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

[SYCL] supprt Flash Attention for fp32/fp16/Q4/Q5/Q8 (#20190)

Browse Source 
* support flash-attention for fp32/fp16/Q4/Q5/Q8

* rm warining

* update for JIT
This commit is contained in:
Neo Zhang
2026-03-08 12:00:07 +08:00
committed by GitHub
parent c5a778891b
commit 213c4a0b81
65 changed files with 20091 additions and 8593 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-sycl/fattn-tile.cpp
+55
View File
@@ -0,0 +1,55 @@
#include <sycl/sycl.hpp>
#include <sycl/ext/oneapi/work_group_static.hpp>
#include "dpct/helper.hpp"
#include "common.hpp"
#include "fattn-common.hpp"
#include "fattn-tile.hpp"
#include <cmath>
#include <float.h>
namespace syclex = sycl::ext::oneapi::experimental;
void ggml_sycl_flash_attn_ext_tile(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
const ggml_tensor * K = dst->src[1];
const ggml_tensor * V = dst->src[2];
switch (K->ne[0]) {
case 40: {
GGML_ASSERT(V->ne[0] == K->ne[0]);
ggml_sycl_flash_attn_ext_tile_case< 40, 40>(ctx, dst);
} break;
case 64: {
GGML_ASSERT(V->ne[0] == K->ne[0]);
ggml_sycl_flash_attn_ext_tile_case< 64, 64>(ctx, dst);
} break;
case 72: {
GGML_ASSERT(V->ne[0] == K->ne[0]);
ggml_sycl_flash_attn_ext_tile_case< 72, 72>(ctx, dst);
} break;
case 80: {
GGML_ASSERT(V->ne[0] == K->ne[0]);
ggml_sycl_flash_attn_ext_tile_case< 80, 80>(ctx, dst);
} break;
case 96: {
GGML_ASSERT(V->ne[0] == K->ne[0]);
ggml_sycl_flash_attn_ext_tile_case< 96, 96>(ctx, dst);
} break;
case 112: {
GGML_ASSERT(V->ne[0] == K->ne[0]);
ggml_sycl_flash_attn_ext_tile_case<112, 112>(ctx, dst);
} break;
case 128: {
GGML_ASSERT(V->ne[0] == K->ne[0]);
ggml_sycl_flash_attn_ext_tile_case<128, 128>(ctx, dst);
} break;
case 256: {
GGML_ASSERT(V->ne[0] == K->ne[0]);
ggml_sycl_flash_attn_ext_tile_case<256, 256>(ctx, dst);
} break;
case 576: {
GGML_ASSERT(V->ne[0] == 512);
ggml_sycl_flash_attn_ext_tile_case<576, 512>(ctx, dst);
} break;
default: {
GGML_ABORT("Unsupported head size");
} break;
}
}