WebGPU (Browser) Inference

Run AI models directly in your browser using your GPU — no server, no API, no internet required after the model is cached.

Experimental: WebGPU inference is available in Chrome 113+ and Edge 113+. Firefox support is in progress.

---

What is WebGPU Inference?

WebGPU allows JavaScript to run compute workloads on your GPU directly in the browser. DevDiff uses this via Transformers.js to run quantized language models in-browser.

Benefits:

• ✅ 100% private — no data leaves your machine
• ✅ No Ollama or server required
• ✅ GPU-accelerated (fast on modern hardware)
• ✅ Works in air-gapped environments

Limitations:

• ❌ Requires Chrome 113+ or Edge 113+ (not Firefox yet)
• ❌ Model quality limited by browser memory constraints
• ❌ First run downloads the model (~500MB–2GB)

---

Browser Requirements

Browser	WebGPU Support	Status
Chrome 113+	✅ Full	Recommended
Edge 113+	✅ Full	Supported
Firefox Nightly	⚠️ Partial	Experimental
Safari 17.4+	⚠️ Partial	Limited

# Check if your browser supports WebGPU
# Open the browser console and run:
navigator.gpu !== undefined
# true = supported

---

Setup

1. Enable WebGPU (if needed)

Chrome: WebGPU is enabled by default in Chrome 113+. No action needed.

Edge: WebGPU is enabled by default in Edge 113+.

Chrome (older): Enable via flags:

1. Open chrome://flags
2. Search "WebGPU"
3. Enable #enable-unsafe-webgpu
4. Restart Chrome

2. Configure DevDiff

// .devdiff.config.js
export default {
  ai: {
    providers: [
      {
        name: "webgpu-browser",
        url: "webgpu://Xenova/Phi-3.5-mini-instruct",
        priority: 1,
      },
    ],
  },
};

3. First Run

On first use, the model downloads and caches in your browser's storage. Subsequent runs are instant.

devdiff generate
# First run: downloads ~600MB model
# Subsequent runs: loads from cache in ~2 seconds

---

Available WebGPU Models

Model	Size	Quality	Best For
Xenova/Phi-3.5-mini-instruct	~600MB	Good	General changelogs
Xenova/Llama-3.2-1B-Instruct	~800MB	Good	Fast analysis
Xenova/Qwen2.5-Coder-1.5B	~900MB	Great	Code-focused diffs

{
  url: "webgpu://Xenova/Phi-3.5-mini-instruct";
}
{
  url: "webgpu://Xenova/Llama-3.2-1B-Instruct";
}
{
  url: "webgpu://Xenova/Qwen2.5-Coder-1.5B";
}

---

DevDiff Playground

The DevDiff Playground uses WebGPU inference by default. Try it in your browser — no installation required.

---

Troubleshooting

"WebGPU not supported"

• Update to Chrome 113+ or Edge 113+
• Check navigator.gpu in browser console

"Out of memory" in browser

• Use a smaller model (Phi-3.5-mini instead of larger variants)
• Close other browser tabs to free GPU memory
• Try Ollama instead for heavier models

"Model download stuck"

• Check browser network tab for errors
• Clear browser cache and retry
• Ensure you have enough disk space (check DevTools → Application → Storage)