[FSDP][RFC] Enforce rank r's current device is cuda:r
#92035
Closed
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
[ghstack-poisoned]
pytorch-bot
bot
added
the
release notes: distributed (fsdp)
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/92035
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 FailuresAs of commit 3e82ae9: NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This was referenced Jan 11, 2023
IIUC, FSDP assumes that rank `r` uses `cuda:r`. In other words, FSDP assumes single-process single-device (SPSD) and CUDA devices only. The SPSD assumption can help us simplify the `compute_device` and `device_id` code since then specifying `device_id` is simply a signal to FSDP whether the user wants FSDP to move parameters on behalf of the user or not. The actual target device is fixed (namely, `cuda:r`). Since the SPSD assumption is deeply embedded into FSDP, I think this is reasonable. We can generalize over non-CUDA devices in the future while still enforcing SPSD. Feel free to let me know your guys' thoughts. If we go with this, we can provide earlier and cleaner error handling in the case the user forgets to set the current CUDA device. See #91661. [ghstack-poisoned]
Contributor
|
I think the SPSD and CUDA device only assumption is current FSDP state, " we can provide earlier and cleaner error handling in the case the user forgets to set the current CUDA device." sounds great to me, thank you!! |
IIUC, FSDP assumes that rank `r` uses `cuda:r`. In other words, FSDP assumes single-process single-device (SPSD) and CUDA devices only. The SPSD assumption can help us simplify the `compute_device` and `device_id` code since then specifying `device_id` is simply a signal to FSDP whether the user wants FSDP to move parameters on behalf of the user or not. The actual target device is fixed (namely, `cuda:r`). Since the SPSD assumption is deeply embedded into FSDP, I think this is reasonable. We can generalize over non-CUDA devices in the future while still enforcing SPSD. Feel free to let me know your guys' thoughts. If we go with this, we can provide earlier and cleaner error handling in the case the user forgets to set the current CUDA device. See #91661. [ghstack-poisoned]
IIUC, FSDP assumes that rank `r` uses `cuda:r`. In other words, FSDP assumes single-process single-device (SPSD) and CUDA devices only. The SPSD assumption can help us simplify the `compute_device` and `device_id` code since then specifying `device_id` is simply a signal to FSDP whether the user wants FSDP to move parameters on behalf of the user or not. The actual target device is fixed (namely, `cuda:r`). Since the SPSD assumption is deeply embedded into FSDP, I think this is reasonable. We can generalize over non-CUDA devices in the future while still enforcing SPSD. Feel free to let me know your guys' thoughts. If we go with this, we can provide earlier and cleaner error handling in the case the user forgets to set the current CUDA device. See #91661. [ghstack-poisoned]
IIUC, FSDP assumes that rank `r` uses `cuda:r`. In other words, FSDP assumes single-process single-device (SPSD) and CUDA devices only. The SPSD assumption can help us simplify the `compute_device` and `device_id` code since then specifying `device_id` is simply a signal to FSDP whether the user wants FSDP to move parameters on behalf of the user or not. The actual target device is fixed (namely, `cuda:r`). Since the SPSD assumption is deeply embedded into FSDP, I think this is reasonable. We can generalize over non-CUDA devices in the future while still enforcing SPSD. Feel free to let me know your guys' thoughts. If we go with this, we can provide earlier and cleaner error handling in the case the user forgets to set the current CUDA device. See #91661. [ghstack-poisoned]
awgu
pushed a commit
to awgu/pytorch
that referenced
this pull request
Jan 19, 2023
ghstack-source-id: df22a11 Pull Request resolved: pytorch#92035
Contributor
|
Looks like this PR hasn't been updated in a while so we're going to go ahead and mark this as |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Stack from ghstack:
r's current device iscuda:r#92035 [FSDP][RFC] Enforce rankr's current device iscuda:ruse_orig_params=True#91767 [FSDP] Do not clean FQNs even foruse_orig_params=Truedevice_id+ CPU offload test #92031 [FSDP][BE] Improvedevice_id+ CPU offload testprefixed_param_names->fqnsfor consolidation #92028 [FSDP][BE] Renameprefixed_param_names->fqnsfor consolidationIIUC, FSDP assumes that rank
rusescuda:r. In other words, FSDP assumes single-process single-device (SPSD) and CUDA devices only.The SPSD assumption can help us simplify the
compute_deviceanddevice_idcode since then specifyingdevice_idis simply a signal to FSDP whether the user wants FSDP to move parameters on behalf of the user or not. The actual target device is fixed (namely,cuda:r). Since the SPSD assumption is deeply embedded into FSDP, I think this is reasonable. We can generalize over non-CUDA devices in the future while still enforcing SPSD.Feel free to let me know your guys' thoughts.
If we go with this, we can provide earlier and cleaner error handling in the case the user forgets to set the current CUDA device. See #91661.