Add VLM-aware model loading to FSDP→HF conversion

[cklxx]

Motivation

• Enable conversion of Vision-Language-Model (VLM) FSDP checkpoints to Hugging Face format by selecting the correct HF model class based on the model config.

Description

• Updated tools/convert_fsdp_to_hf.py to import AutoModelForImageTextToText and added _build_hf_model(config) which prints the detected config.model_type and returns either AutoModelForCausalLM or AutoModelForImageTextToText using trust_remote_code=True.

Testing

• Ran linting and formatting checks: ruff check ., black --check ., and isort --check ., all passed.
• Ran pytest, which failed during collection with ModuleNotFoundError: No module named 'slime' (test environment import issue).

---

Codex Task

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 5bcb474c90

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".