[rollout] fix: use model_dump() for proper Pydantic serialization in token2text

[yurekami]

Summary

• Fixes Verl token2text in mlflow not working #4457 (token2text in MLflow does not work)
• The add_token2text function was using vars(result) which returns a reference to the Pydantic model's internal __dict__
• When this dict is modified to add prompt_text and response_text, it also modifies the original model object
• This can cause serialization issues with MLflow because the model now has extra fields that aren't in its schema

Changes

• Use model_dump() for Pydantic BaseModel objects to get a proper copy
• Fall back to dict(vars(result)) for other objects to ensure a copy

Technical Details

The bug was that vars() on a Pydantic v2 model returns a reference to the model's internal __dict__:

result = AgentLoopOutput(prompt_ids=[1,2,3], response_ids=[4,5,6])
_result = vars(result)
_result['prompt_text'] = 'decoded text'
# Now result.__dict__ also has 'prompt_text'!

This modifies the original model, which can cause MLflow serialization issues. The fix uses model_dump() which returns a proper copy.

Test plan

• Verified syntax is correct
• Tested that model_dump() creates a proper copy that doesn't modify the original
• Tested that the old vars() behavior was indeed problematic

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request correctly addresses a serialization issue with Pydantic models by using model_dump() instead of vars(). This prevents unintended modification of the original model object. My review includes a suggestion to make the fallback logic for non-Pydantic objects more robust to prevent potential crashes with certain object types.

[gemini-code-assist]

The fallback logic dict(vars(result)) is a good improvement for creating a copy, but it's not robust enough for a general utility function. It will raise a TypeError if result is an object that uses __slots__ and does not have a __dict__ attribute. This would cause a crash.

To make this function more robust and prevent potential runtime errors, I suggest explicitly handling objects with __dict__ and __slots__ before falling back to vars().

Suggested change

	if isinstance(result, BaseModel):
	_result = result.model_dump()
	else:
	_result = dict(vars(result))
	if isinstance(result, BaseModel):
	_result = result.model_dump()
	elif hasattr(result, "__dict__"):
	_result = result.__dict__.copy()
	elif hasattr(result, "__slots__"):
	_result = {slot: getattr(result, slot) for slot in result.__slots__}
	else:
	# Fallback for unknown types, which might fail.
	_result = dict(vars(result))