feat: add hf_template tokenize_and_mask method + verl SFTTrainer compat for RLLMSFTDataset

[yifannnwu]

Summary

Two changes to RLLMSFTDataset:

1. verl SFTTrainer compatibility: RLLMSFTDataset.__init__ now accepts processor and
   max_samples kwargs, matching the call signature of verl's create_sft_dataset(). Without
   this, using RLLMSFTDataset as custom_cls in verl's SFTTrainer(config) crashes with
   TypeError: unexpected keyword argument.

2. hf_template tokenization method: The existing cumulative and stepwise methods use
   rLLM's ChatTemplateParser to render messages, which renders tool calls as JSON-in-XML:

   <tool_call>
   {"name": "func", "arguments": {"key": "val"}}
   </tool_call>
   

   But models like Qwen3-Coder expect native XML format:

   <tool_call>
   <function=func>
   <parameter=key>val</parameter>
   </function>
   </tool_call>
   

   The new hf_template method uses tokenizer.apply_chat_template() directly, producing
   the model's native format. It uses an incremental prefix-diff approach to isolate each
   message's tokens for correct loss masking.

   Config: data.rllm.tokenize_and_mask_method: hf_template

Files changed

• rllm/trainer/verl/sft_dataset.py (~40 lines)

Test plan

• Validated on 100 samples — tool calls correctly render in native XML format
• Text-only samples render identically under all methods
• SFT training with verl's SFTTrainer(config) + custom_cls: RLLMSFTDataset works

[listar2000]

@jeffreysijuntan @kylemontgomery1

[kylemontgomery1]

Looks good to me!