[ROCm] cpp_extension allow user to override default flags (pytorch#152432) (#2374)

cherry-pick of
pytorch@e4adf5d

We need -fgpu-rdc for projects such as DeepEP + rocSHMEM. The default of
-no-gpu-rdc doesn't work for such cases.

As per
pytorch#152432 (comment):
"rocshmem shares the same global variable in different files, as deepEP
uses CUDAExtention to build the project
https://github.com/deepseek-ai/DeepEP/blob/65e2a700f0330f3fb1c26f49a0250d1f9d0ac1e3/setup.py#L51
and depends on rocshmem, this -fgpu-rdc is needed. The current logic in
Pytorch prevents users from overriding this flag."

Pull Request resolved: pytorch#152432
Approved by: https://github.com/jeffdaily

Co-authored-by: Jithun Nair <[email protected]>
Co-authored-by: Jeff Daily <[email protected]>

Author: 3 people

torch/utils/cpp_extension.py

@@ -2110,11 +2110,18 @@ def _get_cuda_arch_flags(cflags: Optional[List[str]] = None) -> List[str]:
 
 def _get_rocm_arch_flags(cflags: Optional[List[str]] = None) -> List[str]:
     # If cflags is given, there may already be user-provided arch flags in it
-    # (from `extra_compile_args`)
+    # (from `extra_compile_args`). If user also specified -fgpu-rdc or -fno-gpu-rdc, we
+    # assume they know what they're doing. Otherwise, we force -fno-gpu-rdc default.
+    has_gpu_rdc_flag = False
     if cflags is not None:
+        has_custom_flags = False
         for flag in cflags:
             if 'amdgpu-target' in flag or 'offload-arch' in flag:
-                return ['-fno-gpu-rdc']
+                has_custom_flags = True
+            elif 'gpu-rdc' in flag:
+                has_gpu_rdc_flag = True
+        if has_custom_flags:
+            return [] if has_gpu_rdc_flag else ['-fno-gpu-rdc']
     # Use same defaults as used for building PyTorch
     # Allow env var to override, just like during initial cmake build.
     _archs = os.environ.get('PYTORCH_ROCM_ARCH', None)
@@ -2127,7 +2134,7 @@ def _get_rocm_arch_flags(cflags: Optional[List[str]] = None) -> List[str]:
     else:
         archs = _archs.replace(' ', ';').split(';')
     flags = [f'--offload-arch={arch}' for arch in archs]
-    flags += ['-fno-gpu-rdc']
+    flags += [] if has_gpu_rdc_flag else ['-fno-gpu-rdc']
     return flags
 
 def _get_build_directory(name: str, verbose: bool) -> str:
@@ -2312,8 +2319,8 @@ def _write_ninja_file_to_build_library(path,
 
     if with_cuda and IS_HIP_EXTENSION:
         cuda_flags = ['-DWITH_HIP'] + cflags + COMMON_HIP_FLAGS + COMMON_HIPCC_FLAGS
-        cuda_flags += extra_cuda_cflags
         cuda_flags += _get_rocm_arch_flags(cuda_flags)
+        cuda_flags += extra_cuda_cflags
     elif with_cuda:
         cuda_flags = common_cflags + COMMON_NVCC_FLAGS + _get_cuda_arch_flags()
         if IS_WINDOWS: