NitroInbox Binaries

Pre-built binaries for NitroInbox that aren't available in standard package managers.

Current Binaries

llama-finetune

The LoRA fine-tuning tool from llama.cpp. Used by NitroInbox to create personalized AI adapters based on user corrections.

Building New Versions

When to Build

• llama.cpp releases a new version with improvements/fixes
• Bug fixes needed for the fine-tuning tool
• Adding support for new platforms

How to Build

1. Go to Actions tab → "Build llama-finetune" → "Run workflow"

2. Fill in the parameters:

   • llama_cpp_ref: The llama.cpp git ref to build from (tag, branch, or commit)
     • Use master for latest
     • Use a specific tag like b4600 for reproducible builds
   • release_version: The version tag for this release (e.g., v1.0.2)
     • Must be unique - can't reuse existing version tags
     • Follow semver: v{major}.{minor}.{patch}

3. Click "Run workflow"

4. Wait for build (~3 minutes for Apple Silicon)

5. Verify release was created at: https://github.com/dotnetfactory/nitroinbox-binaries/releases

Updating NitroInbox to Use New Version

After building a new version, update app.config.ts in NitroInbox:

// app.config.ts
fineTuning: {
  binaryUrls: {
    'darwin-arm64':
      'https://github.com/dotnetfactory/nitroinbox-binaries/releases/download/v1.0.2/llama-finetune-macos-arm64.zip',
    // ... other platforms
  },
}

The app automatically detects version changes by comparing:

• Local version: stored in ~/Library/Application Support/NitroInbox/tools/llama-finetune.version
• Config version: extracted from URL (/download/v1.0.2/ → v1.0.2)

When versions differ, the old binary is deleted and the new one is downloaded.

Platform Support

Platform	Runner	Status
macOS Apple Silicon	macos-14	✅ Supported
macOS Intel	macos-15	🔜 Can be added
Linux x64	ubuntu-22.04	🔜 Can be added
Linux ARM64	ubuntu-22.04-arm	🔜 Can be added
Windows	-	❌ Not planned (use WSL)

To add more platforms, update .github/workflows/build-finetune.yml.

Build Configuration

The workflow builds with these CMake flags:

cmake .. \
  -DGGML_METAL=ON \           # Metal GPU acceleration (macOS)
  -DLLAMA_BUILD_EXAMPLES=ON \ # Build example tools
  -DLLAMA_BUILD_TOOLS=ON \    # Build CLI tools
  -DBUILD_SHARED_LIBS=OFF \   # Static linking (self-contained binary)
  -DLLAMA_STATIC=ON           # Static linking

Static linking ensures the binary works without requiring libllama.dylib.

Troubleshooting

Build fails with "No rule to make target"

• The llama.cpp ref might be too old or the API changed
• Try using master or a more recent tag

Binary fails with "Library not loaded: libllama.dylib"

• Build wasn't static - ensure BUILD_SHARED_LIBS=OFF and LLAMA_STATIC=ON

Version not detected by NitroInbox

• Version is extracted from URL pattern: /download/v{version}/
• Ensure release tag follows format: v1.0.0, v1.0.1, etc.