root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-27 20:04:35 +00:00
Code Issues Actions 3 Packages Projects Releases Wiki Activity

llama : fix res_norm offloading

Browse Source
This commit is contained in:
Georgi Gerganov 2023-10-29 09:20:35 +02:00
parent e14aa46151
commit 79617902ea
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
1 changed files with 103 additions and 98 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

31
llama.cpp
View File

@ -5456,13 +5456,15 @@ static struct ggml_cgraph * llama_build_graph(
const int i_gpu_start = n_layer - n_gpu_layers; const int i_gpu_start = n_layer - n_gpu_layers;
// should we offload the final norm? yes if we are not computing embeddings // should we offload the final norm? yes if we are not computing embeddings
const bool off_res_norm = !lctx.embedding.empty(); const bool off_res_norm = lctx.embedding.empty();
// offload functions set the tensor output backend to GPU // offload functions set the tensor output backend to GPU
// tensors are GPU-accelerated if any input or the output has been offloaded // tensors are GPU-accelerated if any input or the output has been offloaded
offload_func_t offload_func_nr = ggml_offload_nop; // nr = non-repeating offload_func_t offload_func_nr = ggml_offload_nop; // nr = non-repeating
offload_func_t offload_func_kq = ggml_offload_nop; offload_func_t offload_func_kq = ggml_offload_nop;
offload_func_t offload_func_v = ggml_offload_nop; offload_func_t offload_func_v = ggml_offload_nop;
offload_func_t offload_func_emb = ggml_offload_nop;
offload_func_t offload_func_out = ggml_offload_nop;
offload_func_t offload_func = ggml_offload_nop; offload_func_t offload_func = ggml_offload_nop;
#ifdef GGML_USE_CUBLAS #ifdef GGML_USE_CUBLAS
@ -5476,10 +5478,20 @@ static struct ggml_cgraph * llama_build_graph(
offload_func_kq = ggml_cuda_assign_buffers_no_alloc; offload_func_kq = ggml_cuda_assign_buffers_no_alloc;
} }
offload_func_emb = off_res_norm ? ggml_cuda_assign_buffers_no_alloc : ggml_offload_nop;
offload_func_out = ggml_offload_nop;
offload_func = ggml_cuda_assign_buffers_no_alloc; offload_func = ggml_cuda_assign_buffers_no_alloc;
#endif // GGML_USE_CUBLAS #endif // GGML_USE_CUBLAS
static const std::unordered_map<std::string, offload_func_t> k_offload_func = { static const std::unordered_map<offload_func_t, std::string> k_offload_func_name = {
{ ggml_offload_nop, "CPU" },
#ifdef GGML_USE_CUBLAS
{ ggml_cuda_assign_buffers_no_alloc, "GPU (CUDA)" },
#endif
};
const std::unordered_map<std::string, offload_func_t> k_offload_func = {
{ "KQ_mask", offload_func_kq }, { "KQ_mask", offload_func_kq },
{ "KQ_pos", offload_func_kq }, { "KQ_pos", offload_func_kq },
{ "K_shift", offload_func_kq }, { "K_shift", offload_func_kq },
@ -5566,15 +5578,8 @@ static struct ggml_cgraph * llama_build_graph(
{ "out_norm_0", offload_func_nr }, { "out_norm_0", offload_func_nr },
{ "out_norm_0_w", offload_func_nr }, { "out_norm_0_w", offload_func_nr },
{ "result_norm", off_res_norm ? offload_func_nr : ggml_offload_nop }, { "result_norm", offload_func_emb },
//{ "result_output", offload_func }, { "result_output", offload_func_out },
};
static const std::unordered_map<offload_func_t, std::string> k_offload_func_name = {
{ ggml_offload_nop, "CPU" },
#ifdef GGML_USE_CUBLAS
{ ggml_cuda_assign_buffers_no_alloc, "GPU (CUDA)" },
#endif
}; };
std::unordered_map<std::string, int> ofn; std::unordered_map<std::string, int> ofn;
@ -5591,7 +5596,7 @@ static struct ggml_cgraph * llama_build_graph(
const auto it = k_offload_func.find(name); const auto it = k_offload_func.find(name);
if (it == k_offload_func.end()) { if (it == k_offload_func.end()) {
// if a tensor that is not view hasn't been offloaded, we warn the user // if a tensor hasn't been offloaded, we warn the user
if (worst_case) { if (worst_case) {
LLAMA_LOG_WARN("%s: node %4d %32s: not offloaded (ref: %s)\n", __func__, LLAMA_LOG_WARN("%s: node %4d %32s: not offloaded (ref: %s)\n", __func__,
i, name.c_str(), "https://github.com/ggerganov/llama.cpp/pull/3837"); i, name.c_str(), "https://github.com/ggerganov/llama.cpp/pull/3837");
@ -5602,7 +5607,7 @@ static struct ggml_cgraph * llama_build_graph(
// count the number of layers and respect the provided n_gpu_layers // count the number of layers and respect the provided n_gpu_layers
offload_func_t f = it->second; offload_func_t f = it->second;
if (f == offload_func) { if (n_gpu_layers < n_layer && f == offload_func) {
if (ofn[name]++ < i_gpu_start) { if (ofn[name]++ < i_gpu_start) {
f = ggml_offload_nop; f = ggml_offload_nop;
} }