[bugfix]Prevent overwriting drafters lm-head and embed_tokens

[HF-001]

What this PR does / why we need it?

Some EAGLE3 drafters might have their own lm_head and/or embed_tokens layers. Existing codebase ignores this.By solving this problem, it can greatly improve acceptance rates. Refer to this pr in vllm: vllm-project/vllm#27737

Does this PR introduce any user-facing change?

How was this patch tested?

export CUDA_VISIBLE_DEVICES=0
export TP=1
export MODEL_PATH=/model/Llama-3.1-8B-Instruct
export MODEL_NAME=Llama-3.1-8B-Instruct
export PORT=10133
python3 -m vllm.entrypoints.openai.api_server --host 0.0.0.0 --port ${PORT} --dtype bfloat16 --model ${MODEL_PATH} --served-model-name ${MODEL_NAME} --tensor-parallel-size ${TP} --gpu-memory-utilization 0.85 --max-model-len 32768 --trust-remote-code --seed 42 --speculative_config '{"method":"eagle3","model":"/model/EAGLE3-LLaMA3.1-Instruct-8B","num_speculative_tokens":5,"draft_tensor_parallel_size":1}'

• vLLM version: v0.11.0
• vLLM main: vllm-project/vllm@83f478b

[github-actions]

👋 Hi! Thank you for contributing to the vLLM Ascend project. The following points will speed up your PR merge:‌‌

• A PR should do only one thing, smaller PRs enable faster reviews.
• Every PR should include unit tests and end-to-end tests ‌to ensure it works and is not broken by other future PRs.
• Write the commit message by fulfilling the PR description to help reviewer and future developers understand.

If CI fails, you can run linting and testing checks locally according Contributing and Testing.

[gemini-code-assist]

Code Review

This pull request aims to prevent overwriting the lm_head and embed_tokens of a drafter model if they are already initialized, which is particularly important for some EAGLE3 drafters. The changes correctly add checks for has_own_embed_tokens and has_own_lm_head flags before sharing these layers from the target model. However, I've identified a critical regression in the logic for handling lm_head for SpecDcodeType.EAGLE. The check that ensures the target model has an lm_head attribute before it's accessed has been removed, which will likely cause an AttributeError at runtime. My review includes a specific code suggestion to fix this issue.

[gemini-code-assist]

The logic for sharing lm_head for SpecDcodeType.EAGLE has been changed in a way that introduces a potential AttributeError. The assignments to self.model.lm_head are no longer guarded by hasattr(model, "lm_head"). If the target model does not have an lm_head attribute, this will cause a crash. The original logic was safer and should be restored.

Suggested change

	if hasattr(model, "lm_head"):
	logger.info("Loading EAGLE LM head weights from the target model.")
	if supports_multimodal(model):
	self.model.lm_head = model.get_language_model().lm_head
	else:
	self.model.lm_head = model.lm_head
	if hasattr(model, "lm_head"):
	logger.info("Loading EAGLE LM head weights from the target model.")
	if supports_multimodal(model):
	self.model.lm_head = model.get_language_model().lm_head
	else:
	self.model.lm_head = model.lm_head

[github-actions]

This pull request has conflicts, please resolve those before we can evaluate the pull request.

[wangxiyuan]

Any progress? If this PR is still alive, please rebase to main and make CI happy. Thanks

[HF-001]

Any progress? If this PR is still alive, please rebase to main and make CI happy. Thanks

@wangxiyuan I will handle this PR soon

[wangxiyuan]

Please rebase and fix the merge conflicts

[wangxiyuan]

no update for long time, close this now. Feel free to reopen if it's still needed.