nicolas
/
ik_llama_opt
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Better barrier (#1456)

This commit is contained in:
Kawrakow 2026-03-19 07:33:00 +01:00 committed by GitHub
parent 9b7db9bc3f
commit b56a3c2dc9
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 25 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
ggml/include/ggml.h
View File

@ -846,6 +846,7 @@ extern "C" {
int size;
int n_nodes;
int n_leafs;
int n_batch;
struct ggml_tensor ** nodes;
struct ggml_tensor ** grads;

31
ggml/src/ggml.c
View File

@ -2783,6 +2783,7 @@ struct ggml_compute_state_shared {
const struct ggml_cplan * cplan;
int n_threads;
int n_batch;
// synchronization primitives
atomic_int n_barrier;
@ -4497,16 +4498,7 @@ inline static void ggml_critical_section_start(void) {
}
}
#ifdef GGML_USE_OPENMP
static void ggml_barrier(struct ggml_compute_state_shared * shared) {
if (shared->n_threads == 1) {
return;
}
#pragma omp barrier
}
#else
static void ggml_barrier(struct ggml_compute_state_shared * shared) {
static inline void ggml_barrier_impl(struct ggml_compute_state_shared * shared) {
if (shared->n_threads == 1) {
return;
}
@ -4539,6 +4531,22 @@ static void ggml_barrier(struct ggml_compute_state_shared * shared) {
}
}
}
#ifdef GGML_USE_OPENMP
static void ggml_barrier(struct ggml_compute_state_shared * shared) {
if (shared->n_threads == 1) {
return;
}
if (shared && shared->n_batch > 32) {
ggml_barrier_impl(shared);
return;
}
#pragma omp barrier
}
#else
static void ggml_barrier(struct ggml_compute_state_shared * shared) {
ggml_barrier_impl(shared);
}
#endif
// TODO: make this somehow automatically executed
@ -25878,6 +25886,7 @@ struct ggml_cgraph * ggml_new_graph_custom(struct ggml_context * ctx, size_t siz
/*.size =*/ size,
/*.n_nodes =*/ 0,
/*.n_leafs =*/ 0,
/*.n_batch =*/ 0,
/*.nodes =*/ nodes_ptr,
/*.grads =*/ grads_ptr,
/*.leafs =*/ leafs_ptr,
@ -25899,6 +25908,7 @@ struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph0, int i0, int i1)
/*.size =*/ 0,
/*.n_nodes =*/ i1 - i0,
/*.n_leafs =*/ 0,
/*.n_batch =*/ cgraph0->n_batch,
/*.nodes =*/ cgraph0->nodes + i0,
/*.grads =*/ cgraph0->grads ? cgraph0->grads + i0 : NULL,
/*.leafs =*/ NULL,
@ -26638,6 +26648,7 @@ enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cpl
/*.cgraph =*/ cgraph,
/*.cgraph_plan =*/ cplan,
/*.n_threads =*/ n_threads,
/*.n_batch =*/ cgraph->n_batch,
/*.n_barrier =*/ 0,
/*.n_barrier_passed =*/ 0,
/*.abort_callback =*/ NULL,

4
src/llama-build-context.cpp
View File

@ -9988,8 +9988,10 @@ ggml_cgraph * llm_build_context::llama_build_graph(
GGML_ABORT("fatal error");
}
result->n_batch = llm.n_tokens;
// add on pooling layer
if (lctx.cparams.mtp_op_type == MTP_OP_NONE && (lctx.cparams.embeddings ||
if (lctx.cparams.mtp_op_type == MTP_OP_NONE && (lctx.cparams.embeddings ||
(lctx.model.hparams.nextn_predict_layers > 0 || lctx.model.mtp))) {
result = llm.append_pooling(result);
}