README.md

metadata

title: QRoute
colorFrom: purple
colorTo: indigo
sdk: docker
app_port: 7860
pinned: false
license: apache-2.0
hardware: cpu-basic
tags:
  - quantum-computing
  - vqc
  - mixture-of-experts
  - pennylane
  - llm
  - routing
  - auto-deploy

QRoute - a quantum MoE router for LLMs

This Space demonstrates QRoute: a variational quantum circuit (VQC) that routes tokens to experts in a Mixture-of-Experts layer. It is project 3 of 3 in the Quantum Co-Processor program (after QVerify and QAgent).

What you can do here

• Explore expert routing. Pick an example token (one per cluster); 8 expert "lamps" light up to show which experts the VQC router and the classical baseline each select (top-2), with the routing-probability bars side by side.
• See training and load balance. Training-loss curves for both routers and an expert-utilization chart with the Gini coefficient (lower = more balanced).
• Read the toy results. Both routers reach 100% validation accuracy at a comparable parameter count.

Design notes

• Precompute and serve. Both routers are trained offline by scripts/precompute_space.py (the README config) and the per-epoch losses, routing probabilities, utilization, and metrics are baked into precomputed/toy_results.json. The Space renders them with matplotlib and does no live training/inference, so torch / pennylane are not in the image.
• Fully open. No IBM path, no quota, no OAuth.
• Honest scope. Simulator only, 3 qubits / 8 experts, a toy task. No quantum advantage is claimed; the VQC router is a trainable, integrable module that matches the classical baseline. The full-scale 7-qubit / 128-expert Gemma plan is in the repo's design.md and roadmap.md.

Auto-deployed from the qroute repo main branch via a GitHub Action (HfApi.upload_folder); see docs/deploy.md.