diff '--color=auto' -ur pytorch-v2.9.0.orig/aten/src/ATen/CMakeLists.txt pytorch-v2.9.0/aten/src/ATen/CMakeLists.txt
--- pytorch-v2.9.0.orig/aten/src/ATen/CMakeLists.txt	2025-10-16 00:08:14.098872822 +0200
+++ pytorch-v2.9.0/aten/src/ATen/CMakeLists.txt	2025-10-16 00:18:27.659279152 +0200
@@ -284,48 +284,64 @@
       "${FBGEMM_GENAI_SRCS}/cutlass_extensions/mx8mx8bf16_grouped/"
     )
 
-    target_include_directories(fbgemm_genai PUBLIC
+    target_include_directories(fbgemm_genai PRIVATE
       ${FBGEMM_THIRD_PARTY}/cutlass/include
       ${FBGEMM_THIRD_PARTY}/cutlass/tools/util/include
       ${fbgemm_genai_mx8mx8bf16_grouped}
       ${FBGEMM_GENAI_SRCS}/common/include/   # includes fbgemm_gpu/quantize/utils.h, fbgemm_gpu/quantize/tuning_cache.hpp
       ${FBGEMM_GENAI_SRCS}/include/          # includes fbgemm_gpu/torch_ops.h
     )
-  else()
-    if(USE_ROCM)
-      # Only include the kernels we want to build to avoid increasing binary size.
-      file(GLOB_RECURSE fbgemm_genai_native_rocm_hip
-        "${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/kernels/fp8_rowwise_grouped*.hip"
-        "${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/fp8_rowwise_grouped_gemm.hip")
-      set_source_files_properties(${fbgemm_genai_native_rocm_hip} PROPERTIES HIP_SOURCE_PROPERTY_FORMAT 1)
-
-      # Add additional HIPCC compiler flags for performance
-      set(FBGEMM_GENAI_EXTRA_HIPCC_FLAGS
-        -mllvm
-        -amdgpu-coerce-illegal-types=1
-        -mllvm
-        -enable-post-misched=0
-        -mllvm
-        -greedy-reverse-local-assignment=1
-        -fhip-new-launch-api)
-
-      hip_add_library(
-        fbgemm_genai STATIC
-        ${fbgemm_genai_native_rocm_hip}
-        HIPCC_OPTIONS ${HIP_HCC_FLAGS} ${FBGEMM_GENAI_EXTRA_HIPCC_FLAGS})
-      set_target_properties(fbgemm_genai PROPERTIES POSITION_INDEPENDENT_CODE ON)
-      target_compile_definitions(fbgemm_genai PRIVATE FBGEMM_GENAI_NO_EXTENDED_SHAPES)
-
-      target_include_directories(fbgemm_genai PUBLIC
-        # FBGEMM version of Composable Kernel is used due to some customizations
-        ${FBGEMM_THIRD_PARTY}/composable_kernel/include
-        ${FBGEMM_THIRD_PARTY}/composable_kernel/library/include
-        ${FBGEMM_THIRD_PARTY}/cutlass/include
-        ${FBGEMM_THIRD_PARTY}/cutlass/tools/util/include
-        ${FBGEMM_GENAI_SRCS}/common/include/   # includes fbgemm_gpu/quantize/utils.h, fbgemm_gpu/quantize/tuning_cache.hpp
-        ${FBGEMM_GENAI_SRCS}/include/          # includes fbgemm_gpu/torch_ops.h
-      )
+
+    # Add FBGEMM_GENAI include directories for torch_ops.h
+    list(APPEND ATen_CUDA_INCLUDE ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/include)
+    list(APPEND ATen_CUDA_INCLUDE ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/common/include)
+  elseif(USE_ROCM)
+    # Only include the kernels we want to build to avoid increasing binary size.
+    file(GLOB_RECURSE fbgemm_genai_native_rocm_hip
+      "${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/kernels/fp8_rowwise_grouped*.hip"
+      "${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/fp8_rowwise_grouped_gemm.hip")
+    set_source_files_properties(${fbgemm_genai_native_rocm_hip} PROPERTIES HIP_SOURCE_PROPERTY_FORMAT 1)
+
+    # Add additional HIPCC compiler flags for performance
+    set(FBGEMM_GENAI_EXTRA_HIPCC_FLAGS
+      -mllvm
+      -amdgpu-coerce-illegal-types=1
+      -mllvm
+      -enable-post-misched=0
+      -mllvm
+      -greedy-reverse-local-assignment=1
+      -fhip-new-launch-api)
+
+    # Only compile for gfx942 for now.
+    # This is rather hacky, I could not figure out a clean solution :(
+    set(HIP_CLANG_FLAGS_ORIGINAL ${HIP_CLANG_FLAGS})
+    string(REGEX REPLACE "--offload-arch=[^ ]*" "" FILTERED_HIP_CLANG_FLAGS "${HIP_CLANG_FLAGS}")
+    if("gfx942" IN_LIST PYTORCH_ROCM_ARCH)
+      list(APPEND FILTERED_HIP_CLANG_FLAGS --offload-arch=gfx942;)
     endif()
+    set(HIP_CLANG_FLAGS ${FILTERED_HIP_CLANG_FLAGS})
+
+    hip_add_library(
+      fbgemm_genai STATIC
+      ${fbgemm_genai_native_rocm_hip}
+      HIPCC_OPTIONS ${HIP_HCC_FLAGS} ${FBGEMM_GENAI_EXTRA_HIPCC_FLAGS})
+    set(HIP_CLANG_FLAGS ${HIP_CLANG_FLAGS_ORIGINAL})
+    set_target_properties(fbgemm_genai PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    target_compile_definitions(fbgemm_genai PRIVATE FBGEMM_GENAI_NO_EXTENDED_SHAPES)
+
+    target_include_directories(fbgemm_genai PRIVATE
+      # FBGEMM version of Composable Kernel is used due to some customizations
+      ${FBGEMM_THIRD_PARTY}/composable_kernel/include
+      ${FBGEMM_THIRD_PARTY}/composable_kernel/library/include
+      ${FBGEMM_THIRD_PARTY}/cutlass/include
+      ${FBGEMM_THIRD_PARTY}/cutlass/tools/util/include
+      ${FBGEMM_GENAI_SRCS}/common/include/   # includes fbgemm_gpu/quantize/utils.h, fbgemm_gpu/quantize/tuning_cache.hpp
+      ${FBGEMM_GENAI_SRCS}/include/          # includes fbgemm_gpu/torch_ops.h
+    )
+
+    # Add FBGEMM_GENAI include directories for torch_ops.h
+    list(APPEND ATen_HIP_INCLUDE ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/include)
+    list(APPEND ATen_HIP_INCLUDE ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/common/include)
   endif()
 endif()
 
@@ -663,12 +679,6 @@
   list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/cutlass/include)
   list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/cutlass/tools/util/include)
 
-  # Add FBGEMM_GENAI include directories for torch_ops.h
-  if(USE_FBGEMM_GENAI)
-    list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/include)
-    list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/common/include)
-  endif()
-
   if($ENV{ATEN_STATIC_CUDA})
     if(CUDA_VERSION VERSION_LESS_EQUAL 12.9)
       list(APPEND ATen_CUDA_DEPENDENCY_LIBS