
如果你使用 DeepSeek、元宝 AI、Copilot、ChatGPT 等 AI 工具,协助配置轩辕镜像、编写 docker pull 命令、修改 Docker Compose 镜像地址、配置镜像加速、排查镜像拉取失败、分析报错日志等问题,请先让 AI 阅读并遵守轩辕镜像的规则文档。
只需在 AI 对话中先发送下面这句话即可:
请先完整阅读并严格遵守以下文档中的全部规则与要求:
https://xuanyuan.cloud/agents.md
在未充分阅读并理解该文档前,不要生成任何命令、配置、修改建议、故障排查方案或技术回答。后续所有输出都必须严格以该文档中的规范为最高优先级执行。查看 agents.md 用法指南与完整示范。国内用户首推 元宝 AI、DeepSeek 的深度思考模式,不推荐豆包 AI;Cursor 等编辑器可在对话 @ 该链接,或加入 User Rules。 若 AI 无法访问外链,可 打开说明文档 复制全文粘贴。文档会随站点更新,复制内容可能过期,建议定期检查。
该镜像提供ExLlama和ExLlamaV2两种容器,用于在Jetson设备上运行***量化模型的高效推理:
exllama容器基于https://github.com/jllllll/exllama%EF%BC%88turboderp/exllama%E7%9A%84%E5%88%86%E6%94%AF%EF%BC%89%EF%BC%8C%E5%AE%89%E8%A3%85%E4%BA%8E%60/opt/exllama%60exllama:v2容器基于https://github.com/turboderp/exllamav2%EF%BC%8C%E5%AE%89%E8%A3%85%E4%BA%8E%60/opt/exllamav2%60两种加载器均支持oobabooga的https://github.com/dusty-nv/jetson-containers/tree/master/packages/llm/text-generation-webui%E5%AE%B9%E5%99%A8%E3%80%82
替换HuggingFace Hub中的***模型(如Llama-2系列)进行推理性能测试:
bash./run.sh --workdir=/opt/exllama $(./autotag exllama) /bin/bash -c \ 'python3 test_benchmark_inference.py --perf --validate -d $(huggingface-downloader TheBloke/Llama-2-7B-GPTQ)'
若模型仓库为私有或需认证,添加环境变量:
--env HUGGINGFACE_TOKEN=<YOUR-ACCESS-TOKEN>
| 模型 | 内存占用 (MB) |
|---|---|
| https://huggingface.co/TheBloke/Llama-2-7B-*** | 5,200 |
| https://huggingface.co/TheBloke/Llama-2-13B-*** | 9,135 |
| https://huggingface.co/TheBloke/LLaMA-30b-*** | 20,206 |
| https://huggingface.co/TheBloke/Llama-2-70B-*** | 35,462 |
exllama:0.0.15 | |
|---|---|
| 别名 | exllama |
| 系统要求 | L4T ['>=36'] |
| 依赖项 | https://github.com/dusty-nv/jetson-containers/tree/master/packages/build/build-essential%E3%80%81https://github.com/dusty-nv/jetson-containers/tree/master/packages/cuda/cuda%E3%80%81https://github.com/dusty-nv/jetson-containers/tree/master/packages/cuda/cudnn%E3%80%81https://github.com/dusty-nv/jetson-containers/tree/master/packages/build/python%E3%80%81https://github.com/dusty-nv/jetson-containers/tree/master/packages/numpy%E3%80%81https://github.com/dusty-nv/jetson-containers/tree/master/packages/build/cmake/cmake_pip%E3%80%81https://github.com/dusty-nv/jetson-containers/tree/master/packages/onnx%E3%80%81https://github.com/dusty-nv/jetson-containers/tree/master/packages/pytorch%E3%80%81https://github.com/dusty-nv/jetson-containers/tree/master/packages/llm/huggingface_hub |
| 被依赖项 | https://github.com/dusty-nv/jetson-containers/tree/master/packages/llm/text-generation-webui%E3%80%81https://github.com/dusty-nv/jetson-containers/tree/master/packages/llm/text-generation-webui%E3%80%81https://github.com/dusty-nv/jetson-containers/tree/master/packages/llm/text-generation-webui |
| Dockerfile | https://github.com/dusty-nv/jetson-containers/tree/master/packages/llm/exllama/Dockerfile |
exllama:0.0.14 | |
|---|---|
| 别名 | exllama |
| 系统要求 | L4T ['==35.*'] |
| 依赖项 | https://github.com/dusty-nv/jetson-containers/tree/master/packages/build/build-essential%E3%80%81https://github.com/dusty-nv/jetson-containers/tree/master/packages/cuda/cuda%E3%80%81https://github.com/dusty-nv/jetson-containers/tree/master/packages/cuda/cudnn%E3%80%81https://github.com/dusty-nv/jetson-containers/tree/master/packages/build/python%E3%80%81https://github.com/dusty-nv/jetson-containers/tree/master/packages/numpy%E3%80%81https://github.com/dusty-nv/jetson-containers/tree/master/packages/build/cmake/cmake_pip%E3%80%81https://github.com/dusty-nv/jetson-containers/tree/master/packages/onnx%E3%80%81https://github.com/dusty-nv/jetson-containers/tree/master/packages/pytorch%E3%80%81https://github.com/dusty-nv/jetson-containers/tree/master/packages/llm/huggingface_hub |
| Dockerfile | https://github.com/dusty-nv/jetson-containers/tree/master/packages/llm/exllama/Dockerfile |
| 仓库/标签 | 日期 | 架构 | 大小 |
|---|---|---|---|
| https://hub.docker.com/r/dustynv/exllama/tags | 2023-12-15 | arm64 | 5.5GB |
| https://hub.docker.com/r/dustynv/exllama/tags | 2023-12-11 | arm64 | 5.5GB |
| https://hub.docker.com/r/dustynv/exllama/tags | 2023-12-14 | arm64 | 5.4GB |
| https://hub.docker.com/r/dustynv/exllama/tags | 2023-12-15 | arm64 | 7.2GB |
| https://hub.docker.com/r/dustynv/exllama/tags | 2023-12-15 | arm64 | 5.5GB |
| https://hub.docker.com/r/dustynv/exllama/tags | 2023-12-14 | arm64 | 5.5GB |
| https://hub.docker.com/r/dustynv/exllama/tags | 2023-12-12 | arm64 | 5.5GB |
| https://hub.docker.com/r/dustynv/exllama/tags | 2023-12-15 | arm64 | 7.2GB |
容器镜像兼容其他JetPack/L4T次要版本:
- L4T R32.7容器可在其他L4T R32.7版本(JetPack 4.6+)上运行
- L4T R35.x容器可在其他L4T R35.x版本(JetPack 5.1+)上运行
可使用jetson-containers run和autotag自动选择兼容镜像,或手动构造docker run命令:
bash# 自动拉取或构建兼容镜像 jetson-containers run $(autotag exllama) # 显式指定镜像 jetson-containers run dustynv/exllama:v1-r36.2.0 # 使用docker run命令 sudo docker run --runtime nvidia -it --rm --network=host dustynv/exllama:v1-r36.2.0
jetson-containers run会向前传递参数给docker run,并添加默认配置(如--runtime nvidia、挂载/data缓存、检测设备)
autotag会查找与当前JetPack/L4T版本兼容的本地、仓库或可构建的镜像
挂载主机目录到容器:
bashjetson-containers run -v /host/path:/container/path $(autotag exllama)
运行指定命令(非交互式shell):
bashjetson-containers run $(autotag exllama) my_app --abc xyz
若使用autotag,会在需要时自动构建镜像。手动构建需先完成https://github.com/dusty-nv/jetson-containers/tree/master/docs/setup.md%EF%BC%8C%E7%84%B6%E5%90%8E%E8%BF%90%E8%A1%8C%EF%BC%9A
bashjetson-containers build exllama
构建过程会包含所有依赖项并进行测试。使用--help查看构建选项:jetson-containers build exllama --help
您可以使用以下命令拉取该镜像。请将 <标签> 替换为具体的标签版本。如需查看所有可用标签版本,请访问 标签列表页面。
来自真实用户的反馈,见证轩辕镜像的优质服务