haraldwolff/llama.cpp
1
0
Fork 0
Code Issues Pull requests Packages Projects Releases Wiki Activity

llama : make tensor_split ptr instead of array (#2272)

Browse source
This commit is contained in:
Georgi Gerganov 2023-07-21 13:10:51 +03:00 committed by GitHub
parent 54e3bc76fe
commit ae178ab46b
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
4 changed files with 8 additions and 4 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
examples/common.cpp
View file

@ -586,7 +586,7 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
lparams.n_batch = params.n_batch;
lparams.n_gpu_layers = params.n_gpu_layers;
lparams.main_gpu = params.main_gpu;
memcpy(lparams.tensor_split, params.tensor_split, LLAMA_MAX_DEVICES*sizeof(float));
lparams.tensor_split = params.tensor_split;
lparams.low_vram = params.low_vram;
lparams.seed = params.seed;
lparams.f16_kv = params.memory_f16;

3
ggml-cuda.cu
View file

@ -2512,6 +2512,9 @@ void ggml_init_cublas() {
}
void ggml_cuda_set_tensor_split(const float * tensor_split) {
if (tensor_split == nullptr) {
return;
}
bool all_zero = true;
for (int i = 0; i < g_device_count; ++i) {
if (tensor_split[i] != 0.0f) {

4
llama.cpp
View file

@ -849,7 +849,7 @@ struct llama_context_params llama_context_default_params() {
/*.n_batch =*/ 512,
/*.gpu_layers =*/ 0,
/*.main_gpu =*/ 0,
/*.tensor_split =*/ {0},
/*.tensor_split =*/ nullptr,
/*.rope_freq_base =*/ 10000.0f,
/*.rope_freq_scale =*/ 1.0f,
/*.progress_callback =*/ nullptr,
@ -1289,7 +1289,7 @@ static bool llama_model_load(
int n_batch,
int n_gpu_layers,
int main_gpu,
float * tensor_split,
const float * tensor_split,
float rope_freq_base,
float rope_freq_scale,
bool low_vram,

3
llama.h
View file

@ -88,7 +88,8 @@ extern "C" {
int32_t n_batch; // prompt processing batch size
int32_t n_gpu_layers; // number of layers to store in VRAM
int32_t main_gpu; // the GPU that is used for scratch and small tensors
float tensor_split[LLAMA_MAX_DEVICES]; // how to split layers across multiple GPUs
const float * tensor_split; // how to split layers across multiple GPUs (size: LLAMA_MAX_DEVICES)
// ref: https://github.com/ggerganov/llama.cpp/pull/2054
float rope_freq_base; // RoPE base frequency