kanshan/TensorRT-LLMs
1
0
Fork 0
mirror of https://github.com/NVIDIA/TensorRT-LLM.git synced 2026-01-22 11:42:41 +08:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity

Merge branch 'user/xiweny/mha_103' into 'feat/b300_cu13'

Browse Source 
update mha cubins and support 103a

See merge request ftp/tekit!9690

Signed-off-by: Xiwen Yu <xiweny@nvidia.com>
This commit is contained in:
Xiwen Yu 2025-09-02 19:26:25 -07:00
commit 1978227bb7
1453 changed files with 8789 additions and 3819 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
3rdparty/cutlass vendored

@ -1 +1 @@
Subproject commit b2dd65dc864e09688245b316ac46c4a6cd07e15c
Subproject commit a49a78ffefc86a87160dfe0ccc3a3a2d1622c918

9
cpp/cmake/modules/cuda_configuration.cmake
View File

@ -138,6 +138,9 @@ function(setup_cuda_architectures)
message(FATAL_ERROR "Unrecognized CUDA architecture: ${CUDA_ARCH}")
endif()
endforeach()
if("103" IN_LIST CMAKE_CUDA_ARCHITECTURES_CLEAN)
list(APPEND CMAKE_CUDA_ARCHITECTURES_CLEAN "100")
endif()
list(REMOVE_DUPLICATES CMAKE_CUDA_ARCHITECTURES_CLEAN)
set(CMAKE_CUDA_ARCHITECTURES_RAW ${CMAKE_CUDA_ARCHITECTURES_CLEAN})
endif()
@ -177,12 +180,6 @@ function(setup_cuda_architectures)
message(STATUS "Excluding SM ${CUDA_ARCH}")
endif()
endforeach()
# deal with SM100/f
if(NOT "100" IN_LIST CMAKE_CUDA_ARCHITECTURES_ORIG
AND NOT "100f" IN_LIST CMAKE_CUDA_ARCHITECTURES_ORIG)
add_definitions("-DEXCLUDE_SM_100")
message(STATUS "Excluding SM 100(f)")
endif()
# -a suffix supported from Hopper (90)
set(MIN_ARCHITECTURE_HAS_ACCEL 90)

1
cpp/tensorrt_llm/kernels/multiHeadAttentionCommon.h
View File

@ -111,6 +111,7 @@ constexpr int32_t kSM_86 = 86;
constexpr int32_t kSM_89 = 89;
constexpr int32_t kSM_90 = 90;
constexpr int32_t kSM_100 = 100;
constexpr int32_t kSM_100f = 10100;
constexpr int32_t kSM_103 = 103;
constexpr int32_t kSM_120 = 120;
constexpr int32_t kSM_121 = 121;

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:b008157f7d6bf0bce51ad66f2924e9a5e6b190254750b85f53a87b1931008697
size 672535

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:e71f661b50c3f2f615a3babe192f924aef33eb315a794e36403a21a92923158f
size 634171

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvCausalP64VarSeqQ128Kv128PersistentContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:349c722e1c175598715fcf9f7ff40ffda9b85fc4f02699adb038aa94023e4f7c
size 651569

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvCausalP64VarSeqQ128Kv128StaticContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:c040724d28bac1cf5034ec987ead0abc3eca29621bfdb02362f6c10c14e81aa5
size 618385

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:2852fa0332933b3d26345ddcbff1abc18cf3111bdc193984d5aa289471df4375
size 676855

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:deee451bca014cfcad6cf4aa15a88dfe7d8bf14e00d9918e5d1fea9b571b348b
size 630307

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:727b8cdc921684e204ad18359350f3478b66fe090a937a0f03d0bdbb3e124511
size 678699

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:4e25de8526e4f8e3ece86814283147edc8246fe2a3f27cae7975d2dec83276d5
size 639255

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ128Kv128PersistentContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:55a1400227f3ee2b7366eaa5b9bf21b6bd03c143ecaa74633cfd3076762e8baf
size 717131

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ128Kv128StaticContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:ced57950db96d94c81a011440b45fc0e3994f47ce1027b274e52feac0f0e0a9b
size 634267

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:14022d08df8e55c360e796644f0fba8499508c58cf3cc037d413c19870ab9945
size 654291

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:502a01513e85095de63486a9602e914656c0cd0ebaad5cdd7d928d5d6898e1fc
size 583118

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:d93d4baf4029784d519cc0cbf165cbb0147b520247dae0d507196fdf1ec6e4d9
size 621999

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:ec6c4e5122da147bf17f1c976816142ec03b2d9870ab06d7c3513258c9b6e993
size 550062

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:4e8ef4599d4c5152fb7dabdd5ad34d511f833f44ef418335b480870432c7ec96
size 657911

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:8973e4673d164413b9041362b25d88d57810c24b4981deccee95acef85c9e35e
size 611362

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:cb708698b0483f01e81de87aa456018241f2226b2b8ac0dde5f42a279df8b5e6
size 659755

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:26706c59e9be4ef51d6f27ac01b5bbfcc5edccc9362f8b5a45130362ab58e177
size 620311

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64VarSeqQ128Kv128PersistentContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:a78bfa03235f26071042c3fb4faae960d5a5e5f3183d334594bb2d1a2f302695
size 696165

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64VarSeqQ128Kv128StaticContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:4a90d1ee0b5b3ab387065795ae837f5619149999a2483d5c4dfa3759ff1c32c1
size 617690

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:b553dc790caab29ae9d9f80f46c1b0e6270f1376fd9b94b5e002e959c7731dc8
size 632189

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:89e55c89df2441bc2a35df4789be5980b03539a925ab2485a0c4c88b0a19a067
size 564174

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:56ae77954f33c37de1f21e2c6ab47b855df8e602f56f5c2fcc86a8cee1b217e6
size 600686

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:609d577a5eb29698b48b4953ac5103051f5893595a25c3fd654ee74182e55360
size 531118

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:8258856933f92d4a34021d0b37d871311e8ad7e53119601d24133ba4f983967e
size 680417

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:9c94ac43674f548280cdaf2a94c32d53d1b38023363fa4dc523d5fe211c6323b
size 630785

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:923396628ebc72517067aca3e9bafa53e6276459afe43860d13e68c0b0bc51d9
size 681471

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:4fdf83cef2002b82171d44c07f6afffec05d181f7aa41f0e714e47082a48544e
size 638945

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:ab41b98e59fb1534ca32ecbe40f660eea103ffe1e3690fc3f8d3eb7f9ffbeab3
size 727131

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:ff6494cbe561c1734b4ef40be238b64daa558003fd54bafa51de9658ad3d7e7b
size 693921

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:a3c9978a03d994080c88365f570d72435f4da32177347144f93ee0de73d90b5b
size 638439

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:04df4666457acee8a7ffb8e8d39bd7a0b61ef3479884544660d010353a45c80d
size 575654

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:7872a32915798ebeb8090ccd12b194d20f56a5fdceb3fd314174f4e9b58f0acc
size 606172

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:3ad9b6c9ae5e7ce08dcb041ae546e0ea8cbacaa0135763f38c32ab465338a89b
size 542598

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:28e2fa67ab200295e105340f068ac59053dd097fd7800fc1f7b192c318bfa458
size 661473

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:8805f5c5d401cf5f1b1e9cf44805fca82e8e70c3f25e9e8d0d1659ac9c489773
size 611840

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:59450c1ee22418af646747820a1ced88bbe3b3fe6f05b9be585b57dc79ad69d3
size 662527

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:d5b098063a4187cb8988877800f2bffaa245e185eb8bcf678e2029da2052283d
size 620001

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64VarSeqQ128Kv128PersistentContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:c91764c5d985686105e55ac4112d4d34b62cf366960a30409246ca24548bbf9e
size 706163

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64VarSeqQ128Kv128StaticContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:cdace05675a733fbb7a5f5a39e537751f570d9e435197399157889efa71d806e
size 678135

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:1c2b504e7484bf76fe5deee8d3f6de60bd366849cba82147c1d331f71e704bb5
size 617126

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:ed81023905f417fb69fbc8773554a8bd367fd88644ae8f7c305030b553a8dd93
size 556710

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:6051bb1b62085f078270ad6c9226dd3718873484f4b83edc3cc21d4c4b42d250
size 584860

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:ef5bb94ceb562a6ca23510bc5573f1ce593c577ef3a47353b21d24ee372c929c
size 524444

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:d3257c7aacfa8ff6de91e816fa15708c1b6f4c84fba1c9172eb0d2eb9ed9c5c7
size 600408

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:f27ce0c4c0eb1ae774e1efdc1e711157eaef1df64831d4a52c2a4c6aacff7b6f
size 565940

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvCausalP64VarSeqQ128Kv128PersistentContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:3e6da1e53115103bba1e67dde7b0986b78d4901302bf9fd2939d2eb7cf5f61da
size 580354

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvCausalP64VarSeqQ128Kv128StaticContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:82d3fad691e48f003b933fc3bd493ab4c64084e6c0653d1f4a5a84c96c5e1160
size 550154

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:88d680ec28eacb402985d353fd1f340922309839ddae097c62ef22ba3d802ec6
size 574462

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:40bbcce84e204bc33b371bc499138230e1fba8c3be53b1b75b5e1928448d967b
size 551692

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:3eb883b27913a3fd3edc671de641ecea6b27fb29b73046b0b967a398a39c82be
size 597592

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:75a5514b55abec8215ad7352a0f6dcbcddea09e959c79cc477eccba7ed04946f
size 569298

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ128Kv128PersistentContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:16447a2a543d46f41106b8732204f71c8f623612cd170baa18c3c9ad53e298a0
size 645793

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ128Kv128StaticContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:e5df6c276ff30cfc907d874dbaad52c3dfe2ef3c070360dd479bf451f8e080ed
size 566036

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:9fbcdf0233e1e90d568e6a65431a2f0518b9a80c6d11104504a7c57f62c872ea
size 590548

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:a1ab7e6ed0f3776fbbfc525109d3b46b5ea5c1b61a2c20b9c23ff38c1e25ccfa
size 503664

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:8bf3e258ae4ec4fe9fcfe6cda40049b602efee7c2e28ae713cb93c0a0a6cc4d8
size 557494

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:548304a8e28396ebe912b24b492397d250b48ffcb87b3aa994db887c1489d1b0
size 473766

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP64MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:40600f5b624b0ca65511b9497f73447ba379eded5b76645b12bcd7ae31ce6aca
size 557886

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP64MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:a399c3304117d755906c4729c9fa7ad5febfccd08c3859a667a60fe751ce12a8
size 534326

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP64MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:1819bfc27bb26deaa75e96f89908e8909ab5496203e90fa47639ef9052cce02a
size 578648

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP64MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:1e09e5cdac9337bef2026bcd65c076454bcf5e24df52b4057604ebe606003de1
size 549564

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP64VarSeqQ128Kv128PersistentContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:1cd12e2534fdbd6b8c547b9d5a2470001f32f2ad5479764ef623d1a840dacc6a
size 625739

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP64VarSeqQ128Kv128StaticContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:8799b2fc6db3aef6cadc81bc808490bfd3b717dd07e98c774c92bbbe5e3e4f8d
size 549460

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP64VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:09419f8222def86bcf5231af3c1c7b2dc31cabc6b6a781dedc7fe0ac13db1dec
size 569236

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP64VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:c3e8f50048cd16b02f51359f3d4c709156421adcd27734cae493d2861c8681c4
size 487088

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP64VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:05cf675571b991a85bd0fc5947d13b6be85c6fc8a06b322023037f3e17421dad
size 535392

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP64VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:c4269a6c9836193305e5ed18c6935efa6c3881e59307a6788effc67ff0e7d947
size 457190

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:0924a29b5dca879d8ac8cfb3ec7fff5851cb2df33f84fd2c0a2f7e58365c9d40
size 591442

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:edafadadc70d17ee6490907f37c7a9f61fa9f0235531ef49995a6e2797c4e955
size 562332

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:8b6e4e37593e07ed92a1142cd72614c818a4722fb932e12a931e1b8ecd8835dd
size 599600

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:0dae16ee22788196ca4951fc123aa47e1cfe2390441e94dd83a8b4aff076b567
size 570516

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:43b8f5eb24e7b546ecd2f7ae7f2f85cae6e76a6d73561323323c2c4d54275db7
size 655003

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:77b9bda1e9f70b502cf47bc2b231b6116f564743f4d1749243c145edc6d3b082
size 619747

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:5bc6ad90456491137c3934f25a1da7c1197fc43f711796be2dd25f55d6121428
size 575512

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:447d21284a61438c9349a628ab346386f37fece8f9f4017517de18e6d3ab3ecd
size 506412

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:b9c841e7f8b6e8e905a599c7a83e5cbd6b03f1b4a7a0d7bb289988a469b9623f
size 540088

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:c841275d836683c14cfaba4e0e4f5f9204141dafb154f534fd6bf18161cd7216
size 474936

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP64MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:0fd309ac9206333a6058cc2ae860647c5dafbaa8f068f67670ccd79863db77b2
size 572498

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP64MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:b1d13dac9c137f97f0e90ad34b00e7ec793d3a66ba1a55f5f9fc3a640f2d228c
size 543388

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP64MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:5b2d8d80a049dfd9d30fad88dfbb5422556960582cb8c24b0b0b8273254400ab
size 580656

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP64MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:bd00c98c4c87d3afdfdadd301f8261d3aff270aa0fbf8216f4aa7fa43bfe7172
size 551572

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP64VarSeqQ128Kv128PersistentContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:7301c79468d8de004f426bb9cdefb7302b8cb77e193fddb996e648adde219d14
size 634159

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP64VarSeqQ128Kv128StaticContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:f09062978682d7add0972aff436486b6b4c82e5c1b0be271d4b1c8f0f8a650b0
size 603170

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP64VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:7cf43248972dcda4fe0334fa016a467e4566350b975fd498343f4535d5a6a278
size 554200

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP64VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:e76cbff768d6c2aff34f15f20035e3d0a9340ed58899efe112b110e5d209b349
size 487468

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP64VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:5a8afe52b247e9c27691025f10fbbca66f3b195a967944a3128b785f2473fb85
size 518776

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP64VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:f524ae3790ab0f8ac026d2c13a8edc9566ab5baa1849d4d5cfa654db91bc2a24
size 455992

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PackedQkvCausalVarSeqQ128Kv128PersistentContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:554e341207184f70b89461fd7dfaf5e4faddd05b01201cad62422c28071641e6
size 802289

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PackedQkvCausalVarSeqQ128Kv128StaticContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:990a64e7dd2326714eccb6af5e9adf8e1e8d64f510961afa127f91a748cf81cd
size 715157

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PackedQkvDenseVarSeqQ128Kv128PersistentContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:f9f8a52ff9d0f6c5ffc98c3a7f21ed2aad30f9d76fe6cac6151783b246af326e
size 804211

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PackedQkvDenseVarSeqQ128Kv128StaticContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:2a8eb61d079eb62c67fc557ac06646e6cd9031d90f42a6b36c6645be062ec453
size 715007

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PackedQkvSlidingOrChunkedCausalVarSeqQ128Kv128PersistentContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:a153012c06a150bd18fe953afd1b98707c1e666ba39dca77040874b8afac9ff2
size 869513

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PackedQkvSlidingOrChunkedCausalVarSeqQ128Kv128StaticContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:f07cec8e71786afa295285c2ad2a5cd752fe308bbc1f0a00d22a74bdec8218a3
size 782579

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PagedKvCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:b6f662c6d889f8588f8952fb1b4b037654e9a2cf42079a922bad1cc43bacd8ae
size 859023

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PagedKvCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:bb4d111e756b1ec729e605fe2eb96ff5c5d3dcf24ea15b1734490d37f05884f8
size 815947

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100fKernel_QkvBfloat16OBfloat16H128PagedKvCausalP64VarSeqQ128Kv128PersistentContext_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:f855a7f0acdeb6744d8b3bb744741884b8bddb6fcdaf4ab4bffdfb1b574b1c42
size 837663

Some files were not shown because too many files have changed in this diff Show More