V3.10.0

Change Log

Feature

• Add support for foreach. (#287, #476, #477)
  • More than 10 optimizers (e.g. AdaFactor, StableAdamW, Lion, AdaBelief, Amos, ...) now support foreach.
  • In most cases, foreach improves training speed by 1.1x to 1.5x, with a moderate increase in memory usage.
  • Like official PyTorch optimizers, the default value of foreach is None. When foreach=None, CUDA paths prefer the foreach implementation over the for-loop implementation.
  • If you need the previous for-loop behavior, set foreach=False explicitly.
• Update the Emo-series optimizers. (#472, #478)
  • Update EmoNavi, EmoFact, and EmoLynx.
  • Begin deprecating EmoNeco and EmoZeal (they are being phased out).
• Implement SpectralSphere optimizer. (#483, #485)
  • Controlled LLM Training on Spectral Sphere
• Support various coefficients for zero_power_via_newton_schulz_5. (#487)
  • Add coefficient presets: original, quintic, polar_express, and polar_express_safer.
  • Support custom coefficient schedules and expose ns_coeffs in Muon, DistributedMuon, AdaMuon, and AdaGO.

Refactor

• Rename and organize type aliases. (#488)

Fix

• Fix misbehavior in AdaFactor optimizer. (#477)
• Fix a potential NaN issue in AdamP optimizer. (#480, #481)
• Fix Lookahead wrapper compatibility with accelerate by normalizing lookahead_state serialization. (#484, #489)

Docs

• Convert the previous docstring style to Google style. (#487)
• Add py.typed to mark distributed typing information. (#487)
• rework the README.md. (#491, #492)

CI/CD

• Introduce uv. (#473)