sumf[N_DST]={0.f}, all_sum;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
const int step = sizeof(block_q4_K) * nb / 2;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
device const float * y4 = y + ix * QK_K + 8 * it;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
uint16_t sc16[4];
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
for (int ib = ix; ib < nb; ib += 8) {
float2 sumy = {0.f, 0.f};
for (int i = 0; i < 8; ++i) {
yl[i] = y4[i+ 0]; sumy[0] += yl[i];
yh[i] = y4[i+32]; sumy[1] += yh[i];
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
}
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
device const uint16_t * sc = (device const uint16_t *)x[ib].scales;
device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 4 * it;
device const half * dh = x[ib].d;
for (int row = 0; row < N_DST; row++) {
sc16[0] = sc[0] & 0x000f;
sc16[1] = sc[0] & 0x0f00;
sc16[2] = sc[0] & 0x00f0;
sc16[3] = sc[0] & 0xf000;
float2 acc1 = {0.f, 0.f};
float2 acc2 = {0.f, 0.f};
for (int i = 0; i < 8; i += 2) {
acc1[0] += yl[i+0] * (qs[i/2] & 0x000F);
acc1[1] += yl[i+1] * (qs[i/2] & 0x0F00);
acc2[0] += yh[i+0] * (qs[i/2] & 0x00F0);
acc2[1] += yh[i+1] * (qs[i/2] & 0xF000);
}
float dall = dh[0];
float dmin = dh[1];
sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc1[1]) * sc16[0] +
(acc2[0] + 1.f/256.f * acc2[1]) * sc16[1] * 1.f/4096.f) -
dmin * 1.f/16.f * (sumy[0] * sc16[2] + sumy[1] * sc16[3] * 1.f/256.f);
qs += step;
sc += step;
dh += step;
}
y4 += 8 * QK_K;
}
for (int row = 0; row < N_DST; ++row) {
all_sum = simd_sum(sumf[row]);
if (tiisg == 0) {
dst[r1*ne0+ r2*ne0*ne1 + first_row + row] = all_sum;
}
}
}
#endif
kernel void kernel_mul_mv_q5_K_f32(
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01[[buffer(4)]],
constant int64_t & ne02[[buffer(5)]],
constant int64_t & ne10[[buffer(9)]],
constant int64_t & ne12[[buffer(11)]],
constant int64_t & ne0[[buffer(15)]],
constant int64_t & ne1[[buffer(16)]],
constant uint & gqa[[buffer(17)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
const int nb = ne00/QK_K;
const int64_t r0 = tgpig.x;
const int64_t r1 = tgpig.y;
const int r2 = tgpig.z;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
const int first_row = (r0 * N_SIMDGROUP + sgitg) * 2;
const uint offset0 = r2/gqa*(nb*ne0);
device const block_q5_K * x = (device const block_q5_K *) src0 + first_row*nb + offset0;
device const float * yy = (device const float *) src1 + r1*ne10 + r2*ne00*ne1;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
float sumf[2]={0.f};
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
const int step = sizeof(block_q5_K) * nb;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
#if QK_K == 256
#
float yl[16], yh[16];
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
const uint16_t kmask1 = 0x3f3f;
const uint16_t kmask2 = 0x0f0f;
const uint16_t kmask3 = 0xc0c0;
const int tid = tiisg/4;
const int ix = tiisg%4;
const int im = tid/4;
const int ir = tid%4;
const int n = 8;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
const int l0 = n*ir;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
const int q_offset = 32*im + l0;
const int y_offset = 64*im + l0;
const uint8_t hm1 = 1u << (2*im);
const uint8_t hm2 = hm1 << 1;
const uint8_t hm3 = hm1 << 4;
const uint8_t hm4 = hm2 << 4;
uint16_t sc16[4];
thread const uint8_t * sc8 = (thread const uint8_t *)sc16;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
device const float * y1 = yy + ix*QK_K + y_offset;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
for (int i = ix; i < nb; i += 4) {
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
device const uint8_t * q1 = x[i].qs + q_offset;
device const uint8_t * qh = x[i].qh + l0;
device const half * dh = &x[i].d;
device const uint16_t * a = (device const uint16_t *)x[i].scales + im;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
device const float * y2 = y1 + 128;
float4 sumy = {0.f, 0.f, 0.f, 0.f};
for (int l = 0; l < 8; ++l) {
yl[l+0] = y1[l+ 0]; sumy[0] += yl[l+0];
yl[l+8] = y1[l+32]; sumy[1] += yl[l+8];
yh[l+0] = y2[l+ 0]; sumy[2] += yh[l+0];
yh[l+8] = y2[l+32]; sumy[3] += yh[l+8];
}
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
for (int row = 0; row < 2; ++row) {
device const uint8_t * q2 = q1 + 64;
sc16[0] = a[0] & kmask1;
sc16[1] = a[2] & kmask1;
sc16[2] = ((a[4] >> 0) & kmask2) | ((a[0] & kmask3) >> 2);
sc16[3] = ((a[4] >> 4) & kmask2) | ((a[2] & kmask3) >> 2);
float4 acc1 = {0.f};
float4 acc2 = {0.f};
for (int l = 0; l < n; ++l) {
uint8_t h = qh[l];
acc1[0] += yl[l+0] * (q1[l] & 0x0F);
acc1[1] += yl[l+8] * (q1[l] & 0xF0);
acc1[2] += yh[l+0] * (q2[l] & 0x0F);
acc1[3] += yh[l+8] * (q2[l] & 0xF0);
acc2[0] += h & hm1 ? yl[l+0] : 0.f;
acc2[1] += h & hm2 ? yl[l+8] : 0.f;
acc2[2] += h & hm3 ? yh[l+0] : 0.f;
acc2[3] += h & hm4 ? yh[l+8] : 0.f;
}
const float dall = dh[0];
const float dmin = dh[1];
sumf[row] += dall * (sc8[0] * (acc1[0] + 16.f*acc2[0]) +
sc8[1] * (acc1[1]/16.f + 16.f*acc2[1]) +
sc8[4] * (acc1[2] + 16.f*acc2[2]) +
sc8[5] * (acc1[3]/16.f + 16.f*acc2[3])) -
dmin * (sumy[0] * sc8[2] + sumy[1] * sc8[3] + sumy[2] * sc8[6] + sumy[3] * sc8[7]);
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
q1 += step;
qh += step;
dh += step/2;
a += step/2;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
}
y1 += 4 * QK_K;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
}
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
#else
float yl[8], yh[8];
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
const int il = 4 * (tiisg/8); // 0, 4, 8, 12
const int ix = tiisg%8;
const int im = il/8; // 0, 0, 1, 1
const int in = il%8; // 0, 4, 0, 4
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
device const float * y = yy + ix*QK_K + il;
for (int i = ix; i < nb; i += 8) {
for (int l = 0; l < 4; ++l) {
yl[l+0] = y[l+ 0];
yl[l+4] = y[l+16];
yh[l+0] = y[l+32];
yh[l+4] = y[l+48];
}
device const half * dh = &x[i].d;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
device const uint8_t * q = x[i].qs + il;
device const uint8_t * h = x[i].qh + in;
device const int8_t * s = x[i].scales;
for (int row = 0; row < 2; ++row) {
const float d = dh[0];
float2 acc = {0.f, 0.f};
for (int l = 0; l < 4; ++l) {
const uint8_t hl = h[l] >> im;
acc[0] += yl[l+0] * s[0] * ((int16_t)(q[l+ 0] & 0x0F) - (hl & 0x01 ? 0 : 16))
+ yl[l+4] * s[1] * ((int16_t)(q[l+16] & 0x0F) - (hl & 0x04 ? 0 : 16));
acc[1] += yh[l+0] * s[2] * ((int16_t)(q[l+ 0] & 0xF0) - (hl & 0x10 ? 0 : 256))
+ yh[l+4] * s[3] * ((int16_t)(q[l+16] & 0xF0) - (hl & 0x40 ? 0 : 256));
}
sumf[row] += d * (acc[0] + 1.f/16.f * acc[1]);
q += step;
h += step;
s += step;
dh += step/2;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
}
y += 8 * QK_K;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
}
#endif
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
for (int row = 0; row < 2; ++row) {
const float tot = simd_sum(sumf[row]);
if (tiisg == 0) {
dst[r1*ne0 + r2*ne0*ne1 + first_row + row] = tot;
}
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
}
}
kernel void kernel_mul_mv_q6_K_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01[[buffer(4)]],
constant int64_t & ne02[[buffer(5)]],
constant int64_t & ne10[[buffer(9)]],
constant int64_t & ne12[[buffer(11)]],
constant int64_t & ne0[[buffer(15)]],
constant int64_t & ne1[[buffer(16)]],
constant uint & gqa[[buffer(17)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
const uint8_t kmask1 = 0x03;
const uint8_t kmask2 = 0x0C;
const uint8_t kmask3 = 0x30;
const uint8_t kmask4 = 0xC0;
const int nb = ne00/QK_K;
const int64_t r0 = tgpig.x;
const int64_t r1 = tgpig.y;
const int r2 = tgpig.z;
const int row = 2 * r0 + sgitg;
const uint offset0 = r2/gqa*(nb*ne0);
device const block_q6_K * x = (device const block_q6_K *) src0 + row * nb + offset0;
device const float * yy = (device const float *) src1 + r1*ne10 + r2*ne00*ne1;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
float sumf = 0;
#if QK_K == 256
const int tid = tiisg/2;
const int ix = tiisg%2;
const int ip = tid/8; // 0 or 1
const int il = tid%8;
const int n = 4;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
const int l0 = n*il;
const int is = 8*ip + l0/16;
const int y_offset = 128*ip + l0;
const int q_offset_l = 64*ip + l0;
const int q_offset_h = 32*ip + l0;
for (int i = ix; i < nb; i += 2) {
device const uint8_t * q1 = x[i].ql + q_offset_l;
device const uint8_t * q2 = q1 + 32;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
device const uint8_t * qh = x[i].qh + q_offset_h;
device const int8_t * sc = x[i].scales + is;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
device const float * y = yy + i * QK_K + y_offset;
const float dall = x[i].d;
float4 sums = {0.f, 0.f, 0.f, 0.f};
for (int l = 0; l < n; ++l) {
sums[0] += y[l+ 0] * ((int8_t)((q1[l] & 0xF) | ((qh[l] & kmask1) << 4)) - 32);
sums[1] += y[l+32] * ((int8_t)((q2[l] & 0xF) | ((qh[l] & kmask2) << 2)) - 32);
sums[2] += y[l+64] * ((int8_t)((q1[l] >> 4) | ((qh[l] & kmask3) << 0)) - 32);
sums[3] += y[l+96] * ((int8_t)((q2[l] >> 4) | ((qh[l] & kmask4) >> 2)) - 32);
}
sumf += dall * (sums[0] * sc[0] + sums[1] * sc[2] + sums[2] * sc[4] + sums[3] * sc[6]);
}
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
#else
const int ix = tiisg/4;
const int il = 4*(tiisg%4);
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
for (int i = ix; i < nb; i += 8) {
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
device const float * y = yy + i * QK_K + il;
device const uint8_t * ql = x[i].ql + il;
device const uint8_t * qh = x[i].qh + il;
device const int8_t * s = x[i].scales;
const float d = x[i].d;
float4 sums = {0.f, 0.f, 0.f, 0.f};
for (int l = 0; l < 4; ++l) {
sums[0] += y[l+ 0] * ((int8_t)((ql[l+ 0] & 0xF) | ((qh[l] & kmask1) << 4)) - 32);
sums[1] += y[l+16] * ((int8_t)((ql[l+16] & 0xF) | ((qh[l] & kmask2) << 2)) - 32);
sums[2] += y[l+32] * ((int8_t)((ql[l+ 0] >> 4) | ((qh[l] & kmask3) >> 0)) - 32);
sums[3] += y[l+48] * ((int8_t)((ql[l+16] >> 4) | ((qh[l] & kmask4) >> 2)) - 32);
}
sumf += d * (sums[0] * s[0] + sums[1] * s[1] + sums[2] * s[2] + sums[3] * s[3]);
}
#endif
const float tot = simd_sum(sumf);
if (tiisg == 0) {
dst[r1*ne0 + r2*ne0*ne1 + row] = tot;
}
}
//============================= templates and their specializations =============================
// NOTE: this is not dequantizing - we are simply fitting the template
template <typename type4x4>
void dequantize_f32(device const float4x4 * src, short il, thread type4x4 & reg) {
float4x4 temp = *(((device float4x4 *)src));
for (int i = 0; i < 16; i++){
reg[i/4][i%4] = temp[i/4][i%4];
}
}
template <typename type4x4>
void dequantize_f16(device const half4x4 * src, short il, thread type4x4 & reg) {
half4x4 temp = *(((device half4x4 *)src));
for (int i = 0; i < 16; i++){
reg[i/4][i%4] = temp[i/4][i%4];
}
}
template <typename type4x4>
void dequantize_q4_0(device const block_q4_0 *xb, short il, thread type4x4 & reg) {
device const uint16_t * qs = ((device const uint16_t *)xb + 1);
const float d1 = il ? (xb->d / 16.h) : xb->d;
const float d2 = d1 / 256.f;
const float md = -8.h * xb->d;
const ushort mask0 = il ? 0x00F0 : 0x000F;
const ushort mask1 = mask0 << 8;
for (int i=0;i<8;i++) {
reg[i/2][2*(i%2)+0] = d1 * (qs[i] & mask0) + md;
reg[i/2][2*(i%2)+1] = d2 * (qs[i] & mask1) + md;
}
}
template <typename type4x4>
void dequantize_q4_1(device const block_q4_1 *xb, short il, thread type4x4 & reg) {
device const uint16_t * qs = ((device const uint16_t *)xb + 2);
const float d1 = il ? (xb->d / 16.h) : xb->d;
const float d2 = d1 / 256.f;
const float m = xb->m;
const ushort mask0 = il ? 0x00F0 : 0x000F;
const ushort mask1 = mask0 << 8;
for (int i=0;i<8;i++) {
reg[i/2][2*(i%2)+0] = ((qs[i] & mask0) * d1) + m;
reg[i/2][2*(i%2)+1] = ((qs[i] & mask1) * d2) + m;
}
}
template <typename type4x4>
void dequantize_q5_0(device const block_q5_0 *xb, short il, thread type4x4 & reg) {
device const uint16_t * qs = ((device const uint16_t *)xb + 3);
const float d = xb->d;
const float md = -16.h * xb->d;
const ushort mask = il ? 0x00F0 : 0x000F;
const uint32_t qh = *((device const uint32_t *)xb->qh);
const int x_mv = il ? 4 : 0;
const int gh_mv = il ? 12 : 0;
const int gh_bk = il ? 0 : 4;
for (int i = 0; i < 8; i++) {
// extract the 5-th bits for x0 and x1
const uint8_t xh_0 = ((qh >> (gh_mv + 2*i )) << gh_bk) & 0x10;
const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
// combine the 4-bits from qs with the 5th bit
const int32_t x0 = ((((qs[i] ) & mask) >> x_mv) | xh_0);
const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
reg[i/2][2*(i%2)+0] = d * x0 + md;
reg[i/2][2*(i%2)+1] = d * x1 + md;
}
}
template <typename type4x4>
void dequantize_q5_1(device const block_q5_1 *xb, short il, thread type4x4 & reg) {
device const uint16_t * qs = ((device const uint16_t *)xb + 4);
const float d = xb->d;
const float m = xb->m;
const ushort mask = il ? 0x00F0 : 0x000F;
const uint32_t qh = *((device const uint32_t *)xb->qh);
const int x_mv = il ? 4 : 0;
const int gh_mv = il ? 12 : 0;
const int gh_bk = il ? 0 : 4;
for (int i = 0; i < 8; i++) {
// extract the 5-th bits for x0 and x1
const uint8_t xh_0 = ((qh >> (gh_mv + 2*i )) << gh_bk) & 0x10;
const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
// combine the 4-bits from qs with the 5th bit
const int32_t x0 = ((((qs[i] ) & mask) >> x_mv) | xh_0);
const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
reg[i/2][2*(i%2)+0] = d * x0 + m;
reg[i/2][2*(i%2)+1] = d * x1 + m;
}
}
template <typename type4x4>
void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg) {
device const int8_t * qs = ((device const int8_t *)xb->qs);
const half d = xb->d;
for (int i=0;i<16;i++) {
reg[i/4][i%4] = (qs[i + 16*il] * d);
}
}
template <typename type4x4>
void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg) {
const half d = xb->d;
const half min = xb->dmin;
device const uint8_t * q = (device const uint8_t *)xb->qs;
half dl, ml;
uint8_t sc = xb->scales[il];
#if QK_K == 256
q = q + 32*(il/8) + 16*(il&1);
il = (il/2)%4;
#endif
half coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
uchar mask = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3);
dl = d * (sc & 0xF) * coef, ml = min * (sc >> 4);
for (int i = 0; i < 16; ++i) {
reg[i/4][i%4] = dl * (q[i] & mask) - ml;
}
}
template <typename type4x4>
void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg) {
const half d_all = xb->d;
device const uint8_t * q = (device const uint8_t *)xb->qs;
device const uint8_t * h = (device const uint8_t *)xb->hmask;
device const int8_t * scales = (device const int8_t *)xb->scales;
#if QK_K == 256
q = q + 32 * (il/8) + 16 * (il&1);
h = h + 16 * (il&1);
uint8_t m = 1 << (il/2);
uint16_t kmask1 = (il/4)>1 ? ((il/4)>2 ? 192 : 48) : \
((il/4)>0 ? 12 : 3);
uint16_t kmask2 = il/8 ? 0xF0 : 0x0F;
uint16_t scale_2 = scales[il%8], scale_1 = scales[8 + il%4];
int16_t dl_int = (il/4)&1 ? (scale_2&kmask2) | ((scale_1&kmask1) << 2)
: (scale_2&kmask2) | ((scale_1&kmask1) << 4);
half dl = il<8 ? d_all * (dl_int - 32.h) : d_all * (dl_int / 16.h - 32.h);
const half ml = 4.h * dl;
il = (il/2) & 3;
const half coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
const uint8_t mask = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3);
dl *= coef;
for (int i = 0; i < 16; ++i) {
reg[i/4][i%4] = dl * (q[i] & mask) - (h[i] & m ? 0 : ml);
}
#else
float kcoef = il&1 ? 1.f/16.f : 1.f;
uint16_t kmask = il&1 ? 0xF0 : 0x0F;
float dl = d_all * ((scales[il/2] & kmask) * kcoef - 8);
float coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
uint8_t mask = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3);
uint8_t m = 1<<(il*2);
for (int i = 0; i < 16; ++i) {
reg[i/4][i%4] = coef * dl * ((q[i] & mask) - ((h[i%8] & (m * (1 + i/8))) ? 0 : 4.f/coef));
}
#endif
}
static inline uchar2 get_scale_min_k4_just2(int j, int k, device const uchar * q) {
return j < 4 ? uchar2{uchar(q[j+0+k] & 63), uchar(q[j+4+k] & 63)}
: uchar2{uchar((q[j+4+k] & 0xF) | ((q[j-4+k] & 0xc0) >> 2)), uchar((q[j+4+k] >> 4) | ((q[j-0+k] & 0xc0) >> 2))};
}
template <typename type4x4>
void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg) {
device const uchar * q = xb->qs;
#if QK_K == 256
short is = (il/4) * 2;
q = q + (il/4) * 32 + 16 * (il&1);
il = il & 3;
const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
const half d = il < 2 ? xb->d : xb->d / 16.h;
const half min = xb->dmin;
const half dl = d * sc[0];
const half ml = min * sc[1];
#else
q = q + 16 * (il&1);
device const uint8_t * s = xb->scales;
device const half2 * dh = (device const half2 *)xb->d;
const float2 d = (float2)dh[0];
const float dl = il<2 ? d[0] * (s[0]&0xF) : d[0] * (s[1]&0xF)/16.h;
const float ml = il<2 ? d[1] * (s[0]>>4) : d[1] * (s[1]>>4);
#endif
const ushort mask = il<2 ? 0x0F : 0xF0;
for (int i = 0; i < 16; ++i) {
reg[i/4][i%4] = dl * (q[i] & mask) - ml;
}
}
template <typename type4x4>
void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg) {
device const uint8_t * q = xb->qs;
device const uint8_t * qh = xb->qh;
#if QK_K == 256
short is = (il/4) * 2;
q = q + 32 * (il/4) + 16 * (il&1);
qh = qh + 16 * (il&1);
uint8_t ul = 1 << (il/2);
il = il & 3;
const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
const half d = il < 2 ? xb->d : xb->d / 16.h;
const half min = xb->dmin;
const half dl = d * sc[0];
const half ml = min * sc[1];
const ushort mask = il<2 ? 0x0F : 0xF0;
const half qh_val = il<2 ? 16.h : 256.h;
for (int i = 0; i < 16; ++i) {
reg[i/4][i%4] = dl * ((q[i] & mask) + (qh[i] & ul ? qh_val : 0)) - ml;
}
#else
q = q + 16 * (il&1);
device const int8_t * s = xb->scales;
const float dl = xb->d * s[il];
uint8_t m = 1<<(il*2);
const float coef = il<2 ? 1.f : 1.f/16.f;
const ushort mask = il<2 ? 0x0F : 0xF0;
for (int i = 0; i < 16; ++i) {
reg[i/4][i%4] = coef * dl * ((q[i] & mask) - (qh[i%8] & (m*(1+i/8)) ? 0.f : 16.f/coef));
}
#endif
}
template <typename type4x4>
void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg) {
const half d_all = xb->d;
device const uint8_t * ql = (device const uint8_t *)xb->ql;
device const uint8_t * qh = (device const uint8_t *)xb->qh;
device const int8_t * scales = (device const int8_t *)xb->scales;
#if QK_K == 256
ql = ql + 64*(il/8) + 32*((il/2)&1) + 16*(il&1);
qh = qh + 32*(il/8) + 16*(il&1);
half sc = scales[(il%2) + 2 * ((il/2))];
il = (il/2) & 3;
#else
ql = ql + 16 * (il&1);
half sc = scales[il];
#endif
const uint16_t kmask1 = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3);
const uint16_t kmask2 = il>1 ? 0xF0 : 0x0F;
const half coef = il>1 ? 1.f/16.h : 1.h;
const half ml = d_all * sc * 32.h;
const half dl = d_all * sc * coef;
for (int i = 0; i < 16; ++i) {
const half q = il&1 ? ((ql[i] & kmask2) | ((qh[i] & kmask1) << 2))
: ((ql[i] & kmask2) | ((qh[i] & kmask1) << 4));
reg[i/4][i%4] = dl * q - ml;
}
}
template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
kernel void kernel_get_rows(
device const void * src0,
device const int * src1,
device float * dst,
constant int64_t & ne00,
constant uint64_t & nb01,
constant uint64_t & nb1,
uint tgpig[[threadgroup_position_in_grid]],
uint tiitg[[thread_index_in_threadgroup]],
uint tptg[[threads_per_threadgroup]]) {
const int i = tgpig;
const int r = ((device int32_t *) src1)[i];
for (int ind = tiitg; ind < ne00/16; ind += tptg) {
float4x4 temp;
dequantize_func(
((device const block_q *) ((device char *) src0 + r*nb01)) + ind/nl, ind%nl, temp);
*(((device float4x4 *) ((device char *) dst + i*nb1)) + ind) = temp;
}
}
#define BLOCK_SIZE_M 64 // 8 simdgroup matrices from matrix A
#define BLOCK_SIZE_N 32 // 4 simdgroup matrices from matrix B
#define BLOCK_SIZE_K 32
#define THREAD_MAT_M 4 // each thread take 4 simdgroup matrices from matrix A
#define THREAD_MAT_N 2 // each thread take 2 simdgroup matrices from matrix B
#define THREAD_PER_BLOCK 128
#define THREAD_PER_ROW 2 // 2 thread for each row in matrix A to load numbers
#define THREAD_PER_COL 4 // 4 thread for each row in matrix B to load numbers
#define SG_MAT_SIZE 64 // simdgroup matrix is of shape 8x8
#define SG_MAT_ROW 8
// each block_q contains 16*nl weights
template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
kernel void kernel_mul_mm(device const uchar * src0,
device const uchar * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne02,
constant int64_t & nb01,
constant int64_t & nb02,
constant int64_t & ne12,
constant int64_t & nb10,
constant int64_t & nb11,
constant int64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & gqa,
threadgroup uchar * shared_memory [[threadgroup(0)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiitg[[thread_index_in_threadgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
threadgroup half * sa = (threadgroup half *)(shared_memory);
threadgroup float * sb = (threadgroup float *)(shared_memory + 4096);
const uint r0 = tgpig.y;
const uint r1 = tgpig.x;
const uint im = tgpig.z;
// if this block is of 64x32 shape or smaller
short n_rows = (ne0 - r0 * BLOCK_SIZE_M < BLOCK_SIZE_M) ? (ne0 - r0 * BLOCK_SIZE_M) : BLOCK_SIZE_M;
short n_cols = (ne1 - r1 * BLOCK_SIZE_N < BLOCK_SIZE_N) ? (ne1 - r1 * BLOCK_SIZE_N) : BLOCK_SIZE_N;
// a thread shouldn't load data outside of the matrix
short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
simdgroup_half8x8 ma[4];
simdgroup_float8x8 mb[2];
simdgroup_float8x8 c_res[8];
for (int i = 0; i < 8; i++){
c_res[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
}
short il = (tiitg % THREAD_PER_ROW);
uint offset0 = im/gqa*nb02;
ushort offset1 = il/nl;
device const block_q * x = (device const block_q *)(src0 + (r0 * BLOCK_SIZE_M + thread_row) * nb01 + offset0) + offset1;
device const float * y = (device const float *)(src1
+ nb12 * im
+ nb11 * (r1 * BLOCK_SIZE_N + thread_col)
+ nb10 * (BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL)));
for (int loop_k = 0; loop_k < ne00; loop_k += BLOCK_SIZE_K) {
// load data and store to threadgroup memory
half4x4 temp_a;
dequantize_func(x, il, temp_a);
threadgroup_barrier(mem_flags::mem_threadgroup);
#pragma unroll(16)
for (int i = 0; i < 16; i++) {
*(sa + SG_MAT_SIZE * ((tiitg / THREAD_PER_ROW / 8) \
+ (tiitg % THREAD_PER_ROW) * 16 + (i / 8) * 8) \
+ (tiitg / THREAD_PER_ROW) % 8 + (i & 7) * 8) = temp_a[i/4][i%4];
}
*(threadgroup float2x4 *)(sb + (tiitg % THREAD_PER_COL) * 8 * 32 + 8 * (tiitg / THREAD_PER_COL)) = *((device float2x4 *)y);
il = (il + 2 < nl) ? il + 2 : il % 2;
x = (il < 2) ? x + (2+nl-1)/nl : x;
y += BLOCK_SIZE_K;
threadgroup_barrier(mem_flags::mem_threadgroup);
// load matrices from threadgroup memory and conduct outer products
threadgroup half * lsma = (sa + THREAD_MAT_M * SG_MAT_SIZE * (sgitg % 2));
threadgroup float * lsmb = (sb + THREAD_MAT_N * SG_MAT_SIZE * (sgitg / 2));
#pragma unroll(4)
for (int ik = 0; ik < BLOCK_SIZE_K / 8; ik++) {
#pragma unroll(4)
for (int i = 0; i < 4; i++) {
simdgroup_load(ma[i],lsma + SG_MAT_SIZE * i);
}
simdgroup_barrier(mem_flags::mem_none);
#pragma unroll(2)
for (int i = 0; i < 2; i++) {
simdgroup_load(mb[i],lsmb + SG_MAT_SIZE * i);
}
lsma += BLOCK_SIZE_M / SG_MAT_ROW * SG_MAT_SIZE;
lsmb += BLOCK_SIZE_N / SG_MAT_ROW * SG_MAT_SIZE;
#pragma unroll(8)
for (int i = 0; i < 8; i++){
simdgroup_multiply_accumulate(c_res[i], mb[i/4], ma[i%4], c_res[i]);
}
}
}
if ((r0 + 1) * BLOCK_SIZE_M <= ne0 && (r1 + 1) * BLOCK_SIZE_N <= ne1) {
device float * C = dst + (BLOCK_SIZE_M * r0 + 32 * (sgitg & 1)) \
+ (BLOCK_SIZE_N * r1 + 16 * (sgitg >> 1)) * ne0 + im*ne1*ne0;
for (int i = 0; i < 8; i++) {
simdgroup_store(c_res[i], C + 8 * (i%4) + 8 * ne0 * (i/4), ne0);
}
} else {
// block is smaller than 64x32, we should avoid writing data outside of the matrix
threadgroup_barrier(mem_flags::mem_threadgroup);
threadgroup float * temp_str = ((threadgroup float *)shared_memory) \
+ 32 * (sgitg&1) + (16 * (sgitg>>1)) * BLOCK_SIZE_M;
for (int i = 0; i < 8; i++) {
simdgroup_store(c_res[i], temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
}
threadgroup_barrier(mem_flags::mem_threadgroup);
device float * C = dst + (BLOCK_SIZE_M * r0) + (BLOCK_SIZE_N * r1) * ne0 + im*ne1*ne0;
if (sgitg == 0) {
for (int i = 0; i < n_rows; i++) {
for (int j = tiitg; j < n_cols; j += BLOCK_SIZE_N) {
*(C + i + j * ne0) = *(temp_str + i + j * BLOCK_SIZE_M);
}
}
}
}
}
#if QK_K == 256
#define QK_NL 16
#else
#define QK_NL 4
#endif
typedef void (get_rows_t)(device const void *, device const int *, device float *, constant int64_t &, \
constant uint64_t &, constant uint64_t &, uint, uint, uint);
template [[host_name("kernel_get_rows_f32")]] kernel get_rows_t kernel_get_rows<float4x4, 1, dequantize_f32>;
template [[host_name("kernel_get_rows_f16")]] kernel get_rows_t kernel_get_rows<half4x4, 1, dequantize_f16>;
template [[host_name("kernel_get_rows_q4_0")]] kernel get_rows_t kernel_get_rows<block_q4_0, 2, dequantize_q4_0>;
template [[host_name("kernel_get_rows_q4_1")]] kernel get_rows_t kernel_get_rows<block_q4_1, 2, dequantize_q4_1>;
template [[host_name("kernel_get_rows_q5_0")]] kernel get_rows_t kernel_get_rows<block_q5_0, 2, dequantize_q5_0>;
template [[host_name("kernel_get_rows_q5_1")]] kernel get_rows_t kernel_get_rows<block_q5_1, 2, dequantize_q5_1>;
template [[host_name("kernel_get_rows_q8_0")]] kernel get_rows_t kernel_get_rows<block_q8_0, 2, dequantize_q8_0>;
template [[host_name("kernel_get_rows_q2_K")]] kernel get_rows_t kernel_get_rows<block_q2_K, QK_NL, dequantize_q2_K>;
template [[host_name("kernel_get_rows_q3_K")]] kernel get_rows_t kernel_get_rows<block_q3_K, QK_NL, dequantize_q3_K>;
template [[host_name("kernel_get_rows_q4_K")]] kernel get_rows_t kernel_get_rows<block_q4_K, QK_NL, dequantize_q4_K>;
template [[host_name("kernel_get_rows_q5_K")]] kernel get_rows_t kernel_get_rows<block_q5_K, QK_NL, dequantize_q5_K>;
template [[host_name("kernel_get_rows_q6_K")]] kernel get_rows_t kernel_get_rows<block_q6_K, QK_NL, dequantize_q6_K>;
typedef void (mat_mm_t)(
device const uchar * src0,
device const uchar * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne02,
constant int64_t & nb01,
constant int64_t & nb02,
constant int64_t & ne12,
constant int64_t & nb10,
constant int64_t & nb11,
constant int64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & gqa,
threadgroup uchar *, uint3, uint, uint);
template [[host_name("kernel_mul_mm_f32_f32")]] kernel mat_mm_t kernel_mul_mm<float4x4, 1, dequantize_f32>;
template [[host_name("kernel_mul_mm_f16_f32")]] kernel mat_mm_t kernel_mul_mm<half4x4, 1, dequantize_f16>;
template [[host_name("kernel_mul_mm_q4_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_0, 2, dequantize_q4_0>;
template [[host_name("kernel_mul_mm_q4_1_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_1, 2, dequantize_q4_1>;
template [[host_name("kernel_mul_mm_q5_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q5_0, 2, dequantize_q5_0>;
template [[host_name("kernel_mul_mm_q5_1_f32")]] kernel mat_mm_t kernel_mul_mm<block_q5_1, 2, dequantize_q5_1>;
template [[host_name("kernel_mul_mm_q8_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q8_0, 2, dequantize_q8_0>;
template [[host_name("kernel_mul_mm_q2_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q2_K, QK_NL, dequantize_q2_K>;
template [[host_name("kernel_mul_mm_q3_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q3_K, QK_NL, dequantize_q3_K>;
template [[host_name("kernel_mul_mm_q4_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_K, QK_NL, dequantize_q4_K>;
template [[host_name("kernel_mul_mm_q5_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q5_K, QK_NL, dequantize_q5_K>;
template [[host_name("kernel_mul_mm_q6_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q6_K, QK_NL, dequantize_q6_K>;