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

* 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
@@ -0,0 +1,5 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-tile.hpp"
DECL_FATTN_TILE_CASE(112, 112);
@@ -0,0 +1,5 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-tile.hpp"
DECL_FATTN_TILE_CASE(128, 128);
@@ -0,0 +1,5 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-tile.hpp"
DECL_FATTN_TILE_CASE(256, 256);
@@ -0,0 +1,5 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-tile.hpp"
DECL_FATTN_TILE_CASE(40, 40);
@@ -0,0 +1,5 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-tile.hpp"
DECL_FATTN_TILE_CASE(576, 512);
@@ -0,0 +1,5 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-tile.hpp"
DECL_FATTN_TILE_CASE(64, 64);
@@ -0,0 +1,5 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-tile.hpp"
DECL_FATTN_TILE_CASE(72, 72);
@@ -0,0 +1,5 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-tile.hpp"
DECL_FATTN_TILE_CASE(80, 80);
@@ -0,0 +1,5 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-tile.hpp"
DECL_FATTN_TILE_CASE(96, 96);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_F16, GGML_TYPE_F16);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_F16, GGML_TYPE_F16);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_0);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q4_0);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_F16, GGML_TYPE_Q4_0);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_1);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q4_1);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_F16, GGML_TYPE_Q4_1);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_0);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q5_0);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_F16, GGML_TYPE_Q5_0);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_1);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q5_1);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_F16, GGML_TYPE_Q5_1);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q8_0);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q8_0);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_F16, GGML_TYPE_Q8_0);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_0, GGML_TYPE_F16);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_F16);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_0, GGML_TYPE_F16);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_0, GGML_TYPE_Q5_0);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_0);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_0, GGML_TYPE_Q5_0);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_0, GGML_TYPE_Q5_1);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_1);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_0, GGML_TYPE_Q5_1);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_1, GGML_TYPE_F16);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_F16);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_1, GGML_TYPE_F16);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_1, GGML_TYPE_Q4_1);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_1);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_1, GGML_TYPE_Q4_1);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_1, GGML_TYPE_Q5_0);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_0);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_1, GGML_TYPE_Q5_0);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_1, GGML_TYPE_Q5_1);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_1);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_1, GGML_TYPE_Q5_1);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_1, GGML_TYPE_Q8_0);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q8_0);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_1, GGML_TYPE_Q8_0);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_0, GGML_TYPE_F16);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_F16);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_0, GGML_TYPE_F16);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_0, GGML_TYPE_Q4_0);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_0);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_0, GGML_TYPE_Q4_0);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_0, GGML_TYPE_Q5_0);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_0);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_0, GGML_TYPE_Q5_0);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_0, GGML_TYPE_Q5_1);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_1);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_0, GGML_TYPE_Q5_1);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_0, GGML_TYPE_Q8_0);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q8_0);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_0, GGML_TYPE_Q8_0);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_1, GGML_TYPE_F16);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_F16);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_1, GGML_TYPE_F16);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_1, GGML_TYPE_Q4_0);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_0);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_1, GGML_TYPE_Q4_0);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_1, GGML_TYPE_Q4_1);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_1);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_1, GGML_TYPE_Q4_1);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_1, GGML_TYPE_Q5_1);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_1);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_1, GGML_TYPE_Q5_1);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q8_0, GGML_TYPE_F16);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_F16);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q8_0, GGML_TYPE_F16);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q8_0, GGML_TYPE_Q4_1);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_1);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q8_0, GGML_TYPE_Q4_1);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q8_0, GGML_TYPE_Q5_0);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_0);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q8_0, GGML_TYPE_Q5_0);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1);
@@ -0,0 +1,7 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../fattn-vec.hpp"
DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0);
DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0);
DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0);