Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

opencl: generalize Adreno MoE kernels on M (#23449)

Browse Source
This commit is contained in:
Shawn Gu
2026-05-22 17:08:41 -07:00
committed by GitHub
parent 1acee6bf89
commit 0f3cb3fc8b
18 changed files with 145 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-opencl/ggml-opencl.cpp
+9 -9
View File
@@ -4693,7 +4693,7 @@ inline bool use_adreno_kernels(const ggml_backend_opencl_context *backend_ctx, c
inline bool use_adreno_moe_kernels(const ggml_backend_opencl_context *backend_ctx, const ggml_tensor *tensor) {
GGML_UNUSED(backend_ctx);
int ne01 = tensor->ne[1];
return (((strstr(tensor->name, "ffn") != NULL) && (strstr(tensor->name, "exps") != NULL)) || (strstr(tensor->name, "as") != NULL)) && (ne01 % 64 == 0);
return (((strstr(tensor->name, "ffn") != NULL) && (strstr(tensor->name, "exps") != NULL)) || (strstr(tensor->name, "as") != NULL)) && (ne01 % 32 == 0);
}
inline bool enable_adreno_trans_weight(const ggml_backend_opencl_context *backend_ctx, const ggml_tensor *tensor) {
@@ -14297,7 +14297,7 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
CL_CHECK(status);
// set thread grid
global_size[0] = static_cast<size_t>(ne01);
global_size[0] = static_cast<size_t>(((ne01 + 63) / 64) * 64);
global_size[1] = 4;
global_size[2] = static_cast<size_t>(ne20);
local_size[1] = 4;
@@ -14513,7 +14513,7 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
CL_CHECK(status);
// set thread grid
global_size[0] = static_cast<size_t>(ne01);
global_size[0] = static_cast<size_t>(((ne01 + 63) / 64) * 64);
global_size[1] = 4;
global_size[2] = static_cast<size_t>(ne20);
local_size[1] = 4;
@@ -14689,7 +14689,7 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
CL_CHECK(status);
// set thread grid
global_size[0] = static_cast<size_t>(ne01);
global_size[0] = static_cast<size_t>(((ne01 + 63) / 64) * 64);
global_size[1] = 4;
global_size[2] = static_cast<size_t>(ne20);
local_size[1] = 4;
@@ -14865,7 +14865,7 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
CL_CHECK(status);
// set thread grid
global_size[0] = static_cast<size_t>(ne01);
global_size[0] = static_cast<size_t>(((ne01 + 63) / 64) * 64);
global_size[1] = 4;
global_size[2] = static_cast<size_t>(ne20);
local_size[1] = 4;
@@ -15118,7 +15118,7 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
CL_CHECK(status);
// set thread grid
global_size[0] = static_cast<size_t>(ne01);
global_size[0] = static_cast<size_t>(((ne01 + 63) / 64) * 64);
global_size[1] = 4;
global_size[2] = static_cast<size_t>(ne20);
local_size[1] = 4;
@@ -15291,7 +15291,7 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
CL_CHECK(status);
// set thread grid
global_size[0] = static_cast<size_t>(ne01);
global_size[0] = static_cast<size_t>(((ne01 + 63) / 64) * 64);
global_size[1] = 4;
global_size[2] = static_cast<size_t>(ne20);
local_size[1] = 4;
@@ -15469,7 +15469,7 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
CL_CHECK(status);
// set thread grid
global_size[0] = static_cast<size_t>(ne01);
global_size[0] = static_cast<size_t>(((ne01 + 63) / 64) * 64);
global_size[1] = 4;
global_size[2] = static_cast<size_t>(ne20);
local_size[1] = 4;
@@ -15644,7 +15644,7 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
CL_CHECK(status);
// set thread grid
global_size[0] = static_cast<size_t>(ne01);
global_size[0] = static_cast<size_t>(((ne01 + 63) / 64) * 64);
global_size[1] = 4;
global_size[2] = static_cast<size_t>(ne20);
local_size[1] = 4;