tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-07-05 17:03:34 +00:00
Code Issues Packages Projects Releases Wiki Activity

cann: support q4_0 model (#8822)

Browse Source
This commit is contained in:
wangshuai09
2024-08-05 12:22:30 +08:00
committed by GitHub
parent 0d6fb52be0
commit c02b0a8a4d
7 changed files with 357 additions and 45 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
ggml/src/ggml-cann.cpp
View File

@ -627,7 +627,6 @@ GGML_CALL static void* ggml_backend_cann_buffer_get_base(
GGML_CALL static void ggml_backend_cann_transform_q4_0(ggml_tensor* tensor,
const void* src,
void* dst) {
GGML_ASSERT(tensor->op == GGML_OP_NONE);
int64_t n_elems = ggml_nelements(tensor);
int64_t groups = n_elems / QK4_0;
@ -679,7 +678,6 @@ GGML_CALL static void ggml_backend_cann_transform_q4_0(ggml_tensor* tensor,
*/
GGML_CALL static void ggml_backend_cann_transform_back_q4_0(
const ggml_tensor* tensor, void* src, void* dst) {
GGML_ASSERT(tensor->op == GGML_OP_NONE);
int64_t n_elems = ggml_nelements(tensor);
int64_t groups = n_elems / QK4_0;
@ -1666,10 +1664,17 @@ GGML_CALL static bool ggml_backend_cann_supports_op(ggml_backend_t backend,
}
case GGML_OP_MUL_MAT: {
switch (op->src[0]->type) {
// case GGML_TYPE_Q4_0:
case GGML_TYPE_F16:
case GGML_TYPE_F32:
case GGML_TYPE_Q8_0:
// TODO: fix me
// Current groupsize should not be greater than k-1 in
// aclnnWeightQuantBatchMatmulV2GetWorkspaceSize().
if (op->src[0]->ne[0]-1 > QK8_0) {
return true;
}
return false;
case GGML_TYPE_Q4_0:
return true;
default:
return false;
@ -1694,6 +1699,7 @@ GGML_CALL static bool ggml_backend_cann_supports_op(ggml_backend_t backend,
case GGML_TYPE_F32:
case GGML_TYPE_F16:
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q4_0:
return true;
default:
return false;