[rl] refactor save and load model weights using DCP #2221
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]
This was referenced Jan 13, 2026
meta-cla
bot
added
the
CLA Signed
wwwjn
changed the title
[ghstack-poisoned]
[ghstack-poisoned]
[ghstack-poisoned]
wwwjn
changed the title
| self.temp_model_dir = os.path.abspath( | ||
| os.path.join(job_config.job.dump_folder, "vllm_temp_model") | ||
| # Load TorchTitan plugin at runtime | ||
| from torchtitan.experiments.rl.unified.plugin import register |
Contributor
There was a problem hiding this comment.
use from torchtitan.experiments.rl.unified import register?
can we move the import statement to the header?
|
|
||
| return self.load_weights_from_state_dict(torchtitan_state_dict) | ||
|
|
||
| def load_weights(self, weights_iter): |
Contributor
There was a problem hiding this comment.
shall we load weights from weights_iter?
Contributor
|
I'm wondering that should we refactor TorchTitan checkpointer so that it can be directly used in this case. While the current PR work, if TorchTitan migrates to a new checkpoint library other use cases need the same updates as well. This is more future work, not blocking this PR. |
tianyu-l
reviewed
Jan 19, 2026
| return logits | ||
|
|
||
| def load_weights(self, weights_iter): | ||
| def load_weights_from_state_dict(self, titan_state_dict): |
Contributor
There was a problem hiding this comment.
titan_state_dict is ambiguous -- both sides should be titan models.
What other name could we use, e.g. trainer_state_dict?
| # We need to split our weights to match the original 2-shard structure | ||
| import glob | ||
| # directly update model weights in place | ||
| load_weights = self._get_model().load_weights_from_state_dict(state_dict) |
Contributor
There was a problem hiding this comment.
- IIUC this only works when trainer and generator are on exactly the same global mesh. Is it right?
- Is it true that this has been the assumption before this PR? I.e. is our current monarch script only allow colocated trainer and generator?
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 (oldest at bottom):