[https://nvbugs/5394685][fix] the bug with spec-decoding + SWA && an accuracy issue related to 2CTA MLA (#6834)

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

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cc91c47e667e39b25840a027fa7ae7d2df40147c39bc978dd8779a91b6b4509a
-size 1539517
+oid sha256:27b1ac149eb002d3a3aaa06193160f667ef7446c6119ea8a7624cf5010e6fa91
+size 716859

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6fc0a1edc16f5c84943061ec75fbd28e90d2b8f6b1c23e1d41aa4d0b097e16b2
-size 1479225
+oid sha256:aa580ccd54b55d86aab06d842953dfa2de1adcd0f08d01fa1d856e6500b42c55
+size 684095

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvCausalP64VarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:307effd5d31b33c91bfa045c4fc5bc2e3369a05adad3c6292ba813002ab5bfcf
-size 1529997
+oid sha256:4b4ac41b6ce52463180d04673e91fac15280ef0cbe3b470307cf01603867072d
+size 698705

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvCausalP64VarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:17e0f2e705585602417ae3320afda128630c1da537254b45e3db701b422e2f67
-size 1466497
+oid sha256:08ed2a80e783dfff875ac4d5dd3d5544707a16bb70af7e290e883741e30a7c97
+size 668307

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d969d5c4f8f15cff58629183ece06c8441d5f477a707cc97900b97fb56333f48
-size 1122185
+oid sha256:46ff8dd8bf7eeae98428cbab52b0beb9acceae29d9fe76c3effdb32ad423a500
+size 709439

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:76545e32a1439f451400df70fffd085b5b38416caec46ed516865f1bea579456
-size 1060271
+oid sha256:bc6e98d50fb43fa50924a8dbfbc0087e8d80e93587e28ed3b0c704ea02012be1
+size 657833

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:806610ff1941e78d62d940bb87a19148a577abdbfa7e392597932a58e3685af7
-size 1123315
+oid sha256:5a75e31f4b04916330fd2b7ed4886c6bdb8fa49852385175d312cdf291ee852c
+size 709703

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:52a7ec3418240849986b11190e0348a58cb79be4713ebf6657b567ac6a24225d
-size 1068405
+oid sha256:779ee33a31a4881f732fff48ebbe76b2a26e99d41dbc860a70a4cfb0d2ba6d57
+size 665203

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9a94337a6e9f3a4d2e427cfa5436a1d260d78fa0e81454a6c032be0354625dac
-size 1605573
+oid sha256:7e6861130024df851504602e4b6482d37674236d80c9316eb855381e0e0c1797
+size 762245

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:419fc190fb35c1b7f9fbe49c6d21100b3b3ef26313908eb807592ae56251a632
-size 1457219
+oid sha256:2e28137d016387ffa13f6c403059a81c733d9574aa5db4aa845129ac7005dedc
+size 682661

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c1da2b06efaaa23769aeaf0a4eb57307ab66151e4efde137718be569566eadfc
-size 1016099
+oid sha256:b9eb2afd61aebdbc44340dc64fd7cea0841e8830f2148e61a62cf24e0524a837
+size 675649

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:48fbd66d38aca6acf449c9634e998ccdea65492e08549cdb594c57b1e3d9537a
-size 1000453
+oid sha256:97eb744d02a4243f1397551924a01380de387941119bc524400ef58135ef29d4
+size 609436

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ad75cd273d6fa02d7cb5cfc948425a2ebe24f5f377a3100ac125de7d4f3076f7
-size 972683
+oid sha256:0489ad85a996e7644c0928607156ed9a5628d3d82abd4bdb1f434afb7a9e62e7
+size 640621

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3cf23a558a2c546cc5cac51521298e61d817357de37bf2d2a4dcdbdaafe41644
-size 957579
+oid sha256:f03e13af5a531d3b85b26f21ce9f0370e04571b3d24a7881bb1c5b621f157bc2
+size 574406

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1f9bdc581d79bf7557286c54cd5378854bc8117d3b7cca9c6f4639206c4fa85b
-size 1109507
+oid sha256:28b66bd75aeaebeb1ddb9be6166369686e6313ac9867ea2ac83ed390d196f421
+size 691283

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f9a1766f14dd5434b2dcc7b4fb1bda2faf30df423b7385c691f746acc085122c
-size 1047543
+oid sha256:4db97c86ad2a6b669e4ad3ccae0ad9baf2ec1817cf9d0baed71f53876794fd89
+size 640469

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8a4dfede411400b59cc61ecd0b62c8ee3b67761095fc25aa2334561b5ec1530d
-size 1110635
+oid sha256:725cb76d009776fc6f24976f56ae7be7ab732e5f0147a6dd6deaa95ffa4fdf52
+size 692339

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:961fba13795c51feff2997a293208d58f7bb4afb4b59f0115d204a40bd45dbeb
-size 1054937
+oid sha256:cfc144f2d52392e50e82335205690bc054203bb9e1ad478e5ef8548c715b124c
+size 648627

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64VarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:05175c99fde2f8be62ef62bc72e559334e4e3343aa21de11f87d39d1cf09d113
-size 1596151
+oid sha256:157a37c3925e884b2f3500709155eab584d7004e275df98384dc5a2f69a7d9d7
+size 743301

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64VarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e8c7fef3a35a67243a777518e9ca236e669b1b15b5bb28ab680f974dce83047b
-size 1444491
+oid sha256:e008ee2db4889f7008c284f4d70d1e226248833191219995509afc427ff694f3
+size 666875

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0650b3cb05e5d4c5537ba569626ba0c94835ccdf8d8bb7a50065ca01f1f9abd6
-size 1006577
+oid sha256:6f2533381882d99e651f99c676626fb82bc94a7655330ddeedb8ea64b063308d
+size 655917

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7209b01d6a4dd75600c38936d9b414d6007d109074b9735c83da8f524af88195
-size 986985
+oid sha256:e139fd02d8968c03a353a1193779d80812b2b5affbbdc66f99ca10545ac937f1
+size 592070

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:57b6676ebb3c1a15681fadd66a7246f78327459e920790344a584ce8771631a6
-size 963853
+oid sha256:344c9ac994de57fc15bb3d9cb00190eff4f5574d76e0d80c95a89c6b91982b62
+size 620099

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvDenseP64VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:473c0bb05a8f21626c5d42012b542e64aa2338e5fb03850062ecff97a9c077ed
-size 944851
+oid sha256:dabd42252c4bd6950b5fc8cf82731d83d012d33b30ac863f8047f3d4325f68e1
+size 556252

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3e420850e12e866eef7501ced4583ba16fb1867e6fe3d821b80fb45c61e935bd
-size 1147775
+oid sha256:a9ab714ab53a270d238a770dffbbd88566afb1e30f06ceebb598c7b9a9fde446
+size 716181

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:744043d2514eeecc4dd3584a0ddeb61808f8a5d60fa298684a4dce12b75a2820
-size 1083639
+oid sha256:7962446880b34073e4719007aa7384dd68461bf0a85625e1a96737ee0712babe
+size 662209

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3988f1739dd7c06534f07300e30d40137fec4b21cbbb0f598f28315aaaf68d69
-size 1148903
+oid sha256:c1807c105d02bc2d5951aa3be505b0a0d727f02ddbb3a4747fd1816b37aaddf6
+size 716447

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:75b49b8f29131122ca058483684be1ca76d39f8306df67467eaf95dbffa6eb0b
-size 1090243
+oid sha256:65c8609b79ee4e130d01129cef89c2ef029b33ba223b58e5ba4cd5e172472989
+size 669579

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:abba6bfaa76eb0730f64787ce70e23e88745f49a42e5b0a931fa390410f18fb2
-size 1831061
+oid sha256:0ff1ea6d051b3a1e333b91276a1f07a42b13d682c7de06841f1d5e47f4321f4e
+size 814177

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:dfdeb552ebec8e741f33d5a9fc4a80c2aab80440d9179ee5efed48be92515725
-size 1771309
+oid sha256:e46798ba011c02938dbd922153aa413723985a35ef05b7d69420da91f57451d2
+size 754969

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3539c079e7da4f223b857a1d705e9cf8be85d543b5963b03ab313d7216f2b913
-size 1038679
+oid sha256:258cc90789698c5372107627acdc637e15a51f6ca56da03ce97c721a7381a15b
+size 664485

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ff95f9469efe14fe2403926e4eb55c5c1fda840214b125edef848299b7c730e8
-size 986673
+oid sha256:2432e8c8bb8857a607077aa2f1317a37555c2d99724bad8804324b43d028392f
+size 605030

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a72a4b20e62e30af78c01d27eb1847271e5602b00babcd6496928b0d249157c5
-size 992797
+oid sha256:3d1a4de8b24d104421be65700e55c7ba65079b3aa0deef8fab76dfbd75507be9
+size 625509

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d6f019c2cb5fcc27c39be6d7a61c898a7a947108ecd575b7a5a46c46ce7b6bb2
-size 941333
+oid sha256:6cc0c71192acc4cf1e3238d6af12c4700bb7abad9cab74f608ee45d0fc5d4f3e
+size 568422

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1f6deace52114082ac75433a256ded21c60bff6ed26cca5ebd5b24a6a9ea11ed
-size 1135095
+oid sha256:b0a8bb2517274b16b60cbbf1c55e292a3a2d6c2828696b74c270dfb8c4c4c255
+size 698817

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ee6f6b20b60360b2623aefab6c2e5a042b0ae2196a762c2af42b9ce50647593d
-size 1070961
+oid sha256:bd09e986f515c4d837e763ecb48998c724bd406ddac83a9d758e92e575d55290
+size 644055

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2af455aea4493096a73ab732779ae345939b4acf14985b4a6e6cee962669db8f
-size 1136175
+oid sha256:08eb909f7d672518a864cb72f7f7b63ad10a64b9f3fe38d5fb3da6e940539327
+size 699081

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:595bafea2997d4f2877ae06e3581a1fd611f36453de9ffcbb4bd2a0386af7181
-size 1077515
+oid sha256:7aff75f98a82c708ece7f646632fdc1890a42fd13ba93c09fd625ab96dd2db0a
+size 653003

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64VarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a8ef516d55dc574173fc8818c2b788d470ae93a087f8ae2b997d213071793381
-size 1822329
+oid sha256:b736f2e778a3675982ca2bf1480f8365b590907b9ac29eb8138b024625c6a43f
+size 796811

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64VarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:26eb5f9b71edabe9724e46c3b59059073e36dfbaf3e1f518253e285b2495e72d
-size 1758631
+oid sha256:3f9643acf43d8dd381d09e0295f2751ca66e97d478a8162f5eae0f9ad4461f9e
+size 739183

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:48a7613a05d2a811ed5171011a3d66d567051295223228d91e562404783a78af
-size 1029157
+oid sha256:891b1c2a60bfbb42eabc9e12b8fe8ec9184dccb24612806577fe6ea3bee8aac2
+size 644751

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:097894eb9983e2ff1ebce28cc90e3f035bc6e502e9889aee88c7467b19b81553
-size 973155
+oid sha256:564a65acc48214153b4a464431d92cff8e45c4106e417ea20780b878300032ca
+size 587664

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:96a3d7e85ed22148a6120dbeeae85fa352954b2ab605a84bd45c6ef2f0c2fd6a
-size 983965
+oid sha256:1a457952848822892441c92db5fa4d8d922584e941a791b5c0adea5cb54d3ecf
+size 605776

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H128PagedKvSlidingOrChunkedCausalP64VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:10d34d87e4eebfb81a7a5f45c171a7d61e76be53208276fc5febfb6721bd591c
-size 928655
+oid sha256:383f10436a3ab958e2d874d57a7b8ea73c05e761796359cca3b6d02ab0efa24e
+size 551846

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5b108e6e1a19f1dfe73c6136602a80e3deb46fd971444f6be1cb6f22f61026b1
-size 1423631
+oid sha256:63eb6a57b257e669aa572f3469a4e86d2d7e3ef00adcc7fba59954a0fdd0ccb3
+size 646483

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:063edcbf92b1397f042faca756a2492d5439680f87d06da4b995b0f59003e735
-size 1413363
+oid sha256:b96345c64c93466fb4aac6cc284b77a6cfbc515650610f480640022f39ab0e95
+size 613520

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvCausalP64VarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:362b195656b812c8321e8793d1b967714191ccfb7293eaae591f1956873ee1ff
-size 1414703
+oid sha256:9f42f1167bb8153983d8b444687753493830074f83dc047503894fb09bed7c62
+size 629119

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvCausalP64VarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8660e0d537c39c9e673cf4b3487f44c773c07f827c6f1bcb1585897ff5bdcd42
-size 1400683
+oid sha256:a900379da740668badc53b4071cde246c4ee6a7349fd583d322f1b1aafa4b613
+size 597734

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d5b90bcaff8daea18a8dcac07ed91962c14a75e7a6d9ec41efe0a2aff3cc9bcb
-size 910741
+oid sha256:d3ebbd4568716506800388ce20cb04631ca97ab03c4cabee1df05b24e1402015
+size 606946

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:13a7750e594f2dfbcd86dea9134a11c448827c19676fc4eeaf8329ce4918296d
-size 878081
+oid sha256:8082d0181f9897682db6c1867eab13984d8d47720d51ecd7bb519478c8fa3a70
+size 580624

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d0c4142b33c683e08b97b66f221e98fbd854641bee03fa7fc871b629625960c3
-size 936685
+oid sha256:b890a2d503020e5995bb55a024603dc5748003715974489f258e93c4d7b37f19
+size 628597

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3e4d219acdf661e52e24d1b8872c981124951c4f01ba0377c522ff2d19920005
-size 900571
+oid sha256:5353e765dcf15f5c77af6c951b2c16b32f07675cbef2dbe2d1a03ba9378f221b
+size 598328

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:794e5fe052b43ee619708d1c5bf108c0033ff7d28bb7a15b7592695a86bea7ca
-size 1489933
+oid sha256:a0eb157a336a566843a5d7b67b219b5078287a27d6ff3f0183f5da42ffaa8711
+size 691869

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:22576cd15f45ae5e6bb8603a7d23487249759cb725dae5594c163f0c44ad518a
-size 1409513
+oid sha256:7ad74ca919d8b6c06008f2c1662c6e7e28d29d69ca8e75d0a1a8905ffc0a44f7
+size 611274

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ceffd9d5f7f5af8708f5abc413760de28a6c73edcf083cfd27097283fdbf8f97
-size 938743
+oid sha256:1129357347e8fc8adda5a2e39819e0e4a730491bfcde0e0764d08e756b8f8d17
+size 611908

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:bfd3c365f46a5bed58c12f3d1b140533c68184a665dadd3b7bc66039930c5050
-size 818411
+oid sha256:fb1311f939b585c94554658af236f2d7ee9d3babc8091f4cd0bcbcc56318860f
+size 531438

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:41743277d38b6ccb41e052e399e281c45a60a8c1816517404953a2f56e403e25
-size 892959
+oid sha256:0c14156039a7b74626ae26b674c5df95784f63496202b57aa9c42a299e8af637
+size 551546

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:719890ecc0038a9d5b52b7d104ade00acb850c6615412e9e29230a116e856a1d
-size 778745
+oid sha256:77410c0853a9d6cbc0b2999d18e939a9407d2cb097f7794b49b38bdffcdfbe4f
+size 497988

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP64MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a79769476d51a3390a3344eb86c6bb65c85d32ab43738a92ca779955f2a90744
-size 897273
+oid sha256:4fff0a3cb5e85385fdb385e010d4922e8d816277c8b826449c41e3d0c0e638d1
+size 590370

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP64MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:79f76f99b36342891cc0f50f0cf1ef3a23f848dac7eba563b8c1749d34498018
-size 865353
+oid sha256:e8d79883401a5ab3435c3309c19e81f9cfc6488acfa33be6062c1033d7b801d4
+size 564048

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP64MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:458bdfbb4a80f473b5f018ef8496788e9a83167b7b53901e53efe4b2502e097c
-size 923957
+oid sha256:66383812d7c619b960df5efbc541038e8730c0f68b0bd3ca64e5396b01d75d56
+size 611232

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP64MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2dae9a76ec3e66a898767c5d5afba1173ecbc28d594fcfc9d319c29db618d829
-size 887103
+oid sha256:908783e8a4965f17c34f4ab7e5fe2a8c1fa3e896e23078eecd68f8ca9d279348
+size 580174

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP64VarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f055299dfd3ed2bf677d98467808a51e27ec7051abf63eae64c7f548e269eeb1
-size 1481251
+oid sha256:b7f7c603148a22daf6555e80b617a5e0c0208f8e1bc71b6d3bff52ccdd63791b
+size 672925

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP64VarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d8dca1dade8864a8f29bc3f6ac4b6d4a48ca167681c5d3544aa1e32ce37dc022
-size 1396833
+oid sha256:40ad18b736cd2614fd7cc28e43e95a58b0a3ca11222a90ec5edbd0b5bc96587f
+size 595486

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP64VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:895076eab0b9314732ff02f2b32dc5d4e5088fe89eb76cfb2385267eba88440d
-size 929123
+oid sha256:7afa6e3a09056ccf180178d99f88b089452ba8e4916ad6f0475bf9aced42bb35
+size 592174

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP64VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cf8bffcfc5c7c9159836ffd3ee6e6e8962febbd6192901e1a19d40195413815d
-size 804943
+oid sha256:452e9f854020241c8d87e03198169dc3b496b459126b7e203c2c05cc965e2fba
+size 514862

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP64VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b42a56996684ca2813a26289f237968accba544469f91c9cd33daecf158037c6
-size 883339
+oid sha256:fbb2155e67f64da9e37ed869961c1f681c5aabb5a02c466577660ffd8bc9828f
+size 534182

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvDenseP64VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2b901eaa5372c329148dc07ef2dcb5d0e09e4a28f9a4eed07db77ad7cd49d541
-size 765227
+oid sha256:8f17a66d53771cc584d83f08d81593b3a78a10d99db91a0d677f446b6f41873c
+size 482200

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3520417f71d65b487e56ed083b46a91cb166dc2b36269381b50361fdc4f0f373
-size 933567
+oid sha256:d38d19571c30ee9b7886c958f1a36e328fc146c6d8534b3ac9967fbdf0b43cb3
+size 627183

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3060f4d447e727cb71366c5936a9e6c9a3e36507bea9920a3f0d2161cc4ba7e6
-size 894395
+oid sha256:490f49e40e6dff2087951f51cf7acb376f0a534289450078eddc768892771834
+size 594520

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a0618e3a2edc1ebdb24a0f95d99e1ddef81283ab0a627b901b9a95afc77bc2c1
-size 962519
+oid sha256:ec407d0f941032e2863f42af54259e7a47a0912dfab15f7953e28bfc7452096a
+size 636131

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a2ee52c9af7304ee1f1a6c7aa29a53904d29830a7a0ee79b556606d736d95e5e
-size 921867
+oid sha256:3a88295931adda37a69245bb95486f1c8c2e8b9028de12ad8ceb55217ef8eca6
+size 602704

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:604edfbd5f9bfec6a07e88fa9e6dff5866c339c6867e779748d0e3ffe7d670c2
-size 1714139
+oid sha256:865ecb0d7f54c04b37acdb228c2b4d0c1e1a6f6324aebfdb9f5a64b5d53c8a15
+size 743801

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f1e965619365ade5feebbce4d7d9dbbc1ec73cde974f8a119ffaffae0127cef1
-size 1704707
+oid sha256:8773f2f05e51a291db0e4752208f54c80b59ea19e0f270c535fab1b14c64eff9
+size 680055

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8000b5be424c5930ab13f8d690c67b790032d3d205554ececdcad1cf98eff8fd
-size 961371
+oid sha256:9c4c516ec54fd4c4f2f1a6f74028efb3837826431a6ab66d37ca93ce81a9f281
+size 600744

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c1e11b8426c15a74a9b7db5d1bef252be7258d3c6a3bfa0e8068c9a2a5de6998
-size 823377
+oid sha256:1afa3135bef82cb681ea28ac42dcd4afd37dac9f9e7d2fa9704a69aa37edd9c4
+size 537342

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1d4b0e97e0940da384cc6ec53c61836953ceed22c3f5ea096fb05fd53d49b19c
-size 911641
+oid sha256:092ca39eb35b5840d1306f83a6e7c3adcf71d31fc852bf6f564062b5342681ae
+size 560978

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP32VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e5f82fb47d0a6d51da6c9b1bc83b9d4b3aee59e763948ea6a991c47342c4f6f5
-size 781393
+oid sha256:055b6501be55680c036698d808989cf124a2ea32752cba221c194b50ce3a3c1c
+size 500734

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP64MultiCtasKvCgaVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c2a2508407cb8ae39d11829dd97341c0964ccfe483e1f828b2288449b78adcf2
-size 920839
+oid sha256:5ed7ac04a950aa6a44979782ebf3db4943bd76e821cf3a5e40402b32d16f35a7
+size 609816

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP64MultiCtasKvCgaVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0f1dcf28bcc3acb1d6eb049c144a2b00c843af64cfc6c403c55c2655981c769c
-size 880927
+oid sha256:2ea0e17ca4f22531f985c8e39d32e3c13b796016aad4da88803dda4e16955e64
+size 577156

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP64MultiCtasKvVarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9111e89b93ec1c0a578c485abcf0f6276d34e0209800a1b07162b1fbf84d9334
-size 949841
+oid sha256:adc910d77c39dd221e1d3a13db23950289d2529cad367c00b1d4936f29e6f3b2
+size 618765

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP64MultiCtasKvVarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5d25c7ba6a73f08bc815bc9fd1b350e4adc558544665582e712d2d33ccf2a1d9
-size 909189
+oid sha256:de788b00d14153d54a35d040be97d5fb186761380d7a4ff27a27615db65828ee
+size 585338

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP64VarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:155d69cafda50992c4cd471f7e3c82887d516f04ebd5a74af6f70d87c217ca44
-size 1705407
+oid sha256:f6953f3004f1eb54baecd7450655ba09672d0c840822d6917a7befd244038ad5
+size 726435

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP64VarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5f2d801c044e47072d2a20abbf1d6cd30954d04d7265a255594fe3655dc6f47c
-size 1692769
+oid sha256:845a68e0343007ea9e560678e13ceb02428cd2ca1e9aedc74949286531d41fdf
+size 664267

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP64VarSeqQ16Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6f45a28bd324cade8f30a1e4ae16d87818ea424761567e254b5d4b7d15cfa8d9
-size 951751
+oid sha256:28f3de585dd46c9458b5ac145dc6554bf6d9af15bc6365079a351ab4d1e128b9
+size 580220

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP64VarSeqQ16Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8bff4a2c7d9e55e73b1b2b4990a02bfb04129392276d185c800f6884faa923a0
-size 810649
+oid sha256:1deaff2efa1eaf5a937449e21af57895657570ee60b65829de775c5d319cb119
+size 519188

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP64VarSeqQ8Kv128PersistentSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:31c8dc0baf42b3768d0143cde663ab7e615fd70a6e9633969cc05ee23926596c
-size 902811
+oid sha256:c9a88d786dffe5a1c72f844e8e2f5a8ac883baf12299668a10f9287d0f2d74f1
+size 540456

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QE4m3KvE2m1OE4m3H64PagedKvSlidingOrChunkedCausalP64VarSeqQ8Kv128StaticSwapsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:874269a179a6cc4500f4e374c8564731a3d2174268b4db065315383454ac098d
-size 768665
+oid sha256:9439b50fabc7333ae49fa0a18a704af7209672520e90375dde7da94d1572be5f
+size 482580

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QkvBfloat16OBfloat16H128PackedQkvCausalVarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:774ae66439acb76385cc6cda417579a3c2cd3f0d3086e644726c5da547fc5d40
-size 1373369
+oid sha256:3b357a6f4c9901fbac1c16a9b328aebc7bd22d36728217adf327d06f636ec3a4
+size 973719

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QkvBfloat16OBfloat16H128PackedQkvCausalVarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7f6e6c25ad3fe8908764c6cf4ed99ee1f49bfc076b1a7078b896bdc34521cfdc
-size 1200991
+oid sha256:04c9edadcdedfc2670ee824caae9bf57c9c95a8dd0584fdf439a954482874197
+size 877413

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QkvBfloat16OBfloat16H128PackedQkvDenseVarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:05629dd08833a7428327022bdc45780a806dbad10d16fc8a6f8921a40a4325f4
-size 1368237
+oid sha256:38442820fae09138f1f7a6868a63076c26075abc113edef04ffadebfbd5857d4
+size 973865

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QkvBfloat16OBfloat16H128PackedQkvDenseVarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b45309cc1c81113ff6fb19ae3aa9d804784cab0ab8ac4e0c31c8d0d0a373de81
-size 1233499
+oid sha256:4f681f620d0f0f9bedc8e0bc7783fe061073235f1c423bbda1ba216e4ae8d827
+size 885353

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QkvBfloat16OBfloat16H128PackedQkvSlidingOrChunkedCausalVarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:933df0e81db4755d9827d308d41fb56e5e5eea6cc636af3b922c00cc1d3ca4fa
-size 1652579
+oid sha256:5e710014ef86cb1d61d16e3c0ba58e4e287f4d8c23a82d48ce9f8a035ec38517
+size 1099551

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QkvBfloat16OBfloat16H128PackedQkvSlidingOrChunkedCausalVarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:49a6fa085bbf78e294f7470e5d870d31e8d9fe9ef32eb1ec8e41ece147a37a57
-size 1481975
+oid sha256:58e7f0ba908cfb60fee905d761c3d15770281f0e63af7a5059d39c1380f7551f
+size 1012963

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QkvBfloat16OBfloat16H128PagedKvCausalP32VarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3b4cc6a2f10fb7283d26bf7442b5702a5956276fe399eaf31c49e40f2506f36d
-size 1390045
+oid sha256:b9b134aec7f60ee1bedf4c60f673f1a5118928feae4159de2c1f1cd83c463a3f
+size 1025621

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QkvBfloat16OBfloat16H128PagedKvCausalP32VarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:96a6c1d407e1b108eb10fb010da27d603ebbde99bc691c80a4256a35248fc20e
-size 1270799
+oid sha256:03d0590aa511ed6b6c33c36cad54180eb2cebd4af66fee03731e11f1f85f7dbb
+size 983729

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QkvBfloat16OBfloat16H128PagedKvCausalP64VarSeqQ128Kv128PersistentContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0b34f428863d011cccf6f5666173b6a55f4ca1f89da97fe633be13599b14fd05
-size 1375739
+oid sha256:60fc1054362b865910bb53074d142bb1307a4bae9bffd74b8bd39072e9dc253e
+size 1006133

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QkvBfloat16OBfloat16H128PagedKvCausalP64VarSeqQ128Kv128StaticContext_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:bcb55db99639de5f2c560d6972a81b6e1ee91d92811cd5e1fbf8362ef681b347
-size 1257281
+oid sha256:cfe5f9a517ba867e0ea930cbab710e53c26ffbea8938e9eb6d60f2b55a7cfb32
+size 963353

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QkvBfloat16OBfloat16H128PagedKvCustomP32MultiCtasKvCgaVarSeqQ128Kv128StaticKeepsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e3f03123919ec56af12cfda16675edee13b5e933185c26a3120089943c89b4bf
-size 2081287
+oid sha256:0ea2a288f05ef2ab28f9d38d4c1f9b0dd064823cb97ff66a21dbace1a49eab3a
+size 1165117

cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/FmhaSm100Kernel_QkvBfloat16OBfloat16H128PagedKvCustomP32MultiCtasKvVarSeqQ128Kv128StaticKeepsAbForGen_cubin.cpp

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:eee84e282e48f89bcc9e963e51e90f6f59d6c939841c18169beed465a7aeac02
-size 2067813
+oid sha256:68da30fa1a1de6c2c2da792f6d2777fac3982c86c35d2958d73b56d2f4196ba6
+size 1147499