[feat] Add P/D attention select for draft model

[Ximingwang-09]

Motivation

#9573 has already enabled compatibility of Hybrid Attention in Speculative Decoding. However, for the draft model, the functionality to select the attention backend separately for the extend mode and decode mode has not been fully implemented—this PR supplements and completes the implementation of this feature.

Modifications

Accuracy Tests

Launch the server

python3 -m sglang.launch_server --model /mnt/Qwen3-8B --trust-remote-code --tp-size 8 --enable-cache-report --dtype bfloat16 --log-level info  --max-running-requests 16 --mem-fraction-static 0.85 --host 0.0.0.0 --port 9122 --speculative-algorithm EAGLE3 --speculative-num-steps 5 --speculative-eagle-topk 2 --speculative-num-draft-tokens 8 --speculative-draft-model-path /mnt/qwen3_8b_eagle3 --prefill-attention-backend fa3 --decode-attention-backend flashinfer

Test accuracy

python3 benchmark/gsm8k/bench_sglang.py  --num-shots 8 --num-questions 1319 --parallel 1319 --port 9122

Accuracy: 0.901

Benchmarking and Profiling

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.

[gemini-code-assist]

Summary of Changes

Hello @Ximingwang-09, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request completes the implementation for selecting distinct attention backends for the draft model within speculative decoding, specifically allowing separate configurations for its prefill (extend) and decode phases. This enhances the flexibility and performance tuning capabilities for the draft model's attention mechanisms.

Highlights

• Granular Attention Backend Control: Enables independent selection of attention backends for the draft model's prefill (draft_extend_attn_backend) and decode (draft_attn_backend) operations, moving beyond a single global attention_backend setting.
• Code Refactoring for Clarity: The init_attention_backend method has been refactored into smaller, more manageable private helper methods (_create_decode_backend, _create_draft_extend_backend, and specific create*_backend functions for each attention type), improving code organization and maintainability.
• Dynamic Backend Initialization: Implements a dynamic mapping system to initialize the correct attention backend based on the decode_attention_backend and prefill_attention_backend arguments provided to the server.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in issue comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request refactors the attention backend initialization for the draft model, allowing separate backends for prefill and decode modes. The changes significantly improve code structure and maintainability by breaking down a large if/elif block into smaller, more focused methods. My review includes a couple of suggestions to further reduce code duplication and improve clarity.

[gemini-code-assist]

The check for use_mla_backend is duplicated in _create_trtllm_mla_decode_backend and _create_trtllm_mla_prefill_backend. To improve maintainability, you can extract this logic into a helper method.

For example, you can add this helper method to the class:

def _check_mla_backend(self):
    if not global_server_args_dict["use_mla_backend"]:
        raise ValueError(
            "trtllm_mla backend requires MLA model (use_mla_backend=True)."
        )

Then you can call self._check_mla_backend() here and in _create_trtllm_mla_prefill_backend.

    def _create_trtllm_mla_decode_backend(self):
        self._check_mla_backend()

        from sglang.srt.layers.attention.trtllm_mla_backend import (
            TRTLLMMLAMultiStepDraftBackend,
        )

        self.has_prefill_wrapper_verify = True
        return TRTLLMMLAMultiStepDraftBackend(
            self.draft_model_runner, self.topk, self.speculative_num_steps
        )