fix(data): prompt-length filtering crashes on VLM dataset with apply_chat_template

[Meihan-chen]

Repro

Prompt-length filtering crashes for multimodal (VLM) datasets when --apply-chat-template is set, making the feature unusable. Reproduced end-to-end on the geo3k VLM example with --apply-chat-template + --rollout-max-prompt-len:

slime/utils/data.py:111, in filter_long_prompt
    multimodal_inputs = process_vision_info(sample.prompt, processor)
TypeError: string indices must be integers, not 'str'

Root cause

In filter_long_prompt (slime/utils/data.py), the multimodal branch re-derived vision info from sample.prompt:

multimodal_inputs = process_vision_info(sample.prompt, processor)
processor_output = processor(text=sample.prompt, **multimodal_inputs)

With apply_chat_template=True, Sample.prompt is the rendered string, but filter_long_prompt passed it to process_vision_info, which expects a conversation list → crash. The vision inputs are already computed and stored in Sample.multimodal_inputs, so this recomputation is both wrong and redundant.

Why it's easy to hit

Setting --rollout-max-context-len (which derives rollout_max_prompt_len), --rollout-max-prompt-len, or --eval-max-prompt-len activates the filter on a VLM dataset and trips the crash.

Fix

Reuse the multimodal inputs already stored on the sample, routed through the same build_processor_kwargs helper the rollout path (sglang_rollout) uses, so the token length measured during filtering matches the real pipeline:

processor_kwargs = build_processor_kwargs(sample.multimodal_inputs)
processor_output = processor(text=sample.prompt, **processor_kwargs)