[None][feat] Update TRT-LLM Gen MoE kernels (#7970)

Signed-off-by: Nikita Korobov <[email protected]>

Author: nekorobov

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

@@ -110,6 +110,20 @@ TrtllmGenBatchedGemmRunner::TrtllmGenBatchedGemmRunner(TrtllmGenBatchedGemmRunne
             && options.mFusedAct == mOptions.fusedAct && options.mIsStaticBatch == mOptions.staticBatch
             && tileSize == mOptions.tileSize)
         {
+            auto sm = configs[i].mSm;
+            if (sm != SmVersion::Sm100f)
+            {
+                int smVersion = tensorrt_llm::common::getSMVersion();
+                if (smVersion == 100 && sm != SmVersion::Sm100a)
+                {
+                    continue;
+                }
+                else if (smVersion == 103 && sm != SmVersion::Sm103a)
+                {
+                    continue;
+                }
+            }
+
             // FIXME: Disable split-k for now.
             if (options.mClusterDimZ != 1)
             {

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

@@ -548,6 +548,7 @@ class BatchedGemmInterface
     // Aligns the pointer to the alignment
     template <typename Dtype>
     inline Dtype* alignPtr(Dtype* ptr, int64_t alignment) const;
+
     // Returns the size of the workspace buffers in bytes
     std::vector<size_t> getWorkspaceSizesInBytes(BatchedGemmConfig const& config, BatchedGemmData const& data) const;