Files
DeepSpeed/scripts/check-torchcuda.py
Junjie Mao 6b731c5c96 scripts: Check .is_cuda only in non-C++ files (#7561)
The check-torchcuda.py today will search for all occurrences of .is_cuda
in the repository when a commit only modifies C++ headers and sources,
which I believe is not intended.

Check usage of .is_cuda only when a commit modifies any non-C++ file.

Signed-off-by: Junjie Mao <junjie.mao@linux.alibaba.com>
2025-09-19 05:01:50 +00:00

81 lines
2.9 KiB
Python
Executable File

#!/usr/bin/env python3
# Copyright (c) Microsoft Corporation.
# SPDX-License-Identifier: Apache-2.0
# DeepSpeed Team
from __future__ import annotations
'''Copyright The Microsoft DeepSpeed Team'''
"""
Checks each file in sys.argv for the string "torch.cuda".
Modified from https://github.com/jlebar/pre-commit-hooks/blob/master/check_do_not_submit.py
"""
import subprocess
import sys
def err(s: str) -> None:
print(s, file=sys.stderr)
print(*sys.argv[1:])
# There are many ways we could search for the string "torch.cuda", but `git
# grep --no-index` is nice because
# - it's very fast (as compared to iterating over the file in Python)
# - we can reasonably assume it's available on all machines
# - unlike plain grep, which is slower and has different flags on MacOS versus
# Linux, git grep is always the same.
res = subprocess.run(
["git", "grep", "-Hn", "--no-index", "-e", r"torch\.cuda", "--and", "--not", "-e", "#ignore-cuda", *sys.argv[1:]],
capture_output=True,
)
if res.returncode == 0:
err('Error: The string "torch.cuda" was found.\nPlease replace all calls to torch.cuda with "get_accelerator()" and add the following import line:\n\n from deepspeed.accelerator import get_accelerator\n\nIf your code is mean to be cuda specific, please add the following comment in the line with torch.cuda:\n\n #ignore-cuda\n'
)
err(res.stdout.decode("utf-8"))
sys.exit(1)
elif res.returncode == 2:
err(f"Error invoking grep on {', '.join(sys.argv[1:])}:")
err(res.stderr.decode("utf-8"))
sys.exit(2)
res = subprocess.run(
["git", "grep", "-Hn", "--no-index", r"\.cuda()", *sys.argv[1:]],
capture_output=True,
)
if res.returncode == 0:
err('Error: The string ".cuda()" was found. This implies convert a tensor to cuda tensor. Please replace all calls to tensor.cuda() with "tensor.to(get_accelerator().device_name())" and add the following import line:\nfrom deepspeed.accelerator import get_accelerator'
)
err(res.stdout.decode("utf-8"))
sys.exit(1)
elif res.returncode == 2:
err(f"Error invoking grep on {', '.join(sys.argv[1:])}:")
err(res.stderr.decode("utf-8"))
sys.exit(2)
files = []
for file in sys.argv[1:]:
if file.endswith(".py"):
files.append(file)
if len(files) > 0:
res = subprocess.run(
["git", "grep", "-Hn", "--no-index", r"\.is_cuda", *files],
capture_output=True,
)
if res.returncode == 0:
err('''
Error: The string ".is_cuda" was found. This implies checking if a tensor is a cuda tensor.
Please replace all calls to "tensor.is_cuda" with "get_accelerator().on_accelerator(tensor)",
and add the following import line:
'from deepspeed.accelerator import get_accelerator'
''')
err(res.stdout.decode("utf-8"))
sys.exit(1)
elif res.returncode == 2:
err(f"Error invoking grep on {', '.join(files)}:")
err(res.stderr.decode("utf-8"))
sys.exit(2)