mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 29 Packages Projects Releases Wiki Activity

ggml-webgpu: Fix bug in FlashAttention support check (#22492)

Browse Source 
* Fix flashattention support check for devices that don't support subgroups

* set path to none if kv_tile doesn't fit
This commit is contained in:
Reese Levine
2026-04-29 00:59:00 -07:00
committed by GitHub
parent 7b95ea5d11
commit d6a5094004
2 changed files with 31 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-webgpu/ggml-webgpu.cpp
+4
View File
@@ -3918,6 +3918,10 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
shader_lib_ctx, ctx->webgpu_global_ctx->capabilities.limits.minStorageBufferOffsetAlignment);
const size_t limit_bytes = ctx->webgpu_global_ctx->capabilities.limits.maxComputeWorkgroupStorageSize;
const bool has_mask = op->src[3] != nullptr;
if (decisions.path == GGML_WEBGPU_FLASH_ATTN_PATH_NONE) {
supports_op = false;
break;
}
if (decisions.path == GGML_WEBGPU_FLASH_ATTN_PATH_VEC) {
const size_t min_bytes =
ggml_webgpu_flash_attn_wg_mem_bytes(decisions.q_tile, decisions.kv_tile, (uint32_t) src0->ne[0],