[None][feat] Add nvfp4 situ moe cubins

cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/KernelRunner.cpp

@@ -45,13 +45,27 @@ using tensorrt_llm::common::fmtstr;
 
 constexpr bool isSMCompatible(int gpuSM, SmVersion kernelSM)
 {
-    if (gpuSM == 103)
+    // Blackwell family. Family-compatible (sm_100f) cubins load on any SM in
+    // [100, 110); arch-specific ones are locked to the SM they were built for.
+    // The batchedGemm drop ships Sm100f, Sm100a and Sm103a only, so family
+    // members other than 100/103 are served by Sm100f alone.
+    // Range kept in sync with tensorrt_llm::common::isSM100Family (cudaUtils.h),
+    // which is not constexpr and so cannot be reused here.
+    if (gpuSM >= 100 && gpuSM < 110)
     {
-        return kernelSM == SmVersion::Sm100f || kernelSM == SmVersion::Sm103a;
-    }
-    else if (gpuSM == 100)
-    {
-        return kernelSM == SmVersion::Sm100f || kernelSM == SmVersion::Sm100a;
+        if (kernelSM == SmVersion::Sm100f)
+        {
+            return true;
+        }
+        if (gpuSM == 100)
+        {
+            return kernelSM == SmVersion::Sm100a;
+        }
+        if (gpuSM == 103)
+        {
+            return kernelSM == SmVersion::Sm103a;
+        }
+        return false;
     }
     else if (gpuSM == 90)
     {

cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/BatchedGemmInterface.h

@@ -958,14 +958,9 @@ class BatchedGemmInterface
         auto options = getOptionsFromConfigAndData(config, data);
 
         // Check options without modifications.
-        //
-        // NOTE(TRT-LLM local fix over producer output a2ad0544-dirty): the generator
-        // inserted the `worldSize` parameter into checkAndUpdateBatchedGemmOptions
-        // without updating this call, so the literal `false` bound positionally to
-        // `worldSize` (=0) and `updateOptions` silently took its default `true`.
-        // Pass both arguments explicitly. Report upstream before the next re-export.
         return checkAndUpdateBatchedGemmOptions(options, config.mSm,
-            /* worldSize */ 1, /* updateOptions */ false);
+            /* worldSize */ 1,
+            /* updateOptions */ false);
     }
 
     //////////////////////////////////////////////////////////////////////////////////////////////////

cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/GemmOptions.h

@@ -2148,8 +2148,8 @@ 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 std::string const tllm_gen_gemm_cubin_path = std::string(TLLM_GEN_GEMM_CUBIN_PATH);
-    std::string const sha256 = config.mHash ? config.mHash : "";
+    static const std::string tllm_gen_gemm_cubin_path = std::string(TLLM_GEN_GEMM_CUBIN_PATH);
+    const std::string sha256 = config.mHash ? config.mHash : "";
     std::string fileName = config.mFunctionName;
     if (!fileName.empty())
     {

cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/KernelParams.h

@@ -232,7 +232,7 @@ 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, int32_t const numEltsPerSf)
+    tg::SfLayout layout, int sfReshapeFactor, const int32_t numEltsPerSf)
 {
 
     // The outer dimension.

cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/KernelTraits.h

@@ -215,6 +215,8 @@ class KernelTraits
         : mMmaKind{mmaKind}
         , mFuseUtccpWithUtcmma{fuseUtccpWithUtcmma}
         , mUseMaxTmemOverlap{useMaxTmemOverlap}
+        , mUseTmaStore{useTmaStore}
+        , mUsePersistentScheduler{usePersistentScheduler}
         , mUseCustomizedMma3xNvFp4{useCustomizedMma3xNvFp4}
         , mFusedBiasShuffleMode{fusedBiasShuffleMode}
         , mNumEpilogueWarps{numEpilogueWarps}
@@ -609,6 +611,10 @@ tg::MmaKind mMmaKind{};
 bool mFuseUtccpWithUtcmma{};
 // Whether use the max TMEM overlap trick.
 bool mUseMaxTmemOverlap{};
+// Whether the epilogue stages output in SMEM for an asynchronous TMA store.
+bool mUseTmaStore{};
+// Whether work tiles are assigned by a persistent scheduler.
+bool mUsePersistentScheduler{};
 // Whether use customized MMA for 3xNvFp4
 bool mUseCustomizedMma3xNvFp4{};
 // Which BiasType::Mn preprocessing steps are fused into the kernel instead of the host.

cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/trtllmGen_bmm_export/config.json

@@ -301,6 +301,11 @@
                     "geglu",
                     "none"
                 ],
+                [
+                    true,
+                    "situ",
+                    "none"
+                ],
                 [
                     false,
                     "swiglu",
@@ -414,6 +419,11 @@
                     "geglu",
                     "none"
                 ],
+                [
+                    true,
+                    "situ",
+                    "none"
+                ],
                 [
                     false,
                     "none",
@@ -497,6 +507,11 @@
                     "geglu",
                     "none"
                 ],
+                [
+                    true,
+                    "situ",
+                    "none"
+                ],
                 [
                     false,
                     "none",
@@ -563,6 +578,11 @@
                     "geglu",
                     "none"
                 ],
+                [
+                    true,
+                    "situ",
+                    "none"
+                ],
                 [
                     false,
                     "none",
@@ -610,6 +630,11 @@
                     "geglu",
                     "none"
                 ],
+                [
+                    true,
+                    "situ",
+                    "none"
+                ],
                 [
                     false,
                     "none",
@@ -663,6 +688,11 @@
                     "geglu",
                     "none"
                 ],
+                [
+                    true,
+                    "situ",
+                    "none"
+                ],
                 [
                     false,
                     "none",
@@ -713,6 +743,11 @@
                     "geglu",
                     "none"
                 ],
+                [
+                    true,
+                    "situ",
+                    "none"
+                ],
                 [
                     false,
                     "none",
@@ -760,6 +795,11 @@
                     "geglu",
                     "none"
                 ],
+                [
+                    true,
+                    "situ",
+                    "none"
+                ],
                 [
                     false,
                     "none",