[XPU] Enable QK Norm + RoPE fusion pass on XPU

[chaojun-zhang]

Performance & Accuracy Test

Commands

Server (fusion_on adds --compilation-config or set , eager adds --enforce-eager instead):

vllm serve Qwen/Qwen3-8B --port 8072 --host 0.0.0.0 --trust-remote-code \
    --gpu-memory-util=0.9 --no-enable-prefix-caching \
    --max-num-batched-tokens=8192 --max-model-len=6000 --block-size 64 -tp=1 \
    --compilation-config '{"pass_config":{"enable_qk_norm_rope_fusion":true}}'

Bench client:

vllm bench serve --model Qwen/Qwen3-8B --dataset-name random \
    --random-input-len=4096 --random-output-len=1024 --ignore-eos \
    --num-prompt 32 --max-concurrency 8 --request-rate inf --backend vllm \
    --trust-remote-code --temperature=0 --port 8072

Benchmark report (input=4096, output=1024, num-prompts=32, max-concurrency=8)

Case	Duration (s)	Output tok/s	Total tok/s	Mean TTFT (ms)	Mean TPOT (ms)
fusion_on	321.33	101.98	509.89	17651.13	53.39
fusion_off	358.83	91.32	456.60	23016.48	54.09
eager[baseline]	321.58	101.90	509.48	16938.86	55.13

fusion_on gives +11.7% throughput and -10.5% duration vs fusion_off, matching eager.

Accuracy report (GSM8K, 5-shot, limit=250)

Case	Exact match (strict)	Exact match (flexible)
fusion_on	0.908	0.916
fusion_off	0.912	0.916
eager	0.904	0.912

All within 1 stderr of each other — no accuracy regression from the fusion pass.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[jikunshang]

cc @Yejing-Lai