root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-26 03:14:35 +00:00
Code Issues Actions 7 Packages Projects Releases Wiki Activity

llama : quantize up to 31% faster on Linux and Windows with mmap (#3206)

Browse Source 
* llama : enable mmap in quantize on Linux -> 31% faster

* also enable mmap on Windows

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Cebtenzzre 2023-09-29 09:48:45 -04:00 committed by GitHub
parent 0a4a4a0982
commit 2777a84be4
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 17 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
llama.cpp
View File

@ -6027,7 +6027,18 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
nthread = std::thread::hardware_concurrency(); nthread = std::thread::hardware_concurrency();
} }
llama_model_loader ml(fname_inp, /*use_mmap*/ false); // mmap consistently increases speed Linux, and also increases speed on Windows with
// hot cache. It may cause a slowdown on macOS, possibly related to free memory.
#if defined(__linux__) || defined(_WIN32)
constexpr bool use_mmap = true;
#else
constexpr bool use_mmap = false;
#endif
llama_model_loader ml(fname_inp, use_mmap);
if (ml.use_mmap) {
ml.mapping.reset(new llama_mmap(&ml.file, /* prefetch */ 0, ggml_is_numa()));
}
llama_model model; llama_model model;
llm_load_arch(ml, model); llm_load_arch(ml, model);
@ -6105,10 +6116,12 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
const std::string name = ggml_get_name(tensor); const std::string name = ggml_get_name(tensor);
if (!ml.use_mmap) {
if (read_data.size() < ggml_nbytes(tensor)) { if (read_data.size() < ggml_nbytes(tensor)) {
read_data.resize(ggml_nbytes(tensor)); read_data.resize(ggml_nbytes(tensor));
} }
tensor->data = read_data.data(); tensor->data = read_data.data();
}
ml.load_data_for(tensor); ml.load_data_for(tensor);
LLAMA_LOG_INFO("[%4d/%4d] %36s - [%s], type = %6s, ", LLAMA_LOG_INFO("[%4d/%4d] %36s - [%s], type = %6s, ",