Skip to main content

mxdeploy

License Python PyPI Platform

国产 GPU(MetaX 曦云 C500 / MXMACA 生态)一键部署 + 评测 + AI 排障 CLI。 一条命令把模型跑起来,出 benchmark 报告,报错自动诊断。

📖 新手教程docs/TUTORIAL.md — 从安装到实战,10 分钟上手。

Features

  • 一键部署:精度检查 → vLLM 配置生成 → 服务拉起 → 健康检查,全程自动化,FP8 等不支持的模型自动拦截
  • 环境体检:自动检测 mx-smi / torch(+metax) / vLLM(+maca) 适配状态,部署前先排雷
  • 性能评测:流式压测吞吐 / 首token延迟 / P95 / TPOT / 显存,输出 table / JSON / Markdown 报告
  • AI 排障:内置真实踩坑知识库(8+ 条规则),喂日志即出根因和修复方案,critical 级别自动失败退出

安装

pip install mxdeploy

要求 Python 3.10+,目标平台为 Linux(模力方舟曦云 C500 实测环境)。

快速开始

# 1. 环境体检 —— 检测 mx-smi / torch(+metax) / vLLM(+maca)
mxdeploy init

# 2. 一键部署模型
mxdeploy deploy Qwen/Qwen2.5-3B-Instruct

# 3. 性能测试,出报告
mxdeploy bench Qwen/Qwen2.5-3B-Instruct

# 4. AI 排障 —— 把报错日志喂给它
mxdeploy doctor deploy.log

命令一览

命令 功能
mxdeploy init 环境体检:mx-smi / torch / vLLM / 系统信息
mxdeploy deploy <model> 一键部署:精度检查、配置生成、服务拉起、健康检查
mxdeploy bench <model> 性能测试:吞吐 / TTFT / P95 / TPOT / 显存
mxdeploy doctor <log> AI 排障:规则引擎 + LLM 分析
mxdeploy version 版本信息

实测数据(模力方舟曦云 C500 16G vGPU,统一:并发8/请求50/max_tokens 256/util 0.8)

模型 精度 模式 吞吐 (t/s) TTFT (ms) TPOT (ms) 成功率
Qwen2.5-1.5B FP16 compile 204.75 38.15 4.77 100%
DeepSeek-R1-Distill-1.5B FP16 compile 200.55 34.54 4.93 100%
Qwen2.5-3B FP16 compile 153.65 48.21 6.40 100%
Qwen2.5-7B-GPTQ-Int8 INT8 compile 159.21 41.67 6.15 100%
GLM-4-9B-GPTQ-Int4 INT4 eager 90.51 45.97 10.79 100%
Qwen2.5-14B-GPTQ-Int4 INT4 eager 46.58 1398.79 14.05 100%

完整矩阵报告见 docs/BENCHMARK_MATRIX_C500_16G.md

64G 整卡(模力方舟曦云 C500 64G,PyTorch 可见 63.59 GiB,同参数)

模型 精度 模式 吞吐 (t/s) TTFT (ms) TPOT (ms) 成功率
Qwen2.5-7B-Instruct FP16 compile 101.78 478.56 8.17 100%
GLM-4-9B-chat FP16 compile 79.84 54.23 12.22 100%
Qwen2.5-14B-Instruct-GPTQ-Int4 INT4 eager 46.57 1408.63 14.15 100%

关键洞察:14B-INT4 eager 在 64G 整卡 46.57 t/s ≈ 16G vGPU 的 46.58 t/s——同一物理卡计算力,vGPU 分片不影响 eager 吞吐;64G 的优势是能跑 compile 模式和更大模型。 完整报告见 docs/BENCHMARK_64G_C500.md

排障知识库

全部规则来自真实部署实测,命中即给出修复方案:

规则 场景
NET-001 huggingface.co 超时 → 自动提示 HF_ENDPOINT 镜像
ENV-001 非交互 SSH 缺 MACA_PATH → vLLM import 崩溃
DEP-001 pip 覆盖官方 torch(+metax) 适配版
PREC-001 曦云 C500 不支持 FP8 → 提示换 FP16/INT8
MEM-001 显存不足 → 给出量化/换卡建议
MEM-002 util 0.9 + torch.compile autotune OOM → 降 0.8
MEM-003 KV cache 不足 → 降 max-model-len(实测 14B-INT4 8192→4096)
MISC-003 模型需 trust-remote-code → 加 --trust-remote-code
MISC-004 量化版缺 chat_template → 从官方仓库补齐 tokenizer
MISC-005 GPTQ fused 层分片精度检查 bug(compile 模式)→ 加 --enforce-eager
... 更多规则见 mxdeploy doctor --list

兼容性

  • 平台:Linux(模力方舟曦云 C500 实例实测)
  • Python:3.10+
  • 依赖:沐曦适配版 torch(+metax)、vllm(+maca)——切勿用 pip 覆盖官方适配版
  • 注意:非交互 SSH 环境需显式 export MACA_PATH=/opt/maca

开发

git clone https://github.com/leo1852098393-blip/mxdeploy
cd mxdeploy
pip install -e ".[dev]"
pytest        # 53 tests

License

Apache-2.0

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

mxdeploy-0.2.1.tar.gz (33.8 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

mxdeploy-0.2.1-py3-none-any.whl (32.4 kB view details)

Uploaded Python 3

File details

Details for the file mxdeploy-0.2.1.tar.gz.

File metadata

  • Download URL: mxdeploy-0.2.1.tar.gz
  • Upload date:
  • Size: 33.8 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/7.0.0 CPython/3.13.5

File hashes

Hashes for mxdeploy-0.2.1.tar.gz
Algorithm Hash digest
SHA256 c65de14da0ff113060e043df51f5b63b77ab4b2fe50e5b2a69e61e8192eb850c
MD5 5a48910555a0b72e0a5862a8f340db22
BLAKE2b-256 bde81d4784bcd652a4878418b264c3304645a0ad8e59b84334789c62f24aba7e

See more details on using hashes here.

File details

Details for the file mxdeploy-0.2.1-py3-none-any.whl.

File metadata

  • Download URL: mxdeploy-0.2.1-py3-none-any.whl
  • Upload date:
  • Size: 32.4 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/7.0.0 CPython/3.13.5

File hashes

Hashes for mxdeploy-0.2.1-py3-none-any.whl
Algorithm Hash digest
SHA256 22005787dd8142d369b33e27b8cef966967e10b442584144ae81d51ea58a2829
MD5 883c33c3816990431bbff47261b4b325
BLAKE2b-256 c6f668ff8242f7221e1fd77cdc149521ad171a2eaff2c32d25c334c1d99cd62d

See more details on using hashes here.

Release history Release notifications | RSS feed

0.2.3

2 files

0.2.2

2 files

This release

0.2.1 This release

2 files

0.2.0

2 files

0.1.2

2 files

0.1.1

2 files

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page