[RFC]: spec decode with draft models

[HF-001]

Motivation.

Enabling draft models for speculative decoding (SD) in vllm achieved good performance improvement，is not yet implemented in vllm-ascend.This type of SD requires no special trained heads (like EAGLE, or Medusa). I want to implement this feature in vllm-ascend.

For detailed information, please refer to vllm-project/vllm#24322

Proposed Change.

A separate script named draft_model.py will be built in the spec_decode directory. And will add draft model related logic will be added to scheduler.py 、model_runner_v1.py、 unit test，etc.

Feedback Period.

No response

CC List.

No response

Any Other Things.

No response

[wxsIcey]

This is a good idea!

[HF-001]

This is a good idea!

@wxsIcey Thank you for your attention and encouragement！