[Feature] [Spec Decode]: Simplify the Use of Eagle Spec Decode

[LiuXiaoxuanPKU]

🚀 The feature, motivation and pitch

As documented here, currently, users need to change the config and checkpoint using the script, which is tedious.
In the ideal case, vllm should load the eagle head automatically from huggingface or locally without extra conversion.
This can be achieved by:

1. Identify this is a eagle head from the model name.
2. Identify the LM head from the original base model.

After 1 and 2, vllm should be able to use eagle speculative decoding without user conversion.

Alternatives

No response

Additional context

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.