vllm-project/vllm

[Performance]: qknorm+rope fusion slower than unfused on H100

Open

#34,391 创建于 2026年2月12日

在 GitHub 查看
 (12 评论) (1 反应) (1 负责人)Python (80,034 star) (16,816 fork)batch import
help wantedperformancetorch.compile

描述

Proposal to improve performance

Running vllm bench sweep serve for -cc.pass_config.enable_qknorm_rope_fusion in {True, False} gives the following results:

# base command
vllm bench sweep serve \
    --serve-cmd "vllm serve --model $MODEL --tensor-parallel-size $TP --port $PORT --no-enable-prefix-caching" \
    --bench-cmd "vllm bench serve --dataset-name random --ignore-eos --model=$MODEL --port $PORT" \
    --bench-params sweep-qps.json \
    --serve-params sweep-qknorm-fusion.json \

# sweep-qps.json
[
  {
    "num-prompts": 120,
    "request-rate": 1
  },{
    "num-prompts": 600,
    "request-rate": 5
  },{
    "num-prompts": 1200,
    "request-rate": 10
  },{
    "num-prompts": 1800,
    "request-rate": 15
  },{
    "num-prompts": 2400,
    "request-rate": 20
  },{
    "num-prompts": 1000,
    "request-rate": "inf"
  }
]

# sweep-qknorm-fusion.json
{
  "fused": {
    "compilation_config": {
      "pass_config": {
        "enable_qknorm_rope_fusion": True
      }
    }
  },
  "unfused": {
    "compilation_config": {
      "pass_config": {
        "enable_qknorm_rope_fusion": False
      }
    }
  }
}

Report of performance regression

No response

Misc discussion on performance

No response

Your current environment (if you think it is necessary)

The output of `python collect_env.py`

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.

贡献者指南

技术栈
python
领域
backendperformance
议题类型
performance
难度面向新贡献者的预计实现难度，1 表示很小改动，5 表示专家级工作。
5
预计时间有经验贡献者完成调查、实现、测试并准备 pull request 的粗略时间范围。
over 1 week
活动状态议题当前的可参与程度：新鲜、活跃、陈旧、阻塞或等待维护者输入。
active
清晰度议题是否清楚说明期望改动、验收标准和下一步。
mostly clear
前置要求
Understanding of CUDA kernel fusionFamiliarity with vLLM's compilation passesExperience with GPU profiling
新手友好度1-100 的估计分数，表示该议题对首次贡献者的友好程度。
15
研究方向
The issue reports that the qknorm rope fusion pass leads to slower performance compared to unfused on H100 GPUs. Investigate the fused kernel implementation in the vLLM codebase, likely in the `csrc/` or `kernels/` directories. Profile both fused and unfused versions using NVIDIA tools like nsight or ncu to identify bottlenecks. The reproduction commands are provided in the issue; use them to generate traces. Look for comments in the pass configuration code (e.g., `enable qknorm rope fusion`) and consider if the fusion introduces unnecessary memory operations or register spills. The goal is to either fix the fusion to match or exceed unfused performance or document the regression.