[Feat][Plugin] Enable MTP for vLLM Plugin

[whx-sjtu]

Motivation

This PR enables MTP feature for running DeepSeekV3 and GLM5 with vLLM+atom.

Technical Details

1. atom_config related bugfix.
2. Fix wrong full_cls_name of different MLA sparse attention backends.
3. Register model architecture and model class for DeepSeek V3 and GLM5 MTP.
4. Add index_buffer for DeepseekMTP.
5. Adapt full graph of main model with mtp enabled.

Test Plan

Comming soon.

Test Result

1. zai-org/GLM-5.1-FP8

Accuracy test commands:

lm_eval --model local-completions \
        --model_args model=/home/models/GLM-5.1-FP8,base_url=http://localhost:8000/v1/completions,num_concurrent=64,max_retries=3 \
        --tasks gsm8k \
        --num_fewshot 20

Accuracy test result with mtp=3:

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|    20|exact_match|↑  |0.9454|±  |0.0063|
|     |       |strict-match    |    20|exact_match|↑  |0.9462|±  |0.0062|

2. deepseek-ai/DeepSeek-R1-0528

Accuracy test commands:

lm_eval --model local-completions \
        --model_args model=/home/models/DeepSeek-R1-0528,base_url=http://localhost:8000/v1/completions,num_concurrent=16,max_retries=3,tokenized_requests=False \
        --tasks gsm8k \
        --num_fewshot 3

Accuracy test result with mtp=3:

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     3|exact_match|↑  |0.9492|±  |0.0060|
|     |       |strict-match    |     3|exact_match|↑  |0.9469|±  |0.0062|

Submission Checklist

• This PR is based on PR [Feat][Plugin] Enable Sparse MLA and GLM-5 for vLLM-ATOM #399 and should be merged after that.

[wuhuikx]

Could you please help attach the accuracy test results on gms8k? Do we support MTP=1 or MTP=1/2/3? How about the acceptance ratio?

[wuhuikx]

I will turn this PR to draft and go through CI after the code review is done.

[whx-sjtu]

Could you please help attach the accuracy test results on gms8k? Do we support MTP=1 or MTP=1/2/3? How about the acceptance ratio?

Sure I will attach the acc results later. Now we support MTP=1/2/3, but the acceptance rate is low (about 20% for first draft token and 0 for other tokens) and I'm working on it.

[zejunchen-zejun]

should be removed?

[whx-sjtu]

done

[zejunchen-zejun]

here should have an assertion to avoid non-vllm backend calling this method
it should only be called by atom-vllm backend

[whx-sjtu]

done

[zejunchen-zejun]

Here maybe a little bit risky. If the forward_atom_config is None, and there is no assertion, it will silent fallback to the global singleton _current_atom_config. Can we add some log here? Or make it more safe
In ideal situation, the lifecycle and ownership forward_atom_config is belong to the model itself and the main model will get its atom config, draft model will get its config. While if draft model cannot get its config, it will fallback to the _current_atom_config, which not be correct

[whx-sjtu]

Do you mean that we should always obtain forward_atom_config from vllm_forward_context? Will there be scenarios that we need to return the default global _current_atom_config?

[zejunchen-zejun]

Let us add some comments here to warn a case: there is no atom config in forward context, so the default global config will be provided. With this warning, we can mitigate the coherent issue for local value and its twins global value.

[zejunchen-zejun]

can we leave some comments here for the position offset

[whx-sjtu]

not needed anymore. removed.