mirror of
https://github.com/zebrajr/pytorch.git
synced 2025-12-06 12:20:52 +01:00
9cc040fef6
Previously: ``` [W Utils.hpp:133] Warning: Environment variable NCCL_ASYNC_ERROR_HANDLING is deprecated; use TORCH_NCCL_ASYNC_ERROR_HANDLING instead (function getCvarInt) [W Utils.hpp:133] Warning: Environment variable NCCL_ASYNC_ERROR_HANDLING is deprecated; use TORCH_NCCL_ASYNC_ERROR_HANDLING instead (function getCvarInt) ``` With this PR, those warnings disappear. They were introduced in #114077 This change was generated with this sed script, applied with `sed -i -f /tmp/x **/*.{py,hpp,cpp,cc}` and hand inspected. ``` s/\bNCCL_BLOCKING_WAIT\b/TORCH_NCCL_BLOCKING_WAIT/g s/\bNCCL_ENABLE_TIMING\b/TORCH_NCCL_ENABLE_TIMING/g s/\bNCCL_DESYNC_DEBUG\b/TORCH_NCCL_DESYNC_DEBUG/g s/\bNCCL_ASYNC_ERROR_HANDLING\b/TORCH_NCCL_ASYNC_ERROR_HANDLING/g s/\bENABLE_NCCL_HEALTH_CHECK\b/TORCH_ENABLE_NCCL_HEALTH_CHECK/g s/\bNCCL_USE_TENSOR_REGISTER_ALLOCATOR_HOOK\b/TORCH_NCCL_USE_TENSOR_REGISTER_ALLOCATOR_HOOK/g ``` Pull Request resolved: https://github.com/pytorch/pytorch/pull/114880 Approved by: https://github.com/kwen2501
24 lines
1.2 KiB
Python
24 lines
1.2 KiB
Python
from torch._C._distributed_c10d import _DEFAULT_PG_TIMEOUT
|
|
from datetime import timedelta
|
|
from typing import Optional
|
|
|
|
__all__ = ['default_pg_timeout', 'default_pg_nccl_timeout']
|
|
|
|
# Default process group wide timeout, if applicable.
|
|
# This only applies to the non-nccl backends
|
|
# To make an attempt at backwards compatibility with THD, we use an
|
|
# extraordinarily high default timeout, given that THD did not have timeouts.
|
|
default_pg_timeout: timedelta = _DEFAULT_PG_TIMEOUT
|
|
# Separate timeout for PGNCCL mainly becuase it's always been that way in the C++ layer, but until recently
|
|
# there was one default that applied across all backends in the python layer.
|
|
# Later, we could consider merging them back together at the c++ layer if we can align on a same value.
|
|
# (only if TORCH_NCCL_BLOCKING_WAIT or TORCH_NCCL_ASYNC_ERROR_HANDLING is set to 1).
|
|
|
|
try:
|
|
from torch._C._distributed_c10d import _DEFAULT_PG_NCCL_TIMEOUT
|
|
default_pg_nccl_timeout: Optional[timedelta] = _DEFAULT_PG_NCCL_TIMEOUT
|
|
except ImportError:
|
|
# if C++ NCCL support is not compiled, we don't have access to the default nccl value.
|
|
# if anyone is actually trying to use nccl in this state, it should error.
|
|
default_pg_nccl_timeout = None
|