Skip to content

Fix linux cpu builds. Resolves nightly build for mac stops on 0422 #2170

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 2 commits into from
May 6, 2025
Merged

Fix linux cpu builds. Resolves nightly build for mac stops on 0422 #2170

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
0b0385d
Revert "[reland][ROCm] preshuffled weight mm (#2044)"
atalman May 5, 2025
53b4b28
Revert "Re-land "Add INT8 SDPA path for CPU" (#2093)"
atalman May 5, 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
125 changes: 29 additions & 96 deletions setup.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -55,10 +55,6 @@ def read_version(file_path="version.txt"):
and platform.system() == "Darwin"
)

use_cpp_avx512 = os.getenv("USE_AVX512", "1") == "1" and platform.system() == "Linux"

from torchao.utils import TORCH_VERSION_AT_LEAST_2_7

version_prefix = read_version()
# Version is version.dev year month date if using nightlies and version if not
version = (
Expand All @@ -83,6 +79,8 @@ def use_debug_mode():
_get_cuda_arch_flags,
)

IS_ROCM = (torch.version.hip is not None) and (ROCM_HOME is not None)


class BuildOptions:
def __init__(self):
Expand Down Expand Up @@ -257,53 +255,35 @@ def get_extensions():
print(
"PyTorch GPU support is not available. Skipping compilation of CUDA extensions"
)
if CUDA_HOME is None and torch.version.cuda:
print("CUDA toolkit is not available. Skipping compilation of CUDA extensions")
if (CUDA_HOME is None and ROCM_HOME is None) and torch.cuda.is_available():
print(
"CUDA toolkit or ROCm is not available. Skipping compilation of CUDA extensions"
)
print(
"If you'd like to compile CUDA extensions locally please install the cudatoolkit from https://anaconda.org/nvidia/cuda-toolkit"
)
if ROCM_HOME is None and torch.version.hip:
print("ROCm is not available. Skipping compilation of ROCm extensions")
print("If you'd like to compile ROCm extensions locally please install ROCm")

use_cuda = torch.version.cuda and CUDA_HOME is not None
use_hip = torch.version.hip and ROCM_HOME is not None
extension = CUDAExtension if (use_cuda or use_hip) else CppExtension

nvcc_args = [
"-DNDEBUG" if not debug_mode else "-DDEBUG",
"-O3" if not debug_mode else "-O0",
"-t=0",
"-std=c++17",
]
hip_args = [
"-DNDEBUG" if not debug_mode else "-DDEBUG",
"-O3" if not debug_mode else "-O0",
"-std=c++17",
]

use_cuda = torch.cuda.is_available() and (
CUDA_HOME is not None or ROCM_HOME is not None
)
extension = CUDAExtension if use_cuda else CppExtension

extra_link_args = []
extra_compile_args = {
"cxx": [f"-DPy_LIMITED_API={PY3_9_HEXCODE}"],
"nvcc": nvcc_args if use_cuda else hip_args,
"nvcc": [
"-DNDEBUG" if not debug_mode else "-DDEBUG",
"-O3" if not debug_mode else "-O0",
"-t=0",
"-std=c++17",
],
}

if not IS_WINDOWS:
extra_compile_args["cxx"].extend(
["-O3" if not debug_mode else "-O0", "-fdiagnostics-color=always"]
)

if use_cpp_avx512 and TORCH_VERSION_AT_LEAST_2_7:
if torch._C._cpu._is_avx512_supported():
extra_compile_args["cxx"].extend(
[
"-DCPU_CAPABILITY_AVX512",
"-march=native",
"-mfma",
"-fopenmp",
]
)

if debug_mode:
extra_compile_args["cxx"].append("-g")
if "nvcc" in extra_compile_args:
Expand All @@ -319,95 +299,48 @@ def get_extensions():
extra_compile_args["nvcc"].append("-g")
extra_link_args.append("/DEBUG")

hip_sparse_marlin_supported = True
if use_hip:
# naive search for hipblalst.h, if any found contain HIPBLASLT_ORDER_COL16 and VEC_EXT
found_col16 = False
found_vec_ext = False
print("ROCM_HOME", ROCM_HOME)
hipblaslt_headers = list(
glob.glob(os.path.join(ROCM_HOME, "include", "hipblaslt", "hipblaslt.h"))
)
print("hipblaslt_headers", hipblaslt_headers)
for header in hipblaslt_headers:
with open(header) as f:
text = f.read()
if "HIPBLASLT_ORDER_COL16" in text:
found_col16 = True
if "HIPBLASLT_MATMUL_DESC_A_SCALE_POINTER_VEC_EXT" in text:
found_vec_ext = True
if found_col16:
extra_compile_args["cxx"].append("-DHIPBLASLT_HAS_ORDER_COL16")
print("hipblaslt found extended col order enums")
else:
print("hipblaslt does not have extended col order enums")
if found_vec_ext:
extra_compile_args["cxx"].append("-DHIPBLASLT_VEC_EXT")
print("hipblaslt found vec ext")
else:
print("hipblaslt does not have vec ext")

# sparse_marlin depends on features in ROCm 6.4, __builtin_amdgcn_global_load_lds
ROCM_VERSION = tuple(int(v) for v in torch.version.hip.split(".")[:2])
hip_sparse_marlin_supported = ROCM_VERSION >= (6, 4)

# Get base directory and source paths
curdir = os.path.dirname(os.path.curdir)
extensions_dir = os.path.join(curdir, "torchao", "csrc")

# Collect C++ source files
sources = list(glob.glob(os.path.join(extensions_dir, "**/*.cpp"), recursive=True))
if IS_WINDOWS:
# Remove csrc/cpu/*.cpp on Windows due to the link issue: unresolved external symbol PyInit__C
excluded_sources = list(
glob.glob(os.path.join(extensions_dir, "cpu/*.cpp"), recursive=True)
)
sources = [s for s in sources if s not in excluded_sources]

# Collect CUDA source files
extensions_cuda_dir = os.path.join(extensions_dir, "cuda")
cuda_sources = list(
glob.glob(os.path.join(extensions_cuda_dir, "**/*.cu"), recursive=True)
)

# Collect HIP source files
extensions_hip_dir = os.path.join(
extensions_dir, "cuda", "tensor_core_tiled_layout"
)
hip_sources = list(
glob.glob(os.path.join(extensions_hip_dir, "*.cu"), recursive=True)
)
if hip_sparse_marlin_supported:
extensions_hip_dir = os.path.join(extensions_dir, "cuda", "sparse_marlin")
hip_sources += list(
glob.glob(os.path.join(extensions_hip_dir, "*.cu"), recursive=True)
)
extensions_hip_dir = os.path.join(extensions_dir, "rocm")
hip_sources += list(
glob.glob(os.path.join(extensions_hip_dir, "**/*.hip"), recursive=True)
)
extensions_hip_dir = os.path.join(extensions_dir, "cuda", "sparse_marlin")
hip_sources += list(
glob.glob(os.path.join(extensions_hip_dir, "**/*.cpp"), recursive=True)
glob.glob(os.path.join(extensions_hip_dir, "*.cu"), recursive=True)
)

# Add CUDA source files if needed
if use_cuda:
# Collect CUDA source files if needed
if not IS_ROCM and use_cuda:
sources += cuda_sources

# TODO: Remove this and use what CUDA has once we fix all the builds.
# Add HIP source files if needed
if use_hip:
# TOOD: Remove this and use what CUDA has once we fix all the builds.
if IS_ROCM and use_cuda:
# Add ROCm GPU architecture check
gpu_arch = torch.cuda.get_device_properties(0).name
if 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 += hip_sources
print(
"Currently only gfx942 is supported. Skipping compilation of ROCm extensions"
)
else:
sources += hip_sources

use_cutlass = False
cutlass_90a_sources = None
if use_cuda and not IS_WINDOWS:
if use_cuda and not IS_ROCM and not IS_WINDOWS:
use_cutlass = True
cutlass_dir = os.path.join(third_party_path, "cutlass")
cutlass_include_dir = os.path.join(cutlass_dir, "include")
Expand Down
Loading
Loading