[Performance] [WIP] Relaxed mtp

[haoyangli0109]

python -m atom.entrypoints.openai_server \
  --model  /shareddata/amd/DeepSeek-R1-0528-MXFP4-MTP-MoEFP4  -tp 8   --method mtp --num-speculative-tokens 3

MODEL=/shareddata/amd/DeepSeek-R1-0528-MXFP4-MTP-MoEFP4
ISL=1024
OSL=1024
CONC=128
PORT=8000
RESULT_FILENAME=Deepseek-R1-result
python -m atom.benchmarks.benchmark_serving \
  --model=$MODEL --backend=vllm --base-url=http://localhost:$PORT \
  --dataset-name=random \
  --random-input-len=$ISL --random-output-len=$OSL \
  --random-range-ratio 0.8 \
  --num-prompts=$(( $CONC * 20 )) \
  --max-concurrency=$CONC \
  --request-rate=inf --ignore-eos \
  --save-result --percentile-metrics="ttft,tpot,itl,e2el" \

mtp=3, concurrency=1, inp=out=1000
	original strict	N=10,delta=0.5
TPOT/ms	3.89	3.31
gsm8k	0.94	0.945
mtp=3, concurrency=128, inp=out=1000
	original strict	N=10,delta=0.5
TPOT/ms	17.39	16.89
gsm8k	0.94	0.95