CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-31 08:51:55 +00:00
Code Issues Packages Projects Releases Wiki Activity

llama : add thread safety test (#14035)

Browse Source 
* llama : add thread safety test

* llamafile : remove global state

* llama : better LLAMA_SPLIT_MODE_NONE logic

when main_gpu < 0 GPU devices are not used

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Diego Devesa
2025-06-16 08:11:43 -07:00
committed by GitHub
parent 0dbcabde8c
commit 6adc3c3ebc
9 changed files with 192 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
ggml/src/ggml-cpu/llamafile/sgemm.cpp
View File

@@ -53,7 +53,6 @@
#include "ggml-cpu-impl.h"
#include "ggml-quants.h"
#include <atomic>
#include <array>
#include <type_traits>
@@ -394,8 +393,6 @@ class tinyBLAS {
template <int RM, int RN, int BM>
NOINLINE void gemm(int64_t m, int64_t n, int64_t BN) {
static std::atomic<int64_t> current_chunk;
GGML_ASSERT(m % (RM * BM) == 0);
const int64_t ytiles = m / (RM * BM);
const int64_t xtiles = (n + RN -1) / RN;
@@ -410,7 +407,7 @@ class tinyBLAS {
if (params->ith == 0) {
GGML_ASSERT( jj_BN * SIZE_BN + (NB_BN - jj_BN) * (SIZE_BN - 1) == xtiles);
// Every thread starts at ith, so the first unprocessed chunk is nth. This save a bit of coordination right at the start.
std::atomic_store_explicit(&current_chunk, (int64_t)params->nth, std::memory_order_relaxed);
ggml_threadpool_chunk_set(params->threadpool, params->nth);
}
ggml_barrier(params->threadpool);
@@ -439,8 +436,7 @@ class tinyBLAS {
GGML_ASSERT(jj == jj2);
}
// next step.
job = std::atomic_fetch_add_explicit(&current_chunk, (int64_t)1, std::memory_order_relaxed);
job = ggml_threadpool_chunk_add(params->threadpool, 1);
}
ggml_barrier(params->threadpool);