EngineX-Ascend/enginex-ascend-910-llama.cpp
10
0
Fork 0
Code Issues Pull Requests Actions 4 Projects Releases Wiki Activity

metal : fuse NORM + MUL + ADD, support non-multiples of 4 (#16220)

Browse Source 
* metal : fuse NORM + MUL + ADD

* metal : support norms of non-multiple of 4

* cont : fix comment [no ci]
This commit is contained in:
Georgi Gerganov
2025-09-25 11:30:16 +03:00
committed by GitHub
parent 4ea00794b8
commit dfcd53f7ec
9 changed files with 206 additions and 232 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
ggml/src/ggml-metal/ggml-metal-impl.h
View File

@@ -428,16 +428,11 @@ typedef struct {
uint64_t nb1;
} ggml_metal_kargs_mul_mv_id;
// NORM
// RMS_NORM
typedef struct {
int32_t ne00;
int32_t ne00_4;
uint64_t nb01;
float eps;
} ggml_metal_kargs_norm;
typedef struct {
int32_t ne00;
int32_t ne00_4;
int32_t ne00_t;
uint64_t nb1;
uint64_t nb2;
uint64_t nb3;
@@ -448,7 +443,7 @@ typedef struct {
uint64_t nbf1[3];
uint64_t nbf2[3];
uint64_t nbf3[3];
} ggml_metal_kargs_rms_norm;
} ggml_metal_kargs_norm;
typedef struct {
int32_t ne00;