kanshan/TensorRT-LLMs
1
0
Fork 0
mirror of https://github.com/NVIDIA/TensorRT-LLM.git synced 2026-01-14 06:27:45 +08:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity

fix missing gemm kernels

Browse Source 
Signed-off-by: Xiwen Yu <13230610+VALLIS-NERIA@users.noreply.github.com>
This commit is contained in:
Xiwen Yu 2025-09-06 17:56:14 +08:00
parent 65f84780bf
commit 3d4f49e508
35 changed files with 2252 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2164
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/KernelMetaInfo.h
View File

File diff suppressed because it is too large Load Diff

1
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/config.json
View File

@ -372,7 +372,6 @@
"_comment": "Tile 8 to 128",
"dtypeC": ["bf16", "fp16", "e4m3"],
"useUnrollLoop2xForMma": [true, false],
"smVersion": ["100a", "103a"],
"mmaN,tileN,epilogueTileN,tileK,numSlicesForSplitK,clusterDimZ": [[8, 8, 8, 512, 2, 2], [16, 16, 16, 512, 2, 2], [32, 32, 32, 512, 2, 2], [64, 64, 64, 512, 2, 2], [128, 128, 128, 256, 1, 1]]
}
]

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_Bfloat16_E2m1E2m1_Fp32_t128x128x256_s3_et128x128_m128x128x64_cga1x1x1_16dp256b_TN_schedS_sm100f_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:cecaa0a37da10747bcd837087ad219d44b1b8e10b131405f22f2e63caac0627b
size 386058

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_Bfloat16_E2m1E4m3_castE4m3_Fp32_sfBlk32_t128x128x256_s2_et128x128_m128x128x32_cga1x1x1_16dp256b_TN_transOut_schedPx3_sm100a_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:f91113ee21e56a59a96af0e4e286c902b033a832c59f5e157fa4b7961a8605dd
size 435250

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_Bfloat16_E2m1E4m3_castE4m3_Fp32_sfBlk32_t128x128x256u2_s2_et128x128_m128x128x32_cga1x1x1_16dp256b_TN_transOut_schedPx3_sm100a_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:5c0a9e7afc31803e49be35cfcaba07b4e168dca080e3aaa06454b3dca78371fd
size 441322

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_Bfloat16_E2m1E4m3_castE4m3_Fp32_sfBlk32_t128x16x512_s2_et128x16_m128x16x32_cga1x1x2_16dp256b_splitK2_TN_transOut_schedPx3_sm100a_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:9ae6cf61be9daa224a2258f49e4a61963a4bc9c91e5bb3283a3d78d06f3335f7
size 539502

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_Bfloat16_E2m1E4m3_castE4m3_Fp32_sfBlk32_t128x16x512u2_s2_et128x16_m128x16x32_cga1x1x2_16dp256b_splitK2_TN_transOut_schedPx3_sm100a_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:8e1af44bc4ca53b5d6a48d11637fcd1cbad0e369eeb9d78f784d29600379936a
size 610644

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_Bfloat16_E2m1E4m3_castE4m3_Fp32_sfBlk32_t128x32x512_s2_et128x32_m128x32x32_cga1x1x2_16dp256b_splitK2_TN_transOut_schedPx3_sm100a_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:173237b9e95f3553b9563a764c871394a5cb444b08a4fbcb42d9d7f234b17e4b
size 561012

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_Bfloat16_E2m1E4m3_castE4m3_Fp32_sfBlk32_t128x32x512u2_s2_et128x32_m128x32x32_cga1x1x2_16dp256b_splitK2_TN_transOut_schedPx3_sm100a_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:f38390a451c61b2ceba2adcaee127ba38b8993a6be4abf1a3396f27cfb39547c
size 632155

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_Bfloat16_E2m1E4m3_castE4m3_Fp32_sfBlk32_t128x64x512_s2_et128x64_m128x64x32_cga1x1x2_16dp256b_splitK2_TN_transOut_schedPx3_sm100a_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:2497d6cedefb1ea185e0a65517f38ce14a7aaba0f63a54a56971ba9bfdd3363d
size 615920

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_Bfloat16_E2m1E4m3_castE4m3_Fp32_sfBlk32_t128x64x512u2_s2_et128x64_m128x64x32_cga1x1x2_16dp256b_splitK2_TN_transOut_schedPx3_sm100a_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:ac4f23168a9124cd2f0ba3fd33954b7b64d56b510e0f0cb513ec2a64855100af
size 686965

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_Bfloat16_E2m1E4m3_castE4m3_Fp32_sfBlk32_t128x8x512_s2_et128x8_m128x8x32_cga1x1x2_16dp256b_splitK2_TN_transOut_schedPx3_sm100a_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:e51d2222b7eeda52c58009763bc3798fdfdc975a2a9695d0744264c4682c9360
size 528742

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_Bfloat16_E2m1E4m3_castE4m3_Fp32_sfBlk32_t128x8x512u2_s2_et128x8_m128x8x32_cga1x1x2_16dp256b_splitK2_TN_transOut_schedPx3_sm100a_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:ce9b923039ea211ebe9b73744004fd1353c98f11f6f1a3b40f96c8a14ec8b5e7
size 599884

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_E4m3_E2m1E4m3_castE4m3_Fp32_sfBlk32_t128x128x256_s2_et128x128_m128x128x32_cga1x1x1_16dp256b_TN_transOut_schedPx3_sm100a_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:e7d80c0cd70c29c70ee73df2738b3cd77f801d35d82a3ff5fdb4ed446523e6af
size 435242

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_E4m3_E2m1E4m3_castE4m3_Fp32_sfBlk32_t128x128x256u2_s2_et128x128_m128x128x32_cga1x1x1_16dp256b_TN_transOut_schedPx3_sm100a_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:9774afd2c2fad59281653d13afdaf0f8400bf44eccf48a5ad77eb71a758b76ad
size 442104

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_E4m3_E2m1E4m3_castE4m3_Fp32_sfBlk32_t128x16x512_s2_et128x16_m128x16x32_cga1x1x2_16dp256b_splitK2_TN_transOut_schedPx3_sm100a_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:ba86b5763c3362fb619cc523aa4622ee1696fc56d60aaa394c5bca26c82b75cb
size 538112

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_E4m3_E2m1E4m3_castE4m3_Fp32_sfBlk32_t128x16x512u2_s2_et128x16_m128x16x32_cga1x1x2_16dp256b_splitK2_TN_transOut_schedPx3_sm100a_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:7110a8ec78303e84d34e2df863a6a884398f0d5dcc887ab1ae1f636d0a689153
size 610044

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_E4m3_E2m1E4m3_castE4m3_Fp32_sfBlk32_t128x32x512_s2_et128x32_m128x32x32_cga1x1x2_16dp256b_splitK2_TN_transOut_schedPx3_sm100a_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:1c55df8697063521bc4eefe2c10b2d35b191f4edb4976cfa63872e86ad0d281d
size 559128

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_E4m3_E2m1E4m3_castE4m3_Fp32_sfBlk32_t128x32x512u2_s2_et128x32_m128x32x32_cga1x1x2_16dp256b_splitK2_TN_transOut_schedPx3_sm100a_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:0c3f3bce6781883773741bfcb02282d69e0d4b407572fa23848d8773ddbe6fc2
size 631851

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_E4m3_E2m1E4m3_castE4m3_Fp32_sfBlk32_t128x64x512_s2_et128x64_m128x64x32_cga1x1x2_16dp256b_splitK2_TN_transOut_schedPx3_sm100a_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:3d17cfcd238d9e486dbc3d4e5dca955b32f69cbb382a5630c9e9e5708ebb4a5e
size 614332

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_E4m3_E2m1E4m3_castE4m3_Fp32_sfBlk32_t128x64x512u2_s2_et128x64_m128x64x32_cga1x1x2_16dp256b_splitK2_TN_transOut_schedPx3_sm100a_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:809c138ec4e2fddec889c2de4bb449f114279bb92fa3f4e0eef6a2d092aa827c
size 686907

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_E4m3_E2m1E4m3_castE4m3_Fp32_sfBlk32_t128x8x512_s2_et128x8_m128x8x32_cga1x1x2_16dp256b_splitK2_TN_transOut_schedPx3_sm100a_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:1b9f0c5237894ddfdb69d172b606f28a922fc9c7390009f9e476d7c0ff677a25
size 526464

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_E4m3_E2m1E4m3_castE4m3_Fp32_sfBlk32_t128x8x512u2_s2_et128x8_m128x8x32_cga1x1x2_16dp256b_splitK2_TN_transOut_schedPx3_sm100a_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:3899c9103de4f49906dd6974d932678fd50eb7f41b57d9a174701dab599aa053
size 598346

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_Fp16_E2m1E2m1_Fp32_t128x128x256_s3_et128x128_m128x128x64_cga1x1x1_16dp256b_TN_schedS_sm100f_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:4696fe15f42eb29905c9b143bb17a60b89b69de07e796551b047531dfa37cabe
size 391626

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_Fp16_E2m1E4m3_castE4m3_Fp32_sfBlk32_t128x128x256_s2_et128x128_m128x128x32_cga1x1x1_16dp256b_TN_transOut_schedPx3_sm100a_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:3294e5f244fc40d4bae1f57cc0fd2276841551b0e0bb675fbcf4f714b40f164f
size 433664

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_Fp16_E2m1E4m3_castE4m3_Fp32_sfBlk32_t128x128x256u2_s2_et128x128_m128x128x32_cga1x1x1_16dp256b_TN_transOut_schedPx3_sm100a_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:87c347dc3d8f950ad90651d303d7fb47b16958676145720995f5be1649ad2ca0
size 439736

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_Fp16_E2m1E4m3_castE4m3_Fp32_sfBlk32_t128x16x512_s2_et128x16_m128x16x32_cga1x1x2_16dp256b_splitK2_TN_transOut_schedPx3_sm100a_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:aedae52dff5215da721d6a7968b39fb66fa14b9b9206e6ffee5fa3cf8aa7775b
size 537916

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_Fp16_E2m1E4m3_castE4m3_Fp32_sfBlk32_t128x16x512u2_s2_et128x16_m128x16x32_cga1x1x2_16dp256b_splitK2_TN_transOut_schedPx3_sm100a_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:3058ec436d960bcca34f31049d3e597cd984e927cd29cef751b4ad0ea846282e
size 609848

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_Fp16_E2m1E4m3_castE4m3_Fp32_sfBlk32_t128x32x512_s2_et128x32_m128x32x32_cga1x1x2_16dp256b_splitK2_TN_transOut_schedPx3_sm100a_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:3bd09b5c517f703e9cc82b2ba89f8aa6c8ef8ea3ff825e8980df6d2b1638e5fa
size 559424

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_Fp16_E2m1E4m3_castE4m3_Fp32_sfBlk32_t128x32x512u2_s2_et128x32_m128x32x32_cga1x1x2_16dp256b_splitK2_TN_transOut_schedPx3_sm100a_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:3712eb9cb0464e88e621daced5e071d7e340c584f0f29decc731941b7b60e575
size 631357

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_Fp16_E2m1E4m3_castE4m3_Fp32_sfBlk32_t128x64x512_s2_et128x64_m128x64x32_cga1x1x2_16dp256b_splitK2_TN_transOut_schedPx3_sm100a_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:7892ceb00d47dfa779dc0dae0dfc2485e99a741bc0c078b75060c0c2211d1cd9
size 614332

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_Fp16_E2m1E4m3_castE4m3_Fp32_sfBlk32_t128x64x512u2_s2_et128x64_m128x64x32_cga1x1x2_16dp256b_splitK2_TN_transOut_schedPx3_sm100a_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:037d4f657b96b557895ba69c14176f598ec4c664fc16ebe9f561a1982f4f4696
size 686167

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_Fp16_E2m1E4m3_castE4m3_Fp32_sfBlk32_t128x8x512_s2_et128x8_m128x8x32_cga1x1x2_16dp256b_splitK2_TN_transOut_schedPx3_sm100a_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:3760ba593cae5977cb4dd27e3cef8b75161fa97a526aab821d696e302c730bd3
size 527154

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_Fp16_E2m1E4m3_castE4m3_Fp32_sfBlk32_t128x8x512u2_s2_et128x8_m128x8x32_cga1x1x2_16dp256b_splitK2_TN_transOut_schedPx3_sm100a_cubin.cpp Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:d3d98c8f3ef2067e4898019969a1a22b357c3982303b745f21082c09695e7ef2
size 599086

3
cpp/tensorrt_llm/kernels/trtllmGenKernels/gemm/trtllmGen_gemm_export/cubins/Gemm_Fp32_E2m1E2m1_Fp32_t128x128x256_s3_et128x128_m128x128x64_cga1x1x1_16dp256b_TN_schedS_sm100f_cubin.cpp
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:853811d158dd1727a6fe9d781ca5ff1be18bcdd2afbb03868c6325836f9102cd
size 390984