LLM-Inference/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-08-12 22:31:11 +04:00
Code Issues Packages Projects Releases Wiki Activity

ggml-webgpu: fix CI errors from #25025 and #25262 (#26566)

Browse Source 
* test new flash_attn test

* rebase and fix to disable subgrou matrices when max_kv_tile == 0

* delete log output

* Add i32 support to cpy and enables the all ops test

* restore the non target ci tests

* comment out of TODO of build-cpu.yml

* fix format
This commit is contained in:
Masashi Yoshimura
2026-08-11 07:10:00 +03:00
committed by GitHub
parent 030ebb558a
commit 84f7129467
3 changed files with 24 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp
+20 -2
View File
@@ -2815,11 +2815,25 @@ class ggml_webgpu_shader_lib {
key.common.v_direct &= decisions.use_sg_matrix && key.common.v_type == GGML_TYPE_F16;
key.use_sg_matrix = decisions.use_sg_matrix;
const uint32_t max_kv_tile = ggml_webgpu_flash_attn_max_kv_tile(
uint32_t max_kv_tile = ggml_webgpu_flash_attn_max_kv_tile(
context.wg_mem_limit_bytes, decisions.q_tile, decisions.use_sg_matrix ? context.sg_mat_n : 1u,
key.common.head_dim_qk, key.common.head_dim_v, key.common.has_mask,
key.common.k_direct || key.common.v_direct);
GGML_ASSERT(max_kv_tile > 0);
// WorkGroup storage size isn't enough for some params with subgroup matrices path (ref. https://github.com/ggml-org/llama.cpp/pull/26566)
if (max_kv_tile == 0) {
GGML_ASSERT(decisions.use_sg_matrix);
// switch to flash_attn_reg_tile path
decisions.use_sg_matrix = false;
decisions.q_tile = GGML_WEBGPU_FLASH_ATTN_TILE_Q_TILE;
key.common.k_direct = false;
key.common.v_direct = false;
key.use_sg_matrix = false;
max_kv_tile = ggml_webgpu_flash_attn_max_kv_tile(
context.wg_mem_limit_bytes, decisions.q_tile, 1u, key.common.head_dim_qk, key.common.head_dim_v,
key.common.has_mask, key.common.k_direct || key.common.v_direct);
GGML_ASSERT(max_kv_tile > 0);
}
decisions.kv_tile = decisions.use_sg_matrix ?
std::min(max_kv_tile, context.sg_mat_n * GGML_WEBGPU_FLASH_ATTN_PREFERRED_KV_SG_TILES) :
@@ -2993,6 +3007,10 @@ class ggml_webgpu_shader_lib {
defines.push_back("SRC_F16");
variant += "_f16";
break;
case GGML_TYPE_I32:
defines.push_back("SRC_I32");
variant += "_i32";
break;
default:
GGML_ABORT("Unsupported src type for cpy shader");
}
ggml/src/ggml-webgpu/ggml-webgpu.cpp
+2 -3
View File
@@ -4283,9 +4283,8 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
break;
case GGML_OP_CPY:
case GGML_OP_CONT:
supports_op = ((op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16) &&
(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16)) ||
(op->type == GGML_TYPE_I32 && src0->type == GGML_TYPE_F32);
supports_op = (op->type == GGML_TYPE_F16 || op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_I32) &&
(src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_I32);
break;
case GGML_OP_SET:
supports_op = src0->type == src1->type && src0->type == op->type &&
ggml/src/ggml-webgpu/wgsl-shaders/cpy.wgsl
+2
View File
@@ -4,6 +4,8 @@ enable f16;
#define SRC_TYPE f32
#elif defined(SRC_F16)
#define SRC_TYPE f16
#elif defined(SRC_I32)
#define SRC_TYPE i32
#endif
#ifdef DST_F32