[NVIDIA] Use sm_100f instead of sm_100a for sgl-kernel and FlashMLA (#33433)

python/sglang/kernels/aot/CMakeLists.txt

@@ -206,14 +206,22 @@ if (ENABLE_BELOW_SM90)
 endif()
 
 if ("${CUDA_VERSION}" VERSION_GREATER_EQUAL "12.8" OR SGL_KERNEL_ENABLE_SM100A)
+    # sm_100f is available in CUDA 12.9+ and covers all sm_100, sm_103, sm_107, sm_10x
+    if ("${CUDA_VERSION}" VERSION_GREATER_EQUAL "12.9")
+        list(APPEND SGL_KERNEL_CUDA_FLAGS
+            "-gencode=arch=compute_100f,code=sm_100f"
+        )
+    else()
+        list(APPEND SGL_KERNEL_CUDA_FLAGS
+            "-gencode=arch=compute_100a,code=sm_100a"
+        )
+    endif()
     list(APPEND SGL_KERNEL_CUDA_FLAGS
-        "-gencode=arch=compute_100a,code=sm_100a"
         "-gencode=arch=compute_120a,code=sm_120a"
     )
     # refer sm_121, sm_110 and sm_101 description  https://github.com/pytorch/pytorch/pull/156176
     if ("${CUDA_VERSION}" VERSION_GREATER_EQUAL "13.0")
         list(APPEND SGL_KERNEL_CUDA_FLAGS
-            "-gencode=arch=compute_103a,code=sm_103a"
             "--compress-mode=size"
         )
         if (CMAKE_SYSTEM_PROCESSOR STREQUAL "aarch64")

python/sglang/kernels/aot/cmake/flashmla.cmake

@@ -35,8 +35,9 @@ if(${CUDA_VERSION} VERSION_GREATER 12.4)
     )
 endif()
 if(${CUDA_VERSION} VERSION_GREATER 12.8)
+    # sm_100f is compatible with all sm_10x
     list(APPEND FLASHMLA_CUDA_FLAGS
-        "-gencode=arch=compute_100a,code=sm_100a"
+        "-gencode=arch=compute_100f,code=sm_100f"
     )
     set(FLASHMLA_ENABLE_SM100 ON)
 endif()
@@ -89,10 +90,6 @@ if(${CUDA_VERSION} VERSION_GREATER_EQUAL "13.0")
     else()
         message(STATUS "cutlass/arch/config.h already patched for SM103a")
     endif()
-
-    list(APPEND FLASHMLA_CUDA_FLAGS
-        "-gencode=arch=compute_103a,code=sm_103a"
-    )
 endif()
 
 

python/sglang/kernels/aot/csrc/moe/fp8_blockwise_moe_kernel.cu

@@ -711,6 +711,9 @@ void fp8_blockwise_scaled_grouped_mm(
   if (sm_version == 100
 #if CUDA_VERSION >= 12090
       || sm_version == 103
+#endif
+#if CUDA_VERSION >= 13040
+      || sm_version == 107
 #endif
   ) {
     if (output.scalar_type() == torch::kBFloat16) {