NVIDIA · dongfengy · Nov 18, 2025 · Nov 17, 2025
diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/GemmOptions.h b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/GemmOptions.h
@@ -1623,8 +1623,8 @@ inline CUresult loadCubinData(CUmodule* module, Config const& config)
     // Trtllm links the cubin into the executable while Flashinfer loads the cubin from storage.
 #ifdef TLLM_GEN_EXPORT_FLASHINFER
 #ifdef TLLM_GEN_GEMM_CUBIN_PATH
-    static const std::string tllm_gen_gemm_cubin_path = std::string(TLLM_GEN_GEMM_CUBIN_PATH);
-    const std::string sha256 = config.mHash ? config.mHash : "";
+    static std::string const tllm_gen_gemm_cubin_path = std::string(TLLM_GEN_GEMM_CUBIN_PATH);
+    std::string const sha256 = config.mHash ? config.mHash : "";
     std::string fileName = config.mFunctionName;
     if (!fileName.empty())
     {

diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/KernelMetaInfo.h b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/KernelMetaInfo.h
diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/KernelParams.h b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/KernelParams.h
@@ -224,7 +224,7 @@ static auto makeTmaShapeStrideAbc(GemmOptions const& options, int sizeM, int siz
 
 // Create the TMA shape/stride for A/B block scaling factors.
 static auto makeTmaShapeStrideSfAb(int mM, int mN, int mK, MatrixType matrixType, int tileM, int tileN, int tileK,
-    tg::SfLayout layout, int sfReshapeFactor, const int32_t numEltsPerSf)
+    tg::SfLayout layout, int sfReshapeFactor, int32_t const numEltsPerSf)
 {
 
     // The outer dimension.

diff --git a/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/config.json b/cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/config.json
@@ -39,7 +39,8 @@
             "dtypeA": "e2m1",
             "dtypeC": "e2m1",
             "mmaK": 64,
-            "tileK": 512
+            "tileK": 512,
+            "biasType": "m"
         },
         "BatchedGemmFp4LowLatency": {
             "_template": "BatchedGemmFp4Base",
@@ -53,8 +54,7 @@
             "mmaM": 256,
             "clusterDimX": 2,
             "sfLayoutB": "128x4",
-            "sfLayoutC": "128x4",
-            "biasType": "m"
+            "sfLayoutC": "128x4"
         },
 
         "BatchedGemmPerTensorScalingFp8Base": {
@@ -156,7 +156,6 @@
             "dtypeB": "e4m3",
             "dtypeMmaA": "e4m3",
             "sfBlockSizeA": 32,
-            "biasType": "m",
             "act": "swiglu"
         }
     },

diff --git a/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...4_cga1x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...4_cga1x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...28x16x64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...28x16x64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...4_cga1x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...4_cga1x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...28x16x64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...28x16x64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...4_cga2x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...4_cga2x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...56x16x64_cga2x1x1_16dp256b_rM_TN_transOut_schedS_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...56x16x64_cga2x1x1_16dp256b_rM_TN_transOut_schedS_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...64_cga2x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...64_cga2x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...4_cga2x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...4_cga2x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...56x16x64_cga2x1x1_16dp256b_rM_TN_transOut_schedS_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...56x16x64_cga2x1x1_16dp256b_rM_TN_transOut_schedS_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...64_cga2x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...64_cga2x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/..._16dp256b_rM_TN_transOut_schedPx3_biasM_fCp_tmOv_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/..._16dp256b_rM_TN_transOut_schedPx3_biasM_fCp_tmOv_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...4_cga1x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...4_cga1x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...28x32x64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...28x32x64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...4_cga1x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...4_cga1x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...28x32x64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...28x32x64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...4_cga2x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...4_cga2x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...56x32x64_cga2x1x1_16dp256b_rM_TN_transOut_schedS_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...56x32x64_cga2x1x1_16dp256b_rM_TN_transOut_schedS_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...64_cga2x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...64_cga2x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...4_cga2x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...4_cga2x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...56x32x64_cga2x1x1_16dp256b_rM_TN_transOut_schedS_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...56x32x64_cga2x1x1_16dp256b_rM_TN_transOut_schedS_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...64_cga2x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...64_cga2x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...4_cga1x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...4_cga1x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...128x8x64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...128x8x64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...4_cga1x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...4_cga1x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...128x8x64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...128x8x64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...et128x8_m128x8x64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_bN_tmaOpt_clmp_sm100f_cubin.cpp b/...et128x8_m128x8x64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_bN_tmaOpt_clmp_sm100f_cubin.cpp
diff --git a/...8_m128x8x64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_sm100f_cubin.cpp b/...8_m128x8x64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_sm100f_cubin.cpp
diff --git a/...4_cga1x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...4_cga1x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...128x8x64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...128x8x64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...ga1x1x2_16dp256b_rM_splitK2_TN_transOut_schedPx3_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...ga1x1x2_16dp256b_rM_splitK2_TN_transOut_schedPx3_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...2_16dp256b_rM_splitK2_TN_transOut_schedPx3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...2_16dp256b_rM_splitK2_TN_transOut_schedPx3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...ga1x1x3_16dp256b_rM_splitK3_TN_transOut_schedPx3_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...ga1x1x3_16dp256b_rM_splitK3_TN_transOut_schedPx3_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...3_16dp256b_rM_splitK3_TN_transOut_schedPx3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...3_16dp256b_rM_splitK3_TN_transOut_schedPx3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...ga1x1x4_16dp256b_rM_splitK4_TN_transOut_schedPx3_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...ga1x1x4_16dp256b_rM_splitK4_TN_transOut_schedPx3_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...4_16dp256b_rM_splitK4_TN_transOut_schedPx3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...4_16dp256b_rM_splitK4_TN_transOut_schedPx3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...et128x8_m128x8x64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_bN_tmaOpt_clmp_sm100f_cubin.cpp b/...et128x8_m128x8x64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_bN_tmaOpt_clmp_sm100f_cubin.cpp
diff --git a/...8_m128x8x64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_sm100f_cubin.cpp b/...8_m128x8x64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_sm100f_cubin.cpp
diff --git a/...4_cga1x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...4_cga1x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...128x8x64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...128x8x64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp b/...64_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100f_cubin.cpp
diff --git a/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100a_cubin.cpp b/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100a_cubin.cpp
diff --git a/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm103a_cubin.cpp b/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm103a_cubin.cpp
diff --git a/...32_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100a_cubin.cpp b/...32_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100a_cubin.cpp
diff --git a/...32_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm103a_cubin.cpp b/...32_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm103a_cubin.cpp
diff --git a/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100a_cubin.cpp b/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100a_cubin.cpp
diff --git a/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm103a_cubin.cpp b/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm103a_cubin.cpp
diff --git a/...32_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100a_cubin.cpp b/...32_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100a_cubin.cpp
diff --git a/...32_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm103a_cubin.cpp b/...32_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm103a_cubin.cpp
diff --git a/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100a_cubin.cpp b/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100a_cubin.cpp
diff --git a/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm103a_cubin.cpp b/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm103a_cubin.cpp
diff --git a/...32_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100a_cubin.cpp b/...32_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100a_cubin.cpp
diff --git a/...32_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm103a_cubin.cpp b/...32_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm103a_cubin.cpp
diff --git a/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100a_cubin.cpp b/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm100a_cubin.cpp
diff --git a/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm103a_cubin.cpp b/...x1x1_16dp256b_rM_TN_transOut_schedP2x1x2x3_biasM_bN_tmaOpt_clmp_dynBatch_sm103a_cubin.cpp
diff --git a/...32_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100a_cubin.cpp b/...32_cga1x1x1_16dp256b_rM_TN_transOut_schedS_biasM_bN_tmaOpt_clmp_dynBatch_sm100a_cubin.cpp