[Plugin] [Feature] Supoort MLA q/k norm-quant fusion with SGLang + ATOM plugin for Deepseek

[qichu-yun]

Motivation

DeepSeek MLA preprocessing in the SGLang + ATOM plugin was still doing q/k RMSNorm and q quantization in separate steps, leaving unnecessary kernel and memory overhead in a hot path. Since ATOM already provides a gated fused norm-quant implementation for DeepSeek, this PR integrates that path into the plugin so supported workloads can benefit from the fusion while unsupported cases continue to use the existing fallback path.

before :

after :

Test Plan

lauch server:

export AITER_QUICK_REDUCE_QUANTIZATION=INT4
export SGLANG_AITER_FP8_PREFILL_ATTN=0
export SGLANG_USE_AITER=1
export ATOM_ENABLE_DS_QKNORM_QUANT_FUSION=1

model_path=/shared/data/amd_int/models/DeepSeek-R1-0528

export SGLANG_EXTERNAL_MODEL_PACKAGE=atom.plugin.sglang.models

TORCHINDUCTOR_COMPILE_THREADS=128 python3 -m sglang.launch_server \
    --model-path $model_path \
    --host localhost \
    --port 9000 \
    --trust-remote-code \
    --tensor-parallel-size 4 \
    --kv-cache-dtype fp8_e4m3 \
    --mem-fraction-static 0.9 \
    --page-size 1 \
    --disable-radix-cache \

client:

model_path=/shared/data/amd_int/models/DeepSeek-R1-0528-MXFP4

ISL=8000
OSL=1000
CON=4
NUM=$(( CON * 2 ))
RANGE_RATIO=1.0

PYTHONDONTWRITEBYTECODE=1 python "/home/qichu_qle/my_sgl/bench_serving/benchmark_serving.py" \
  --model=$model_path \
  --backend=sglang \
  --base-url=http://127.0.0.1:9000 \
  --dataset-name=random \
  --random-input-len="${ISL}" \
  --random-output-len="${OSL}" \
  --random-range-ratio "${RANGE_RATIO}" \
  --num-prompts="${NUM}" \
  --max-concurrency="${CON}" \
  --trust-remote-code \
  --request-rate=inf \
  --num-warmups="$(( 2 * CON ))" \
  --ignore-eos \
  --save-result \
  --percentile-metrics="ttft,tpot,itl,e2el" \
  --result-dir="./tmp/oot-benchmark-results" \
  --result-filename="${ISL}_${OSL}_${CON}.json" \
  --profile

Test Result

============ Serving Benchmark Result ============
Successful requests:                     8         
Benchmark duration (s):                  97.66     
Total input tokens:                      64000     
Total generated tokens:                  8000      
Request throughput (req/s):              0.08      
Output token throughput (tok/s):         81.92     
Total Token throughput (tok/s):          737.26    
---------------Time to First Token----------------
Mean TTFT (ms):                          1330.96   
Median TTFT (ms):                        1457.61   
P99 TTFT (ms):                           1891.41   
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          20.20     
Median TPOT (ms):                        20.08     
P99 TPOT (ms):                           21.02     
---------------Inter-token Latency----------------
Mean ITL (ms):                           20.20     
Median ITL (ms):                         19.68     
P99 ITL (ms):                            20.15     
----------------End-to-end Latency----------------
Mean E2EL (ms):                          21514.63  
Median E2EL (ms):                        21514.56  
P99 E2EL (ms):                           21516.52  
==================================================

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[qichu-yun]

Could you please kindly help review this PR? @valarLip @wuhuikx @zejunchen-zejun

[qichu-yun]

The failed case is irrelevant to this PR change， please review the code when you're free, thanks! @valarLip