[v3.2] deep_gemm::fp8_mqa_logits performs poorly on the H200

[thqq479]

[thqq479]

env : python3.10 cuda12.4 deep_gemm=2.1.1.post3

server : python -m sglang.launch_server --disable-overlap-schedule --model-path "$MODEL_PATH" --port "$PORT" --schedule-conservativeness 0.2 --tp 8 --chunked-prefill-size 8192 --mem-fraction-static 0 .83 --tool-call-parser deepseekv32 --reasoning-parser deepseek-v3 --decode-log-interval 1 --disable-radix-cache

client : python3 -m sglang.bench_serving --backend sglang --dataset-name random-ids --num-prompts 1 --random-input 64000 --random-output 100 --random-range-ratio 0.9 --model /DeepSeek-V3.2 --max-concurrency 1 --host 127.0.0.1 --port 30000 --request-rate 1 --profier

[thqq479]

The performance of sglang running v3.2 is significantly worse than v3.1. Profiler analysis shows that the fp8_mqa_logits() function accounts for half of the total attention time, which is not what we expected, right? @LyricZhao

[thqq479]