Skip to content

Default to allow_tf32=True for GPU Devices #1275

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
Jul 12, 2022
Merged

Default to allow_tf32=True for GPU Devices #1275

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
62d505f
Default to `allow_tf32=True` for GPU Devices
ravi-mosaicml Jul 12, 2022
50c3644
Force Jenkins
ravi-mosaicml Jul 12, 2022
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
17 changes: 15 additions & 2 deletions composer/trainer/devices/device_gpu.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,6 +8,8 @@
from typing import Any, Dict, Optional, TypeVar

import torch
import torch.backends.cuda
import torch.backends.cudnn
import torch.cuda.amp
import torch.utils.data

Expand All @@ -24,18 +26,29 @@ class DeviceGPU(Device):

Args:
device_id (int, optional): Integer ID of a GPU device to train with. If not specified, the local rank
of the current process is used. Default: None.
of the current process is used. Default: None.
allow_tf32 (bool, optional): Whether to allow TF32 matrix multiplications. Defaults to True.
For more information, see :ref:`torch:tf32_on_ampere`.
"""
dist_backend = 'nccl'

def __init__(self, device_id: Optional[int] = None):
def __init__(
self,
device_id: Optional[int] = None,
*,
allow_tf32: bool = True,
):
if not torch.cuda.is_available():
raise ValueError('DeviceGPU cannot be created as torch.cuda is not available.')
if not device_id:
device_id = dist.get_local_rank()
self._device = torch.device(f'cuda:{device_id}')
torch.cuda.set_device(self._device)
assert torch.cuda.current_device() == device_id
torch.backends.cuda.matmul.allow_tf32 = allow_tf32
# pyright error: "allow_tf32" is not a known member of module
# however, this flag exists on pytorch 1.9+: https://pytorch.org/docs/1.9.0/backends.html
torch.backends.cudnn.allow_tf32 = allow_tf32 # type: ignore

def module_to_device(self, module: T_nnModule) -> T_nnModule:
return module.to(self._device)
Expand Down