Remote Model Viewer

Remote Model Viewer is a VS Code extension for inspecting machine learning model files without leaving the editor.
Remote Model Viewer 是一个用于在 VS Code 内查看机器学习模型文件的扩展。

It provides two viewers:
它提供两种查看方式：

• Remote Model Viewer: launches Netron on the current extension host and embeds the Netron UI inside VS Code.
  Remote Model Viewer：在当前 extension host 所在机器上启动 Netron，并将 Netron UI 嵌入到 VS Code 中。
• Safetensors Explorer: a built-in viewer for .safetensors and .safetensors.index.json with tensor browsing and paged value preview.
  Safetensors Explorer：内置的 .safetensors / .safetensors.index.json 查看器，支持 tensor 浏览和分页预览。

It works both locally and with Remote SSH.
它既支持本地工作区，也支持 Remote SSH 工作区。

Features / 功能

• Open model files in an embedded Netron view from the Explorer, editor title, or tab context menu.
  可从资源管理器、编辑器标题栏或标签页右键菜单中，以内嵌 Netron 方式打开模型文件。
• Open .safetensors and .safetensors.index.json in a dedicated explorer.
  可使用专用查看器打开 .safetensors 和 .safetensors.index.json。
• Browse tensors by hierarchical name prefix.
  支持按 tensor name 前缀层级浏览。
• Filter tensors by keyword or prefix-only matching.
  支持关键字过滤和前缀过滤。
• Inspect tensor metadata such as shape, dtype, shard, element count, and byte size.
  可查看 tensor 的 shape、dtype、分片、元素数量和字节大小等元信息。
• Preview tensor values with manual offsets and paged loading.
  支持手动 offset 和分页加载的 tensor 值预览。
• Preview quantized safetensors dtypes including FP8 and packed MXFP4/MXFP6 values.
  支持预览 FP8 以及打包 MXFP4/MXFP6 等量化 safetensors dtype。

Requirements / 依赖要求

For the Netron-based viewer:
对于基于 Netron 的查看器：

• Python must be available on the extension host.
  extension host 所在机器需要有 Python。
• The netron Python package must be installed on the extension host.
  extension host 所在机器需要安装 netron Python 包。

pip install netron

In a Remote SSH session, install netron on the remote server, not on the local machine.
在 Remote SSH 场景下，应当在远程服务器上安装 netron，而不是本地机器。

The Safetensors Explorer does not depend on Python or the netron package.
Safetensors Explorer 不依赖 Python，也不依赖 netron。

Commands / 命令

This extension contributes the following commands:
此扩展提供以下命令：

• Open in Remote Model Viewer
• Open Active File in Remote Model Viewer
• Open in Safetensors Explorer

These commands are available from:
这些命令可以从以下位置使用：

• Explorer context menu
  资源管理器右键菜单
• Editor title actions
  编辑器标题栏
• Editor tab context menu
  编辑器标签页右键菜单
• Command Palette
  命令面板

Usage / 使用方式

Open With Netron / 使用 Netron 打开

Use Open in Remote Model Viewer for model formats supported by Netron, for example:
对于 Netron 支持的模型格式，可使用 Open in Remote Model Viewer，例如：

• .onnx
• .pt
• .pth
• .pb
• .tflite
• .mlmodel
• .h5
• .keras
• .safetensors

The extension starts Netron on the extension host, exposes the local service through vscode.env.asExternalUri(), and renders the Netron page inside a webview.
扩展会在 extension host 所在机器启动 Netron，通过 vscode.env.asExternalUri() 暴露本地服务，并在 webview 中渲染 Netron 页面。

Open With Safetensors Explorer / 使用 Safetensors Explorer 打开

Use Open in Safetensors Explorer for:
对于以下文件，可使用 Open in Safetensors Explorer：

• .safetensors
• .safetensors.index.json

Symbolic links are supported. For a linked sharded index, shard files are resolved relative to both the link location and the real index location.

支持软链接。对于软链接指向的分片索引，扩展会同时从链接所在目录和真实索引所在目录解析分片文件。

The Safetensors Explorer provides:
Safetensors Explorer 提供：

• Tensor tree grouped by name prefix
  按名称前缀分组的 tensor 树
• Tensor metadata panel
  tensor 元信息面板
• Offset-based value preview
  基于 offset 的值预览
• Quantized dtype preview for F8_E5M2, F8_E4M3, F8_E8M0, F4, F6_E2M3, and F6_E3M2, with automatic dequantized preview when related scale / zero-point tensors can be matched.
  支持 F8_E5M2、F8_E4M3、F8_E8M0、F4、F6_E2M3、F6_E3M2 等量化 dtype 预览；当能匹配到关联的 scale / zero-point tensor 时，会自动显示反量化预览。
• Load Tensor Values and Load More pagination
  Load Tensor Values 与 Load More 分页加载

Tensor values are loaded only when requested from the detail pane.
只有在详情面板里主动请求时，才会读取 tensor 值。

For quantized tensors, the preview always shows the stored values. If a related scale tensor such as weight_scale, weight_scale_inv, weight.scale, or weight.scales can be mapped to the selected weight, Safetensors Explorer also shows a dequantized preview. Matching zero-point tensors such as weight_zero_point are applied when present.
对于量化 tensor，预览始终会显示存储值。如果能把 weight_scale、weight_scale_inv、weight.scale、weight.scales 等关联 scale tensor 映射到当前权重，Safetensors Explorer 还会显示反量化预览；如果存在 weight_zero_point 等 zero-point tensor，也会一起应用。

Sibling block-scale tensors are supported. For example, mtp.0.attn.wo_a.weight with shape [16384, 4096] and mtp.0.attn.wo_a.scale with shape [128, 32] is treated as a [128, 32] scale grid over [128, 128] weight blocks.
支持同前缀的 block-scale tensor。例如 mtp.0.attn.wo_a.weight 的 shape 是 [16384, 4096]，mtp.0.attn.wo_a.scale 的 shape 是 [128, 32] 时，会按 [128, 32] 的 scale 网格映射到 [128, 128] 的权重 block。

For I8 weights with F8_E8M0 block scales, select the *.weight tensor and click Load Tensor Values. Selecting the *.scale tensor shows the scale tensor's stored values instead.
对于 I8 权重配 F8_E8M0 block scale 的情况，需要选中 *.weight tensor 并点击 Load Tensor Values；如果选中的是 *.scale tensor，则显示的是 scale tensor 自己的存储值。

Remote SSH Behavior / Remote SSH 行为

When used with Remote SSH:
在 Remote SSH 场景下：

• The extension host runs on the remote machine.
  extension host 运行在远程机器上。
• Netron is launched on the remote machine.
  Netron 运行在远程机器上。
• Safetensors are read directly from the remote filesystem.
  Safetensors 文件直接从远程文件系统读取。
• You do not need to manually copy the model file to local disk before opening it.
  不需要先手动把模型文件复制到本地磁盘再打开。

For the Netron viewer, the UI is still rendered in the local VS Code client.
对于 Netron 查看器，UI 仍然是在本地 VS Code 客户端中渲染的。

Extension Settings / 扩展配置

This extension contributes the following settings:
此扩展提供以下配置项：

• remoteModelViewer.pythonCommand: Python executable used to launch the Netron helper process. Default: python3
remoteModelViewer.pythonCommand：启动 Netron 辅助进程所使用的 Python 命令，默认值为 python3
• remoteModelViewer.startupTimeoutMs: Timeout for waiting until Netron becomes reachable. Default: 15000
remoteModelViewer.startupTimeoutMs：等待 Netron 可访问的超时时间，默认值为 15000
• remoteModelViewer.enableVerboseServerLog: Enables verbose logging when starting Netron. Default: false
remoteModelViewer.enableVerboseServerLog：启动 Netron 时启用详细日志，默认值为 false

Installation / 安装

Install from a packaged .vsix:
通过打包后的 .vsix 安装：

1. Run:
   执行：

npm install
npm run package

2. In VS Code, run Extensions: Install from VSIX...
   在 VS Code 中执行 Extensions: Install from VSIX...
3. Select the generated .vsix file, for example:
   选择生成的 .vsix 文件，例如：

remote-model-viewer-0.0.10.vsix

Development / 开发

This repository is a plain CommonJS extension and does not require a build step.
这个仓库是一个普通的 CommonJS 扩展，不需要额外构建步骤。

Useful commands:
常用命令：

npm run check

npm test

npm run package

To launch an Extension Development Host:
启动扩展开发宿主的方法：

1. Open this folder in VS Code
   用 VS Code 打开当前目录
2. Press F5
   按 F5
3. Choose the Run Extension configuration if prompted
   如果出现配置选择，使用 Run Extension

Known Limitations / 已知限制

• The Netron viewer depends on the upstream netron Python package and does not bundle it.
  Netron 查看器依赖上游 netron Python 包，本扩展不会内置它。
• The extension embeds the Netron UI rather than re-implementing it.
  扩展是嵌入 Netron UI，而不是重写 Netron。
• Safetensors Explorer focuses on metadata browsing and paged preview, not full tensor export.
  Safetensors Explorer 目前侧重元信息浏览和分页预览，不是完整 tensor 导出工具。
• Large safetensors files or large sharded indexes may still take noticeable time to open.
  较大的 safetensors 文件或大型分片索引在打开时仍可能有明显延迟。
• Dequantized preview is intentionally conservative. It applies directly stored quantized values with related scale / zero-point tensors, but it does not unpack format-specific qweight layouts such as GPTQ/AWQ without an explicit, reliable recipe in the safetensors file.
  反量化预览会保持保守：它会对直接存储的量化值应用关联的 scale / zero-point tensor，但不会在 safetensors 文件缺少可靠规则时强行解包 GPTQ/AWQ 等格式专用的 qweight 布局。

Related Projects / 相关项目

• Netron
• vtemplier/vscode-netron
• NetronInVSCode