[NVIDIA] Fix breakage of using trtllm-gen fp8 moe

[kaixih]

This PR fixes a breakage caused by this PR.

cc. @kushanam

[gemini-code-assist]

Summary of Changes

Hello @kaixih, 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 addresses a breakage in the "trtllm-gen fp8 moe" functionality. I've adjusted the expected input to the "forward" methods within the relevant MoE layers to correctly handle the data format required by "TRTLLM" mode, ensuring the system operates as intended.

Highlights

• Method Signature Update: Modified the "forward" method signature in "FlashInferEPMoE" and "FlashInferFusedMoE" classes to accept a "topk_output" tuple instead of "router_logits".
• Input Validation and Extraction: Introduced a check to ensure "topk_output" is a tuple of length 2 and extracts "router_logits" from it, aligning with the "TRTLLM" mode's expectations.

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 or fill out our survey to provide feedback.

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 addresses a breakage in trtllm-gen fp8 moe by updating the forward method signature in FlashInferEPMoE and FlashInferFusedMoE. The change correctly adapts to a new API where topk_output is passed as a tuple containing router logits. My main feedback is to refactor the duplicated validation logic into a shared helper function to improve code reuse and maintainability.

[gemini-code-assist]

This logic for validating and unpacking topk_output is duplicated in python/sglang/srt/layers/moe/fused_moe_triton/layer.py. To improve maintainability and reduce code duplication, consider extracting this logic into a shared helper function.

For example, you could create a helper function like this in a utility file or at the module level in fused_moe_triton/layer.py:

def _unpack_topk_output_for_trtllm(topk_output: tuple, class_name: str) -> torch.Tensor:
    """Unpacks topk_output tuple for TRTLLM mode and returns router_logits."""
    if not isinstance(topk_output, tuple) or len(topk_output) != 2:
        raise ValueError(
            f"{class_name} expects (TopK_config, router_logits) tuple, got {type(topk_output)}"
        )
    _, router_logits = topk_output
    return router_logits

Then you could call it from both FlashInferEPMoE.forward and FlashInferFusedMoE.forward.

[gemini-code-assist]

This validation and unpacking logic for topk_output is also present in FlashInferEPMoE.forward. To improve code reuse and maintainability, you could extract this logic into a helper function.

For example, you could define this function in this file:

def _unpack_topk_output_for_trtllm(topk_output: tuple, class_name: str) -> torch.Tensor:
    """Unpacks topk_output tuple for TRTLLM mode and returns router_logits."""
    if not isinstance(topk_output, tuple) or len(topk_output) != 2:
        raise ValueError(
            f"{class_name} expects (TopK_config, router_logits) tuple, got {type(topk_output)}"
        )
    _, router_logits = topk_output
    return router_logits

Then, you can simplify the forward method here and in FlashInferEPMoE by calling this helper.

[kaixih]

Talked with @ch-wan and learned that the EPMoE and FusedMoE are sharing backends now. So, I removed FlashInferEPMoE and now everything (TP or EP MoE) should use FlashInferFusedMoE for trtllm-gen backends.

[kaixih]

The accuracy tests are conducted as well:

for TP:
sglang (pretrained=/model/models--deepseek-ai--DeepSeek-R1-0528/snapshots/4236a6af538feda4548eca9ab308586007567f52/,trust_remote_code=True,tp_size=8,max_model_len=32768,add_bos_token=True,enable_flashinfer_trtllm_moe=True,disable_shared_experts_fusion=True), gen_kwargs: (None), limit: None, num_fewshot: 5, batch_size: 512
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.9575|±  |0.0056|
|     |       |strict-match    |     5|exact_match|↑  |0.9553|±  |0.0057|

for EP:
sglang (pretrained=/model/models--deepseek-ai--DeepSeek-R1-0528/snapshots/4236a6af538feda4548eca9ab308586007567f52/,trust_remote_code=True,tp_size=8,ep_size=8,max_model_len=32768,add_bos_token=True,enable_flashinfer_trtllm_moe=True,disable_shared_experts_fusion=True), gen_kwargs: (None), limit: None, num_fewshot: 5, batch_size: 512
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.9545|±  |0.0057|
|     |       |strict-match    |     5|exact_match|↑  |0.9522|±  |0.0059|

[zhyncs]

Thanks @kaixih ! May you also help test fp4 ckpt accuracy like this #8552

[kaixih]

Sure. Just done the fp4 accuracy test with repro commands in here.

> python3 -m sglang.launch_server --model-path nvidia/DeepSeek-R1-FP4 --trust-remote-code --tp-size 4 --quantization modelopt_fp4 --enable-flashinfer-trtllm-moe --disable-shared-experts-fusion &
> echo $! > sglang_server.pid
> echo "Server PID: $(cat sglang_server.pid)"
> python3 benchmark/gsm8k/bench_sglang.py \
  --num-questions 900 \
  --parallel 32 \
  --num-shots 8
Downloading from https://raw.githubusercontent.com/openai/grade-school-math/master/grade_school_math/data/test.jsonl to /tmp/test.jsonl
/tmp/test.jsonl: 732kB [00:00, 26.8MB/s]
100%|███████████████████████████████████████████████████████| 900/900 [01:55<00:00,  7.78it/s]
Accuracy: 0.957
Invalid: 0.000
Latency: 116.863 s
Output throughput: 786.992 token/s

@zhyncs ^^