[https://nvbugs/5727952][fix] a pdl bug in trtllm-gen fmha kernels (#9913)

Signed-off-by: Perkz Zheng <67892460+PerkzZheng@users.noreply.github.com>

.pre-commit-config.yaml

@@ -1395,6 +1395,8 @@ repos:
     -   id: check-symlinks
     -   id: detect-private-key
     -   id: end-of-file-fixer
+        exclude: |
+            (?x)^(.*cubin.cpp | .*cubin.h)$
     -   id: check-yaml
         args: [--allow-multiple-documents]
         exclude: ".*/gitlab/.*.yml"

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:93da3db51dc2c74c43549079c18247be8c423165c425be80c9b4171624c8a1d7
-size 646191
+oid sha256:0afc687e183286972166696ff33bedbe8dfa8bc6bdf3213d25ac909e5f3040a9
+size 609323

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d49a0d4280f06012123d03c1f19b2c21eb0884b4621530741c23255c5b1857b9
-size 590830
+oid sha256:ec9b5930e50e96a57b1c4cdb147dcbc54a484d2a8c8b25fb3587f4bc0378c12c
+size 545131

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvCausalP64VarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:abf634eaf82d7c72b7202706c41bbbf1027c049bcd78e899207c77b344b5d47e
-size 626655
+oid sha256:2ae138ce71c60752175fb660cb30f1633028a0cc3592a30ea2601f50394f87af
+size 598271

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvCausalP64VarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:79083440bb77f1423d1b3c33e6ff6e8434a9e95fb803f7a2dfaea1f972eb62c4
-size 574254
+oid sha256:8dded9944d41cabd7857a6b5ef229e05a843c191774997662029dbf09026a2a3
+size 534081

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:56be9d2c764a380d45775cb93ebdc59cf6840267a695ad7910306c316d6e5a6d
-size 490794
+oid sha256:0759014bf3c82064f8be83eb50850b8da2e90b267f41cfdc969224620c16c6a4
+size 460315

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:27fa306c15c04506f2eeb49456ad1920dc3a0ec06cf142f82dab0d3b6d8a5e51
-size 457936
+oid sha256:3eac32eb1df49520c891b1183e152875e9e22e36fbda48e5708aeaadfd4751a3
+size 429011

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f05b47b2541185087b3d3502a98b6e1fcc50f10ea4f5fbcb173f0a7063392a7b
-size 481636
+oid sha256:090271f2b5789f2814042c9bd0544d4cf52860fca9d295535d1c5813766da6c9
+size 449579

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5b7ad324102fa900c47b2cb8feaca95cd332b84eea0e6ced413d7da86103ff8e
-size 454304
+oid sha256:e48501fbee77407818bd30591048da1c6ec16bb69e358b0482ae023c34a3f015
+size 425403

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a140c2e2014fba58b2789fc18349a9d8ccd361448ff59179ae5839daa414e477
-size 638789
+oid sha256:221938cece8855eb304e324d4178be760c242d5d79688c1dfdf733a32dbcae64
+size 605373

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:154807c5077953c2a5585410562d2a4f5ebf0b1b014f4a32ff70e1b65b193a1c
-size 586906
+oid sha256:40cac8f1f30e356ff10e785dbb8c221928d0458499032ccf00b4e01d5be23010
+size 543625

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:983c3043a48bae92ac6e564c62041d37fbeb5995bfa5e1cf87030dcee8f63f17
-size 447238
+oid sha256:a3f6fa1bfda76ab86d3efb46f8a1e12862661a59e4b2f66f158552fe8c449a02
+size 428597

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:712e5ec1c49ccfc014430442ab3944544bd32724f3372956afef4eab55548bec
-size 396366
+oid sha256:01c6243daedace544fa3503b27fed667b312946c91c0546a5c11f9bad1a26193
+size 369833

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6badebbedf2079dc2f50340cb401545a1592de7093c79d3e08c3d27fa1191391
-size 429376
+oid sha256:0e1285d9312d0903da950160e1f4b7da9f77a72d6c1483a71c09566c32f7be22
+size 412315

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0480d9b5c03425111629b47a6660b071e2b49b70417a56106d01d077a4aa7a4d
-size 378506
+oid sha256:811b96b34e0922e2e87716e2ecead2d5827245e73793092aac0216a7e37f6cdf
+size 353551

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e7051e49c82e0d10102950a887d639c218b16e20d266cc3d7bb702154a299339
-size 474218
+oid sha256:72673296d19f9a875c19be8bd47a17d2e6d93889fac3ead8ae5754fe4b6100bb
+size 449263

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:79a038482959da988e9ccf405591b1e5bc2f02213f3bd3590f44f32ca1eb36ce
-size 441360
+oid sha256:1548dfb89b0ebbba33b02c73a6ae2ae79a34768ff7bbd09a78e5d1123cdac990
+size 417959

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ce5c3c5fc0a318747990414fb172712b83c28d0ed501309a7cd2fbe786b67f9f
-size 465060
+oid sha256:b556c9490648423f0408a2f33461a73a53694ff3cb353ac3f7b897126e3ae0e8
+size 438527

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a0643f9373aa0a579efbd7b01b967e535534f6c123057e954a7eff2bef1d2933
-size 437728
+oid sha256:480b93ea21a7b8fae2a1e0e859812de21f6e31c93eba1ac48fa5e94cce9e91ff
+size 414353

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64VarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:40bde6ad280f45eb133eee398de57aa841d438ae72ada024a77374126d30faa7
-size 621917
+oid sha256:59ce7a5a8f3310ded18a2c1f06d7209f378bbe33d2c896b74ff39be782bc6ef2
+size 594323

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64VarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:72f04b54572591e5cf0539628ba89b2ae45fc43f38f764590b60dfed291a7415
-size 571910
+oid sha256:5454594d90ce70e3e307c1e13c64f9172447efb1fcc9121ae30febfdcf309a7e
+size 532575

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:133585e8799982a224109cd76c24057f64c570c5815a163c75a391ac2b2ff6e6
-size 430662
+oid sha256:b08910e66cccb8f148cf5915198ded637fd1d114d5eb6e8fecb9181f58d52a59
+size 418337

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:10dd0e900165da3946a290fa7644b3965c679b85e4512c836b3cb6f88e53a53a
-size 379790
+oid sha256:05eb3611265e95f2b90110c1b891b5ce634ba7e00436f333a18e21638f92eea3
+size 359573

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:97bf0b35d7adcae83e2614b287eec3477407910591747fb9bf7a8f9d2e4100dd
-size 412800
+oid sha256:6cbbb2adba9cd8c5f344146b8aa736e386e1349b2f6a885ec9f16e47d2babd70
+size 401265

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:38c475483ad75da7821f3a907a4f2f0cf0bf57630f8908ce035daa75b92d79e9
-size 361140
+oid sha256:cf89b20f4c91ae06d7835c8293763bfbcb80fda4aa1272fc36236a70879f5632
+size 343291

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:065d4ceb129aee11bced17a64c72d8bdbc49213705e70f6e2fbbf9fa038533cb
-size 505826
+oid sha256:e2b3f95727f9d0b67f23bcfe3fc6e5a70f29165572216da801563d40569fc45b
+size 479293

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f8089c7c268bc3a172f4cb14a6ca3e670fa50ca3b3fcbd329ba86f9f9c9e175a
-size 471388
+oid sha256:46db4b56250d0c57cb59ce3c63219b84e9829663941c395cfe7e8e27c3ffbc85
+size 448777

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8a1f079dd99ee5f17825c71799f6a7a0b36e89398e16a9b93ee23c9923ad05ca
-size 495878
+oid sha256:7a623c03909951ebe246ce9bca4fedb8f9b51fb6b5cd4789f84052bf83c6ec6d
+size 469345

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ed86e9c733c048bd520786a10cf1620bc5c3071d4d84532ff2c904790cd60b02
-size 466968
+oid sha256:8cc2f9a0cc088cf90ad3c8dbd7d2e3d7afe727566f3f51e2c003713e147e7ef3
+size 443591

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:96f9657094acc07b98d2ee41bf79e7e397a2ff1a66254a2d70040d1f12894a92
-size 701575
+oid sha256:6fd8dbe74d29b1e5823d39ffa5cb970796100e83dfbe1b8f57fd53059c700378
+size 647342

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:51c90f7a06bde3f283e4fa78ede68f390a2d74a1b742827ed28bfa53ba0af38f
-size 644193
+oid sha256:e718f24c5621e72fd733ab92317e5865d233ed9c6abd8a23e9aa1710ac5ead3b
+size 583595

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8d5f6eca62c7644bcce981f107b09dd82e95d3e1f8b5a1dda38fd626db0eb411
-size 481878
+oid sha256:4368ec5ecfb8fc89c992eb1a5f7044b35b3cda654fe4da7fbeeb7c220bcb8296
+size 452213

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1aabf34bc8b83d18d0c34f14bbb905247e02e38bb5e423a1e040fb3458764e8b
-size 400224
+oid sha256:5e8c2d9ff07f0342eda303d760666a14aebc726455e1584b5b618a970eaeeef4
+size 383163

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:327df4bcaf3e7bff22742d558bfba0e0fecb7ab3da6c79a82a74665d357385ec
-size 457704
+oid sha256:a22afbc2465200fbedf95628d2f1bc6aefad7bcdccca912901540792f1735297
+size 432773

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:57df5394b6e47321c21553a330b45b4003e127196103eecc23921ac165f83d13
-size 380784
+oid sha256:b9b1831e5b9d3fb25a8d11fe005dbaa8f6cb722260c5eec24c7ceb327d0b98d9
+size 365303

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:60f3586c65a1c05a88e0729da3bb6e6cbc48f6bebd2566efefe82b4bbe1e2cc3
-size 489250
+oid sha256:c1925ff51f32cd31e21c81d8ccd1722a5e45178c06240d13b0bcb83e3a94fc5e
+size 468241

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0af68c71338072381aa01878db5a816156b518109a058331928d26d0e100548b
-size 454812
+oid sha256:c902c68ffd17c3c51d21b5b118021418210e0bb0db24d4320e4aa4b1e4e13af5
+size 437727

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:95e0c23f48a26672e87787326ec4c21b2d88203a7ae621da3db7eb4c0e5c38ff
-size 479302
+oid sha256:2faa1170c9ee7d6fcdc71b49ecad3d89466d25cd45b81e878875bd953ba82a73
+size 458295

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e9d4b8fc32aadbd4ac6a3e379fb72cd5b6be96977205b301c0dc6468a3fd5274
-size 450392
+oid sha256:cbee7ef245739a75f79c846a74a6fc55119c664f03f4545a41f6ceebdaa33573
+size 432541

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64VarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4593f7e2d25319baac5ee97eba18cdbe22c5a3dda77c6fafc0fbc4c16f163988
-size 682039
+oid sha256:c1795aeff12231948d225fe6bfb39a98a459b8517ed767353a6b175e2c598b02
+size 636292

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64VarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1b812ec9d0a8259418346cab34054a359eef4bad1d97abc6ae2f34c854244f92
-size 626827
+oid sha256:6d004a3c852e3024ea7646b04fee6ddb86a46aa2b2076652b780ed59db9110b5
+size 572543

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8afe73f4880ce3db60376c3da8e23b7d1d7b9d43243b8f4dc7242fb6f8b0c56c
-size 460566
+oid sha256:ceb815e908297617c3cc4277b60ccc97c03796e0e810c36c3bc6d43755dcf471
+size 439583

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:165f4d266232b757878fb4ddfa5e9ca8b6eaf859de430d9db34aa94bcc26753d
-size 385226
+oid sha256:d3a0454ef56f20bd64689b03152e6cc62abc7c3dc9ddc93d763bd2727902a3f9
+size 372113

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:be8d1f839c20ac0404f901064a74e1edb22a9a4133f8291799321849b19038f9
-size 436392
+oid sha256:844b18b29e3f4bec4f50cdb9182d1b7a318143fc6eaf8c54df08ed7613713f22
+size 419355

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:70ef01714ae6a43a4ec4ea46810196cdf3c8a3fa0de254a21f86709761da3c74
-size 365788
+oid sha256:14fe3afb9af41be719b48de9737c94dcce14f3e79508d64e436c5f0c062155f8
+size 354251

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c5332e744cc896ed0806d76a401ea1d6375d04543cdda4da6cdc39d3bb3ca756
-size 687015
+oid sha256:c1173104b88f6b54b8b1998fae637cfa007e2a75a4caab87a4134b6ffa495ffb
+size 638998

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1bad078994fbc0138bd20118fe1a58762e6bda82d853a12d8d57cd7494884ff2
-size 596282
+oid sha256:bb18f305f16a28295c83579c36310280a165dba85c8b117fe4e93269af3cb936
+size 553617

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvCausalP64VarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a797733dbbe12b8ce9e88ea899218cccc835fa9e1aa7a59ced0cb2c24b4e180d
-size 648437
+oid sha256:32cd965d5de74d0422d160b581f24e682968d6d532db82cdafcb90e4bd80e600
+size 617685

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvCausalP64VarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2954719c224c9f41c4d49846ec419ef65914b32f33d3549487d6c7cbfa9b218e
-size 560762
+oid sha256:7455e79134289fbd542b7f7abee3c7ae1273745352e7e3ea53740591007eb08c
+size 532305

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d6a1ec3025b9d8e4383aa0a5d6ab4aabb7b4722960a18a3b4570998a9d506236
-size 624981
+oid sha256:9f067d5344d1bdfef2e86308b188c2ffeecc756193ec0091b89d2ae1fdb47c40
+size 570821

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6984c53944d4a207d95acc63936d535443561ae6b957be87607410896cc6ca49
-size 597648
+oid sha256:7f3fefa50f011446b3273bf09ada4d63fc15786b649b938c2454acd1e90f2f1f
+size 545831

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a37fbe08db8771b4c234c2ecd1a74655fe71a821c9438eb648afbc25b68e0ca7
-size 614244
+oid sha256:4ee27b32516fde1d060bdf0aa640277f2d261c16db874a733be071d4860bbacc
+size 557717

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3243364468a0d44768d2144f2cb11e50ff408cb8b1a86e4c5a6150841c8042df
-size 591648
+oid sha256:09b3ff8afc266cbb50a99bbbc97731500d0dd2d413290fac302a4977d700a0c2
+size 537489

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32VarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9a9a4821030a08be606feb8b5c581561b7287e4aad21daf5aae286596e24d8b7
-size 691255
+oid sha256:fec44028cfe46baa466c2a0e6b1a36a21e4959bc17ce0047e0d81eaaaf7f975b
+size 639588

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32VarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e1d02f535560d7918d1da51a43a17a98713c9e9712c1af075606684efcebab62
-size 637499
+oid sha256:632f92da0d61586cd4bb0a780ca01a10eb82745a8b7c53ad02c1689f573db536
+size 587679

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ad2341ceaccad4f99a9fa77077c73ea12d51d782ae87f2c7657b297ce3a7e0ae
-size 572742
+oid sha256:7405065c7c7dd6386c40328b674202dfb0f5d7a336e795647e17be1586a75178
+size 534369

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:59b0d398d37e92de77f29095a0a32658ba48387a87d791313c1470a25c125eb5
-size 523450
+oid sha256:08a03bc885d1b000358e43bdf79aad6785b1ac9d9f9ee058aa7a2919d7fb000a
+size 477183

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:338c0f4bbb814dac0a15e39cbbff9039bdf869a231212d57110cc7c0eb051609
-size 552512
+oid sha256:de8703173c3277546fca9eaa2a78bd8a9864ad3a3633cc8deedd31ba88985fa3
+size 515719

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:102eb7cc918b90a7830e2c0e6931312d63618a5f4c60342a30e9dcc81a4f7fdb
-size 503220
+oid sha256:f2ebe23bf99ef53c48f8d868d7346323cb76416259b10201e254bc6fe849a699
+size 459323

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP64MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cc11fc72935a9f154c0fde530c448d30c3888ce5c490073fb951fdd9bdef7535
-size 587092
+oid sha256:06f08ee0709815162c8f36b4d9db4498dbcab01d74df002f2c483eb1df9b8499
+size 549509

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP64MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f6942c1f3f93e68c4e7cdcdecf6c5d8dae8bba5f2349ae106fa11fc9c82df6df
-size 559760
+oid sha256:48f0343461f9424db63327346be6d760f99428a5627a5da5652b22a1e96b2963
+size 524519

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP64MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:12c9e9c5de52e5957734c58fea720e44a8a8b6727342efcd7bffdf84080f392e
-size 575568
+oid sha256:e71fd1ef2676d0c8f3787c805b1d180da49efdbaa1b676d0239a723ee8b804f0
+size 536405

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP64MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:30f31afd6adbdaf917014bbfd67d3f2f0b2c78612b43ec687aa15152034ab2b8
-size 552970
+oid sha256:9e71e703fa71c56074ff88720e1073a12526247b76f570b5d4e6e8aa6eb009cf
+size 516177

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP64VarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2dbd0a4fdda0f1b27752494513515ff6216cde3e6dfe1564848a4ff38f4464d3
-size 649519
+oid sha256:1baa033be11890063bb73b85eadacda7ab091eb0d7cdcdcb0953ed67d1a556af
+size 618276

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP64VarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5a4479ee73e40e0fc7fdc2d9a99caaa5277e4bbc5801d73ee130c6e9d10fa4df
-size 602768
+oid sha256:4584edb326b07a05ef0c58028c3b45e88e8bb5e020586c709a4ea2af4a8d8f6a
+size 567157

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP64VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8467693421a3b6e1f35c49968dc15bb4e73b46646a9af2a00bf5e2a4027c44c3
-size 536432
+oid sha256:6a1914c862dd23dc77f71070c7d323b4dec4467dec31149b0075de776e529d70
+size 513057

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP64VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:11e8aee556153ad65fad19236f50cb87c1b0c9c518a56612b4a7579bc177bd74
-size 486350
+oid sha256:3d7a75c385d355c4713b5e5858f43aeda62863d3f104d71cf78a3d3f73cc81ee
+size 455871

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP64VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d7d5362499c29096ee4e0c7d1fbdfa2c1f836ebdbfeb8092a9ef18096a054035
-size 516204
+oid sha256:53e6c7dc373eae865ea65ad6251cb44970b6a83f62da9aa2170a70f4c20a95ad
+size 494407

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvDenseP64VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ed9a9f8b7761028007007937e2775e299670804075a4cb8cc795dcf5f49f3ed8
-size 466122
+oid sha256:924a56dfb6f59402cf82fc674408ddaab843b04e31700cf973c18303c3c46413
+size 438799

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a2cc2a41306c8a19875e2951b0689ed3f92a5021b0df2e844f4461a11ff084f7
-size 640803
+oid sha256:f1150d780f008ec7e3b602117dca13fe048407fb3035818525c9d2f7e1d19142
+size 590589

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8009feb523958aca6a46e40d7d3e2ba7b6a4b773941ec4c1c8b17583948f33bf
-size 611890
+oid sha256:e26f4db8db5693119d677100c08d7d1aaa86a7ee38e671bd5f1228f8a6e213d8
+size 564809

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4e6b358b90095f24f7679b8c4a0ed3a197c532b36999c4a27a9e490b6d665775
-size 629277
+oid sha256:97f705e6962eca45a27300ba02f79eb81eea9b1ebea4e1d68f26c00e147a1a89
+size 577485

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d42910094ae7245c01f2a1735bdfa394a63e0ac57915871e27b43345cf460ace
-size 605100
+oid sha256:067a89e42058a1e38d309a68b6363af2d1981963849c808ea7ff9cb986662ca5
+size 555677

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f33752bd17738a5a5651368efab153d542301fb864e87dfbbf3ba7df0ef89b0c
-size 749009
+oid sha256:c4d9bdc67d9bd2eed43d35ef189b03cbeebde7ef3bba70be4d8599b805b6e8fe
+size 677016

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9de61de96614d58c1d8c3a9c72ab3e355fc59caefe1474dc2bd60c82f340c163
-size 633315
+oid sha256:1100f6174f982fe1764f1d519fc9fe38aa4bcdd00e667171176ab08e97874ab7
+size 592425

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:495bd3ebcdf6f6f2270ab5c10825517b065303dde45553f10b65ab333a740126
-size 606594
+oid sha256:85031dc34441a75463fde5db1f25ec625460ff495e3872e48206c10ec7b87d05
+size 557195

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8396e31640be20e49854d7e061dd9aa451516796c4d1b8b4236186a26f258240
-size 523360
+oid sha256:6baf149c4058560d1b580122d829c8dec65f25552b21825ec7e6bc8c5fee4645
+size 484987

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:04ad51d8836ce6cdf64bf267731bc07b340a205f3bcf6ccb7842085042151396
-size 581628
+oid sha256:63fd9cb67162865d01f06aa677f528de0f255f8988e9d9bac5ae64ba8e5f9e6a
+size 536177

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6fbdb691bb14f6b7bd7bc0188171cb14155f0be066cda3687ec7e0947d633157
-size 503920
+oid sha256:854270c05f23fd8996cb24332d52592d60a9922d9c0e48e096d6b305ca1853f3
+size 466337

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP64MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e4dcbb1fd7db30c312e7ff37754e87b7f022b52973359092dd6fafe05767b0b5
-size 602124
+oid sha256:d64b8964521d3479c2bf9ea6c73daa0ec70ab58ee62cbe94d0f0a4bd58f9f062
+size 569277

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP64MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b89e910fce8afe4a1650259e5628c2c79535fdfadde7063a3c4cc36ba98e6dae
-size 573212
+oid sha256:2485e712013bc555685f145decf7b439be95cc7218a41f0a1ab7ad5d9b41d1e1
+size 543497

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP64MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b416a3295f7fa46b5ef9cf3b2fddbfdb924e191eb6758ecd722b35805d59d2fa
-size 590598
+oid sha256:5430dd5303eaa83e2d42d18af80cf465e8931edff54f6a8623ee5d028c6aec9f
+size 556961

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP64MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:61c360e55f8a5dd89601bec6a1047b3c1d7269cd4980d34790277d5291479897
-size 566424
+oid sha256:dbb0fc5e8ac3843fe0b4cd925beb64040e5a10554b8c0e4ec93b65e48e0eeb48
+size 534365

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP64VarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:267e6057b6bd4fe3a94adc24dbe0495388976874a6397c7438365b9482973219
-size 707273
+oid sha256:f12fa046ddbc44cc59819330e4167ff8c1543a64fcc70eee2e07c590e984a42f
+size 655704

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP64VarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:56b034f9974955e092c407ed23bbf427593b43d2c5a4de352d1b5fe2d1c13b01
-size 597004
+oid sha256:94574d7d32795256b765e3048c8b2bfdfff67e54ff7a968eb216caed9790608d
+size 571903

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP64VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2cbd4199b9a5929907cc48e68ac53cd0c16cd3e30d4a2d8bce8b4bb7dd722755
-size 566338
+oid sha256:f028034495299162e677eca9d3634662a365b356d70d3ccbea8639a815ce92c3
+size 534303

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP64VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:80b25b6631be05309453f6fea1c94f03afa22ac794966406958d5ef0c54d5947
-size 487840
+oid sha256:d365f379ceec42a6dbb4d465dc50262413b60c00b0a3e62601f6bbf5f20c1171
+size 463675

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP64VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d13c8a5ff0a64d243d633c04bd9feb6c43e8369fd0a748b949dbcf09e6acb798
-size 542162
+oid sha256:21f3444b7938619f6eb9b4f49f9bddc7fa259ab5b1e2929702c5ef56835cf7b3
+size 513285

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H256PagedKvSlidingOrChunkedCausalP64VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ebf364a63f6b8b18bee09a555fb9a80e80768862c21ff11e540ec7649a15b592
-size 468400
+oid sha256:c25c9dbfb3cc1b4aec205d19bc28c7ca5fc1c0305fc2d9b15ef1ddbbe50f6bbb
+size 445025

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:af793d48315c7f635641e4977269c2220c90210d3e4c06c85f170f9b3a263109
-size 581660
+oid sha256:4dcbb77c074ef7a1c4345c33f481f353e12f243722c1809694e9913a11ffb713
+size 540673

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a5575d32f19de2bc79e0019c460dd9cfdacadab04f3f7d1423795e596e106cf4
-size 521096
+oid sha256:657335d1b467345921db997a17815f88f6aff345b7243c881498d41d0d3e25a6
+size 482131

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvCausalP64VarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:81af6008028a6b1182d5e9919d2aa7bcfe1a70184e51487994f993c816d010ac
-size 560818
+oid sha256:61f596a041ef16d34986e1cc8f095eb30993caed469f81cc5ace92d72f46a966
+size 529623

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvCausalP64VarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6d9900baa4d38bb1a962febc14c5748ed693603e0b9e3d5353d07f8b248f0d97
-size 503730
+oid sha256:66637af56df4410a9eb388153abae05dce6b2672818ec29c0e464b93b868b9c1
+size 471081

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:12fe0a99bb49acc6effaf982ac94cca4ef8d63e96e6e8bfb0b9ffba31b24d2a2
-size 464942
+oid sha256:848bef441ff7edeb3dcee0fcea6f0428cdb2ac1bc3229212e01e64a6b25a9a28
+size 433697

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:545a970d35d1a1b72fe6110e82e072ba1ebe1286b9dfcce53665e8a054171dfc
-size 449252
+oid sha256:10790c4d0f5759b4178188d3c8f1a87964da400e9bd93ddf229805f90550894a
+size 418747

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b1278ea9908cce2b4095e8f2312aa97d2b94eaec477457203013e75855a77b3f
-size 459730
+oid sha256:a61eeb34960ad5ce4b5bfdf13a26cb2c8ba49506ccf52ff5d5dd91e1bdf71e86
+size 428461

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f291d926e676e4c8dbfe11171ba85931cf743634693d77393fab74c2cf5b266a
-size 444040
+oid sha256:9fc43f0fbc5f404ca40cd8a96ec5f2f7b6017cc8a681d7d958f2eec874b5c148
+size 413561

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8188f3cbd205d1122c1d2e1cffef8d6b35c599c39e47a47e7642f9288ed31098
-size 574234
+oid sha256:b710411c2a2e5b1c5cbd2edd2e483267239f258f5ec47e8f604f296572b3de30
+size 535145

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:58e60b3efd13ad6926596b6a474d6a4b3d379668b191a972e3252599d90ea328
-size 517222
+oid sha256:f121b9bf126b1a1832066abdaa484c74e282d9a0f9e2e2355ed273fec1a99fa6
+size 480625

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100aKernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:547233ce34185393dd425eb4aa164b6712966b891afc8e16a22d40fc47cdae1f
-size 434804
+oid sha256:1a085f128d810ae87aaf05ea72b3498059060c4257afc97f2940b8b12c3fa58e
+size 417743