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

Description

# 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
      }
    }
  }
}

The output of `python collect_env.py`

Contributor guide

Tech stack

python

Domain

backendperformance

Issue type

performance

DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.

5

Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.

over 1 week

Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.

active

ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.

mostly clear

Prerequisites

Understanding of CUDA kernel fusionFamiliarity with vLLM's compilation passesExperience with GPU profiling

Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.

15

Research direction

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.