fix: preserve multi_modal_data in generate_opt_level=0 path

[sanmuf]

Summary

This PR fixes VLM generation in generate_scheduler.py by preserving multi_modal_data when constructing gen_batch.

Previously, request_data.pop(...) only kept tensor fields such as input_ids, attention_mask, and position_ids. For multimodal generation, multi_modal_data is stored in non_tensor_batch, so it was dropped before calling actor_cluster.generate(...).

As a result, VLM backends such as vLLM/SGLang could receive text-only prompts without image payloads.

For Qwen3-VL in vLLM, this could fail during M-RoPE position initialization with an error like:

IndexError: list index out of range

File ".../vllm/model_executor/models/qwen3_vl.py", line 1521, in get_mrope_input_positions
image_grid_thw[image_index][0]

Changes

• Preserve multi_modal_data from request_data.non_tensor_batch when building gen_batch.
• Keep behavior unchanged for text-only models.

Why

For VLM models, multi_modal_data contains image payloads and prompt token ids required by generation backends. Dropping it causes multimodal sampling to fail or silently degrade to text-only generation.

Test

• Verified with Qwen3-VL RLVR pipeline using generate_opt_level: 0.
• Confirmed that multi_modal_data is preserved and passed to the generation backend.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

fusen seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}