root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-11 03:01:45 +00:00
Code Issues Actions 11 Packages Projects Releases Wiki Activity

Workrounnd to set node->backend

Browse Source
This commit is contained in:
mqy 2023-06-15 08:28:39 +08:00
parent 9106232260
commit bb590f1482
3 changed files with 15 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
ggml-opencl.cpp
View File

@ -1599,8 +1599,8 @@ bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tens
// TODO: find the optimal values for these // TODO: find the optimal values for these
if ((src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && if ((src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) &&
src1->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 &&
dst->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32 /*&&
((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_GPU)) { ((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_GPU)*/) {
return true; return true;
} }

13
ggml.c
View File

@ -15938,6 +15938,18 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
struct ggml_task_stage *stages = node->task_profile.stages; struct ggml_task_stage *stages = node->task_profile.stages;
// Workrounnd to set node->backend.
for (int j = 0; j < 3; j++) {
if (node->backend == GGML_BACKEND_CPU &&
(stages[j].backend & GGML_TASK_BACKEND_GPU)) {
if (ggml_cpu_has_cublas() || ggml_cpu_has_clblast()) {
node->backend = GGML_BACKEND_GPU;
} else {
GGML_ASSERT(false);
}
}
}
// compute stage n_tasks. // compute stage n_tasks.
int n_tasks = stages[1].parallel ? n_threads : 1; int n_tasks = stages[1].parallel ? n_threads : 1;
@ -16008,6 +16020,7 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
if (comp_backend == GGML_TASK_BACKEND_GPU_CL) { if (comp_backend == GGML_TASK_BACKEND_GPU_CL) {
#if defined(GGML_USE_CLBLAST) #if defined(GGML_USE_CLBLAST)
GGML_ASSERT(ggml_cl_can_mul_mat(node->src0, node->src1, node));
cur = ggml_cl_mul_mat_get_wsize(node->src0, node->src1, node); cur = ggml_cl_mul_mat_get_wsize(node->src0, node->src1, node);
#else #else
GGML_ASSERT(false); GGML_ASSERT(false);

1
tests/test-ggml-tune.c
View File

@ -85,7 +85,6 @@ ggml_task_profiles_mock_qxx_provider(struct ggml_tensor *node,
struct ggml_task_profile *profiles) { struct ggml_task_profile *profiles) {
UNUSED(node); UNUSED(node);
profiles[0].stages[0].backend = GGML_TASK_BACKEND_CPU; profiles[0].stages[0].backend = GGML_TASK_BACKEND_CPU;
profiles[0].stages[0].backend = GGML_TASK_BACKEND_CPU;
profiles[0].stages[1].backend = GGML_TASK_BACKEND_CPU; profiles[0].stages[1].backend = GGML_TASK_BACKEND_CPU;
profiles[1].stages[0].backend = GGML_TASK_BACKEND_CPU; profiles[1].stages[0].backend = GGML_TASK_BACKEND_CPU;
profiles[1].stages[1].backend = GGML_TASK_BACKEND_CPU_BLAS; profiles[1].stages[1].backend = GGML_TASK_BACKEND_CPU_BLAS;