经 AI Skill Hub 精选评估,vLLM诊断工具 获评「推荐使用」。这款AI工具在功能完整性、社区活跃度和易用性方面表现出色,AI 评分 7.5 分,适合有一定技术背景的用户使用。
vLLM诊断工具 是一款基于 Python 开发的开源工具,专注于 llm、vllm、python 等核心功能。作为 GitHub 开源项目,它拥有活跃的社区支持和持续的版本迭代,代码完全透明可审计,支持本地部署以保护数据隐私。无论是个人使用还是集成到企业工作流,都能提供稳定可靠的解决方案。
vLLM诊断工具 是一款基于 Python 开发的开源工具,专注于 llm、vllm、python 等核心功能。作为 GitHub 开源项目,它拥有活跃的社区支持和持续的版本迭代,代码完全透明可审计,支持本地部署以保护数据隐私。无论是个人使用还是集成到企业工作流,都能提供稳定可靠的解决方案。
# 方式一:pip 安装(推荐)
pip install vllm-doctor
# 方式二:虚拟环境安装(推荐生产环境)
python -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install vllm-doctor
# 方式三:从源码安装(获取最新功能)
git clone https://github.com/aminalaee/vllm-doctor
cd vllm-doctor
pip install -e .
# 验证安装
python -c "import vllm_doctor; print('安装成功')"
# 命令行使用
vllm-doctor --help
# 基本用法
vllm-doctor input_file -o output_file
# Python 代码中调用
import vllm_doctor
# 示例
result = vllm_doctor.process("input")
print(result)
# vllm-doctor 配置文件示例(config.yml) app: name: "vllm-doctor" debug: false log_level: "INFO" # 运行时指定配置文件 vllm-doctor --config config.yml # 或通过环境变量配置 export VLLM_DOCTOR_API_KEY="your-key" export VLLM_DOCTOR_OUTPUT_DIR="./output"
<img src="https://raw.githubusercontent.com/aminalaee/vllm-doctor/main/docs/assets/wordmark.svg" alt="vLLM Doctor" width="360">
<p> <a href="https://pypi.org/project/vllm-doctor/"> <img src="https://badge.fury.io/py/vllm-doctor.svg" alt="Package version"> </a> <a href="https://pypi.org/project/vllm-doctor/"> <img src="https://img.shields.io/pypi/pyversions/vllm-doctor.svg?color=%2334D058" alt="Supported Python versions"> </a> </p>
Diagnose vLLM server bottlenecks from live metrics.

vLLM Doctor reads vLLM server metrics and turns them into diagnostic findings: what looks unhealthy, why it may be happening, and which vLLM settings are worth checking first.
vllm-doctor diagnose http://localhost:8000/metrics
vLLM Doctor is not a dashboard replacement or benchmark runner. It is a fast server-side diagnostic snapshot for a single vLLM server or Prometheus target.
With pip:
pip install vllm-doctor
With uv:
uv tool install vllm-doctor
A prebuilt image is published to GitHub Container Registry:
docker run --rm ghcr.io/aminalaee/vllm-doctor diagnose <url>
<url> is your vLLM /metrics or Prometheus endpoint — the same argument the CLI takes — reachable from inside the container.
GuideLLM is a good fit for generating workloads and measuring endpoint behavior. vLLM Doctor is a good fit for explaining server-side symptoms from vLLM metrics.
Used together, GuideLLM can create or replay load while vLLM Doctor helps explain bottlenecks such as queue pressure, KV cache pressure, high TTFT, or high TPOT.
Direct scrape:
vllm-doctor diagnose http://localhost:8000/metrics
Prometheus:
vllm-doctor diagnose http://localhost:9090
─────────────────────────────────── vLLM Doctor · Health: CRITICAL · Since: now ────────────────────────────────────
╭─ ✖ KV cache pressure [high confidence] ─────────────────────────────────────────────────────────────────────────────╮
│ GPU KV cache usage: 94% (threshold: 90%) · Waiting requests: 7 (blocked by full cache) │
│ │
│ → Reduce max_num_seqs to limit concurrent sequences │
│ → Reduce max_num_batched_tokens to cap memory per step │
│ → Increase gpu_memory_utilization if GPU memory headroom exists │
│ → Route long-context requests to a dedicated replica │
╰──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯
╭─ ⚠ High time to first token (TTFT) [high confidence] ───────────────────────────────────────────────────────────────╮
│ TTFT p95: 3.200s · TPOT p95: 0.050s · Waiting requests: 7 │
│ │
│ → Enable or tune chunked prefill (--enable-chunked-prefill) │
│ → Reduce max prompt length or filter long requests │
│ → Inspect queue depth — consider adding replicas │
│ → Separate long-context traffic to dedicated instances │
╰──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯
╭─ ⚠ Replica imbalance [high confidence] ─────────────────────────────────────────────────────────────────────────────╮
│ meta-llama/Llama-3.1-8B: running vllm-1=10 vs vllm-0=2; cache 94% vs 41%; waiting vllm-1=7 vs vllm-0=0 │
│ │
│ → Check the load balancer / service routing and session affinity settings │
│ → Verify readiness probes — an unready replica receives no traffic │
│ → Compare per-replica latency and restart any unhealthy replica │
│ → Confirm newly added replicas are registered with the load balancer │
╰──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯
╭─ ⚠ Queue pressure [low confidence] ─────────────────────────────────────────────────────────────────────────────────╮
│ Waiting requests: 7 (threshold: 5) │
│ │
│ → Add replicas or increase concurrency limits │
│ → Inspect autoscaling thresholds │
│ → Separate long-context traffic to a dedicated replica │
│ → Reduce incoming request rate │
╰──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯
KV Cache Pressure ✖ critical [high]
High TTFT ⚠ warning [high]
Replica Imbalance ⚠ warning [high]
Queue Pressure ⚠ warning [low]
Queue Latency ✓ ok
Preemption Pressure ✓ ok
Low Throughput ✓ ok
Error Rate ✓ ok
High TPOT ✓ ok
Prefix Cache Efficiency ✓ ok
─────────────────────────────────────────────────── Observed Metrics ───────────────────────────────────────────────────
Summary
Requests Running 12
Requests Waiting 7
GPU Cache Usage ███████████████████░ 94%
Prefill Tokens/s 390.0
Decode Tokens/s 252.0
Requests Success 114
Requests Error 0
Requests Aborted 0
TTFT p95 (s) 3.200
TPOT p95 (s) 0.050
Queue Time p95 (s) 0.800
Preemptions Total 0
Prefix Cache Hit Rate 50%
─────────────────────────────────────────────── Observed Metrics per pod ───────────────────────────────────────────────
vllm-1 vllm-0
Requests Running 10 2
Requests Waiting 7 0
GPU Cache Usage 94% 41%
Prefill Tokens/s 80.0 310.0
Decode Tokens/s 42.0 210.0
Requests Success 30 84
Requests Error 0 0
Requests Aborted 0 0
Preemptions Total 0 0
功能齐全,易于使用,但社区支持度较低
该工具未明确声明开源协议,商业使用前请联系原作者确认授权范围,避免侵权风险。
AI Skill Hub 为第三方内容聚合平台,本页面信息基于公开数据整理,不对工具功能和质量作任何法律背书。
建议在沙箱或测试环境中充分验证后,再部署至生产环境,并做好必要的安全评估。
AI Skill Hub 点评:vLLM诊断工具 的核心功能完整,质量良好。对于AI 技术爱好者来说,这是一个值得纳入个人工具库的选择。建议先在非生产环境试用,再逐步推广。
| 原始名称 | vllm-doctor |
| 原始描述 | 开源AI工具:Diagnostic tool for vLLM inference servers。⭐11 · Python |
| Topics | llmvllmpython |
| GitHub | https://github.com/aminalaee/vllm-doctor |
| 语言 | Python |
收录时间:2026-06-11 · 更新时间:2026-06-11 · License:未公布 · AI Skill Hub 不对第三方内容的准确性作法律背书。