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

feat(256k-4k-probe): verify SGLang/vLLM can serve 256k+4k without OOM

Browse Source
This commit is contained in:
yy-fighting 2026-07-08 09:59:05 +00:00
parent 0f9399baf6
commit cd2ba29d49
12 changed files with 1033 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

85
experiments/dsv4_h200_256k_4k_probe/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()

32
experiments/dsv4_h200_256k_4k_probe/config.env Normal file
View File

@ -0,0 +1,32 @@
# OOM probe for 256k input + 4k output on DeepSeek-V4-Flash (H200).
# Tests whether SGLang and vLLM can serve this context length without OOM
# at concurrency=2 before adding it to the full long-context matrix.
EXPERIMENT="dsv4_h200_256k_4k_probe"
MODEL_NAME="DeepSeek-V4-Flash"
MODEL_PATH="/data/models/DeepSeek-V4-Flash"
SERVED_MODEL_NAME="deepseek-v4-flash"
SGLANG_PORT="${SGLANG_PORT:-30006}"
VLLM_PORT="${VLLM_PORT:-30005}"
VENV_SGLANG="${VENV_SGLANG:-/data/user1/yy/envs/sglang}"
VENV_VLLM="${VENV_VLLM:-/data/user1/yy/envs/vllm}"
export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
TP=8
# 256k input + 4k output + padding -> use 270000.
MAX_MODEL_LEN=270000
MAX_NUM_SEQS=8
MAX_RUNNING=8
# "concurrency input_len output_len num_prompts"
declare -a SCENARIOS=(
"2 262144 4096 4"
)
VENV_CLIENT="${VENV_CLIENT:-$VENV_SGLANG}"
SGLANG_START_SCRIPT="${SCRIPT_DIR:-.}/start_sglang.sh"
VLLM_START_SCRIPT="${SCRIPT_DIR:-.}/start_vllm.sh"

219
experiments/dsv4_h200_256k_4k_probe/parse_backend.py Executable file
View File

@ -0,0 +1,219 @@
#!/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 slo_status(metrics: dict) -> dict:
"""Check S2-tier SLO: TTFT P95 < 3s, TPOT mean < 50ms."""
ttft_ok = metrics["ttft_ms"]["p95"] < 3000.0
tpot_ok = metrics["tpot_ms"]["mean"] < 50.0
if ttft_ok and tpot_ok:
mark = "✅"
elif ttft_ok or tpot_ok:
mark = "⚠️"
else:
mark = "❌"
return {
"ttft_p95_ok": ttft_ok,
"tpot_mean_ok": tpot_ok,
"overall": mark,
}
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) | SLO |\n")
f.write("|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|\n")
for s in scenarios:
cfg = s["config"]
m = s["metrics"]
slo = s.get("slo_status", {}).get("overall", "")
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} | {slo} |\n"
)
f.write("\n")
f.write("SLO: S2 tier — TTFT P95 < 3000ms, TPOT mean < 50ms. ✅ pass, ⚠️ partial, ❌ fail.\n\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")
scenarios = []
for jsonl_path in sorted(raw_dir.glob(f"{backend}_*.jsonl")):
parts = jsonl_path.stem.split("_")
if len(parts) < 5:
continue
label = parts[1]
if label == "sharegpt":
phase = "sharegpt"
dataset = "sharegpt"
else:
phase = label
dataset = "random"
# The last three numeric tokens are: concurrency, input_len/output_ctx, output_len.
try:
concurrency, input_len, output_len = int(parts[-3]), int(parts[-2]), int(parts[-1])
except ValueError:
continue
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": phase,
"concurrency": concurrency,
"input_len": input_len,
"output_len": output_len,
"dataset": dataset,
"num_prompts": metrics["success"] + metrics["failed"],
},
"metrics": metrics,
"slo_status": slo_status(metrics),
"raw_file": str(jsonl_path),
}
scenarios.append(scenario)
if not scenarios:
print("No benchmark outputs found to parse")
return
if results_json.exists():
with open(results_json, "r", encoding="utf-8") as f:
data = json.load(f)
data["scenarios"] = scenarios
with open(results_json, "w", encoding="utf-8") as f:
json.dump(data, f, indent=2, ensure_ascii=False)
generate_report(result_root, backend, scenarios)
print(f"Parsed {len(scenarios)} scenarios into {result_root}/report.md")
if __name__ == "__main__":
main()

21
experiments/dsv4_h200_256k_4k_probe/results/20260708-094838/comparison.md Normal file
View File

@ -0,0 +1,21 @@
# SGLang vs vLLM on DeepSeek-V4-Flash (H200, TP=8)
## Summary
- Model: `/data/models/DeepSeek-V4-Flash`
- Hardware: 8x NVIDIA H200 143GB
- Tensor Parallelism: 8
- Benchmark client: `sglang.bench_serving`
- SLO reference: S2 tier (TTFT P95 < 3s, TPOT < 50ms)
## Side-by-side results
| 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 |
|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
| c2_i262144_o4096 | sglang | 2 | 262144 | 4096 | 0.07 | 165.58 | 9734.97 | 10056.38 | 9.32 | 12.82 | 13.41 | 38983.24 | ⚠️ |
| c2_i262144_o4096 | vllm | 2 | 262144 | 4096 | 0.08 | 180.27 | 8742.36 | 9023.10 | 8.64 | 11.56 | 12.05 | 35853.05 | ⚠️ |
## Notes
- SLO check uses TTFT P95 and TPOT mean (the same criteria as `scripts/SLO_STANDARDS.md`).
- A ⚠️ indicates one of the two metrics is out of target; ❌ indicates both are out.

15
experiments/dsv4_h200_256k_4k_probe/results/20260708-094838/sglang/report.md Normal file
View File

@ -0,0 +1,15 @@
# H200 SGLANG Comparison Benchmark Report
- Result root: `/data/user1/yy/experiments/dsv4_h200_256k_4k_probe/results/20260708-094838/sglang`
- Model: `/data/models/DeepSeek-V4-Flash`
- Backend: SGLANG (TP=8)
- Benchmark client: `sglang.bench_serving --backend sglang`
## Results
| 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) | SLO |
|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
| c2_i262144_o4096 | 64k | 2 | 262144 | 4096 | 54.92 | 4 | 0.07 | 9185.73 | 165.58 | 9351.31 | 6239.82 | 9734.97 | 10056.38 | 9.32 | 12.82 | 13.41 | 26116.39 | 38290.97 | 38983.24 | ⚠️ |
SLO: S2 tier — TTFT P95 < 3000ms, TPOT mean < 50ms. ✅ pass, ⚠️ partial, ❌ fail.

84
experiments/dsv4_h200_256k_4k_probe/results/20260708-094838/sglang/results.json Normal file
View File

@ -0,0 +1,84 @@
{
"metadata": {
"experiment": "dsv4_h200_256k_4k_probe_sglang",
"run_id": "20260708-094838",
"timestamp": "2026-07-08T09:48:42+00:00",
"model": "/data/models/DeepSeek-V4-Flash",
"backend": "sglang",
"engine": "sglang",
"hardware": "8x NVIDIA H200 143GB",
"accelerator": "NVIDIA H200",
"chip": "nvidia_h200",
"script": "experiments/dsv4_h200_256k_4k_probe/run_bench.sh",
"env": "/data/user1/yy/envs/sglang",
"git_commit": "a9d87f5",
"git_dirty": "dirty",
"description": "H200 64k-context sglang TP=8 comparison for DeepSeek-V4-Flash"
},
"config": {
"tp": 8,
"cuda_visible_devices": "0,1,2,3,4,5,6,7",
"max_model_len": 70000,
"backend": "sglang",
"server_start_script": "experiments/dsv4_h200_256k_4k_probe/start_sglang.sh",
"server_args": "sglang serve --trust-remote-code --model-path /data/models/DeepSeek-V4-Flash --tp 8 --moe-runner-backend marlin --context-length 270000 --max-running-requests 8 --mem-fraction-static 0.88 --host 0.0.0.0 --port 30006"
},
"scenarios": [
{
"name": "c2_i262144_o4096",
"config": {
"phase": "64k",
"concurrency": 2,
"input_len": 262144,
"output_len": 4096,
"dataset": "random",
"num_prompts": 4
},
"metrics": {
"success": 4,
"failed": 0,
"duration_s": 54.917225173005136,
"request_throughput": 0.07283689202065187,
"input_token_throughput": 9185.733591069486,
"output_token_throughput": 165.57646478594688,
"total_token_throughput": 9351.310055855432,
"total_input_tokens": 504455,
"total_output_tokens": 9093,
"e2e_ms": {
"mean": 26116.388653255854,
"p50": 24796.823772507196,
"p90": 37425.63136690442,
"p95": 38290.972255454835,
"p99": 38983.24496629517
},
"ttft_ms": {
"mean": 6239.822703748359,
"p50": 6964.9924714976805,
"p90": 9333.200673892861,
"p95": 9734.969091441597,
"p99": 10056.383825480587
},
"tpot_ms": {
"mean": 9.319386968312292,
"p50": 8.102992777237152,
"p90": 12.085884372263013,
"p95": 12.818917980381144,
"p99": 13.40534486687565
},
"itl_ms": {
"mean": 8.747509787324509,
"p50": 7.522233994677663,
"p90": 7.900186805636622,
"p95": 8.008256598259322,
"p99": 8.26424767961726
}
},
"slo_status": {
"ttft_p95_ok": false,
"tpot_mean_ok": true,
"overall": "⚠️"
},
"raw_file": "/data/user1/yy/experiments/dsv4_h200_256k_4k_probe/results/20260708-094838/sglang/raw_outputs/sglang_64k_0708_2_262144_4096.jsonl"
}
]
}

15
experiments/dsv4_h200_256k_4k_probe/results/20260708-094838/vllm/report.md Normal file
View File

@ -0,0 +1,15 @@
# H200 VLLM Comparison Benchmark Report
- Result root: `/data/user1/yy/experiments/dsv4_h200_256k_4k_probe/results/20260708-094838/vllm`
- Model: `/data/models/DeepSeek-V4-Flash`
- Backend: VLLM (TP=8)
- Benchmark client: `sglang.bench_serving --backend vllm`
## Results
| 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) | SLO |
|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
| c2_i262144_o4096 | 64k | 2 | 262144 | 4096 | 50.44 | 4 | 0.08 | 10000.93 | 180.27 | 10181.20 | 5506.23 | 8742.36 | 9023.10 | 8.64 | 11.56 | 12.05 | 24057.58 | 35278.29 | 35853.05 | ⚠️ |
SLO: S2 tier — TTFT P95 < 3000ms, TPOT mean < 50ms. ✅ pass, ⚠️ partial, ❌ fail.

84
experiments/dsv4_h200_256k_4k_probe/results/20260708-094838/vllm/results.json Normal file
View File

@ -0,0 +1,84 @@
{
"metadata": {
"experiment": "dsv4_h200_256k_4k_probe_vllm",
"run_id": "20260708-094838",
"timestamp": "2026-07-08T09:48:42+00:00",
"model": "/data/models/DeepSeek-V4-Flash",
"backend": "vllm",
"engine": "vllm",
"hardware": "8x NVIDIA H200 143GB",
"accelerator": "NVIDIA H200",
"chip": "nvidia_h200",
"script": "experiments/dsv4_h200_256k_4k_probe/run_bench.sh",
"env": "/data/user1/yy/envs/vllm",
"git_commit": "a9d87f5",
"git_dirty": "dirty",
"description": "H200 64k-context vllm TP=8 comparison for DeepSeek-V4-Flash"
},
"config": {
"tp": 8,
"cuda_visible_devices": "0,1,2,3,4,5,6,7",
"max_model_len": 70000,
"backend": "vllm",
"server_start_script": "experiments/dsv4_h200_256k_4k_probe/start_vllm.sh",
"server_args": "vllm serve /data/models/DeepSeek-V4-Flash --trust-remote-code --tensor-parallel-size 8 --kv-cache-dtype fp8 --max-model-len 270000 --max-num-seqs 8 --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 30005"
},
"scenarios": [
{
"name": "c2_i262144_o4096",
"config": {
"phase": "64k",
"concurrency": 2,
"input_len": 262144,
"output_len": 4096,
"dataset": "random",
"num_prompts": 4
},
"metrics": {
"success": 4,
"failed": 0,
"duration_s": 50.4408314510074,
"request_throughput": 0.07930083396593401,
"input_token_throughput": 10000.92554957131,
"output_token_throughput": 180.27062081305948,
"total_token_throughput": 10181.19617038437,
"total_input_tokens": 504455,
"total_output_tokens": 9093,
"e2e_ms": {
"mean": 24057.57891399844,
"p50": 22897.005692502717,
"p90": 34559.83010400087,
"p95": 35278.28712599876,
"p99": 35853.052743597065
},
"ttft_ms": {
"mean": 5506.22531299814,
"p50": 6125.462318494101,
"p90": 8391.44238250883,
"p95": 8742.361543760488,
"p99": 9023.096872761815
},
"tpot_ms": {
"mean": 8.641820628886816,
"p50": 7.61968608874154,
"p90": 10.948930343003886,
"p95": 11.562458186103475,
"p99": 12.053280460583151
},
"itl_ms": {
"mean": 8.202191201391031,
"p50": 7.1586980047868565,
"p90": 7.325935200788081,
"p95": 7.377560806344263,
"p99": 7.590774544223677
}
},
"slo_status": {
"ttft_p95_ok": false,
"tpot_mean_ok": true,
"overall": "⚠️"
},
"raw_file": "/data/user1/yy/experiments/dsv4_h200_256k_4k_probe/results/20260708-094838/vllm/raw_outputs/vllm_64k_0708_2_262144_4096.jsonl"
}
]
}

275
experiments/dsv4_h200_256k_4k_probe/run_bench.sh Normal file
View File

@ -0,0 +1,275 @@
#!/usr/bin/env bash
# 64k-context SGLang vs vLLM 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}"
log "max_model_len=${MAX_MODEL_LEN}"
log "scenarios=${#SCENARIOS[@]}"
# ---------------------------------------------------------------------------
# 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="$1"
local pid_file="/data/user1/yy/dsv4_h200_256k_4k_probe_${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="$1"
local start_script
if [[ "$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 with ${start_script}"
bash "${start_script}" >> "${log_dir_global}/${backend}.server.outer.log" 2>&1
if ! is_server_healthy "$port"; then
log "error: ${backend} server failed to become healthy"
return 1
fi
log "${backend} server is healthy on port ${port}"
}
run_warmup() {
local backend="$1"
local port input_len output_len num
if [[ "$backend" == "sglang" ]]; then
port="$SGLANG_PORT"
else
port="$VLLM_PORT"
fi
# Warmup with a moderate 64k request to exercise the long-context path.
input_len=65536
output_len=256
num=1
log "warming up ${backend} (input=${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}.warmup.log" 2>&1
log "warmup for ${backend} completed"
}
scenario_already_completed() {
local output_file="$1"
local expected="$2"
[[ -s "$output_file" ]] || return 1
local completed
completed="$(${VENV_CLIENT}/bin/python -c "
import json, sys
path = sys.argv[1]
try:
with open(path, 'r', encoding='utf-8') as f:
for line in f:
line = line.strip()
if line:
data = json.loads(line)
print(data.get('completed', 0))
break
except Exception:
print(0)
" "$output_file")"
[[ "${completed:-0}" -ge "$expected" ]]
}
run_benchmark() {
local backend="$1"
local result_root="${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 port
if [[ "$backend" == "sglang" ]]; then
port="$SGLANG_PORT"
else
port="$VLLM_PORT"
fi
log "===== ${backend} START (max_model_len=${MAX_MODEL_LEN}) ====="
stop_server "$backend"
start_server "$backend"
run_warmup "$backend"
for scenario in "${SCENARIOS[@]}"; do
read -r concurrency input_len output_len num_prompts <<< "$scenario"
output_file="${raw_dir}/${backend}_64k_$(date '+%m%d')_${concurrency}_${input_len}_${output_len}.jsonl"
detail_log="${phase_log_dir}/${backend}_64k_c${concurrency}_i${input_len}_o${output_len}.log"
if scenario_already_completed "$output_file" "$num_prompts"; then
log "skipping already-completed ${backend} scenario: c=${concurrency} i=${input_len} o=${output_len}"
continue
fi
log "running ${backend} scenario: c=${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: ${backend} scenario c=${concurrency} i=${input_len} o=${output_len} failed; see ${detail_log}"
continue
}
log "finished ${backend} scenario: output=${output_file}"
done
stop_server "$backend"
log "===== ${backend} 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="$VENV_VLLM"
fi
# Capture the exact server launch args for reproducibility.
local server_args
if [[ "$backend" == "sglang" ]]; then
server_args="sglang serve --trust-remote-code --model-path $MODEL_PATH --tp $TP --moe-runner-backend marlin --context-length $MAX_MODEL_LEN --max-running-requests $MAX_RUNNING --mem-fraction-static 0.88 --host 0.0.0.0 --port $SGLANG_PORT"
else
server_args="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 $VLLM_PORT"
fi
write_metadata_json \
"$meta_json" \
"${EXPERIMENT_NAME}_${backend}" \
"$RUN_ID" \
"$MODEL_PATH" \
"$backend" \
"$backend" \
"$HARDWARE" \
"$ACCELERATOR" \
"$CHIP" \
"experiments/${EXPERIMENT_NAME}/run_bench.sh" \
"$env_path" \
"H200 64k-context ${backend} TP=8 comparison for DeepSeek-V4-Flash"
# Embed config and server args.
jq --arg backend "$backend" \
--arg server_args "$server_args" \
'.config = {
"tp": 8,
"cuda_visible_devices": "0,1,2,3,4,5,6,7",
"max_model_len": 70000,
"backend": $backend,
"server_start_script": "experiments/dsv4_h200_256k_4k_probe/start_\($backend).sh",
"server_args": $server_args
}' "$meta_json" > "${meta_json}.tmp" && mv "${meta_json}.tmp" "$meta_json"
}
# ---------------------------------------------------------------------------
# Main
# ---------------------------------------------------------------------------
# Cleanup any leftovers.
stop_server sglang
stop_server vllm
write_backend_metadata sglang
write_backend_metadata vllm
# Run SGLang.
run_benchmark sglang
parse_backend sglang
# Run vLLM.
run_benchmark vllm
parse_backend vllm
# Generate comparison.
log "generating comparison report"
"${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}"

62
experiments/dsv4_h200_256k_4k_probe/start_sglang.sh Executable file
View File

@ -0,0 +1,62 @@
#!/bin/bash
# Start SGLang server for the 64k-context comparison.
set -e
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
# shellcheck source=/dev/null
source "${SCRIPT_DIR}/config.env"
cd /data/user1/yy
mkdir -p logs
export PATH="${VENV_SGLANG}/bin:$PATH"
export PYTHONUNBUFFERED=1
export SGLANG_LOG_LEVEL=info
export TMPDIR=/data/user1/yy/tmp
export CUDA_VISIBLE_DEVICES="${CUDA_VISIBLE_DEVICES}"
LOG="/data/user1/yy/logs/dsv4_h200_256k_4k_probe_sglang_$(date +%Y%m%d_%H%M%S).log"
PID_FILE="/data/user1/yy/dsv4_h200_256k_4k_probe_sglang.pid"
rm -f "$PID_FILE"
echo "=== Starting SGLang server (TP=$TP, max_model_len=$MAX_MODEL_LEN, max_running=$MAX_RUNNING) ==="
echo "Model: $MODEL_PATH"
echo "Port: $SGLANG_PORT"
echo "Log: $LOG"
nohup sglang serve \
--trust-remote-code \
--model-path "$MODEL_PATH" \
--tp "$TP" \
--moe-runner-backend marlin \
--context-length "$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

65
experiments/dsv4_h200_256k_4k_probe/start_vllm.sh Executable file
View File

@ -0,0 +1,65 @@
#!/bin/bash
# Start vLLM server for the 64k-context comparison.
set -e
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
# shellcheck source=/dev/null
source "${SCRIPT_DIR}/config.env"
cd /data/user1/yy
mkdir -p logs
VENV="${VENV_VLLM}"
export PATH="$VENV/bin:$PATH"
export PYTHONUNBUFFERED=1
export TMPDIR=/data/user1/yy/tmp
export CUDA_VISIBLE_DEVICES="${CUDA_VISIBLE_DEVICES}"
LOG="/data/user1/yy/logs/dsv4_h200_256k_4k_probe_vllm_$(date +%Y%m%d_%H%M%S).log"
PID_FILE="/data/user1/yy/dsv4_h200_256k_4k_probe_vllm.pid"
rm -f "$PID_FILE"
echo "=== Starting vLLM server (TP=$TP, max_model_len=$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_256k_4k_probe/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()