llama.cpp/ggml-metal.m

812 lines
36 KiB
Mathematica
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
#import "ggml-metal.h"
#import "ggml.h"
#import <Foundation/Foundation.h>
#import <Metal/Metal.h>
#import <MetalPerformanceShaders/MetalPerformanceShaders.h>
#ifdef GGML_METAL_NDEBUG
#define metal_printf(...)
#else
#define metal_printf(...) fprintf(stderr, __VA_ARGS__)
#endif
#define UNUSED(x) (void)(x)
struct ggml_metal_buffer {
const char * name;
void * data;
size_t size;
id<MTLBuffer> metal;
};
struct ggml_metal_context {
float * logits;
id<MTLDevice> device;
id<MTLCommandQueue> queue;
id<MTLLibrary> library;
int n_buffers;
struct ggml_metal_buffer buffers[GGML_METAL_MAX_BUFFERS];
// custom kernels
#define GGML_METAL_DECL_KERNEL(name) \
id<MTLFunction> function_##name; \
id<MTLComputePipelineState> pipeline_##name
GGML_METAL_DECL_KERNEL(add);
GGML_METAL_DECL_KERNEL(mul);
GGML_METAL_DECL_KERNEL(mul_row); // TODO: avoid this extra kernel, instead extend the "mul" kernel to support broadcast
GGML_METAL_DECL_KERNEL(scale);
GGML_METAL_DECL_KERNEL(silu);
GGML_METAL_DECL_KERNEL(relu);
GGML_METAL_DECL_KERNEL(gelu);
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
GGML_METAL_DECL_KERNEL(soft_max);
GGML_METAL_DECL_KERNEL(diag_mask_inf);
2023-06-06 19:16:57 +02:00
GGML_METAL_DECL_KERNEL(get_rows_f16);
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
GGML_METAL_DECL_KERNEL(get_rows_q4_0);
GGML_METAL_DECL_KERNEL(get_rows_q4_1);
GGML_METAL_DECL_KERNEL(get_rows_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
GGML_METAL_DECL_KERNEL(get_rows_q3_k);
GGML_METAL_DECL_KERNEL(get_rows_q4_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
GGML_METAL_DECL_KERNEL(get_rows_q5_k);
GGML_METAL_DECL_KERNEL(get_rows_q6_k);
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
GGML_METAL_DECL_KERNEL(rms_norm);
GGML_METAL_DECL_KERNEL(mul_mat_f16_f32);
2023-06-06 19:16:57 +02:00
GGML_METAL_DECL_KERNEL(mul_mat_q4_0_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q4_1_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q2_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
GGML_METAL_DECL_KERNEL(mul_mat_q3_k_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q4_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
GGML_METAL_DECL_KERNEL(mul_mat_q5_k_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q6_k_f32);
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
GGML_METAL_DECL_KERNEL(rope);
GGML_METAL_DECL_KERNEL(cpy_f32_f16);
GGML_METAL_DECL_KERNEL(cpy_f32_f32);
#undef GGML_METAL_DECL_KERNEL
};
// MSL code
// TODO: move the contents here when ready
// for now it is easier to work in a separate file
static NSString * const msl_library_source = @"see metal.metal";
// Here to assist with NSBundle Path Hack
@interface GGMLMetalClass : NSObject
@end
@implementation GGMLMetalClass
@end
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
struct ggml_metal_context * ggml_metal_init(void) {
fprintf(stderr, "%s: allocating\n", __func__);
struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
ctx->device = MTLCreateSystemDefaultDevice();
ctx->queue = [ctx->device newCommandQueue];
ctx->n_buffers = 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
// determine if we can use MPS
if (MPSSupportsMTLDevice(ctx->device)) {
fprintf(stderr, "%s: using MPS\n", __func__);
} else {
fprintf(stderr, "%s: not using MPS\n", __func__);
GGML_ASSERT(false && "MPS not supported");
}
#if 0
// compile from source string and show compile log
{
NSError * error = nil;
ctx->library = [ctx->device newLibraryWithSource:msl_library_source options:nil error:&error];
if (error) {
fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
exit(1);
}
}
#else
UNUSED(msl_library_source);
// read the source from "ggml-metal.metal" into a string and use newLibraryWithSource
{
NSError * error = nil;
//NSString * path = [[NSBundle mainBundle] pathForResource:@"../../examples/metal/metal" ofType:@"metal"];
NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
NSString * path = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
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
fprintf(stderr, "%s: loading '%s'\n", __func__, [path UTF8String]);
NSString * src = [NSString stringWithContentsOfFile:path encoding:NSUTF8StringEncoding error:&error];
if (error) {
fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
exit(1);
}
ctx->library = [ctx->device newLibraryWithSource:src options:nil error:&error];
if (error) {
fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
exit(1);
}
}
#endif
// load kernels
{
#define GGML_METAL_ADD_KERNEL(name) \
ctx->function_##name = [ctx->library newFunctionWithName:@"kernel_"#name]; \
ctx->pipeline_##name = [ctx->device newComputePipelineStateWithFunction:ctx->function_##name error:nil]; \
fprintf(stderr, "%s: loaded %-32s %16p\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name);
GGML_METAL_ADD_KERNEL(add);
GGML_METAL_ADD_KERNEL(mul);
GGML_METAL_ADD_KERNEL(mul_row);
GGML_METAL_ADD_KERNEL(scale);
GGML_METAL_ADD_KERNEL(silu);
GGML_METAL_ADD_KERNEL(relu);
GGML_METAL_ADD_KERNEL(gelu);
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
GGML_METAL_ADD_KERNEL(soft_max);
GGML_METAL_ADD_KERNEL(diag_mask_inf);
2023-06-06 19:16:57 +02:00
GGML_METAL_ADD_KERNEL(get_rows_f16);
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
GGML_METAL_ADD_KERNEL(get_rows_q4_0);
GGML_METAL_ADD_KERNEL(get_rows_q4_1);
GGML_METAL_ADD_KERNEL(get_rows_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
GGML_METAL_ADD_KERNEL(get_rows_q3_k);
GGML_METAL_ADD_KERNEL(get_rows_q4_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
GGML_METAL_ADD_KERNEL(get_rows_q5_k);
GGML_METAL_ADD_KERNEL(get_rows_q6_k);
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
GGML_METAL_ADD_KERNEL(rms_norm);
GGML_METAL_ADD_KERNEL(mul_mat_f16_f32);
2023-06-06 19:16:57 +02:00
GGML_METAL_ADD_KERNEL(mul_mat_q4_0_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q4_1_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q2_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
GGML_METAL_ADD_KERNEL(mul_mat_q3_k_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q4_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
GGML_METAL_ADD_KERNEL(mul_mat_q5_k_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q6_k_f32);
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
GGML_METAL_ADD_KERNEL(rope);
GGML_METAL_ADD_KERNEL(cpy_f32_f16);
GGML_METAL_ADD_KERNEL(cpy_f32_f32);
#undef GGML_METAL_ADD_KERNEL
}
return ctx;
}
void ggml_metal_free(struct ggml_metal_context * ctx) {
fprintf(stderr, "%s: deallocating\n", __func__);
free(ctx);
}
// finds the Metal buffer that contains the tensor data on the GPU device
// the assumption is that there is 1-to-1 mapping between the host and device memory buffers, so we can find the
// Metal buffer based on the host memory pointer
//
static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_metal_context * ctx, struct ggml_tensor * t, size_t * offs) {
//fprintf(stderr, "%s: data tensor '%16s', offs_data = %8ld, offs_eval = %8ld, offs_cach = %8ld\n", __func__, t->name, offs_data, offs_eval, offs_cach);
for (int i = 0; i < ctx->n_buffers; ++i) {
const int64_t ioffs = (int64_t) t->data - (int64_t) ctx->buffers[i].data;
if (ioffs >= 0 && ioffs < (int64_t) ctx->buffers[i].size) {
*offs = (size_t) ioffs;
//fprintf(stderr, "%s: '%s' tensor '%16s', offs = %8ld\n", __func__, ctx->buffers[i].name, t->name, *offs);
return ctx->buffers[i].metal;
}
}
fprintf(stderr, "%s: error: buffer is nil\n", __func__);
return nil;
}
bool ggml_metal_add_buffer(
struct ggml_metal_context * ctx,
const char * name,
void * data,
size_t size) {
if (ctx->n_buffers >= GGML_METAL_MAX_BUFFERS) {
fprintf(stderr, "%s: too many buffers\n", __func__);
return false;
}
if (data) {
// verify that the buffer does not overlap with any of the existing buffers
for (int i = 0; i < ctx->n_buffers; ++i) {
const int64_t ioffs = (int64_t) data - (int64_t) ctx->buffers[i].data;
if (ioffs >= 0 && ioffs < (int64_t) ctx->buffers[i].size) {
fprintf(stderr, "%s: error: buffer '%s' overlaps with '%s'\n", __func__, name, ctx->buffers[i].name);
return false;
}
}
size_t page_size = getpagesize();
size_t aligned_size = size;
if ((aligned_size % page_size) != 0) {
aligned_size += (page_size - (aligned_size % page_size));
}
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
ctx->buffers[ctx->n_buffers].name = name;
ctx->buffers[ctx->n_buffers].data = data;
ctx->buffers[ctx->n_buffers].size = size;
if (ctx->device.maxBufferLength < aligned_size) {
fprintf(stderr, "%s: buffer '%s' size %zu is larger than buffer maximum of %zu\n", __func__, name, aligned_size, ctx->device.maxBufferLength);
return false;
}
ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:data length:aligned_size options:MTLResourceStorageModeShared deallocator:nil];
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 (ctx->buffers[ctx->n_buffers].metal == nil) {
fprintf(stderr, "%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, aligned_size / 1024.0 / 1024.0);
return false;
} else {
fprintf(stderr, "%s: allocated '%-16s' buffer, size = %8.2f MB\n", __func__, name, aligned_size / 1024.0 / 1024.0);
}
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
++ctx->n_buffers;
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
}
return true;
}
void ggml_metal_set_tensor(
struct ggml_metal_context * ctx,
struct ggml_tensor * t) {
metal_printf("%s: set input for tensor '%s'\n", __func__, t->name);
size_t offs;
id<MTLBuffer> id_dst = ggml_metal_get_buffer(ctx, t, &offs);
memcpy((void *) ((uint8_t *) id_dst.contents + offs), t->data, ggml_nbytes(t));
}
void ggml_metal_get_tensor(
struct ggml_metal_context * ctx,
struct ggml_tensor * t) {
metal_printf("%s: extract results for tensor '%s'\n", __func__, t->name);
size_t offs;
id<MTLBuffer> id_src = ggml_metal_get_buffer(ctx, t, &offs);
memcpy(t->data, (void *) ((uint8_t *) id_src.contents + offs), ggml_nbytes(t));
}
void ggml_metal_graph_compute(
struct ggml_metal_context * ctx,
struct ggml_cgraph * gf) {
metal_printf("%s: evaluating graph\n", __func__);
size_t offs_src0 = 0;
size_t offs_src1 = 0;
size_t offs_dst = 0;
id<MTLCommandBuffer> command_buffer = [ctx->queue commandBuffer];
id<MTLComputeCommandEncoder> encoder = nil;
for (int i = 0; i < gf->n_nodes; ++i) {
//metal_printf("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
struct ggml_tensor * src0 = gf->nodes[i]->src0;
struct ggml_tensor * src1 = gf->nodes[i]->src1;
struct ggml_tensor * dst = gf->nodes[i];
const int64_t ne00 = src0 ? src0->ne[0] : 0;
const int64_t ne01 = src0 ? src0->ne[1] : 0;
const int64_t ne02 = src0 ? src0->ne[2] : 0;
const int64_t ne03 = src0 ? src0->ne[3] : 0;
const uint64_t nb00 = src0 ? src0->nb[0] : 0;
const uint64_t nb01 = src0 ? src0->nb[1] : 0;
const uint64_t nb02 = src0 ? src0->nb[2] : 0;
const uint64_t nb03 = src0 ? src0->nb[3] : 0;
const int64_t ne10 = src1 ? src1->ne[0] : 0;
const int64_t ne11 = src1 ? src1->ne[1] : 0;
const int64_t ne12 = src1 ? src1->ne[2] : 0;
const int64_t ne13 = src1 ? src1->ne[3] : 0; UNUSED(ne13);
const uint64_t nb10 = src1 ? src1->nb[0] : 0;
const uint64_t nb11 = src1 ? src1->nb[1] : 0;
const uint64_t nb12 = src1 ? src1->nb[2] : 0;
const uint64_t nb13 = src1 ? src1->nb[3] : 0; UNUSED(nb13);
const int64_t ne0 = dst ? dst->ne[0] : 0;
const int64_t ne1 = dst ? dst->ne[1] : 0;
const int64_t ne2 = dst ? dst->ne[2] : 0;
const int64_t ne3 = dst ? dst->ne[3] : 0;
const uint64_t nb0 = dst ? dst->nb[0] : 0;
const uint64_t nb1 = dst ? dst->nb[1] : 0;
const uint64_t nb2 = dst ? dst->nb[2] : 0;
const uint64_t nb3 = dst ? dst->nb[3] : 0;
const enum ggml_type src0t = src0 ? src0->type : GGML_TYPE_COUNT;
const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
const enum ggml_type dstt = dst ? dst->type : GGML_TYPE_COUNT;
id<MTLBuffer> id_src0 = src0 ? ggml_metal_get_buffer(ctx, src0, &offs_src0) : nil;
id<MTLBuffer> id_src1 = src1 ? ggml_metal_get_buffer(ctx, src1, &offs_src1) : nil;
id<MTLBuffer> id_dst = dst ? ggml_metal_get_buffer(ctx, dst, &offs_dst) : nil;
//metal_printf("%s: op - %s\n", __func__, ggml_op_name(dst->op));
//if (src0) {
// metal_printf("%s: src0 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src0t), ne00, ne01, ne02,
// ggml_is_contiguous(src0), src0->name);
//}
//if (src1) {
// metal_printf("%s: src1 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src1t), ne10, ne11, ne12,
// ggml_is_contiguous(src1), src1->name);
//}
//if (dst) {
// metal_printf("%s: dst - %4s [%5lld, %5lld, %5lld], 1, %s\n", __func__, ggml_type_name(dstt), ne0, ne1, ne2,
// dst->name);
//}
switch (dst->op) {
case GGML_OP_RESHAPE:
case GGML_OP_VIEW:
case GGML_OP_TRANSPOSE:
case GGML_OP_PERMUTE:
{
// noop
} break;
case GGML_OP_ADD:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoder];
}
[encoder setComputePipelineState:ctx->pipeline_add];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
const int64_t n = ggml_nelements(dst);
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_OP_MUL:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoder];
}
if (ggml_nelements(src1) == ne10) {
// src1 is a row
[encoder setComputePipelineState:ctx->pipeline_mul_row];
} else {
[encoder setComputePipelineState:ctx->pipeline_mul];
}
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
const int64_t n = ggml_nelements(dst);
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_OP_SCALE:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoder];
}
const float scale = *(const float *) src1->data;
[encoder setComputePipelineState:ctx->pipeline_scale];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&scale length:sizeof(scale) atIndex:2];
const int64_t n = ggml_nelements(dst);
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_OP_SILU:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoder];
}
[encoder setComputePipelineState:ctx->pipeline_silu];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
const int64_t n = ggml_nelements(dst);
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_OP_RELU:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoder];
}
[encoder setComputePipelineState:ctx->pipeline_relu];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
const int64_t n = ggml_nelements(dst);
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_OP_GELU:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoder];
}
[encoder setComputePipelineState:ctx->pipeline_gelu];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
const int64_t n = ggml_nelements(dst);
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
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
case GGML_OP_SOFT_MAX:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoder];
}
const int nth = 32;
[encoder setComputePipelineState:ctx->pipeline_soft_max];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
[encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_DIAG_MASK_INF:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoder];
}
const int n_past = ((int32_t *)(src1->data))[0];
[encoder setComputePipelineState:ctx->pipeline_diag_mask_inf];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
[encoder setBytes:&n_past length:sizeof(int) atIndex:4];
[encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_OP_MUL_MAT:
{
// TODO: needs to be updated after PR: https://github.com/ggerganov/ggml/pull/224
GGML_ASSERT(ne00 == ne10);
GGML_ASSERT(ne02 == ne12);
if (ggml_is_contiguous(src0) &&
ggml_is_contiguous(src1) &&
(src0t == GGML_TYPE_F32 || src0t == GGML_TYPE_F16) && ne11 > 1) {
if (encoder != nil) {
[encoder endEncoding];
encoder = nil;
}
MPSDataType src0dt = src0t == GGML_TYPE_F32 ? MPSDataTypeFloat32 : MPSDataTypeFloat16;
MPSDataType src1dt = src1t == GGML_TYPE_F32 ? MPSDataTypeFloat32 : MPSDataTypeFloat16;
// for F32 x F32 we use MPS
MPSMatrixDescriptor * desc0 = [MPSMatrixDescriptor
matrixDescriptorWithRows:ne01 columns:ne00 rowBytes:src0->nb[1] dataType:src0dt];
MPSMatrixDescriptor * desc1 = [MPSMatrixDescriptor
matrixDescriptorWithRows:ne11 columns:ne10 rowBytes:src1->nb[1] dataType:src1dt];
MPSMatrixDescriptor * desc = [MPSMatrixDescriptor
matrixDescriptorWithRows:ne1 columns:ne0 rowBytes:dst->nb[1] dataType:MPSDataTypeFloat32];
MPSMatrixMultiplication * mul = [[MPSMatrixMultiplication alloc]
initWithDevice:ctx->device transposeLeft:false transposeRight:true
resultRows:ne11 resultColumns:ne01 interiorColumns:ne00 alpha:1.0 beta:0.0];
// we need to do ne02 multiplications
// TODO: is there a way to do this in parallel - currently very slow ..
// TODO: might be possible to offload part of the computation to ANE using Accelerate's CBLAS
for (int64_t i02 = 0; i02 < ne02; ++i02) {
size_t offs_src0_cur = offs_src0 + i02*nb02;
size_t offs_src1_cur = offs_src1 + i02*nb12;
size_t offs_dst_cur = offs_dst + i02*nb2;
MPSMatrix * mat_src0 = [[MPSMatrix alloc] initWithBuffer:id_src0 offset:offs_src0_cur descriptor:desc0];
MPSMatrix * mat_src1 = [[MPSMatrix alloc] initWithBuffer:id_src1 offset:offs_src1_cur descriptor:desc1];
MPSMatrix * mat_dst = [[MPSMatrix alloc] initWithBuffer:id_dst offset:offs_dst_cur descriptor:desc ];
[mul encodeToCommandBuffer:command_buffer leftMatrix:mat_src1 rightMatrix:mat_src0 resultMatrix:mat_dst];
}
} else {
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoder];
}
int nth0 = 32;
int nth1 = 1;
// use custom matrix x vector kernel
switch (src0t) {
2023-06-06 19:16:57 +02:00
case GGML_TYPE_F16:
{
GGML_ASSERT(ne02 == ne12);
nth0 = 64;
nth1 = 1;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32];
} break;
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
case GGML_TYPE_Q4_0:
{
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);
nth0 = 8;
nth1 = 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
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_0_f32];
} break;
case GGML_TYPE_Q4_1:
{
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);
nth0 = 8;
nth1 = 8;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_1_f32];
} break;
case GGML_TYPE_Q2_K:
{
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);
nth0 = 4;
nth1 = 16;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q2_k_f32];
} break;
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
case GGML_TYPE_Q3_K:
{
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);
nth0 = 4;
nth1 = 16;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q3_k_f32];
} break;
case GGML_TYPE_Q4_K:
{
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);
nth0 = 4;
nth1 = 16;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_k_f32];
} break;
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
case GGML_TYPE_Q5_K:
{
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);
nth0 = 4;
nth1 = 16;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q5_k_f32];
} break;
case GGML_TYPE_Q6_K:
{
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);
nth0 = 4;
nth1 = 16;
[encoder setComputePipelineState:ctx->pipeline_mul_mat_q6_k_f32];
} break;
default:
{
fprintf(stderr, "Asserting on type %d\n",(int)src0t);
GGML_ASSERT(false && "not implemented");
}
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
};
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
[encoder setBytes:&nb00 length:sizeof(nb00) atIndex:5];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:6];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:7];
[encoder setBytes:&ne10 length:sizeof(ne10) atIndex:8];
[encoder setBytes:&ne11 length:sizeof(ne11) atIndex:9];
[encoder setBytes:&nb10 length:sizeof(nb10) atIndex:10];
[encoder setBytes:&nb11 length:sizeof(nb11) atIndex:11];
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:12];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:13];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:14];
if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_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
[encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
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
}
else if (src0t == GGML_TYPE_Q2_K ||
src0t == GGML_TYPE_Q3_K ||
src0t == GGML_TYPE_Q4_K ||
src0t == GGML_TYPE_Q5_K ||
src0t == GGML_TYPE_Q6_K) {
[encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 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
} else {
[encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
}
} break;
case GGML_OP_GET_ROWS:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoder];
}
switch (src0->type) {
2023-06-06 19:16:57 +02:00
case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_get_rows_f16]; break;
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
case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_0]; break;
case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_1]; break;
case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q2_k]; break;
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
case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q3_k]; break;
case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_k]; break;
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
case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q5_k]; break;
case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q6_k]; break;
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
default: GGML_ASSERT(false && "not implemented");
}
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&(src0->ne[0]) length:sizeof( int64_t) atIndex:3];
[encoder setBytes:&(src0->nb[1]) length:sizeof(uint64_t) atIndex:4];
[encoder setBytes:&(dst->nb[1]) length:sizeof(uint64_t) atIndex:5];
const int64_t n = ggml_nelements(src1);
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_OP_RMS_NORM:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoder];
}
const float eps = 1e-6f;
const int nth = 256;
[encoder setComputePipelineState:ctx->pipeline_rms_norm];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
[encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3];
[encoder setBytes:&eps length:sizeof( float) atIndex:4];
[encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0];
const int64_t nrows = ggml_nrows(src0);
[encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_ROPE:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoder];
}
const int n_dims = ((int32_t *) src1->data)[1];
const int mode = ((int32_t *) src1->data)[2];
const int n_past = ((int32_t *)(src1->data))[0];
[encoder setComputePipelineState:ctx->pipeline_rope];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
[encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3];
[encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4];
[encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5];
[encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6];
[encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7];
[encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8];
[encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9];
[encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:10];
[encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:11];
[encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:12];
[encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:13];
[encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:14];
[encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:15];
[encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:16];
[encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:17];
[encoder setBytes:&n_past length:sizeof( int) atIndex:18];
[encoder setBytes:&n_dims length:sizeof( int) atIndex:19];
[encoder setBytes:&mode length:sizeof( int) atIndex:20];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_OP_CPY:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoder];
}
const int nth = 32;
switch (src0t) {
case GGML_TYPE_F32:
{
switch (dstt) {
case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_f16]; break;
case GGML_TYPE_F32: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_f32]; break;
default: GGML_ASSERT(false && "not implemented");
};
} break;
default: GGML_ASSERT(false && "not implemented");
}
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
[encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3];
[encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4];
[encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5];
[encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6];
[encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7];
[encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8];
[encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9];
[encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:10];
[encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:11];
[encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:12];
[encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:13];
[encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:14];
[encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:15];
[encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:16];
[encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:17];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
default:
fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
GGML_ASSERT(false);
}
}
if (encoder != nil) {
[encoder endEncoding];
encoder = nil;
}
[command_buffer commit];
[command_buffer waitUntilCompleted];
{
const double time_elapsed = [command_buffer GPUEndTime] - [command_buffer GPUStartTime];
UNUSED(time_elapsed);
metal_printf("%s: time elapsed = %f ms\n", __func__, time_elapsed * 1000.0);
}
}