support qwen3vl-235b

[liusy58]

Motivation

This PR adds support assets for training Eagle3 on Qwen3-VL-235B by introducing:

• a new draft model config for Qwen3-VL-235B
• an offline training example script for hidden state preparation and Eagle3 training

Why

Qwen3-VL-235B needs a dedicated Eagle3 configuration and a reproducible offline example for training.

The example script was also updated to avoid machine-specific absolute paths in command arguments, which makes it easier to reuse across different environments and reduces configuration mistakes.

Modifications

Related Issues

Accuracy Test

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]

Code Review

This pull request adds a configuration file for the Qwen3-VL-235B model with Eagle3 and an offline training script. The feedback identifies that the rope_type should be set to mrope for multimodal rotary embeddings and that the training script should use the correct model path variable to load weights.

[gemini-code-assist]

For Qwen3-VL models, the rope_type should be set to mrope to correctly utilize multimodal rotary embeddings. The current default setting will cause the model to use standard Llama rotary embeddings, which are incompatible with the 3D rope indices used in VL models. The mrope_section and mrope_interleaved parameters are only active when rope_type is mrope in the LlamaAttention implementation.

Suggested change

	"rope_type": "default"
	"rope_type": "mrope"

[gemini-code-assist]

The --target-model-path in the training step should use $MODEL_PATH instead of "$MODEL_NAME". $MODEL_NAME is a descriptive string (e.g., Qwen3-VL-235B-A22B-Instruct-FP8), whereas $MODEL_PATH contains the actual path to the model weights required for loading embeddings and the LM head during offline training. This is consistent with the first torchrun command in this script.

Suggested change

	--target-model-path "$MODEL_NAME" \
	--target-model-path "$MODEL_PATH" \