feat: record exact server launch args in results.json for reproducibility

- sglang_vs_vllm/run_bench.sh: add phase1/phase2 server_args per backend
- dspark_vs_default/run_bench.sh: add server_args for dspark/default
- READMEs document where to find the recorded server args

experiments/dsv4_h200_sglang_vs_vllm/README.md

@@ -9,6 +9,7 @@
 - 场景矩阵、并发、输入/输出长度、请求数对两个 backend 完全一致。
 - 每个 phase 启动 server 后先做 warmup，再跑正式压测。
 - 短上下文和长上下文分阶段启动 server，避免 `max-model-len` / `max-num-seqs` 显存冲突。
+- 每次 server 启动的**完整命令行参数**会写入 `results/<RUN_ID>/<backend>/results.json` 的 `config.phase1_server_args` / `config.phase2_server_args`，方便在 P800 等其他平台复现时保持参数一致。
 
 ## 场景矩阵
 

experiments/dsv4_h200_sglang_vs_vllm/run_bench.sh

@@ -251,6 +251,16 @@ write_backend_metadata() {
     env_path="$VENV_VLLM"
   fi
 
+  # Capture the exact server launch args for reproducibility.
+  local phase1_args phase2_args
+  if [[ "$backend" == "sglang" ]]; then
+    phase1_args="sglang serve --trust-remote-code --model-path $MODEL_PATH --tp 8 --moe-runner-backend marlin --context-length $PHASE1_MAX_MODEL_LEN --max-running-requests $PHASE1_MAX_RUNNING --mem-fraction-static 0.88 --host 0.0.0.0 --port $SGLANG_PORT"
+    phase2_args="sglang serve --trust-remote-code --model-path $MODEL_PATH --tp 8 --moe-runner-backend marlin --context-length $PHASE2_MAX_MODEL_LEN --max-running-requests $PHASE2_MAX_RUNNING --mem-fraction-static 0.88 --host 0.0.0.0 --port $SGLANG_PORT"
+  else
+    phase1_args="vllm serve $MODEL_PATH --trust-remote-code --tensor-parallel-size 8 --kv-cache-dtype fp8 --max-model-len $PHASE1_MAX_MODEL_LEN --max-num-seqs $PHASE1_MAX_NUM_SEQS --block-size 256 --gpu-memory-utilization 0.90 --tokenizer-mode deepseek_v4 --reasoning-parser deepseek_v4 --no-disable-hybrid-kv-cache-manager --disable-uvicorn-access-log --port $VLLM_PORT"
+    phase2_args="vllm serve $MODEL_PATH --trust-remote-code --tensor-parallel-size 8 --kv-cache-dtype fp8 --max-model-len $PHASE2_MAX_MODEL_LEN --max-num-seqs $PHASE2_MAX_NUM_SEQS --block-size 256 --gpu-memory-utilization 0.90 --tokenizer-mode deepseek_v4 --reasoning-parser deepseek_v4 --no-disable-hybrid-kv-cache-manager --disable-uvicorn-access-log --port $VLLM_PORT"
+  fi
+
   write_metadata_json \
     "$meta_json" \
     "${EXPERIMENT_NAME}_${backend}" \
@@ -265,12 +275,12 @@ write_backend_metadata() {
     "$env_path" \
     "H200 ${backend} TP=8 comparison benchmark for DeepSeek-V4-Flash"
 
-  # Embed config.
+  # Embed config and server args.
-  "${VENV_CLIENT}/bin/python" - "$meta_json" "$backend" <<'PY'
+  "${VENV_CLIENT}/bin/python" - "$meta_json" "$backend" "$phase1_args" "$phase2_args" <<'PY'
 import json
 import sys
 
-path, backend = sys.argv[1], sys.argv[2]
+path, backend, phase1_args, phase2_args = sys.argv[1], sys.argv[2], sys.argv[3], sys.argv[4]
 with open(path, "r", encoding="utf-8") as f:
     data = json.load(f)
 
@@ -280,6 +290,9 @@ data["config"] = {
     "phase1_max_model_len": 32768,
     "phase2_max_model_len": 210000,
     "backend": backend,
+    "server_start_script": f"experiments/dsv4_h200_sglang_vs_vllm/start_{backend}.sh",
+    "phase1_server_args": phase1_args,
+    "phase2_server_args": phase2_args,
 }
 with open(path, "w", encoding="utf-8") as f:
     json.dump(data, f, indent=2, ensure_ascii=False)

experiments/dsv4_h200_vllm_dspark_vs_default/README.md

@@ -6,6 +6,8 @@
 
 你之前的观察是：开启 DSpark 后 TTFT 比默认 vLLM 更长。这个实验把两个配置放在同一套场景下重测，确认现象并量化差异，为后续分析根因提供数据。
 
+> 每次 server 启动的**完整命令行参数**会写入 `results/<RUN_ID>/<backend>/results.json` 的 `config.server_args`，方便在 P800 等其他平台复现时保持参数一致。
+
 ## 控制变量
 
 | 维度 | DSpark | Default |

experiments/dsv4_h200_vllm_dspark_vs_default/run_bench.sh

@@ -178,6 +178,14 @@ write_backend_metadata() {
     engine="vllm-default"
   fi
 
+  # Capture exact server launch args for reproducibility.
+  local server_args
+  if [[ "$backend" == "dspark" ]]; then
+    server_args="vllm serve $model_path --trust-remote-code --tensor-parallel-size 8 --kv-cache-dtype fp8 --max-model-len $MAX_MODEL_LEN --max-num-seqs $MAX_NUM_SEQS --block-size 256 --gpu-memory-utilization 0.90 --tokenizer-mode deepseek_v4 --reasoning-parser deepseek_v4 --spec-method dspark --spec-model $model_path --spec-tokens 5 --no-disable-hybrid-kv-cache-manager --disable-uvicorn-access-log --port $DSPARK_PORT"
+  else
+    server_args="vllm serve $model_path --trust-remote-code --tensor-parallel-size 8 --kv-cache-dtype fp8 --max-model-len $MAX_MODEL_LEN --max-num-seqs $MAX_NUM_SEQS --block-size 256 --gpu-memory-utilization 0.90 --tokenizer-mode deepseek_v4 --reasoning-parser deepseek_v4 --no-disable-hybrid-kv-cache-manager --disable-uvicorn-access-log --port $DEFAULT_PORT"
+  fi
+
   write_metadata_json \
     "$meta_json" \
     "${EXPERIMENT_NAME}_${backend}" \
@@ -192,12 +200,12 @@ write_backend_metadata() {
     "$env_path" \
     "H200 ${engine} TP=8 benchmark for TTFT comparison"
 
-  # Embed config.
+  # Embed config and server args.
-  "${VENV_CLIENT}/bin/python" - "$meta_json" "$backend" <<'PY'
+  "${VENV_CLIENT}/bin/python" - "$meta_json" "$backend" "$server_args" <<'PY'
 import json
 import sys
 
-path, backend = sys.argv[1], sys.argv[2]
+path, backend, server_args = sys.argv[1], sys.argv[2], sys.argv[3]
 with open(path, "r", encoding="utf-8") as f:
     data = json.load(f)
 
@@ -207,6 +215,8 @@ data["config"] = {
     "max_model_len": 32768,
     "max_num_seqs": 256,
     "backend": backend,
+    "server_start_script": f"experiments/dsv4_h200_vllm_dspark_vs_default/start_{backend}.sh",
+    "server_args": server_args,
 }
 with open(path, "w", encoding="utf-8") as f:
     json.dump(data, f, indent=2, ensure_ascii=False)