[🐯+GRPO] Support FSDP + Fix bug when using LigerGRPO with DDP

[shivam15s]

What does this PR do?

This PR aims to do two things:

1. The recent integration of LigerGRPO had a bug: when using DDP and performing a forward pass through a submodule of the unwrapped model, the necessary hooks weren't registered correctly. This caused the model weights across GPUs to fall out of sync. To fix this, the PR introduces a Forward Redirection mechanism—a workaround that ensures hooks are properly registered (compatible with both DDP and FSDP) and enables the custom forward pass required by Liger.
2. Add support for FSDP to GRPO Trainer. We leverage summon_full_params to make model.generate work with FSDP.

Experiment Script: https://gist.github.com/shivam15s/08a9bccd0d72dd0d29bdb912cb9885be

DDP: Liger (blue) v Non-liger (black)

FSDP: Liger (green) v Non-liger (Purple)

Known Limitations with FSDP (can add support in subsequent PR(s))

1. sync_ref_model not supported currently
2. create_reference_model not supported currently

Benchmarking:
Dist Strategy: DDP
7 policy workers, 1 vllm worker (8 h100)

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a GitHub issue? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[kashif]

testing using:

import torch
from datasets import load_dataset
from trl import GRPOConfig, GRPOTrainer
import torch.distributed as dist
from torch.profiler import profile, record_function, ProfilerActivity
from transformers import TrainerCallback
import os
# from torch.distributed.fsdp import FSDPConfig, AutoWrapPolicy
# dataset = load_dataset("trl-internal-testing/zen", "standard_prompt_only", split="train")
dataset = load_dataset("trl-lib/ultrafeedback-gpt-3.5-turbo-helpfulness", split="train")
# only keep the prompt column
dataset = dataset.map(lambda x: {"prompt": x["prompt"]}, remove_columns=dataset.column_names)

training_args = GRPOConfig(
    output_dir="./scratch_dir",
    learning_rate=0.001,  # increase the learning rate to speed up the test
    per_device_train_batch_size=3,  # reduce the batch size to reduce memory usage
    num_generations=3,  # reduce the number of generations to reduce memory usage
    report_to=["tensorboard"],
    max_completion_length=256,  # reduce the completion length to reduce memory usage
    logging_steps=1,
    save_strategy="no",
    max_steps=50,
    use_liger_loss=True,
)
trainer = GRPOTrainer(
    model="trl-internal-testing/tiny-Qwen2ForCausalLM-2.5",
    reward_funcs="trl-internal-testing/tiny-Qwen2ForSequenceClassification-2.5",
    args=training_args,
    train_dataset=dataset,
)

class ProfCallback(TrainerCallback):
    def __init__(self, prof):
        self.prof = prof

    def on_step_end(self, args, state, control, **kwargs):
        self.prof.step()

# Create directory for profiling outputs
os.makedirs("profiling_results", exist_ok=True)

# Define profiling context manager
def train_with_profiling(enable_profiling=True):
    if enable_profiling:
        with profile(
            activities=[ProfilerActivity.CPU, ProfilerActivity.CUDA],
            record_shapes=True,
            profile_memory=True,
            with_stack=True,
            with_flops=True,
            on_trace_ready=torch.profiler.tensorboard_trace_handler("profiling_results") if trainer.accelerator.is_main_process else None,
            schedule=torch.profiler.schedule(
                wait=1,
                warmup=1,
                active=2,
                repeat=1),
        ) as prof:
            trainer.add_callback(ProfCallback(prof))
            trainer.train()
        # Print profiling results summary
        # print(prof.key_averages().table(sort_by="cpu_time_total", row_limit=20))
    else:
        trainer.train()

# trainer.train()
train_with_profiling(enable_profiling=False)

# destroy process group
if dist.is_initialized():
    dist.destroy_process_group()

[HuggingFaceDocBuilderDev]

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.

[shivam15s]

Python 3.9 test is failing because of a backward incompatible initialization. Created a PR in liger to fix this

[hjh0119]

Awesome work! Could you share any memory footprint comparisons (with vs. without Liger Loss enabled)?

[shivam15s]

Hi @hjh0119
I was able to achieve about a 20% reduction in peak memory usage.

Peak Memory:
• Without Liger: 71.1 GB
• With Liger: 56.7 GB
here's my code: https://gist.github.com/shivam15s/08a9bccd0d72dd0d29bdb912cb9885be