[Perf](bench): refactor benchmark scripts for unified format

scripts/performance/benchmark.sh → scripts/benchmark/models/benchmark.sh

@@ -1,20 +1,12 @@
 #!/bin/bash
 set -euo pipefail
 
-# step 1: please pip install pandas openpyxl
-# step 3: change the variables as below
-# step 4: bash benchmark.sh
 
-##################### need changed #######################
-MODEL_PATH="/data/models/deepseek-ai/DeepSeek-V3.2-Exp/"
-ISL_LIST=(1023 2047 4093 6143)
-CONC_LIST=(8 4 1)
-OSL=1024
-PORT=8888
-##########################################################
+if [ ! -e "$MODEL_PATH" ]; then
+    echo "model '$MODEL_PATH' does not exist, please set MODEL_PATH firstly."
+    exit 1
+fi
 
-LOG_FILE="result.txt"
-# clean old file
 if [ -f "$LOG_FILE" ]; then
     rm "$LOG_FILE"
 fi
@@ -25,9 +17,8 @@ curl -sf "http://localhost:$PORT/health" > /dev/null || {
     exit 1
 }
 
-for ISL in "${ISL_LIST[@]}"; do
-
-    for CONC in "${CONC_LIST[@]}"; do
+for ISL in "${ISL_LOOP[@]}"; do
+    for CONC in "${CONC_LOOP[@]}"; do
     echo "=========================================" | tee -a "$LOG_FILE"
     echo "Start Test ISL=$ISL, OSL=$OSL, CONC=$CONC" | tee -a "$LOG_FILE"
     echo "=========================================" | tee -a "$LOG_FILE"
@@ -50,6 +41,27 @@ for ISL in "${ISL_LIST[@]}"; do
     done
 done
 
-if [ -f "$LOG_FILE" ]; then
-    python save_csv.py $LOG_FILE
-fi
+for COMBINATION in "${ISL_OSL_CONC[@]}"; do
+
+    IFS=':' read -r ISL OSL CONC <<< "$COMBINATION"
+
+    echo "=========================================" | tee -a "$LOG_FILE"
+    echo "Start Test ISL=$ISL, OSL=$OSL, CONC=$CONC" | tee -a "$LOG_FILE"
+    echo "=========================================" | tee -a "$LOG_FILE"
+
+    python -m atom.benchmarks.benchmark_serving \
+        --model="$MODEL_PATH" \
+        --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 * 10)) \
+        --max-concurrency="$CONC" \
+        --request-rate=inf \
+        --ignore-eos \
+        --percentile-metrics="ttft,tpot,itl,e2el"  2>&1 | tee -a "$LOG_FILE"
+
+    echo -e "\n" | tee -a "$LOG_FILE"
+done

scripts/benchmark/models/deepseek32/run_bench.sh

@@ -0,0 +1,25 @@
+#!/bin/bash
+set -euo pipefail
+
+# step 1: please pip install openpyxl
+# step 3: change the configs as below
+# step 4: bash benchmark.sh
+
+CUR_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )"
+BENCHMARK_SCRIPT="$CUR_DIR/../benchmark.sh"
+SAVE_CSV_FILE="$CUR_DIR/../save_csv.py"
+
+(
+######################### CONFIGS ###########################
+MODEL_PATH="/mnt/raid0/models/deepseek-ai/DeepSeek-V3.2-Exp/"
+ISL_LOOP=(1023 8191)
+CONC_LOOP=(1 8 64)
+OSL=1024
+ISL_OSL_CONC=(65535:1024:1 65535:1024:8 1023:1024:128 1023:8192:64)
+PORT=8888
+LOG_FILE="deepseek32"
+########################## CONFIGS ##########################
+
+source $BENCHMARK_SCRIPT
+python $SAVE_CSV_FILE $LOG_FILE
+)

scripts/benchmark/models/deepseek32/run_server.sh

@@ -0,0 +1,22 @@
+#!/bin/bash
+set -euo pipefail
+
+######################### CONFIGS ###########################
+MODEL_PATH="/mnt/raid0/models/deepseek-ai/DeepSeek-V3.2-Exp/"
+######################### CONFIGS ###########################
+
+export ATOM_ENABLE_ALLREDUCE_RMSNORM_FUSION=0
+export ATOM_ENABLE_DS_QKNORM_QUANT_FUSION=0
+
+if [ ! -e "$MODEL_PATH" ]; then
+    echo "model '$MODEL_PATH' does not exist, please set MODEL_PATH firstly."
+    exit 1
+fi
+
+python3 -m atom.entrypoints.openai_server \
+        --model /mnt/raid0/models/deepseek-ai/DeepSeek-V3.2-Exp/ \
+        --gpu-memory-utilization 0.8 \
+        --max-num-batched-tokens 70000 \
+        -tp 8 \
+        --server-port 8888 \
+        --block-size 16

scripts/performance/save_csv.py → scripts/benchmark/models/save_csv.py

@@ -223,7 +223,7 @@ def main():
         print(f"file not exist - {log_file_path}")
         sys.exit(1)
 
-    output_file = "benchmark.xlsx"
+    output_file = args.input_file + "_benchmark.xlsx"
 
     try:
         # Parse log file