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

sycl: add fused top-k MoE (#25217)

Browse Source 
* sycl: add fused top-k MoE

* sycl: address review: GGML_SYCL_ENABLE_FUSION env, move fusion dispatch to topk-moe

* sycl: print GGML_SYCL_ENABLE_FUSION at startup like other env vars

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
Frosty40
2026-07-13 09:56:41 +03:00
committed by GitHub
co-authored by Claude Fable 5
parent e474bba7af
commit efb3036c18
6 changed files with 645 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-sycl/ggml-sycl.cpp
+10 -1
View File
@@ -85,6 +85,7 @@ int g_ggml_sycl_enable_optimize = 1;
int g_ggml_sycl_enable_graph = 0;
int g_ggml_sycl_enable_dnn = 1;
int g_ggml_sycl_enable_vmm = 1;
int g_ggml_sycl_enable_fusion = 1;
int g_ggml_sycl_prioritize_dmmv = 0;
int g_ggml_sycl_use_async_mem_op = 0;
int g_ggml_sycl_use_async_mem_op_requested = 1;
@@ -285,6 +286,7 @@ static void ggml_check_sycl() try {
g_ggml_sycl_enable_graph = ggml_sycl_get_env("GGML_SYCL_ENABLE_GRAPH", 0);
g_ggml_sycl_enable_dnn = ggml_sycl_get_env("GGML_SYCL_ENABLE_DNN", 1);
g_ggml_sycl_enable_vmm = ggml_sycl_get_env("GGML_SYCL_ENABLE_VMM", 1);
g_ggml_sycl_enable_fusion = ggml_sycl_get_env("GGML_SYCL_ENABLE_FUSION", 1);
g_ggml_sycl_prioritize_dmmv = ggml_sycl_get_env("GGML_SYCL_PRIORITIZE_DMMV", 0);
g_ggml_sycl_dev2dev_memcpy = ggml_sycl_get_env("GGML_SYCL_DEV2DEV_MEMCPY", DEV2DEV_MEMCPY_SYCL);
@@ -353,7 +355,6 @@ static void ggml_check_sycl() try {
#else
GGML_LOG_INFO(" GGML_SYCL_ENABLE_DNN: DNN disabled by compile flag\n");
#endif
#ifdef SYCL_FLASH_ATTN
GGML_LOG_INFO(" GGML_SYCL_ENABLE_FLASH_ATTN: %d\n", g_ggml_sycl_enable_flash_attention);
#else
@@ -375,6 +376,8 @@ static void ggml_check_sycl() try {
GGML_LOG_INFO(" GGML_SYCL_ENABLE_VMM: virtual memory extension is not available\n");
#endif
GGML_LOG_INFO(" GGML_SYCL_ENABLE_FUSION: %d\n", g_ggml_sycl_enable_fusion);
GGML_LOG_INFO(" GGML_SYCL_PRIORITIZE_DMMV: %d\n", g_ggml_sycl_prioritize_dmmv);
g_ggml_sycl_use_async_mem_op_requested = ggml_sycl_get_env("GGML_SYCL_USE_ASYNC_MEM_OP", 1);
@@ -5370,6 +5373,12 @@ static void ggml_backend_sycl_graph_compute_impl(ggml_backend_sycl_context * syc
if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
continue;
}
const int nodes_to_skip = ggml_sycl_fuse(*sycl_ctx, cgraph, i);
if (nodes_to_skip != 0) {
i += nodes_to_skip;
continue;
}
#ifndef NDEBUG
assert(node->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device));
for (int j = 0; j < GGML_MAX_SRC; j++) {