[Megatron] Add checkpointing support#298
Merged
erictang000 merged 16 commits intoNovaSky-AI:mainfrom Sep 18, 2025
Merged
Conversation
tyler-griggs
commented
Sep 15, 2025
Member
Author
There was a problem hiding this comment.
Note: These trainer updates may need to be changed after #297 is merged
tyler-griggs
commented
Sep 15, 2025
| self.init_weight_sync_state() | ||
|
|
||
| # Load policy model to GPU before loading checkpoint. | ||
| if self.cfg.trainer.placement.colocate_all: |
Member
Author
There was a problem hiding this comment.
Note: Policy model needs to be on GPU for Megatron load_checkpoint as required by Megatron's dist_checkpoint library
tyler-griggs
commented
Sep 15, 2025
| print("Phase 3: Verify state consistency") | ||
|
|
||
| # Compare captured states | ||
| for key in state_before: |
Member
Author
There was a problem hiding this comment.
Note: this was only confirming global_step, which we already do
17 tasks
erictang000
reviewed
Sep 17, 2025
Collaborator
erictang000
left a comment
erictang000
left a comment
There was a problem hiding this comment.
pretty much LGTM! thanks. Let's just merge main after #297 is merged to make the trainer changes consistent
for multi-node checkpointing and save_hf_model, we'll want these but I can help test out multi-node as I get model training running, and we can rely on external scripts for converting megatron to HF for a little bit.
| # All ranks wait for the checkpoint directory to be created before saving. | ||
| dist.barrier() | ||
|
|
||
| # Collect the sharded state dicts for model and optimizer, and full state dict for the scheduler. |
Collaborator
There was a problem hiding this comment.
so other than the scheduler there's no other additional memory load or communication here?
Member
Author
There was a problem hiding this comment.
If I understand your question correctly, no! We should just be loading the sharded model and optimizer state dicts.
erictang000
approved these changes
Sep 18, 2025
SumanthRH
added a commit
that referenced
this pull request
Sep 19, 2025
# What does this PR do? #298 broke GPU CI a bit: 1. Megatron related dependencies have not been resolved properly on `main` yet, and this test should be skipped. 2. We use a simple 4xL4 instance, but then the test was modified in #298 to request 8 GPUs (non-colocated training) --------- Signed-off-by: SumanthRH <sumanthrh99@gmail.com>
SumanthRH
added a commit
that referenced
this pull request
Sep 21, 2025
dzorlu
pushed a commit
to fleet-ai/SkyRL
that referenced
this pull request
Feb 4, 2026
## What does this PR do? This PR implements support for `save_checkpoint` and `load_checkpoint` for the Megatron training backend. We use Megatron's `dist_checkpointing` library to perform checkpointing in parallel across ranks, which also allows for reloading the checkpoints in a different parallelism scheme. Other minor changes: * Rename `save/load_ckpt` to `save/load_checkpoint` * Removed unused arguments to `save/load_checkpoint`, primarily the inclusion of non-backend specific state (`tag`, `client_state`, `global_step`). This change keeps training backend checkpointing logic focused on the training backend's state. ## Testing * Extended two GPU checkpointing tests to cover Megatron * Also moves `test_save_load_checkpoint.py` into `gpu_ci`. Note, however, that Megatron tests are disabled in CI because they currently require a different `flash-attn` install. * Manually save-and-resumed several times: <img width="426" height="292" alt="Screenshot 2025-09-15 at 3 30 17 PM" src="https://github.com/user-attachments/assets/2a4170fe-fddd-4086-961f-ca3170e654ab" /> ## What's next? - [ ] Test multi-node checkpointing - [ ] Implement `save_hf_model`
dzorlu
pushed a commit
to fleet-ai/SkyRL
that referenced
this pull request
Feb 4, 2026
# What does this PR do? NovaSky-AI#298 broke GPU CI a bit: 1. Megatron related dependencies have not been resolved properly on `main` yet, and this test should be skipped. 2. We use a simple 4xL4 instance, but then the test was modified in NovaSky-AI#298 to request 8 GPUs (non-colocated training) --------- Signed-off-by: SumanthRH <sumanthrh99@gmail.com>
dzorlu
pushed a commit
to fleet-ai/SkyRL
that referenced
this pull request
Feb 4, 2026
# What does this PR do? Fixes async trainer example after NovaSky-AI#298. We renamed `setup_policy_and_generator` to `init_weight_sync_state` but missed the update in some places. --------- Signed-off-by: SumanthRH <sumanthrh@anyscale.com>
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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.
What does this PR do?
This PR implements support for
save_checkpointandload_checkpointfor the Megatron training backend. We use Megatron'sdist_checkpointinglibrary to perform checkpointing in parallel across ranks, which also allows for reloading the checkpoints in a different parallelism scheme.Other minor changes:
save/load_ckpttosave/load_checkpointsave/load_checkpoint, primarily the inclusion of non-backend specific state (tag,client_state,global_step). This change keeps training backend checkpointing logic focused on the training backend's state.Testing
test_save_load_checkpoint.pyintogpu_ci. Note, however, that Megatron tests are disabled in CI because they currently require a differentflash-attninstall.What's next?
save_hf_model