-
Notifications
You must be signed in to change notification settings - Fork 26.3k
Make torch.cuda.gds APIs public #147120
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.
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
Make torch.cuda.gds APIs public #147120
Conversation
[ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/147120
Note: Links to docs will display an error until the docs builds have been completed. ⏳ No Failures, 3 PendingAs of commit 5768d7f with merge base f95bdf5 ( UNSTABLE - The following jobs are marked as unstable, possibly due to flakiness on trunk:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
Follow up to #145748 [ghstack-poisoned]
Follow up to #145748 [ghstack-poisoned]
Follow up to #145748 [ghstack-poisoned]
albanD
left a comment
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Sounds good, only small nits
|
|
||
|
|
||
| def _gds_register_buffer(s: Storage) -> None: | ||
| def gds_register_buffer(s: Storage) -> None: |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Any user of the old private APIs? Do you want to keep them to avoid breaking any previous user by doing something like _gds_register_buffer = gds_register_buffer ?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
0 hits on github from repos that are not just copy pastes of pytorch so I feel ok about breaking the BC here
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Ok!
| gds_register_buffer | ||
| gds_deregister_buffer | ||
| GdsFile | ||
|
|
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Can you add here or link from here to an example on how to use these APIs?
Follow up to #145748 that turned USE_CUFILE on for CUDA 12.6 and 12.8 binaries [ghstack-poisoned]
albanD
left a comment
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Moving API to public sounds good.
Let's plan on having a tutorial (including serialization config needs) before 2.7 as an E2E example.
Follow up to #145748 that turned USE_CUFILE on for CUDA 12.6 and 12.8 binaries [ghstack-poisoned]
Follow up to #145748 that turned USE_CUFILE on for CUDA 12.6 and 12.8 binaries [ghstack-poisoned]
Follow up to #145748 that turned USE_CUFILE on for CUDA 12.6 and 12.8 binaries [ghstack-poisoned]
|
@pytorchbot merge |
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
Follow up after : #147120 Cufile was enabled only on Linux: https://pypi.org/project/nvidia-cufile-cu12/#files Fixes validation workflow failues: https://github.com/pytorch/test-infra/actions/runs/13558218752/job/37896578837 ``` File "C:\Jenkins\Miniconda3\envs\conda-env-13558218752\lib\site-packages\torch\cuda\gds.py", line 105, in __init__ raise RuntimeError("GdsFile is not supported on this platform.") RuntimeError: GdsFile is not supported on this platform. Exception ignored in: <function GdsFile.__del__ at 0x000001772B5003A0> Traceback (most recent call last): File "C:\Jenkins\Miniconda3\envs\conda-env-13558218752\lib\site-packages\torch\cuda\gds.py", line 113, in __del__ if self.handle is not None: AttributeError: 'GdsFile' object has no attribute 'handle' ``` Pull Request resolved: #148060 Approved by: https://github.com/mikaylagawarecki
Follow up after : #147120 Cufile was enabled only on Linux: https://pypi.org/project/nvidia-cufile-cu12/#files Fixes validation workflow failues: https://github.com/pytorch/test-infra/actions/runs/13558218752/job/37896578837 ``` File "C:\Jenkins\Miniconda3\envs\conda-env-13558218752\lib\site-packages\torch\cuda\gds.py", line 105, in __init__ raise RuntimeError("GdsFile is not supported on this platform.") RuntimeError: GdsFile is not supported on this platform. Exception ignored in: <function GdsFile.__del__ at 0x000001772B5003A0> Traceback (most recent call last): File "C:\Jenkins\Miniconda3\envs\conda-env-13558218752\lib\site-packages\torch\cuda\gds.py", line 113, in __del__ if self.handle is not None: AttributeError: 'GdsFile' object has no attribute 'handle' ``` Pull Request resolved: #148060 Approved by: https://github.com/mikaylagawarecki
Follow up to #145748 that turned USE_CUFILE on for CUDA 12.6 and 12.8 binaries
Stack from ghstack (oldest at bottom):