This example runs a Burn model in the browser while executing every tensor operation on a remote compute peer reached over Iroh. The browser holds only the model definition and weights; convolutions, matrix multiplies and everything else run on the peer's backend (CPU or GPU). Only the 28x28 input and the 10 output probabilities cross the wire.

It mirrors the mnist-inference-web demo, but swaps the local WebAssembly backend for the burn-remote Iroh client.

Why this is interesting

Web scripting and experimentation against real GPUs. A browser tab can drive a CUDA/Metal/Vulkan backend without shipping a native build to the user.
Models larger than the browser sandbox. Memory and compute live on the peer, so the browser is not bound by WebGPU limits or tab memory.
One model, many backends. The same client code targets whatever backend the peer hosts.

How it works

The browser and the compute peer find each other through a shared topic string. Both sides hash the topic to derive the peer's Iroh identity, so no node id has to be copied around:

topic --blake3--> secret key --> peer endpoint identity
   (peer binds the secret key; the browser derives its public half)

The client then binds its own Iroh endpoint, opens an authenticated QUIC session to the peer (through a relay when a direct path is not available), and ships operations as they are submitted.

Running it

1. Start a compute peer (native)

CPU backend:

cargo run -p remote-compute-peer -- burn-web

GPU backend (wgpu):

cargo run -p remote-compute-peer --features wgpu -- burn-web

The argument (burn-web) is the topic; it must match what you type in the browser.

2. Build the web client

cd examples/remote-inference-web
./build-for-web.sh

3. Serve and open

./run-server.sh

Open http://localhost:8000, enter the same topic, click Connect, then draw a digit. The probabilities are computed on the peer and streamed back.

Notes

The compute peer is model-agnostic; it just executes operations. Swapping the model on the client side requires no change to the peer.
model.bpk is the trained MNIST model from the mnist example, identical to the one used by mnist-inference-web.
Connecting through public relays requires outbound network access from both the browser and the peer. For a fully local setup, configure both endpoints with a self-hosted relay or direct addressing through the Iroh Endpoint builder.