diff options
| author | Kawrakow <iwankawrakow@gmail.com> | 2025-03-05 07:27:49 +0200 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2025-03-05 07:27:49 +0200 |
| commit | 7bdbf99bbdbfe46b01f7783a7c98a30a1558e2c3 (patch) | |
| tree | ecd49fd2b7bccc1b3d70819241d0eaf9c4dddede /ggml/src/ggml-cuda/template-instances | |
| parent | a87e54db6ec2409284a55f029d4abe9e50990064 (diff) | |
DeepSeek CUDA Flash Attention (#241)
* WIP CUDA FA with Dk != Dv
* WIP
* CUDA FA WIP - It actually works!
No TG yet, but for PP I can run FA with fp16 cache and it gets
the same answer.
* CUDA FA WIP - it now works for Q8_0 + Q8_0 for KV cache
* CUDA FA WIP - TG, not working yet.
* CUDA FA with Dk != Dv: it works now for DeepSeek
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Diffstat (limited to 'ggml/src/ggml-cuda/template-instances')
9 files changed, 30 insertions, 0 deletions
diff --git a/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs192-f16-f16.cu b/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs192-f16-f16.cu new file mode 100644 index 00000000..7dda0133 --- /dev/null +++ b/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs192-f16-f16.cu @@ -0,0 +1,5 @@ +// This file has been autogenerated by generate_cu_files.py, do not edit manually. + +#include "../fattn-vec-f16.cuh" + +DECL_FATTN_VEC_F16_CASE_DKDV(192, 128, GGML_TYPE_F16, GGML_TYPE_F16); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs192-q8_0-q8_0.cu b/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs192-q8_0-q8_0.cu new file mode 100644 index 00000000..740ac37d --- /dev/null +++ b/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs192-q8_0-q8_0.cu @@ -0,0 +1,5 @@ +// This file has been autogenerated by generate_cu_files.py, do not edit manually. + +#include "../fattn-vec-f16.cuh" + +DECL_FATTN_VEC_F16_CASE_DKDV(192, 128, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs192-f16-f16.cu b/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs192-f16-f16.cu new file mode 100644 index 00000000..1ea24302 --- /dev/null +++ b/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs192-f16-f16.cu @@ -0,0 +1,5 @@ +// This file has been autogenerated by generate_cu_files.py, do not edit manually. + +#include "../fattn-vec-f32.cuh" + +DECL_FATTN_VEC_F32_CASE_DKDV(192, 128, GGML_TYPE_F16, GGML_TYPE_F16); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs192-q8_0-q8_0.cu b/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs192-q8_0-q8_0.cu new file mode 100644 index 00000000..6be4d042 --- /dev/null +++ b/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs192-q8_0-q8_0.cu @@ -0,0 +1,5 @@ +// This file has been autogenerated by generate_cu_files.py, do not edit manually. + +#include "../fattn-vec-f32.cuh" + +DECL_FATTN_VEC_F32_CASE_DKDV(192, 128, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqfloat-cpb16.cu b/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqfloat-cpb16.cu index 2d94e65c..334e1deb 100644 --- a/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqfloat-cpb16.cu +++ b/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqfloat-cpb16.cu @@ -8,3 +8,5 @@ DECL_FATTN_WMMA_F16_CASE(96, 16, float); DECL_FATTN_WMMA_F16_CASE(112, 16, float); DECL_FATTN_WMMA_F16_CASE(128, 16, float); DECL_FATTN_WMMA_F16_CASE(256, 16, float); + +DECL_FATTN_WMMA_F16_CASE_DKDV(192, 128, 16, float); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqfloat-cpb32.cu b/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqfloat-cpb32.cu index c3d9df3c..1faf3c9b 100644 --- a/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqfloat-cpb32.cu +++ b/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqfloat-cpb32.cu @@ -7,3 +7,5 @@ DECL_FATTN_WMMA_F16_CASE(80, 32, float); DECL_FATTN_WMMA_F16_CASE(96, 32, float); DECL_FATTN_WMMA_F16_CASE(112, 32, float); DECL_FATTN_WMMA_F16_CASE(128, 32, float); + +DECL_FATTN_WMMA_F16_CASE_DKDV(192, 128, 32, float); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb16.cu b/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb16.cu index bb680e40..48973618 100644 --- a/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb16.cu +++ b/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb16.cu @@ -8,3 +8,5 @@ DECL_FATTN_WMMA_F16_CASE(96, 16, half); DECL_FATTN_WMMA_F16_CASE(112, 16, half); DECL_FATTN_WMMA_F16_CASE(128, 16, half); DECL_FATTN_WMMA_F16_CASE(256, 16, half); + +DECL_FATTN_WMMA_F16_CASE_DKDV(192, 128, 16, half); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb32.cu b/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb32.cu index 073f71b1..ed92963e 100644 --- a/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb32.cu +++ b/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb32.cu @@ -8,3 +8,5 @@ DECL_FATTN_WMMA_F16_CASE(96, 32, half); DECL_FATTN_WMMA_F16_CASE(112, 32, half); DECL_FATTN_WMMA_F16_CASE(128, 32, half); DECL_FATTN_WMMA_F16_CASE(256, 32, half); + +DECL_FATTN_WMMA_F16_CASE_DKDV(192, 128, 32, half); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb8.cu b/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb8.cu index d30710c5..4e221003 100644 --- a/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb8.cu +++ b/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb8.cu @@ -6,3 +6,5 @@ DECL_FATTN_WMMA_F16_CASE(64, 8, half); DECL_FATTN_WMMA_F16_CASE(96, 8, half); DECL_FATTN_WMMA_F16_CASE(128, 8, half); DECL_FATTN_WMMA_F16_CASE(256, 8, half); + +DECL_FATTN_WMMA_F16_CASE_DKDV(192, 128, 8, half); |