[train] Save HF processor on checkpoint export for VLMs

[dinhxuanvu]

Summary

VLM checkpoints exported via DistributedStrategy.save_hf_configs are missing preprocessor_config.json (and the image/video processor configs). Only the model config, tokenizer, and generation config are written. As a result, AutoProcessor.from_pretrained() cannot reload the exported checkpoint and downstream vLLM serving crashes on load.

This adds processor saving to the HF export path:

• Adds check_is_vlm() and get_processor() helpers to skyrl/utils/tok.py.
• In save_hf_configs, when the model is a VLM, resolves the processor from model_config.name_or_path and writes it alongside the other configs. This mirrors the existing generation_config handling, which already resolves from name_or_path.
• No-op for text-only models (no vision_config), and wrapped in try/except so a processor-resolution failure cannot abort a checkpoint save.

Both export paths funnel through save_hf_configs (FSDP passes the live model .config; Megatron passes hf_config loaded via AutoConfig.from_pretrained(model_path)), so the fix covers both backends.

Test plan

• New CPU unit tests in tests/backends/skyrl_train/distributed/test_save_hf_configs_processor.py:
  • processor saved for VLM
  • processor not resolved for text-only
  • processor-resolution failure does not raise
  • skipped when name_or_path is empty
• End-to-end: export a VLM checkpoint and confirm preprocessor_config.json is present and AutoProcessor.from_pretrained() reloads it.

[gemini-code-assist]

Code Review

This pull request introduces support for saving Hugging Face processor configurations (such as preprocessor_config.json) when exporting vision-language model (VLM) checkpoints, preventing potential loading crashes in downstream environments like vLLM. It adds utility functions to check if a model is a VLM and to retrieve its processor, along with corresponding unit tests. Feedback suggests wrapping the VLM check in a try-except block and passing the already loaded model configuration directly to avoid redundant disk/network I/O. Additionally, defensive checks should be added to handle processors that lack a tokenizer, and the unit tests should be updated accordingly.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.