Speedtronic v2
Version 2.0.0 adds optimization and systems techniques that can be enabled independently while preserving CPU, CUDA, and MPS fallbacks.
What is new
| Area | Feature | Default |
|---|---|---|
| Optimizer | Hybrid Muon + AdamW routing | AdamW |
| Optimizer | Muon+ post-orthogonal normalization | Off |
| Optimizer | Cautious sign-aligned update wrapper | Off |
| Systems | CUDA stream-backed out-of-order backprop (conservative) | Off |
| Systems | Startup shape-alignment warnings | On when applicable |
| Distributed | Asynchronous delta dispatch and global polling | On |
Design rules
- Muon's Newton–Schulz iteration uses only PyTorch operations.
- CPU and MPS never require CUDA streams; they use the standard path.
- Shape findings are warnings, not startup failures.
- Distributed Hub work is dispatched away from the training thread by default.
- Features compose, but
torch.compileand stream scheduling are mutually exclusive in this release. - Deferred techniques are documented explicitly rather than exposed as dead configuration flags.