feat: add Qwen3.5 Dense Model EAGLE3 training support

[36330]

• Add qwen3_5_eagle_patch.py: Monkey patch for SGLang's Qwen3.5 models

  • Captures aux_hidden_states from 3 layers (layer 2, mid, layer-3)
  • Supports both Qwen3_5ForCausalLM and Qwen3_5ForConditionalGeneration
  • Environment variable QWEN35_EAGLE3_ENABLE=1 to enable

• Add qwen3.5-4b-eagle3.json: Draft model config for Qwen3.5-4B

• Modify train_eagle3.py: Auto-apply patch on initialization

• Modify eagle3_target_model.py: Auto-detect and patch Qwen3.5 models

• Modify llama3_eagle.py: Handle 'default' RoPE scaling type for Qwen3.5

Tested:

• Hidden states generation: ✓
• Training with TTT: ✓
• Qwen3.5-4B Dense model: ✓

example

# generate hidden states
torchrun --nproc_per_node=1 scripts/prepare_hidden_states.py \
    --target-model-path Qwen/Qwen3.5-4B \
    --enable-aux-hidden-states

# train
torchrun --nproc_per_node=4 scripts/train_eagle3.py \
    --target-model-path Qwen/Qwen3.5-4B \
    --draft-model-config configs/qwen3.5-4b-eagle3.json

[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!

[36330]

[jiapingW]

We have implemented the qwen3.5 eagle3 training in PR.