qqtang/sskj
1
2
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: add SGLang vs vLLM controlled comparison experiment

Browse Source 
- New experiments/dsv4_h200_sglang_vs_vllm/ with TP=8 on all 8 H200 cards
- Phase1 short-context throughput + Phase2 long context up to 200k
- Unified scenario matrix, warmup, parsing, and side-by-side comparison report
- H200 SGLang vs vLLM entry added to root README
This commit is contained in:
yy-fighting 2026-07-08 07:34:34 +00:00
parent 8b82f700e4
commit e6d9e99b6f
9 changed files with 903 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
README.md
View File

@ -28,6 +28,7 @@
| DSV4 H200 DSpark | `experiments/dsv4_h200_dspark/run_bench.sh` | NVIDIA H200 + vllm-dspark + DeepSeek-V4-Flash-DSpark | | DSV4 H200 DSpark | `experiments/dsv4_h200_dspark/run_bench.sh` | NVIDIA H200 + vllm-dspark + DeepSeek-V4-Flash-DSpark |
| DSV4 H200 vLLM baseline | `experiments/dsv4_h200_vllm/run_bench.sh` | NVIDIA H200 + vLLM + DeepSeek-V4-Flash | | DSV4 H200 vLLM baseline | `experiments/dsv4_h200_vllm/run_bench.sh` | NVIDIA H200 + vLLM + DeepSeek-V4-Flash |
| DSV4 H200 SGLang baseline | `experiments/dsv4_h200_sglang/run_bench.sh` | NVIDIA H200 + native SGLang + DeepSeek-V4-Flash | | DSV4 H200 SGLang baseline | `experiments/dsv4_h200_sglang/run_bench.sh` | NVIDIA H200 + native SGLang + DeepSeek-V4-Flash |
| DSV4 H200 SGLang vs vLLM | `experiments/dsv4_h200_sglang_vs_vllm/run_bench.sh` | NVIDIA H200 上 SGLang 与 vLLM 控制变量对比（TP=8，最长 200k 上下文） |
### 旧结构（scripts/ + bench_results/） ### 旧结构（scripts/ + bench_results/）
@ -62,6 +63,12 @@ bash experiments/dsv4_h200_dspark/run_bench.sh
bash experiments/dsv4_h200_sglang/run_bench.sh bash experiments/dsv4_h200_sglang/run_bench.sh
``` ```
### H200 SGLang vs vLLM 对比
```bash
bash experiments/dsv4_h200_sglang_vs_vllm/run_bench.sh
```
### DSpark grid（旧结构） ### DSpark grid（旧结构）
```bash ```bash

72
experiments/dsv4_h200_sglang_vs_vllm/README.md Normal file
View File

@ -0,0 +1,72 @@
# DSV4 H200 SGLang vs vLLM 对比实验
在 8x NVIDIA H200 上，用统一控制变量对比 SGLang 与 vLLM serving DeepSeek-V4-Flash。
## 设计原则
- 两个 backend 都使用 **TP=8**，绑定全部 8 张卡。
- 压测客户端统一使用 `sglang.bench_serving`。
- 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。
## 场景矩阵
### Phase 1：短上下文吞吐（max_model_len=32k）
| Concurrency | Input | Output | Num prompts |
|---:|---:|---:|---:|
| 1 | 512 | 256 | 32 |
| 32 | 512 | 256 | 128 |
| 128 | 512 | 256 | 128 |
| 1 | 4000 | 512 | 32 |
| 32 | 4000 | 512 | 64 |
### Phase 2：长上下文（max_model_len=210k）
| Concurrency | Input | Output | Num prompts |
|---:|---:|---:|---:|
| 1 | 32768 | 1024 | 8 |
| 1 | 65536 | 1024 | 4 |
| 1 | 131072 | 1024 | 2 |
| 1 | 200000 | 1024 | 1 |
## 快速运行
```bash
bash experiments/dsv4_h200_sglang_vs_vllm/run_bench.sh
```
结果保存在 `experiments/dsv4_h200_sglang_vs_vllm/results/<RUN_ID>/`：
```
results/<RUN_ID>/
├── sglang/
│ ├── results.json
│ ├── report.md
│ └── raw_outputs/ # gitignored
├── vllm/
│ ├── results.json
│ ├── report.md
│ └── raw_outputs/ # gitignored
├── comparison.md
└── logs/
```
## 文件说明
| 文件 | 作用 |
|---|---|
| `config.env` | 公共配置：模型路径、端口、场景矩阵、虚拟环境 |
| `start_sglang.sh` | 按 phase 启动 SGLang server |
| `start_vllm.sh` | 按 phase 启动 vLLM server |
| `run_bench.sh` | 总 orchestrator：跑 SGLang → 跑 vLLM → 生成对比报告 |
| `warmup.py` | 用 `sglang.bench_serving` 发送少量预热请求 |
| `parse_backend.py` | 解析单个 backend 的 raw jsonl 为 `results.json` + `report.md` |
| `compare.py` | 读取两个 backend 的 `results.json`，生成 `comparison.md` |
## 环境
- SGLM server: `/data/user1/yy/envs/sglang`
- vLLM server: `/data/user1/yy/envs/vllm`
- Benchmark client: `/data/user1/yy/envs/sglang`（统一用该环境的 `sglang.bench_serving`）

85
experiments/dsv4_h200_sglang_vs_vllm/compare.py Executable file
View File

@ -0,0 +1,85 @@
#!/usr/bin/env python3
"""Generate a side-by-side comparison of SGLang and vLLM results.
Usage:
python3 compare.py --sglang <sglang_result_root> --vllm <vllm_result_root> \
[--output comparison.md]
"""
import argparse
import json
from collections import defaultdict
from pathlib import Path
def load_result(result_root: Path) -> dict:
path = result_root / "results.json"
with open(path, "r", encoding="utf-8") as f:
return json.load(f)
def slo_status(ttft_p95_ms: float, tpot_mean_ms: float) -> str:
# S2 tier targets from scripts/SLO_STANDARDS.md.
ttft_ok = ttft_p95_ms < 3000.0
tpot_ok = tpot_mean_ms < 50.0
if ttft_ok and tpot_ok:
return "✅"
if ttft_ok or tpot_ok:
return "⚠️"
return "❌"
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--sglang", type=Path, required=True)
parser.add_argument("--vllm", type=Path, required=True)
parser.add_argument("-o", "--output", type=Path, default=Path("comparison.md"))
args = parser.parse_args()
sglang_data = load_result(args.sglang)
vllm_data = load_result(args.vllm)
by_scenario = defaultdict(dict)
for data in (sglang_data, vllm_data):
backend = data["metadata"]["engine"]
for s in data.get("scenarios", []):
key = s["name"]
by_scenario[key][backend] = s
with open(args.output, "w", encoding="utf-8") as f:
f.write("# SGLang vs vLLM on DeepSeek-V4-Flash (H200, TP=8)\n\n")
f.write("## Summary\n\n")
f.write("- Model: `/data/models/DeepSeek-V4-Flash`\n")
f.write("- Hardware: 8x NVIDIA H200 143GB\n")
f.write("- Tensor Parallelism: 8\n")
f.write("- Benchmark client: `sglang.bench_serving`\n")
f.write("- SLO reference: S2 tier (TTFT P95 < 3s, TPOT < 50ms)\n\n")
f.write("## Side-by-side results\n\n")
f.write("| Scenario | Backend | Conc | Input | Output | Req/s | OutTok/s | TTFT P95(ms) | TTFT P99(ms) | TPOT Mean(ms) | TPOT P95(ms) | TPOT P99(ms) | E2E P99(ms) | SLO |\n")
f.write("|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|\n")
for scenario_name in sorted(by_scenario.keys()):
for backend in ("sglang", "vllm"):
s = by_scenario[scenario_name].get(backend)
if s is None:
continue
cfg = s["config"]
m = s["metrics"]
status = slo_status(m["ttft_ms"]["p95"], m["tpot_ms"]["mean"])
f.write(
f"| {scenario_name} | {backend} | {cfg['concurrency']} | {cfg['input_len']} | {cfg['output_len']} | "
f"{m['request_throughput']:.2f} | {m['output_token_throughput']:.2f} | "
f"{m['ttft_ms']['p95']:.2f} | {m['ttft_ms']['p99']:.2f} | "
f"{m['tpot_ms']['mean']:.2f} | {m['tpot_ms']['p95']:.2f} | {m['tpot_ms']['p99']:.2f} | "
f"{m['e2e_ms']['p99']:.2f} | {status} |\n"
)
f.write("\n## Notes\n\n")
f.write("- SLO check uses TTFT P95 and TPOT mean (the same criteria as `scripts/SLO_STANDARDS.md`).\n")
f.write("- A ⚠️ indicates one of the two metrics is out of target; ❌ indicates both are out.\n")
print(f"Wrote comparison to {args.output}")
if __name__ == "__main__":
main()

55
experiments/dsv4_h200_sglang_vs_vllm/config.env Normal file
View File

@ -0,0 +1,55 @@
# Common configuration for the SGLang vs vLLM comparison on H200.
# All values can be overridden via environment variables.
EXPERIMENT="dsv4_h200_sglang_vs_vllm"
MODEL_NAME="DeepSeek-V4-Flash"
MODEL_PATH="/data/models/DeepSeek-V4-Flash"
SERVED_MODEL_NAME="deepseek-v4-flash"
# Ports must differ so both backends can be tested independently.
SGLANG_PORT="${SGLANG_PORT:-30006}"
VLLM_PORT="${VLLM_PORT:-30005}"
# Virtual environments.
VENV_SGLANG="${VENV_SGLANG:-/data/user1/yy/envs/sglang}"
VENV_VLLM="${VENV_VLLM:-/data/user1/yy/envs/vllm}"
# Hardware: use all 8 H200 cards for both backends.
export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
TP=8
# Phase 1: short-context throughput (max_model_len 32k).
# Each element: "concurrency input_len output_len num_prompts"
if [[ -z "${PHASE1_SCENARIOS:-}" ]]; then
declare -a PHASE1_SCENARIOS=(
"1 512 256 32"
"32 512 256 128"
"128 512 256 128"
"1 4000 512 32"
"32 4000 512 64"
)
fi
PHASE1_MAX_MODEL_LEN=32768
PHASE1_MAX_NUM_SEQS=256
PHASE1_MAX_RUNNING=256
# Phase 2: long context (max_model_len 210k to leave room for output + padding).
if [[ -z "${PHASE2_SCENARIOS:-}" ]]; then
declare -a PHASE2_SCENARIOS=(
"1 32768 1024 8"
"1 65536 1024 4"
"1 131072 1024 2"
"1 200000 1024 1"
)
fi
PHASE2_MAX_MODEL_LEN=210000
PHASE2_MAX_NUM_SEQS=2
PHASE2_MAX_RUNNING=2
# Benchmark client always runs from the sglang env so that the same
# sglang.bench_serving version is used for both backends.
VENV_CLIENT="${VENV_CLIENT:-$VENV_SGLANG}"
# Server start scripts bundled with this experiment.
SGLANG_START_SCRIPT="${SCRIPT_DIR:-.}/start_sglang.sh"
VLLM_START_SCRIPT="${SCRIPT_DIR:-.}/start_vllm.sh"

187
experiments/dsv4_h200_sglang_vs_vllm/parse_backend.py Executable file
View File

@ -0,0 +1,187 @@
#!/usr/bin/env python3
"""Parse raw sglang.bench_serving JSONL outputs for one backend.
Reads JSONL files like {sglang|vllm}_phase1_MMDD_concurrency_inputlen_outputlen.jsonl
and generates results.json + report.md in the given result root.
Usage:
python3 parse_backend.py <result_root> [--backend sglang|vllm]
"""
import argparse
import json
import re
import sys
from pathlib import Path
def parse_jsonl(path: Path) -> dict | None:
with open(path, "r", encoding="utf-8") as f:
for line in f:
line = line.strip()
if not line:
continue
try:
return json.loads(line)
except json.JSONDecodeError:
continue
return None
def compute_metrics(data: dict) -> dict:
completed = data.get("completed", 0)
total = len(data.get("input_lens", []))
failed = total - completed if total > 0 else 0
duration_s = data.get("duration", 0.0)
return {
"success": completed,
"failed": failed,
"duration_s": duration_s,
"request_throughput": data.get("request_throughput", 0.0),
"input_token_throughput": data.get("input_throughput", 0.0),
"output_token_throughput": data.get("output_throughput", 0.0),
"total_token_throughput": data.get("total_throughput", 0.0),
"total_input_tokens": data.get("total_input_tokens", 0),
"total_output_tokens": data.get("total_output_tokens", 0),
"e2e_ms": {
"mean": data.get("mean_e2e_latency_ms", 0.0),
"p50": data.get("median_e2e_latency_ms", 0.0),
"p90": data.get("p90_e2e_latency_ms", 0.0),
"p95": data.get("p95_e2e_latency_ms", 0.0),
"p99": data.get("p99_e2e_latency_ms", 0.0),
},
"ttft_ms": {
"mean": data.get("mean_ttft_ms", 0.0),
"p50": data.get("median_ttft_ms", 0.0),
"p90": data.get("p90_ttft_ms", 0.0),
"p95": data.get("p95_ttft_ms", 0.0),
"p99": data.get("p99_ttft_ms", 0.0),
},
"tpot_ms": {
"mean": data.get("mean_tpot_ms", 0.0),
"p50": data.get("median_tpot_ms", 0.0),
"p90": data.get("p90_tpot_ms", 0.0),
"p95": data.get("p95_tpot_ms", 0.0),
"p99": data.get("p99_tpot_ms", 0.0),
},
"itl_ms": {
"mean": data.get("mean_itl_ms", 0.0),
"p50": data.get("median_itl_ms", 0.0),
"p90": data.get("p90_itl_ms", 0.0),
"p95": data.get("p95_itl_ms", 0.0),
"p99": data.get("p99_itl_ms", 0.0),
},
}
def scenario_name(concurrency: int, input_len: int, output_len: int) -> str:
return f"c{concurrency}_i{input_len}_o{output_len}"
def append_scenario(results_json: Path, scenario: dict) -> None:
with open(results_json, "r", encoding="utf-8") as f:
data = json.load(f)
data["scenarios"].append(scenario)
with open(results_json, "w", encoding="utf-8") as f:
json.dump(data, f, indent=2, ensure_ascii=False)
def generate_report(result_root: Path, backend: str, scenarios: list[dict]) -> None:
report_path = result_root / "report.md"
with open(report_path, "w", encoding="utf-8") as f:
f.write(f"# H200 {backend.upper()} Comparison Benchmark Report\n\n")
f.write(f"- Result root: `{result_root}`\n")
f.write(f"- Model: `/data/models/DeepSeek-V4-Flash`\n")
f.write(f"- Backend: {backend.upper()} (TP=8)\n")
f.write(f"- Benchmark client: `sglang.bench_serving --backend {backend}`\n\n")
f.write("## Results\n\n")
f.write("| Scenario | Phase | Concurrency | Input | Output | Duration(s) | Success | Req/s | In tok/s | Out tok/s | Total tok/s | Mean TTFT(ms) | P95 TTFT(ms) | P99 TTFT(ms) | Mean TPOT(ms) | P95 TPOT(ms) | P99 TPOT(ms) | Mean E2E(ms) | P95 E2E(ms) | P99 E2E(ms) |\n")
f.write("|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|\n")
for s in scenarios:
cfg = s["config"]
m = s["metrics"]
f.write(
f"| {s['name']} | {cfg['phase']} | {cfg['concurrency']} | {cfg['input_len']} | {cfg['output_len']} | "
f"{m['duration_s']:.2f} | {m['success']} | {m['request_throughput']:.2f} | "
f"{m['input_token_throughput']:.2f} | {m['output_token_throughput']:.2f} | "
f"{m['total_token_throughput']:.2f} | "
f"{m['ttft_ms']['mean']:.2f} | {m['ttft_ms']['p95']:.2f} | {m['ttft_ms']['p99']:.2f} | "
f"{m['tpot_ms']['mean']:.2f} | {m['tpot_ms']['p95']:.2f} | {m['tpot_ms']['p99']:.2f} | "
f"{m['e2e_ms']['mean']:.2f} | {m['e2e_ms']['p95']:.2f} | {m['e2e_ms']['p99']:.2f} |\n"
)
f.write("\n")
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("result_root", type=Path)
parser.add_argument("--backend", default=None, choices=["sglang", "vllm"])
args = parser.parse_args()
result_root = args.result_root
raw_dir = result_root / "raw_outputs"
results_json = result_root / "results.json"
if not raw_dir.exists():
raise SystemExit(f"raw_outputs directory not found: {raw_dir}")
backend = args.backend
if backend is None:
# Infer from filenames if not provided.
for p in raw_dir.iterdir():
if p.name.startswith("sglang_"):
backend = "sglang"
break
if p.name.startswith("vllm_"):
backend = "vllm"
break
if backend is None:
raise SystemExit("Could not infer backend from raw outputs")
# Filename: {backend}_phase{1|2}_MMDD_concurrency_inputlen_outputlen.jsonl
pattern = re.compile(rf"^{backend}_phase(\d+)_(\d+)_(\d+)_(\d+)_(\d+)\.jsonl$")
scenarios = []
for jsonl_path in sorted(raw_dir.glob(f"{backend}_phase*.jsonl")):
m = pattern.match(jsonl_path.name)
if not m:
continue
phase_num, concurrency, input_len, output_len = m.groups()
concurrency, input_len, output_len = int(concurrency), int(input_len), int(output_len)
data = parse_jsonl(jsonl_path)
if data is None:
continue
metrics = compute_metrics(data)
scenario = {
"name": scenario_name(concurrency, input_len, output_len),
"config": {
"phase": f"phase{phase_num}",
"concurrency": concurrency,
"input_len": input_len,
"output_len": output_len,
"dataset": "random",
"num_prompts": metrics["success"] + metrics["failed"],
},
"metrics": metrics,
"raw_file": str(jsonl_path),
}
scenarios.append(scenario)
if not scenarios:
print("No benchmark outputs found to parse")
return
if results_json.exists():
for s in scenarios:
append_scenario(results_json, s)
generate_report(result_root, backend, scenarios)
print(f"Parsed {len(scenarios)} scenarios into {result_root}/report.md")
if __name__ == "__main__":
main()

266
experiments/dsv4_h200_sglang_vs_vllm/run_bench.sh Executable file
View File

@ -0,0 +1,266 @@
#!/usr/bin/env bash
# SGLang vs vLLM controlled comparison on H200.
set -Eeuo pipefail
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
EXPERIMENT_NAME="$(basename "$SCRIPT_DIR")"
# shellcheck source=/dev/null
source "${SCRIPT_DIR}/../../scripts/common/lib.sh"
# shellcheck source=/dev/null
source "${SCRIPT_DIR}/../../scripts/common/platform.sh"
# shellcheck source=/dev/null
source "${SCRIPT_DIR}/config.env"
RUN_ID="${RUN_ID:-$(date '+%Y%m%d-%H%M%S')}"
RESULT_BASE="${SCRIPT_DIR}/results"
log_dir_global="${RESULT_BASE}/${RUN_ID}/logs"
mkdir -p "$log_dir_global"
log_init "${log_dir_global}/orchestrator.log"
log "experiment=${EXPERIMENT_NAME}"
log "run_id=${RUN_ID}"
log "platform=${PLATFORM}"
log "hardware=${HARDWARE}"
log "model=${MODEL_PATH}"
# ---------------------------------------------------------------------------
# Helpers
# ---------------------------------------------------------------------------
is_server_healthy() {
local port="$1"
curl --fail --silent --show-error --max-time 5 "http://127.0.0.1:${port}/health" >/dev/null 2>&1
}
stop_server() {
local backend="$1"
local pid_file="/data/user1/yy/dsv4_h200_sglang_vs_vllm_${backend}.pid"
if [[ -f "$pid_file" ]]; then
local pid
pid="$(cat "$pid_file")"
if kill -0 "$pid" 2>/dev/null; then
log "stopping ${backend} server pid=${pid}"
kill "$pid" 2>/dev/null || true
sleep 5
kill -9 "$pid" 2>/dev/null || true
fi
rm -f "$pid_file"
fi
# Fallback cleanup.
if [[ "$backend" == "sglang" ]]; then
pkill -9 -f "sglang serve.*DeepSeek-V4-Flash" 2>/dev/null || true
else
pkill -9 -f "vllm serve.*DeepSeek-V4-Flash" 2>/dev/null || true
fi
sleep 2
}
start_server() {
local backend="$1"
local phase="$2"
local start_script
if [[ "$backend" == "sglang" ]]; then
start_script="${SGLANG_START_SCRIPT}"
local port="$SGLANG_PORT"
else
start_script="${VLLM_START_SCRIPT}"
local port="$VLLM_PORT"
fi
log "starting ${backend} server (phase=${phase}) with ${start_script}"
bash "${start_script}" "$phase" >> "${log_dir_global}/${backend}_${phase}.server.outer.log" 2>&1
if ! is_server_healthy "$port"; then
log "error: ${backend} server (phase=${phase}) failed to become healthy"
return 1
fi
log "${backend} server is healthy on port ${port}"
}
run_warmup() {
local backend="$1"
local phase="$2"
local port input_len output_len num
if [[ "$backend" == "sglang" ]]; then
port="$SGLANG_PORT"
else
port="$VLLM_PORT"
fi
if [[ "$phase" == "phase1" ]]; then
input_len=4000
output_len=512
num=2
else
input_len=200000
output_len=1024
num=1
fi
log "warming up ${backend} (phase=${phase}, input=${input_len}, output=${output_len}, num=${num})"
"${VENV_CLIENT}/bin/python" "${SCRIPT_DIR}/warmup.py" \
--backend "$backend" \
--port "$port" \
--input-len "$input_len" \
--output-len "$output_len" \
--num "$num" \
--env-python "${VENV_CLIENT}/bin/python" \
>> "${log_dir_global}/${backend}_${phase}.warmup.log" 2>&1
log "warmup for ${backend} ${phase} completed"
}
run_phase() {
local backend="$1"
local phase="$2"
local result_root="${RESULT_BASE}/${RUN_ID}/${backend}"
local raw_dir="${result_root}/raw_outputs"
local phase_log_dir="${result_root}/logs"
mkdir -p "$raw_dir" "$phase_log_dir"
local -n scenarios
local max_model_len
if [[ "$phase" == "phase1" ]]; then
scenarios=PHASE1_SCENARIOS
max_model_len="$PHASE1_MAX_MODEL_LEN"
else
scenarios=PHASE2_SCENARIOS
max_model_len="$PHASE2_MAX_MODEL_LEN"
fi
local port
if [[ "$backend" == "sglang" ]]; then
port="$SGLANG_PORT"
else
port="$VLLM_PORT"
fi
log "===== ${backend} ${phase} START (max_model_len=${max_model_len}) ====="
stop_server "$backend"
start_server "$backend" "$phase"
run_warmup "$backend" "$phase"
for scenario in "${scenarios[@]}"; do
read -r concurrency input_len output_len num_prompts <<< "$scenario"
output_file="${raw_dir}/${backend}_${phase}_$(date '+%m%d')_${concurrency}_${input_len}_${output_len}.jsonl"
detail_log="${phase_log_dir}/${backend}_${phase}_c${concurrency}_i${input_len}_o${output_len}.log"
log "running ${backend} ${phase} scenario: c=${concurrency} i=${input_len} o=${output_len} n=${num_prompts}"
"${VENV_CLIENT}/bin/python" -m sglang.bench_serving \
--backend "$backend" \
--host 127.0.0.1 \
--port "$port" \
--dataset-name random \
--random-input-len "$input_len" \
--random-output-len "$output_len" \
--num-prompts "$num_prompts" \
--max-concurrency "$concurrency" \
--request-rate 10000 \
--output-file "$output_file" \
--output-details \
> "$detail_log" 2>&1 || {
log "ERROR: ${backend} ${phase} scenario c=${concurrency} i=${input_len} o=${output_len} failed; see ${detail_log}"
continue
}
log "finished ${backend} ${phase} scenario: output=${output_file}"
done
stop_server "$backend"
log "===== ${backend} ${phase} DONE ====="
}
parse_backend() {
local backend="$1"
local result_root="${RESULT_BASE}/${RUN_ID}/${backend}"
log "parsing ${backend} results in ${result_root}"
"${VENV_CLIENT}/bin/python" "${SCRIPT_DIR}/parse_backend.py" "$result_root" --backend "$backend" \
>> "${result_root}/logs/parse.log" 2>&1 || {
log "WARNING: parser failed for ${backend}; see ${result_root}/logs/parse.log"
}
}
write_backend_metadata() {
local backend="$1"
local result_root="${RESULT_BASE}/${RUN_ID}/${backend}"
ensure_result_root "$result_root"
local meta_json="${result_root}/results.json"
local env_path
if [[ "$backend" == "sglang" ]]; then
env_path="$VENV_SGLANG"
else
env_path="$VENV_VLLM"
fi
write_metadata_json \
"$meta_json" \
"${EXPERIMENT_NAME}_${backend}" \
"$RUN_ID" \
"$MODEL_PATH" \
"$backend" \
"$backend" \
"$HARDWARE" \
"$ACCELERATOR" \
"$CHIP" \
"experiments/${EXPERIMENT_NAME}/run_bench.sh" \
"$env_path" \
"H200 ${backend} TP=8 comparison benchmark for DeepSeek-V4-Flash"
# Embed config.
"${VENV_CLIENT}/bin/python" - "$meta_json" "$backend" <<'PY'
import json
import sys
path, backend = sys.argv[1], sys.argv[2]
with open(path, "r", encoding="utf-8") as f:
data = json.load(f)
data["config"] = {
"tp": 8,
"cuda_visible_devices": "0,1,2,3,4,5,6,7",
"phase1_max_model_len": 32768,
"phase2_max_model_len": 210000,
"backend": backend,
}
with open(path, "w", encoding="utf-8") as f:
json.dump(data, f, indent=2, ensure_ascii=False)
PY
}
# ---------------------------------------------------------------------------
# Main
# ---------------------------------------------------------------------------
# Cleanup any leftovers.
stop_server sglang
stop_server vllm
write_backend_metadata sglang
write_backend_metadata vllm
# Run SGLang.
run_phase sglang phase1
run_phase sglang phase2
parse_backend sglang
# Run vLLM.
run_phase vllm phase1
run_phase vllm phase2
parse_backend vllm
# Generate comparison.
log "generating comparison report"
"${VENV_CLIENT}/bin/python" "${SCRIPT_DIR}/compare.py" \
--sglang "${RESULT_BASE}/${RUN_ID}/sglang" \
--vllm "${RESULT_BASE}/${RUN_ID}/vllm" \
--output "${RESULT_BASE}/${RUN_ID}/comparison.md" \
>> "${log_dir_global}/compare.log" 2>&1 || {
log "WARNING: comparison script failed; see ${log_dir_global}/compare.log"
}
log "all results saved to ${RESULT_BASE}/${RUN_ID}"

76
experiments/dsv4_h200_sglang_vs_vllm/start_sglang.sh Executable file
View File

@ -0,0 +1,76 @@
#!/bin/bash
# Start SGLang server for the comparison, in either phase1 or phase2 mode.
set -e
PHASE="${1:-phase1}"
if [[ "$PHASE" != "phase1" && "$PHASE" != "phase2" ]]; then
echo "Usage: $0 {phase1|phase2}"
exit 1
fi
cd /data/user1/yy
mkdir -p logs
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
# shellcheck source=/dev/null
source "${SCRIPT_DIR}/config.env"
export PATH="${VENV_SGLANG}/bin:$PATH"
export PYTHONUNBUFFERED=1
export SGLANG_LOG_LEVEL=info
export TMPDIR=/data/user1/yy/tmp
export CUDA_VISIBLE_DEVICES="${CUDA_VISIBLE_DEVICES}"
if [[ "$PHASE" == "phase1" ]]; then
MAX_MODEL_LEN="$PHASE1_MAX_MODEL_LEN"
MAX_RUNNING="$PHASE1_MAX_RUNNING"
else
MAX_MODEL_LEN="$PHASE2_MAX_MODEL_LEN"
MAX_RUNNING="$PHASE2_MAX_RUNNING"
fi
LOG="/data/user1/yy/logs/dsv4_h200_sglang_vs_vllm_sglang_${PHASE}_$(date +%Y%m%d_%H%M%S).log"
PID_FILE="/data/user1/yy/dsv4_h200_sglang_vs_vllm_sglang.pid"
rm -f "$PID_FILE"
echo "=== Starting SGLang server (TP=$TP, phase=$PHASE, max_model_len=$MAX_MODEL_LEN) ==="
echo "Model: $MODEL_PATH"
echo "Port: $SGLANG_PORT"
echo "Log: $LOG"
nohup sglang serve \
--trust-remote-code \
--model-path "$MODEL_PATH" \
--tp "$TP" \
--moe-runner-backend marlin \
--max-model-len "$MAX_MODEL_LEN" \
--max-running-requests "$MAX_RUNNING" \
--mem-fraction-static 0.88 \
--host 0.0.0.0 \
--port "$SGLANG_PORT" \
> "$LOG" 2>&1 &
PID=$!
echo $PID > "$PID_FILE"
echo "PID: $PID"
echo "Waiting for health..."
for i in $(seq 1 240); do
if curl --fail --silent --show-error --max-time 5 "http://127.0.0.1:${SGLANG_PORT}/health" >/dev/null 2>&1; then
echo "SGLang server is ready at http://127.0.0.1:${SGLANG_PORT}"
echo "Log: $LOG"
exit 0
fi
if ! kill -0 $PID 2>/dev/null; then
echo "ERROR: SGLang server exited early"
tail -200 "$LOG"
exit 1
fi
echo "Waiting... ($i/240)"
sleep 5
done
echo "ERROR: SGLang server not healthy after 240 retries"
tail -200 "$LOG"
exit 1

79
experiments/dsv4_h200_sglang_vs_vllm/start_vllm.sh Executable file
View File

@ -0,0 +1,79 @@
#!/bin/bash
# Start vLLM server for the comparison, in either phase1 or phase2 mode.
set -e
PHASE="${1:-phase1}"
if [[ "$PHASE" != "phase1" && "$PHASE" != "phase2" ]]; then
echo "Usage: $0 {phase1|phase2}"
exit 1
fi
cd /data/user1/yy
mkdir -p logs
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
# shellcheck source=/dev/null
source "${SCRIPT_DIR}/config.env"
VENV="${VENV_VLLM}"
export PATH="$VENV/bin:$PATH"
export PYTHONUNBUFFERED=1
export TMPDIR=/data/user1/yy/tmp
export CUDA_VISIBLE_DEVICES="${CUDA_VISIBLE_DEVICES}"
if [[ "$PHASE" == "phase1" ]]; then
MAX_MODEL_LEN="$PHASE1_MAX_MODEL_LEN"
MAX_NUM_SEQS="$PHASE2_MAX_NUM_SEQS"
else
MAX_MODEL_LEN="$PHASE2_MAX_MODEL_LEN"
MAX_NUM_SEQS="$PHASE2_MAX_NUM_SEQS"
fi
LOG="/data/user1/yy/logs/dsv4_h200_sglang_vs_vllm_vllm_${PHASE}_$(date +%Y%m%d_%H%M%S).log"
PID_FILE="/data/user1/yy/dsv4_h200_sglang_vs_vllm_vllm.pid"
rm -f "$PID_FILE"
echo "=== Starting vLLM server (TP=$TP, phase=$PHASE, max_model_len=$MAX_MODEL_LEN, max_num_seqs=$MAX_NUM_SEQS) ==="
echo "Model: $MODEL_PATH"
echo "Port: $VLLM_PORT"
echo "Log: $LOG"
nohup vllm serve "$MODEL_PATH" \
--trust-remote-code \
--tensor-parallel-size "$TP" \
--kv-cache-dtype fp8 \
--max-model-len "$MAX_MODEL_LEN" \
--max-num-seqs "$MAX_NUM_SEQS" \
--block-size 256 \
--gpu-memory-utilization 0.90 \
--tokenizer-mode deepseek_v4 \
--reasoning-parser deepseek_v4 \
--no-disable-hybrid-kv-cache-manager \
--disable-uvicorn-access-log \
--port "$VLLM_PORT" \
> "$LOG" 2>&1 &
PID=$!
echo $PID > "$PID_FILE"
echo "PID: $PID"
echo "Waiting for health..."
for i in $(seq 1 240); do
if curl --fail --silent --show-error --max-time 5 "http://127.0.0.1:${VLLM_PORT}/health" >/dev/null 2>&1; then
echo "vLLM server is ready at http://127.0.0.1:${VLLM_PORT}"
echo "Log: $LOG"
exit 0
fi
if ! kill -0 $PID 2>/dev/null; then
echo "ERROR: vLLM server exited early"
tail -200 "$LOG"
exit 1
fi
echo "Waiting... ($i/240)"
sleep 5
done
echo "ERROR: vLLM server not healthy after 240 retries"
tail -200 "$LOG"
exit 1

76
experiments/dsv4_h200_sglang_vs_vllm/warmup.py Executable file
View File

@ -0,0 +1,76 @@
#!/usr/bin/env python3
"""Send a small number of warmup requests to a running backend.
Uses sglang.bench_serving with a single request so that the same code path
(prefill / decode kernels, CUDA graphs, etc.) is exercised before the real
benchmark begins. Discards the output.
"""
import argparse
import subprocess
import sys
import tempfile
from pathlib import Path
def run_warmup(backend: str, host: str, port: int, input_len: int, output_len: int, num: int, env_python: Path) -> None:
with tempfile.NamedTemporaryFile(mode="w", suffix=".jsonl", delete=True) as tmp:
cmd = [
str(env_python),
"-m",
"sglang.bench_serving",
"--backend",
backend,
"--host",
host,
"--port",
str(port),
"--dataset-name",
"random",
"--random-input-len",
str(input_len),
"--random-output-len",
str(output_len),
"--num-prompts",
str(num),
"--max-concurrency",
"1",
"--request-rate",
"10000",
"--output-file",
tmp.name,
"--output-details",
]
print(f"[warmup] {' '.join(cmd)}", flush=True)
result = subprocess.run(cmd, capture_output=True, text=True)
if result.returncode != 0:
print("[warmup] FAILED", file=sys.stderr)
print(result.stdout, file=sys.stderr)
print(result.stderr, file=sys.stderr)
sys.exit(1)
print(f"[warmup] OK: backend={backend} port={port} input={input_len} output={output_len} num={num}")
def main():
parser = argparse.ArgumentParser(description="Warmup a serving backend.")
parser.add_argument("--backend", required=True, choices=["sglang", "vllm"])
parser.add_argument("--port", type=int, required=True)
parser.add_argument("--input-len", type=int, required=True)
parser.add_argument("--output-len", type=int, required=True)
parser.add_argument("--num", type=int, default=1)
parser.add_argument("--host", default="127.0.0.1")
parser.add_argument("--env-python", default="/data/user1/yy/envs/sglang/bin/python")
args = parser.parse_args()
run_warmup(
backend=args.backend,
host=args.host,
port=args.port,
input_len=args.input_len,
output_len=args.output_len,
num=args.num,
env_python=Path(args.env_python),
)
if __name__ == "__main__":
main()