kanshan/TensorRT-LLMs
1
0
Fork 0
mirror of https://github.com/NVIDIA/TensorRT-LLM.git synced 2026-01-26 05:32:57 +08:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity

Update SM100f cubins

Browse Source 
Signed-off-by: Tian Zheng <29906817+Tom-Zheng@users.noreply.github.com>
This commit is contained in:
Tian Zheng 2025-07-02 07:43:45 +00:00 committed by Xiwen Yu
parent 1b846046dd
commit 3a94d80839
1273 changed files with 1444 additions and 1850 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
cpp/tensorrt_llm/common/attentionOp.cpp
View File

@ -2402,7 +2402,7 @@ int AttentionOp::initialize() noexcept
if (mFP8ContextFMHA)
{
TLLM_CHECK_WITH_INFO(mEnableContextFMHA, "FP8 FMHA cannot be enabled because Context FMHA is not supported.");
TLLM_CHECK_WITH_INFO(mSM == 89 || mSM == 90 || mSM == 100 || mSM == 120 || mSM == 121,
TLLM_CHECK_WITH_INFO(mSM == 89 || mSM == 90 || mSM == 100 || mSM == 103 || mSM == 120 || mSM == 121,
"FP8 FMHA can only be enabled on sm_89, sm_90, sm_100, sm_120 or sm_121.");
}
@ -2410,14 +2410,14 @@ int AttentionOp::initialize() noexcept
if (mFP8GenerationMLA)
{
TLLM_CHECK_WITH_INFO(mIsMLAEnabled, "FP8 Generation MLA cannot be enabled because MLA is not supported.");
TLLM_CHECK_WITH_INFO(mSM == 89 || mSM == 90 || mSM == 100 || mSM == 120 || mSM == 121,
TLLM_CHECK_WITH_INFO(mSM == 89 || mSM == 90 || mSM == 100 || mSM == 103 || mSM == 120 || mSM == 121,
"FP8 Generation MLA is supported on Ada, Hopper or Blackwell architecture.");
}
// Check requirements for FP4 output.
TLLM_CHECK_WITH_INFO(!mFuseFp4Quant || mEnableContextFMHA, "Context FMHA must enable if fuse_fp4_quant is enabled");
TLLM_CHECK_WITH_INFO(!mFuseFp4Quant || mSM == 100 || mSM == 120 || mSM == 121,
"fuse_fp4_quant only supports SM100 or SM120 or SM121 devices.");
TLLM_CHECK_WITH_INFO(!mFuseFp4Quant || (mSM == 100 || mSM == 103) || mSM == 120 || mSM == 121,
"fuse_fp4_quant only supports SM100f or SM120 or SM121 devices.");
TLLM_CHECK(isRoPE() == (mRotaryEmbeddingDim != 0));
TLLM_CHECK_WITH_INFO((mSM >= 80) || (mType != nvinfer1::DataType::kBF16),

6
cpp/tensorrt_llm/kernels/contextFusedMultiHeadAttention/fmhaRunner.cpp
View File

@ -84,7 +84,7 @@ FusedMHARunnerV2::FusedMHARunnerV2(MHARunnerFixedParams fixedParams)
: mFixedParams(fixedParams)
{
TLLM_CHECK_WITH_INFO((mSM == kSM_80 || mSM == kSM_86 || mSM == kSM_89 || mSM == kSM_90 || mSM == kSM_100
|| mSM == kSM_120 || mSM == kSM_121),
|| mSM == kSM_103 || mSM == kSM_120 || mSM == kSM_121),
"Unsupported architecture");
TLLM_CHECK_WITH_INFO((mFixedParams.dataType == DATA_TYPE_FP16 || mFixedParams.dataType == DATA_TYPE_BF16
|| mFixedParams.dataType == DATA_TYPE_E4M3),
@ -313,7 +313,7 @@ void FusedMHARunnerV2::setupLaunchParams(MHARunnerParams runnerParams)
bool const isSm8x = (mSM == kSM_86 || mSM == kSM_89);
bool const isSm80 = (mSM == kSM_80);
bool const isSm89 = (mSM == kSM_89);
bool const isSm100 = (mSM == kSM_100);
bool const isSm100f = (mSM == kSM_100 || mSM == kSM_103);
bool const isSm120f = (mSM == kSM_120 || mSM == kSM_121);
// Sliding_or_chunked_causal mask.
@ -382,7 +382,7 @@ void FusedMHARunnerV2::setupLaunchParams(MHARunnerParams runnerParams)
// flash attention tiled kernel is faster on Ada and Ampere derivatives when head_size>=256
mLaunchParams.granular_tiling = false;
}
else if (isSm80 || isSm8x || isSm100 || isSm120f)
else if (isSm80 || isSm8x || isSm100f || isSm120f)
{
// otherwise, choose tiled kernel for Ampere/Ada/Gb20x
mLaunchParams.granular_tiling = true;

2
cpp/tensorrt_llm/kernels/fmhaDispatcher.cpp
View File

@ -42,7 +42,7 @@ QkvLayout AttentionInputLayoutToQkvLayout(AttentionInputLayout layout)
FmhaDispatcher::FmhaDispatcher(MHARunnerFixedParams fixedParams)
: mFixedParams(fixedParams)
, mUseTllmGen(tensorrt_llm::common::getSMVersion() == 100)
, mUseTllmGen(tensorrt_llm::common::getSMVersion() == 100 || tensorrt_llm::common::getSMVersion() == 103)
{
if (mUseTllmGen)
{

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_103 = 103;
constexpr int32_t kSM_120 = 120;
constexpr int32_t kSM_121 = 121;

2
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/CMakeLists.txt
View File

@ -18,7 +18,7 @@
file(GLOB_RECURSE SRC_CPP *.cpp)
file(GLOB_RECURSE SRC_CU *.cu)
filter_cuda_archs("100" SRC_CPP)
# filter_cuda_archs("100" SRC_CPP)
add_library(trtllm_gen_fmha OBJECT ${SRC_CPP} ${SRC_CU})
set_property(TARGET trtllm_gen_fmha PROPERTY POSITION_INDEPENDENT_CODE ON)

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:e06195ad4d3acd230b8bc1d66b689f1a58dd48370c3b7cf6f055f9ef34fb47c9
size 1577653

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:788a02f89f4cd29ea535d4d6b8c5c0395a86ab84554b229620123b1030a3d99c
size 1477991

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvCausalP64VarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:07822a0005c071db6f7d504d8c1b159913cc5aefc7b617ba280efa5e1b2e28cf
size 1568033

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvCausalP64VarSeqQ128Kv128StaticContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:f04d446486d7d0c20a06a954d5844b3693f03ce3dd89de4c49b8e0b36edb6e7f
size 1465263

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:db0e02eb9c68788c4ebea760b2010e79b5d13e06733ac1ed3f04c43738904b8f
size 1118485

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:b8972ae5deaef930cab70df016f0555dbbdece0309c7e58a67b69c9b957b7006
size 1055583

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:dc0b0c64136781dc1e5c9bf35a0340b9316a1d6a660d73567630523ea6fe2d55
size 1119417

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:9bc2b2256618c802dbe58d08ec929361a12bdd43e6e43c6565d081e7c0f8c125
size 1063717

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:1a62ded81591bb72ed93c9e6e7245d55b372e13325a788a59b4b9bd89df7b9cd
size 1643905

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ128Kv128StaticContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:3e1d14b3722762fc24eab015e69dbd63988854f1e503088ecc7ac5d5209de925
size 1458601

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:77b29f9d36821556fc0276ea3cf1a0e137138ae23da00fc00c24f54affea328f
size 1038447

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:3bff1274df8545acfe965366a58492498a7ea0d27840c66cee72c31771c6e54d
size 994877

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:3e3698ee6f1c0c12605faa20556841fa6e42dc8fa2ea979cf4e143d0db64f4bb
size 997251

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:861c4eeb7673da56809ed671b3ba72ccc9697ce255add3339209726a1ba0362b
size 955063

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:62ecce1491ebb4f7f26763547e504e941e56063ec10d13d5739fb58eb28c290a
size 1105807

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:a3bf50dbbf2954c63b5d4d1ddc103f247d5270d2a330c04b649f4741020e9d4d
size 1042855

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:5b51e99e1e17f8066712011716b7919163d1b4778822b16edda363da6b211fb3
size 1106739

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:f43d9b28fe24fd45b11af21b87aae3ddbcfebf8c36d5172abb827dee0f2930ec
size 1051039

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64VarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:9fd087f9b858d133491899e71749455c98cc46a45e47adcb62c35ab08c79489b
size 1635271

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64VarSeqQ128Kv128StaticContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:03b282c6fdcc15e8aac9c17638cdfa55bd37392d3bbbee0969270f795a29fa4e
size 1445923

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:41c82230d693b60525150cf0beb2e42138b90b207879b198e02568c9799a8bfe
size 1028827

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:deeb830670035939846b3d30015b340987e0369453b6dc7cc87ac0e06c45a006
size 982199

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:2898d04970fabca72c62bd866f7f6b98bf41f4b18f2091e28e5b250877589529
size 988421

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:2f7fb4ed37994a43b9026bd36fbf71415675a77f5e2d205036190cf35a13ea6a
size 941547

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:f0ac29509b7cbab6ca7423c840c6bbc4327ca2fe55a51c05bfa95c2151697c20
size 1152115

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:694aad35ea06d09a371e61b77323196bcfdc31cdeef1843015ffeb47e0e82e31
size 1087585

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:06c858ee356f26823de82481687d13f8935cebdedcadc0b20ac295c6b5e44a34
size 1153047

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:418863319fdfef15b568d6232f1de36e3ac0a34bdf687f5bf6d1d74c221853f3
size 1094979

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:e418038e657da4d89251c46e1c634868ec49f57804c55df7c2089a6aa54f2a3b
size 1874375

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:60688f6c0b1f05ebddc1ce3ad3c57a800edc60c14ad93d19e21e773af9f0b287
size 1771753

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:78ad6c400bb6363f20b4030db6bfb13a000409bc2e70182a9af3136ce12b6eb4
size 1060631

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:88f89996338c6588dcf8bc4cc62cc97b790e8d8be13533b5ac9d3f69686fc4a9
size 988547

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:f0d484bf43e73b6d85b5775231fa41ffb90ea6c54b1a1ddc4ac10ab7d574773b
size 1018401

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:ddf2175f5c980646fc9f7a56dbd9602402d64394ef6d5b208595b459bb8dac1b
size 945675

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:8f3d1afa542c0a25dad12f634d6ebcc20d246736b88aa65411d71afe9b212b90
size 1139437

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:eb7553fc02e30f70efb13045eb3ea010cc4dc1f5e54be0b6be9d3aeda8188702
size 1074907

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:0b8cad736bcf5303d65a896b795a5933effbfb1d7d781acaddc537c1e0f064af
size 1140319

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:0b5e81443bb67983eb8b346672bc742574aa6620e7956eeeeab08dd1208b31c3
size 1082251

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64VarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:493fa76b2ac5e1560f87d8c287abc00c750be790e5ac7f7118e7a2dd3377149a
size 1865593

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64VarSeqQ128Kv128StaticContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:b2ff066dc8cfd93f50b7e8819bfc00bcc67be126e178c190719011f0c7c5f51f
size 1759075

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:1a962367078720a56c9d5b2ab3ddb04470f43741bb99c13fd1fe7604f39aecbd
size 1051899

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:1f0c2bfbc2da95afbd8749795c172b6b525612b2ef85b12dc2dcf63ede90b86b
size 975819

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:899df5fc4d618c374300552927817fe2da06ff6325a09aef1e631d4251fe31bf
size 1008879

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:287213dae3d07b7a831df298d2273f02b504a05b0c4321e1378ae09fa243f5bc
size 932207

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:3a547e6fa5ba7af37f0d4b6402642ced535517c88bbcba612de877a387ea251a
size 1451455

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:a5d0cda417dadebd9d6a480f337a932b083974e298701bc30203019f2f166e0f
size 1409071

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvCausalP64VarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:2cec68eade631c09f7678671461bea10938f265e8c28bf7fa64a6ba205b846ed
size 1441835

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvCausalP64VarSeqQ128Kv128StaticContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:59a5d52c62236962537fa76e491b6e4d78569d54c0587d8dc6662f7697f48f1d
size 1397181

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:f928ca9b335066e3ee55b8dbc5458217353ea2deb776a633cb55f83ef431ffd1
size 897963

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:9519caf4fadef08536696d8a5b9688929c9b81ef904f6c7471de97bafa3032b8
size 861109

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:cc6e1288714eb1fa66d58a41b4adac33c1546d9f1ad98b46b5788c07821fcf16
size 924301

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:c019f739acc0e41f849ecd59864858e4a365a9f27ae33a760a3cc01cb9134c02
size 886609

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:3f9bb909fc8684ec0e9980a15e02c7708f3ca72ea850a123adf2e4842a1057fb
size 1517313

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ128Kv128StaticContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:e6c33ebe1e64cc85f4bf1651533d31672a4edde2d1dfb10b15d2659967eb4b25
size 1405121

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:2bc3ec487d845371a43502d85dc79a286eb63385627bd62aadb4520851ee6ec7
size 949299

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:515554943613218a7b860837e9391f3dad62009806998106d1a4588fa6eec3a0
size 804351

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:7a2292f450077c1c47f19103144292dd4ef251545c756e377e018a3f2a035d2e
size 905489

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:5a5922553ad55cbd5b844dde8b4f5dcea80e260c4280a0b7e4a8d77909b664bb
size 766953

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP64MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:4981bf68d89348d819dd8b403f40b66b3b8efd2fb982d2bca8587ea8835576f5
size 885235

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP64MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:0bf319e9ba8550360f5b6fc7b950c34f794d358a1ff5abf6ae7d6faec4aa19ae
size 848431

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP64MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:a920f53c8c45330e5a499219eb0885bd65c34aac0d083e520f457f801a8f97ea
size 910785

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP64MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:ba131b816304ce2c3e699da3d60425b47e20cdafe52d8377e8b154fbb3d1f3f4
size 873881

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP64VarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:6a5848029ec2a1ac39e05ffefc30cbf119b09b20cf48739d0c57fdf69e85662f
size 1508483

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP64VarSeqQ128Kv128StaticContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:76d2e5db59b62a1e71f6c19eee9d53c3e5aa54140f2327a35c80031929975169
size 1392443

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP64VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:dcf749a1da698564bbbb5b56c1ba09b2fd4c87d31e159d27ee8ebc372eea6a30
size 939779

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP64VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:c709448d3f06076f603e57de9423e47da538b452569739739c81a7339378cce3
size 791623

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP64VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:4775a18bc7608c762aca645f9c8ff8a93ed3b55d8d9c50f144f4d976ee730ec7
size 895869

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP64VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:f9dd89f0fecad3a6c1b89ddbe4c3bc77d175bd38f1f023105363f11ca85788cd
size 754225

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:e950ae2b1e9988bbef846d8b14956ad8f380f4ced8cc68a56a38c2faddefbe3f
size 927203

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:73fcfc8a1a4d3a66d05c9fe4d5c3f2456453b12343defab9ac10ff1120fc3cba
size 887191

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:bb9f3e56b3bebfc14cadf048e8ec80118d3a0262c1ccb0c75d9b30af6c94c92d
size 956995

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:6a0d75e1b8ce22903de2c079a931b0a5122ca12d25149646b6137a5f322f88af
size 916145

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:1e7649c8fda314b4f912c8ade15359d4bb76974ecc04091ec9a1fd55c2321960
size 1744133

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:c224e54833ddffa82f6435e37c1ed3a1e4ac69a11c763e6a58b1e2c2c99b7f1b
size 1702883

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:e83f92f8f11a7af5d1b7badb0bce88ab06793de9f0b844fea400cd3d8aee332e
size 973705

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:d51e86537c3a662387b424daf61e9c5f6743dc097886114c32c8ec15dd47c944
size 818297

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:6b397bc5be8e05479f944cb4eb4bddf53c0a04f9fab703c29f131e6927a2c1c6
size 926491

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:0c40260319ba8665d205091f2630332cbdf7a7e641b59153723b9c9cd33d965f
size 776953

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP64MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:7921441bbd2d86c753df503ade50070cd3161b3397d9fd4bce34fbdbe0aab324
size 913735

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP64MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:a60a75f389734340b793d822371c082e5dfa3c2eb20765eb2b6721fb00c704c7
size 874513

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP64MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:1bc8bda594ef8e5c7f14fea6c89fa9b1dfd665e71f7427e7fa0f74fa3a05ed72
size 943477

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP64MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:b5002fe1507f69a376f6349497df6b0af7861d17b6f38643bea4054432e85ad5
size 903417

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP64VarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:7dfb3b2146b9d738f13b2afb1357f6ed993b356968bd90597122b13e581824f4
size 1734513

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP64VarSeqQ128Kv128StaticContext_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:6d0bafc1ce666303be6e95407e7b133d3eb45f0f26c691dd0ce6ab7dd663e061
size 1690993

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP64VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:d4a040f2b919816549f16ff42fd08372f4c410f473b27703156c4e10d54c5d31
size 964973

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP64VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:e7d25ade9216b396f229f363ba31fee9b80fd526dab279d40d80a17933650798
size 804829

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP64VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:670fc4574b61bf1f47be254ab5d8b4dd5f24807bfcad8faf6dd1228066bc7220
size 917659

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP64VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:a1bca38ead9b35d0fdc755b757dd63b19e4c82aef0469d36139208d0bff23eb5
size 763435

4
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QkvBfloat16OBfloat16H128PackedQkvCausalVarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:84ab0e7ff561025efc2522cdd9199c4090a694ee4f5b76b5489c3ba8d1d00ae9
size 1325071
oid sha256:008fff06e1c656ddd2fd9e31d2d6fd3a27bb0a60fc98286914cb545c4b873241
size 1325861

2
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QkvBfloat16OBfloat16H128PackedQkvCausalVarSeqQ128Kv128StaticContext_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:064665f69939210694ee1c8826381d11993503f58e17471306acc2955337f137
oid sha256:d8a0e87c9ac534472a73eb22170159bbf53108a1bc5b6e6bae4200460e4ef62f
size 1194873

2
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QkvBfloat16OBfloat16H128PackedQkvDenseVarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:c2242a3a3860ca517aafaacd92a64cc5f793cf1f7a17bb26e3fd4b699994f5f4
oid sha256:65ad7a0e3272713469a244a36cf27d8158d20ca1d4193ab44d41c2b8ada29a76
size 1321665

2
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QkvBfloat16OBfloat16H128PackedQkvDenseVarSeqQ128Kv128StaticContext_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:cc09f17fda106c34d50c6554ae4fed44ecb7a304b4ed9b9042b7e6482872e024
oid sha256:8d5de9e0e756c983ec789f72a3acd1e09898c3d589ff93273257b60bda61c7af
size 1228763

2
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QkvBfloat16OBfloat16H128PackedQkvSlidingOrChunkedCausalVarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:4d609266b5d33a5520a7226c86466abcce56c8c6ba9540842eb54caeacd457b8
oid sha256:07ab803a1c52dcf6565cb0fbcac6905dc05f3d3fcead47ba9b5299df8ded9bb1
size 1606895

2
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QkvBfloat16OBfloat16H128PackedQkvSlidingOrChunkedCausalVarSeqQ128Kv128StaticContext_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:d8fa687c4827ef1169d3e31ef8840f7e3c666df2ebaa7d88a713bc10ae0b823d
oid sha256:c2e0290c42041816bca27cb316195d94f2ba87e890aa714aea5c524c16754a96
size 1476253

2
cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QkvBfloat16OBfloat16H128PagedKvCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp
View File

@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:5f974de4939f6e2a51d9fb3af9de07ce4192e1eed6e0c8dbe42251de20a7c42d
oid sha256:fad5e6b2fe34e5d4fa85bcbdf7e6de8f48126447861699ecdfeef2aef80c5e00
size 1442339

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