How to assign ROCm architecture during pip installing

[deeptimhe]

I am installing flash-attn rocm version (2.7.0.post2), but I only have a A100 GPU machine. It gives error when pip install --no-cache-dir flash-attn --no-build-isolation:

Using envvar MAX_JOBS (12) as the number of workers...
      [1/7] /opt/rocm-6.2.4/bin/hipcc  -I/tmp/pip-install-2zfjkyvz/flash-attn_fd8e4c61763a4b06b464c6b98ad7105f/csrc/composable_kernel/include -I/tmp/pip-install-2zfjkyvz/flash-attn_fd8e4c61763a4b06b464c6b98ad7105f/csrc/composable_kernel/library/include -I/tmp/pip-install-2zfjkyvz/flash-attn_fd8e4c61763a4b06b464c6b98ad7105f/csrc/composable_kernel/example/ck_tile/01_fmha -I/usr/local/lib/python3.10/dist-packages/torch/include -I/usr/local/lib/python3.10/dist-packages/torch/include/torch/csrc/api/include -I/usr/local/lib/python3.10/dist-packages/torch/include/TH -I/usr/local/lib/python3.10/dist-packages/torch/include/THC -I/usr/local/lib/python3.10/dist-packages/torch/include/THH -I/opt/rocm-6.2.4/include -I/usr/include/python3.10 -c -c /tmp/pip-install-2zfjkyvz/flash-attn_fd8e4c61763a4b06b464c6b98ad7105f/csrc/flash_attn_ck/mha_bwd.hip -o /tmp/pip-install-2zfjkyvz/flash-attn_fd8e4c61763a4b06b464c6b98ad7105f/build/temp.linux-x86_64-3.10/csrc/flash_attn_ck/mha_bwd.o -fPIC -D__HIP_PLATFORM_AMD__=1 -DUSE_ROCM=1 -DHIPBLAS_V2 -DCUDA_HAS_FP16=1 -D__HIP_NO_HALF_OPERATORS__=1 -D__HIP_NO_HALF_CONVERSIONS__=1 --offload-arch=native -O3 -std=c++17 -DCK_TILE_FMHA_FWD_FAST_EXP2=1 -fgpu-flush-denormals-to-zero -DCK_ENABLE_BF16 -DCK_ENABLE_BF8 -DCK_ENABLE_FP16 -DCK_ENABLE_FP32 -DCK_ENABLE_FP64 -DCK_ENABLE_FP8 -DCK_ENABLE_INT8 -DCK_USE_XDL -DUSE_PROF_API=1 -D__HIP_PLATFORM_HCC__=1 -DCK_TILE_FLOAT_TO_BFLOAT16_DEFAULT=3 -fno-offload-uniform-block -mllvm -enable-post-misched=0 -mllvm -amdgpu-early-inline-all=true -mllvm -amdgpu-function-calls=false -mllvm -amdgpu-coerce-illegal-types=1 -DTORCH_API_INCLUDE_EXTENSION_H '-DPYBIND11_COMPILER_TYPE="_gcc"' '-DPYBIND11_STDLIB="_libstdcpp"' '-DPYBIND11_BUILD_ABI="_cxxabi1011"' -DTORCH_EXTENSION_NAME=flash_attn_2_cuda -D_GLIBCXX_USE_CXX11_ABI=0 -fno-gpu-rdc
      FAILED: /tmp/pip-install-2zfjkyvz/flash-attn_fd8e4c61763a4b06b464c6b98ad7105f/build/temp.linux-x86_64-3.10/csrc/flash_attn_ck/mha_bwd.o
      /opt/rocm-6.2.4/bin/hipcc  -I/tmp/pip-install-2zfjkyvz/flash-attn_fd8e4c61763a4b06b464c6b98ad7105f/csrc/composable_kernel/include -I/tmp/pip-install-2zfjkyvz/flash-attn_fd8e4c61763a4b06b464c6b98ad7105f/csrc/composable_kernel/library/include -I/tmp/pip-install-2zfjkyvz/flash-attn_fd8e4c61763a4b06b464c6b98ad7105f/csrc/composable_kernel/example/ck_tile/01_fmha -I/usr/local/lib/python3.10/dist-packages/torch/include -I/usr/local/lib/python3.10/dist-packages/torch/include/torch/csrc/api/include -I/usr/local/lib/python3.10/dist-packages/torch/include/TH -I/usr/local/lib/python3.10/dist-packages/torch/include/THC -I/usr/local/lib/python3.10/dist-packages/torch/include/THH -I/opt/rocm-6.2.4/include -I/usr/include/python3.10 -c -c /tmp/pip-install-2zfjkyvz/flash-attn_fd8e4c61763a4b06b464c6b98ad7105f/csrc/flash_attn_ck/mha_bwd.hip -o /tmp/pip-install-2zfjkyvz/flash-attn_fd8e4c61763a4b06b464c6b98ad7105f/build/temp.linux-x86_64-3.10/csrc/flash_attn_ck/mha_bwd.o -fPIC -D__HIP_PLATFORM_AMD__=1 -DUSE_ROCM=1 -DHIPBLAS_V2 -DCUDA_HAS_FP16=1 -D__HIP_NO_HALF_OPERATORS__=1 -D__HIP_NO_HALF_CONVERSIONS__=1 --offload-arch=native -O3 -std=c++17 -DCK_TILE_FMHA_FWD_FAST_EXP2=1 -fgpu-flush-denormals-to-zero -DCK_ENABLE_BF16 -DCK_ENABLE_BF8 -DCK_ENABLE_FP16 -DCK_ENABLE_FP32 -DCK_ENABLE_FP64 -DCK_ENABLE_FP8 -DCK_ENABLE_INT8 -DCK_USE_XDL -DUSE_PROF_API=1 -D__HIP_PLATFORM_HCC__=1 -DCK_TILE_FLOAT_TO_BFLOAT16_DEFAULT=3 -fno-offload-uniform-block -mllvm -enable-post-misched=0 -mllvm -amdgpu-early-inline-all=true -mllvm -amdgpu-function-calls=false -mllvm -amdgpu-coerce-illegal-types=1 -DTORCH_API_INCLUDE_EXTENSION_H '-DPYBIND11_COMPILER_TYPE="_gcc"' '-DPYBIND11_STDLIB="_libstdcpp"' '-DPYBIND11_BUILD_ABI="_cxxabi1011"' -DTORCH_EXTENSION_NAME=flash_attn_2_cuda -D_GLIBCXX_USE_CXX11_ABI=0 -fno-gpu-rdc
      clang++: error: cannot determine amdgcn architecture: /opt/rocm-6.2.4/lib/llvm/bin/amdgpu-arch: ; consider passing it via '--offload-arch'
      failed to execute:/opt/rocm-6.2.4/lib/llvm/bin/clang++  --offload-arch=native  -I/tmp/pip-install-2zfjkyvz/flash-attn_fd8e4c61763a4b06b464c6b98ad7105f/csrc/composable_kernel/include -I/tmp/pip-install-2zfjkyvz/flash-attn_fd8e4c61763a4b06b464c6b98ad7105f/csrc/composable_kernel/library/include -I/tmp/pip-install-2zfjkyvz/flash-attn_fd8e4c61763a4b06b464c6b98ad7105f/csrc/composable_kernel/example/ck_tile/01_fmha -I/usr/local/lib/python3.10/dist-packages/torch/include -I/usr/local/lib/python3.10/dist-packages/torch/include/torch/csrc/api/include -I/usr/local/lib/python3.10/dist-packages/torch/include/TH -I/usr/local/lib/python3.10/dist-packages/torch/include/THC -I/usr/local/lib/python3.10/dist-packages/torch/include/THH -I/opt/rocm-6.2.4/include -I/usr/include/python3.10 -c -c -x hip /tmp/pip-install-2zfjkyvz/flash-attn_fd8e4c61763a4b06b464c6b98ad7105f/csrc/flash_attn_ck/mha_bwd.hip -o "/tmp/pip-install-2zfjkyvz/flash-attn_fd8e4c61763a4b06b464c6b98ad7105f/build/temp.linux-x86_64-3.10/csrc/flash_attn_ck/mha_bwd.o" -fPIC -D__HIP_PLATFORM_AMD__=1 -DUSE_ROCM=1 -DHIPBLAS_V2 -DCUDA_HAS_FP16=1 -D__HIP_NO_HALF_OPERATORS__=1 -D__HIP_NO_HALF_CONVERSIONS__=1 -O3 -std=c++17 -DCK_TILE_FMHA_FWD_FAST_EXP2=1 -fgpu-flush-denormals-to-zero -DCK_ENABLE_BF16 -DCK_ENABLE_BF8 -DCK_ENABLE_FP16 -DCK_ENABLE_FP32 -DCK_ENABLE_FP64 -DCK_ENABLE_FP8 -DCK_ENABLE_INT8 -DCK_USE_XDL -DUSE_PROF_API=1 -D__HIP_PLATFORM_HCC__=1 -DCK_TILE_FLOAT_TO_BFLOAT16_DEFAULT=3 -fno-offload-uniform-block -mllvm -enable-post-misched=0 -mllvm -amdgpu-early-inline-all=true -mllvm -amdgpu-function-calls=false -mllvm -amdgpu-coerce-illegal-types=1 -DTORCH_API_INCLUDE_EXTENSION_H -DPYBIND11_COMPILER_TYPE=\"_gcc\" -DPYBIND11_STDLIB=\"_libstdcpp\" -DPYBIND11_BUILD_ABI=\"_cxxabi1011\" -DTORCH_EXTENSION_NAME=flash_attn_2_cuda -D_GLIBCXX_USE_CXX11_ABI=0 -fno-gpu-rdc

Seems I should assign amdgpu-arch when installing?