Merge branch 'main' into user/fanrongl/mtp3_support_for_ds32

Author: lfr-0531

cpp/tensorrt_llm/kernels/cutlass_kernels/fp8_rowwise_gemm/fp8_rowwise_gemm_template.h

@@ -686,7 +686,7 @@ size_t CutlassFp8RowwiseGemmRunner<T>::dispatchToArch(void* D, void const* A, vo
         return dispatchGemmToCutlassSm90<T>(D, A, B, C_bias, quantOption, m, n, k, scale_d0, scale_d1, gemmConfig,
             workspace, workspaceBytes, stream, occupancy);
     }
-    else if (mSm == 100)
+    else if (mSm == 100 || mSm == 103)
     {
         return dispatchGemmToCutlassSm100<T>(D, A, B, C_bias, quantOption, m, n, k, scale_d0, scale_d1, gemmConfig,
             workspace, workspaceBytes, stream, occupancy);
@@ -758,7 +758,7 @@ std::vector<tkc::CutlassGemmConfig> CutlassFp8RowwiseGemmRunner<T>::getConfigs()
             }
         }
     }
-    else if (mSm == 100)
+    else if (mSm == 100 || mSm == 103)
     {
         std::vector<tkc::CutlassTileConfigSM100> tilesSm100 = {
             tkc::CutlassTileConfigSM100::CtaShape64x32x128B,

security_scanning/docs/poetry.lock

@@ -10,7 +10,7 @@ python = ">=3.10,<3.13"
 sphinx = ">=7.0"
 sphinx-argparse = "^0.5.2"
 sphinx-click = "^6.1.0"
-nvidia-sphinx-theme = "^0.0.8"
+nvidia-sphinx-theme = "^0.0.9.post1"
 myst-parser = "^4.0.1"
 breathe = "^4.36.0"
 sphinx-copybutton = "^0.5.2"