Skip to main content

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​

AreaFeatureDefault
OptimizerHybrid Muon + AdamW routingAdamW
OptimizerMuon+ post-orthogonal normalizationOff
OptimizerCautious sign-aligned update wrapperOff
SystemsCUDA stream-backed out-of-order backprop (conservative)Off
SystemsStartup shape-alignment warningsOn when applicable
DistributedAsynchronous delta dispatch and global pollingOn

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.compile and stream scheduling are mutually exclusive in this release.
  • Deferred techniques are documented explicitly rather than exposed as dead configuration flags.

Pages​