Skip to content

Conditional ROCm kernel build #2839

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Aug 26, 2025
Merged

Conditional ROCm kernel build #2839

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
abd108e
conditional kernel build
Aug 21, 2025
fbb24b5
lint
Aug 21, 2025
4eb0357
Remove GPU architecture check for ROCm in setup.py and add a TODO for…
Aug 26, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
14 changes: 6 additions & 8 deletions setup.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -433,6 +433,7 @@ def get_extensions():
extra_link_args.append("/DEBUG")

rocm_sparse_marlin_supported = False
rocm_tiled_layout_supported = False
if use_rocm:
# naive search for hipblalst.h, if any found contain HIPBLASLT_ORDER_COL16 and VEC_EXT
found_col16 = False
Expand Down Expand Up @@ -488,8 +489,11 @@ def get_extensions():
# Define ROCm source directories
rocm_source_dirs = [
os.path.join(extensions_dir, "rocm", "swizzle"),
os.path.join(extensions_dir, "cuda", "tensor_core_tiled_layout"),
]
if rocm_tiled_layout_supported:
rocm_source_dirs.append(
os.path.join(extensions_dir, "cuda", "tensor_core_tiled_layout")
)
if rocm_sparse_marlin_supported:
rocm_source_dirs.extend([os.path.join(extensions_dir, "cuda", "sparse_marlin")])

Expand All @@ -512,14 +516,8 @@ def get_extensions():
sources = [s for s in sources if s not in mxfp8_sources_to_exclude]

# TOOD: Remove this and use what CUDA has once we fix all the builds.
# TODO: Add support for other ROCm GPUs
if use_rocm:
# Add ROCm GPU architecture check
gpu_arch = None
if torch.cuda.is_available():
gpu_arch = torch.cuda.get_device_properties(0).name
if gpu_arch and gpu_arch != "gfx942":
print(f"Warning: Unsupported ROCm GPU architecture: {gpu_arch}")
print("Currently only gfx942 is supported. Compiling only for gfx942.")
extra_compile_args["nvcc"].append("--offload-arch=gfx942")
sources += rocm_sources
else:
Expand Down
Loading