[fsdp] fix: add missing mixed precision configuration to FSDPEngineConfig

[xxrjun]

What does this PR do?

The FSDPEngineConfig dataclass was missing the mixed_precision field that the runtime code expected. By adding:

mixed_precision: Optional[dict[str, Any]] = None

The dataclass now properly supports the mixed precision configuration that the FSDP workers code uses with fsdp_config.get("mixed_precision", None).

verl/verl/workers/fsdp_workers.py

Line 371 in 55e3c5b

mixed_precision_config = fsdp_config.get("mixed_precision", None)

Otherwise, if we run with:

python3 -m verl.trainer.main_ppo \
    actor_rollout_ref.actor.fsdp_config.mixed_precision.param_dtype=bf16 \
    actor_rollout_ref.actor.fsdp_config.mixed_precision.reduce_dtype=fp32 \
    actor_rollout_ref.actor.fsdp_config.mixed_precision.buffer_dtype=fp32 \
    # ... other parameters

The following error may occur:

raise InstantiationException(msg) from e
hydra.errors.InstantiationException: Error in call to target 'verl.workers.config.engine.FSDPEngineConfig':
TypeError("FSDPEngineConfig.__init__() got an unexpected keyword argument 'mixed_precision'")
full_key: actor_rollout_ref.actor.fsdp_config

Backward compatibility

No behavior change for existing configs (default remains None).

Checklist Before Starting

• Search for similar PRs. Paste at least one query link here: https://github.com/volcengine/verl/pulls?q=mixed_precision
• 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

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).)

[gemini-code-assist]

Code Review

This pull request correctly adds the missing mixed_precision configuration to FSDPEngineConfig, fixing a runtime TypeError. My review focuses on improving the maintainability and robustness of this change. I've suggested using a dedicated dataclass for the new configuration for better type safety and have pointed out the need for adding unit tests to cover this fix and prevent future regressions.

[gemini-code-assist]

For better type safety and maintainability, consider defining a dedicated dataclass for the mixed_precision configuration instead of using dict[str, Any]. This makes the configuration more explicit, self-documenting, and allows centralizing default values.

You could define a new dataclass, for example FSDPMixedPrecisionConfig, and use it here. This would also involve updating verl/workers/engine/fsdp/engine_impl.py to use attribute access instead of get(). This change would make the configuration structure more robust and less prone to errors from typos in dictionary keys.

[gemini-code-assist]

This bug fix introduces a new configuration field, but it's not accompanied by a unit test. The contribution checklist also indicates that tests have not been added.

Please add a unit test to tests/workers/config/test_engine_config_on_cpu.py to verify that:

1. FSDPEngineConfig can be instantiated with the mixed_precision argument.
2. The default value of mixed_precision is None.
3. A custom dictionary can be passed and is stored correctly.

Adding tests is crucial for preventing future regressions and ensuring the stability of the codebase.