Checklist Before Starting

• Search for similar PR(s).

What does this PR do?

Add one-line overview of what this PR aims to achieve or accomplish.

Resolved the tool formatting issue: Previously, arguments were stored as strings, causing iterative addition of \\ due to multiple calls to json.dumps.

Fixed the response_position_ids mismatch between generate_sequences and generate_sequences_with_tools: In the earlier implementation, generate_sequences_with_tools used zero padding for positions where attention mask == 0, which resulted in NaN values during the training phase.

Specific Changes

List the specific changes.

• Introduced a new schema, OpenAIFunctionCallSchema, to store converted tool calls.
• Updated the AsyncSGLangRollout tool to skip non-dict type arguments instead of handling any string at the arguments position.
• Aligned response_position_ids in generate_sequences_with_tools with the behavior of generate_sequences.
• Enhanced tool descriptions to prevent misleading parse errors, as returning 0.0 caused the model to incorrectly modify answers.

API

Demonstrate how the API changes if any.

• Revise the execute interface of the tool to directly accept dict[str, Any] instead of a JSON string.

Usage Example

Provide usage example(s) for easier usage.

# Add code snippet or script demonstrating how to use this 

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, evaluatuion results, etc.

Additional Info.

• Issue Number: Fixes issue # or discussion # if any.
• Training: [Note which backend this PR will affect: FSDP, Megatron, both, or none]
• Inference: [Note which backend this PR will affect: vLLM, SGLang, both, or none]

Checklist Before Submitting

• Read the Contribute Guide.
• Apply pre-commit checks.
• Add [BREAKING] to the PR title if it breaks any API.
• Update the documentation about your changes in the docs.
• Add CI test(s) if neccessary.