mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-04-16 03:26:08 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CANN]MUL_MAT optimization (#12382)

Browse Source
This commit is contained in:
Chenguang Li 2025-03-15 09:31:08 +08:00 committed by GitHub
parent 9f2250ba72
commit 92a391327e
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 6 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
ggml/src/ggml-cann/aclnn_ops.cpp
View File

@ -2790,10 +2790,14 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
(char*)output_buffer + batch1 * output_stride, ACL_FLOAT16,
output_elem_size, output_ne, output_nb, 2, ACL_FORMAT_ND,
output_ne_offset);
int64_t antiquantGroupSize = 0;
if (src0->ne[0] > QK8_0) {
antiquantGroupSize = QK8_0;
}
ACL_CHECK(aclnnWeightQuantBatchMatmulV2GetWorkspaceSize(
acl_input_tensor, acl_weight_tensor, acl_scale_tensor, nullptr,
nullptr, nullptr, nullptr, QK8_0, acl_output_tensor,
nullptr, nullptr, nullptr, antiquantGroupSize, acl_output_tensor,
&workspaceSize, &executor));
if (workspaceAddr == nullptr) {
workspaceAddr = workspace_allocator.alloc(workspaceSize);
@ -2833,7 +2837,7 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
ACL_CHECK(aclnnWeightQuantBatchMatmulV2GetWorkspaceSize(
acl_input_tensor, acl_weight_tensor, acl_scale_tensor,
nullptr, nullptr, nullptr, nullptr, QK8_0,
nullptr, nullptr, nullptr, nullptr, antiquantGroupSize,
acl_output_tensor, &workspaceSize, &executor));
ACL_CHECK(aclnnWeightQuantBatchMatmulV2(
workspaceAddr, workspaceSize, executor, ctx.stream()));

5
ggml/src/ggml-cann/ggml-cann.cpp
View File

@ -1689,11 +1689,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
case GGML_OP_MUL_MAT: {
switch (op->src[0]->type) {
case GGML_TYPE_Q8_0:
// Current groupsize should not be greater than k-1 in
// aclnnWeightQuantBatchMatmulV2GetWorkspaceSize
if (op->src[0]->ne[0] <= QK8_0) {
return false;
}
case GGML_TYPE_F16:
case GGML_TYPE_F32:
case GGML_TYPE_Q4_0: