⏯️ Fix: handle None inputs when resuming GRPO Trainer from checkpoint#3148
Merged
Conversation
Member
|
Thanks!! Can share a simple piece of code that would fail without your fix? |
Contributor
Author
|
@qgallouedec here's the sample code: from datasets import load_dataset
from trl import GRPOTrainer, GRPOConfig
def reward_fn(completions, **_):
return [1.0 for _ in completions]
# Normal training
trainer = GRPOTrainer(
model="facebook/opt-125m",
args=GRPOConfig(
output_dir="save/test",
num_generations=2,
per_device_train_batch_size=2,
num_iterations=4,
save_steps=1,
max_steps=10,
max_prompt_length=1,
max_completion_length=1,
),
reward_funcs=reward_fn,
train_dataset=load_dataset("trl-lib/tldr", split="train"),
)
trainer.train()
# Simulating a fresh new trainer instance after interruption
trainer = GRPOTrainer(
model="facebook/opt-125m",
args=GRPOConfig(
output_dir="save/test",
num_generations=2,
per_device_train_batch_size=2,
num_iterations=4,
save_steps=1,
max_steps=10,
max_prompt_length=1,
max_completion_length=1,
),
reward_funcs=reward_fn,
train_dataset=load_dataset("trl-lib/tldr", split="train"),
)
# Resume from checkpoint at step which is not divisible by num_iterations
trainer.train(resume_from_checkpoint="save/test/checkpoint-6")error traceback: |
qgallouedec
reviewed
Mar 31, 2025
qgallouedec
approved these changes
Mar 31, 2025
qgallouedec
left a comment
Member
There was a problem hiding this comment.
Nice, thank you @PenutChen! In this case, the results won't be exactly the same as if we hadn't interrupted the training (we would have to save and load this buffer), but that's not a big deal.
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
kashif
pushed a commit
to kashif/trl
that referenced
this pull request
Mar 31, 2025
…huggingface#3148) Co-authored-by: Quentin Gallouédec <45557362+qgallouedec@users.noreply.github.com>
BjarniHaukur
pushed a commit
to ASSERT-KTH/trl
that referenced
this pull request
Apr 15, 2025
…ggingface#3131) Co-authored-by: Quentin Gallouédec <45557362+qgallouedec@users.noreply.github.com> Co-authored-by: Quentin Gallouédec <gallouedec.quentin@gmail.com> log answer key to wandb all Table HTML logging table bump patch hmm formatting html esacape reward isnt string [Liger] Liger KTO support (huggingface#2812) Co-authored-by: Kashif Rasul <kashif.rasul@gmail.com> Co-authored-by: Quentin Gallouédec <gallouedec.quentin@gmail.com> 🏃 Migrate CI to self-hosted runners (huggingface#3174) ❤️🩹 [CI] fix transformers dev CI failure (huggingface#3176) Co-authored-by: Quentin Gallouédec <45557362+qgallouedec@users.noreply.github.com> ⏯️ Fix: handle None inputs when resuming GRPO Trainer from checkpoint (huggingface#3148) Co-authored-by: Quentin Gallouédec <45557362+qgallouedec@users.noreply.github.com> 📎 Fix is_clipped to compute the effective clip_ratio (huggingface#3175) Co-authored-by: Quentin Gallouédec <45557362+qgallouedec@users.noreply.github.com> Co-authored-by: Quentin Gallouédec <gallouedec.quentin@gmail.com> Fix breaking typo for flash_attention reducing_memory_usage.md (huggingface#3190) Show unique prompts in GRPO WandB tables (huggingface#3191) 🐗 [CI] Fix trufflehog false positives (huggingface#3192) [GRPO] Improve completion length logging (huggingface#3188) preliminary openai compatible endpoint early concept, needs refining dedupe debug print some slop to work on unslop, missing hist almost valid pseudocode middle-ware monkey patch in mp.Pool()... remove unused More accurate .md need gpu renting lambda again much nicer small aider-chat and datasets conflict risky reqs change should work, but hacky some insights, but monkeypatching probably wont suffice refactor: Rewrite test script to use SWE-bench dataset with MultiProcessAider refactor: Remove logging statements from test.py one step closer finally, the correct abstraction doc todo unslop unslop undo accidental black cleaner abstraction new abstraction
yxliu-TAMU
pushed a commit
to mincheolseong/ECEN743-GRPO-Project-Proposal
that referenced
this pull request
Apr 20, 2025
…huggingface#3148) Co-authored-by: Quentin Gallouédec <45557362+qgallouedec@users.noreply.github.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
3 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 refactors the
_prepare_inputsmethod to ensure it never returnsNonewhen resuming training from checkpoints inGRPOTrainer.Previously, if
self._buffered_inputs[...]wasNoneduring the resume process, andself.state.global_step % self.num_iterations != 0, the method would returnNone. This caused issues downstream where non-Noneinputs were expected.To fix this, the logic has been updated so that if the buffered input is
None, it always falls back to generating new inputs via_generate_and_score_completions(inputs), regardless of the step count. This ensures stability and correctness when resuming from checkpoints.In addition, the modulo expression
self._step % self.args.gradient_accumulation_stepshas been refactored into a dedicated variableaccumulation_indexto improve readability and reduce redundancy.Fixes
No corresponding issue was filed, but this change addresses a potential silent failure when using
resume_from_checkpointwithGRPOTrainer.Motivation and context
Users resuming training from checkpoints may encounter
Noneinputs in_prepare_inputs, leading to errors in the training loop. This fix ensures robustness by avoiding returningNonein any case and improves the maintainability of the code through refactoring.Before submitting
make precommitto ensure code style consistency.Who can review?
Anyone familiar with
GRPOTrainer, buffered input logic, or checkpoint resume behavior in TRL.