support train qwen2.5-vl-32b eagle3 model

[gerayking]

Motivation

I noticed SpecForge currently supports Qwen2.5-VL-7B via Transformers. I think we should instead add systematic VL support through sglang, rather than a model-specific Transformers integration.

Modifications

• Dataset preprocessing: Use transformers to generate pixel_values and image_grid_thw (already available).

• Request packing (SGLang): Wrap the data into an sglang Request. When unpacking/splitting into per-request chunks, segment pixel_values by offset based on image_grid_thw, and ensure compatibility with mRoPE.

• Draft model: During forward(), align mRoPE behavior with the main model.

• Initialize mmCache: Set up the multimodal cache during initialization.

Related Issues

Fixes #403

Accuracy Test

TODO

Benchmark & Profiling

Checklist

• Format your code according to the Code Formatting with Pre-Commit.
• Add unit tests as outlined in the Running Unit Tests.
• Update documentation / docstrings / example tutorials as needed, according to Writing Documentation.
• Provide throughput / latency benchmark results and accuracy evaluation results as needed, according to Benchmark and Profiling and Accuracy Results.
• For reviewers: If you haven't made any contributions to this PR and are only assisting with merging the main branch, please remove yourself as a co-author when merging the PR.
• Please feel free to join our Slack channel at https://sgl-fru7574.slack.com/archives/C09784E3EN6 to discuss your PR.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!