diff --git a/experiments/dsv4_h200_sglang_vs_vllm/README.md b/experiments/dsv4_h200_sglang_vs_vllm/README.md index ec71558..1b594c6 100644 --- a/experiments/dsv4_h200_sglang_vs_vllm/README.md +++ b/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///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 diff --git a/experiments/dsv4_h200_sglang_vs_vllm/run_bench.sh b/experiments/dsv4_h200_sglang_vs_vllm/run_bench.sh index 5c4614e..ab26636 100755 --- a/experiments/dsv4_h200_sglang_vs_vllm/run_bench.sh +++ b/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 ${backend} TP=8 comparison benchmark for DeepSeek-V4-Flash" - # Embed config. - "${VENV_CLIENT}/bin/python" - "$meta_json" "$backend" <<'PY' + # Embed config and server args. + "${VENV_CLIENT}/bin/python" - "$meta_json" "$backend" "$phase1_args" "$phase2_args" <<'PY' import json import sys -path, backend = sys.argv[1], sys.argv[2] +path, backend, phase1_args, phase2_args = sys.argv[1], sys.argv[2], sys.argv[3], sys.argv[4] with open(path, "r", encoding="utf-8") as f: data = json.load(f) @@ -280,6 +290,9 @@ data["config"] = { "phase1_max_model_len": 32768, "phase2_max_model_len": 210000, "backend": backend, + "server_start_script": f"experiments/dsv4_h200_sglang_vs_vllm/start_{backend}.sh", + "phase1_server_args": phase1_args, + "phase2_server_args": phase2_args, } with open(path, "w", encoding="utf-8") as f: json.dump(data, f, indent=2, ensure_ascii=False) diff --git a/experiments/dsv4_h200_vllm_dspark_vs_default/README.md b/experiments/dsv4_h200_vllm_dspark_vs_default/README.md index 4dcb34a..1fbb519 100644 --- a/experiments/dsv4_h200_vllm_dspark_vs_default/README.md +++ b/experiments/dsv4_h200_vllm_dspark_vs_default/README.md @@ -6,6 +6,8 @@ 你之前的观察是：开启 DSpark 后 TTFT 比默认 vLLM 更长。这个实验把两个配置放在同一套场景下重测，确认现象并量化差异，为后续分析根因提供数据。 +> 每次 server 启动的**完整命令行参数**会写入 `results///results.json` 的 `config.server_args`，方便在 P800 等其他平台复现时保持参数一致。 + ## 控制变量 | 维度 | DSpark | Default | diff --git a/experiments/dsv4_h200_vllm_dspark_vs_default/run_bench.sh b/experiments/dsv4_h200_vllm_dspark_vs_default/run_bench.sh index bcf7cf2..3cd9019 100755 --- a/experiments/dsv4_h200_vllm_dspark_vs_default/run_bench.sh +++ b/experiments/dsv4_h200_vllm_dspark_vs_default/run_bench.sh @@ -178,6 +178,14 @@ write_backend_metadata() { engine="vllm-default" fi + # Capture exact server launch args for reproducibility. + local server_args + if [[ "$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 ${engine} TP=8 benchmark for TTFT comparison" - # Embed config. - "${VENV_CLIENT}/bin/python" - "$meta_json" "$backend" <<'PY' + # Embed config and server args. + "${VENV_CLIENT}/bin/python" - "$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)