[fsdp] fix: patch fsdp2 to support hf transformer==4.54.0 and above

[weifengpy]

What does this PR do?

@ETOgaosion found that fsdp2 cannot work with GenericForTokenClassification in transformer==4.54.0+ pytorch/pytorch#160068
#2947

FSDP2 complains about object layerout mismatch when constructing FSDPGenericForTokenClassification. The solution is to let FSDPModule inherit ABC as well

this should unblock @vermouth1992 on deprecating fsdp1

Add concise overview of what this PR aims to achieve or accomplish. Reference related GitHub issues and PRs that help with the review.

Checklist Before Starting

• Search for similar PRs. Paste at least one query link here: ...
• Format the PR title as [{modules}] {type}: {description} (This will be checked by the CI)
  • {modules} include fsdp, megatron, sglang, vllm, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data
  • If this PR involves multiple modules, separate them with , like [megatron, fsdp, doc]
  • {type} is in feat, fix, refactor, chore, test
  • If this PR breaks any API (CLI arguments, config, function signature, etc.), add [BREAKING] to the beginning of the title.
  • Example: [BREAKING][fsdp, megatron] feat: dynamic batching

Test

add following script to apply_fsdp2

# pip install transformers==4.54.0
from transformers import AutoModelForTokenClassification
model = AutoModelForTokenClassification.from_pretrained("Qwen/Qwen2-0.5B")
with maybe_patch_fsdp_module(model):
    fully_shard(model, **fsdp_kwargs)
return

PYTHONUNBUFFERED=1 python3 -m verl.trainer.main_ppo \
 data.train_files=$HOME/data/gsm8k/train.parquet \
 data.val_files=$HOME/data/gsm8k/test.parquet \
 data.train_batch_size=256 \
 data.max_prompt_length=512 \
 data.max_response_length=256 \
 actor_rollout_ref.model.path=Qwen/Qwen2.5-0.5B-Instruct \
 actor_rollout_ref.actor.optim.lr=1e-6 \
 actor_rollout_ref.actor.ppo_mini_batch_size=64 \
 actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=4 \
 actor_rollout_ref.rollout.name=vllm \
 actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8 \
 actor_rollout_ref.rollout.tensor_model_parallel_size=1 \
 actor_rollout_ref.rollout.gpu_memory_utilization=0.4 \
 actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=4 \
 critic.optim.lr=1e-5 \
 critic.model.path=Qwen/Qwen2.5-0.5B-Instruct \
 critic.ppo_micro_batch_size_per_gpu=4 \
 algorithm.kl_ctrl.kl_coef=0.001 \
 trainer.logger=console \
 trainer.val_before_train=False \
 trainer.n_gpus_per_node=2 \
 trainer.nnodes=1 \
 trainer.save_freq=10 \
 trainer.test_freq=10 \
 trainer.total_epochs=15 \
 critic.strategy=fsdp2 \
 actor_rollout_ref.actor.strategy=fsdp2

For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc.

API and Usage Example

Demonstrate how the API changes if any, and provide usage example(s) if possible.

# Add code snippet or script demonstrating how to use this

Design & Code Changes

Demonstrate the high-level design if this PR is complex, and list the specific changes.

Checklist Before Submitting

Important

Please check all the following items before requesting a review, otherwise the reviewer might deprioritize this PR for review.

• Read the Contribute Guide.
• Apply pre-commit checks: pre-commit install && pre-commit run --all-files --show-diff-on-failure --color=always
• Add / Update the documentation.
• Add unit or end-to-end test(s) to the CI workflow to cover all the code. If not feasible, explain why: ...
• Once your PR is ready for CI, send a message in the ci-request channel in the verl Slack workspace. (If not accessible, please try the Feishu group (飞书群).)

[weifengpy]

cc @ETOgaosion @vermouth1992 to kick off CI

[gemini-code-assist]

Code Review

This pull request addresses a metaclass conflict issue when using fsdp2 with recent versions of the Hugging Face transformers library. The solution involves monkey-patching PyTorch's FSDPModule to ensure compatibility. While the approach is generally sound, I have identified a critical thread-safety issue in the patching mechanism that could lead to race conditions in a concurrent environment. Additionally, the use of private PyTorch APIs introduces a maintainability risk. I have provided detailed comments and code suggestions to address these points.

[gemini-code-assist]

Accessing private PyTorch modules like _fully_shard and _composable is risky as their implementation can change without warning in future releases, which could break this functionality. While this appears necessary for the fix, it's good practice to add a comment explaining the rationale and acknowledging the maintenance risk. This will help future developers understand this dependency.

[weifengpy]

this is gated by 2.4. should be fine

[weifengpy]

this is for existing torch stable. for torch 2.9+, we will land a fix at pytorch side

[wuxibin89]

@weifengpy Please format code with:
https://github.com/volcengine/verl/blob/main/CONTRIBUTING.md#code-linting-and-formatting

[weifengpy]

@wuxibin89 formatted with pre-commit run --all-files