Skip to content

v0.235.0

Latest
Latest

Choose a tag to compare

View all tags
@ooples ooples released this 29 Oct 12:35
v0.235.0
8542dd4
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

0.235.0 (2026-10-29)

Features

  • nn: the base class owns every network's training step, on one shared fused session (#1804) (#2295) (6786764)

Bug Fixes

  • CLAP as LAION-CLAP's network (#2313), BLIP grounded in the image (#2297), and net462/netstandard2.0 targets (#2326) (#2323) (8964274)
  • copy a nested options object instead of aliasing it, and test which inference components engage (#1993, #1632) (#2317) (d87617f)
  • decoder-only block (#2325), vgg back in the gate (#1394), heavy-lane defects (#2087); #2322 diagnosed (#2331) (f171c7e)
  • diagnostics: build against current Tensors; diffusion probe scaling modes (Tensors #653) (#2336) (5752a0f)
  • five defects (culture parsing, console exceptions, stable hashes, registry races, httpclient lifetime) (#2364) (7186337)
  • five defects on master (random, regex timeouts, swallowed exceptions, null-forgiving initializers, options seed) (#2357) (a4511f3)
  • fused attention training, namesake model coverage, cancellation reporting, clap fidelity and factory architecture (#2312, #2138, #2131, #2313, #2147) (#2316) (1b8c757)
  • generated test families for every excluded base (#2139), the bugs they found, and the #2087/#2147 slices (#2345) (6b5c085)
  • nn: apply options.seed for models that keep their options to themselves (#2350) (663459f)
  • nn: redraw the t5 bias table when the model's seed arrives (#2367) (bc71bc1)
  • OpenSora nightly fixture, 40x default-inference slowdown, detector mAP overfit test (#2365, #1632, #2170) (#2366) (8987611)
  • playground: recover Stripe webhooks and isolate code execution (#2377) (2ff2052)

Performance

  • benchmarks: run the PyTorch parity harness over 11 model families, with fair inputs and loss evidence (#2381) (0d2b2bf)
  • diffusion: reuse a tensorarena for eager inference; fix(generators): stop persisting dispose-assigned state (#2368) (701415b)
  • extend fused recurrence and reusable state across model foundations (#2380) (d42d6b6)
  • nn: skip per-layer arena recycling when a forward rented no host scratch (#2370) (28ce9d2)
  • timeseries: fused lstm for deepar, device-side best-epoch checkpoint for all six forecasters (#2378) (04e44e8)

Refactoring

  • gpu: run gpu inference through forward and delete the forwardgpu overrides (#2332) (f836269)
  • nn: let the base own every network's options (#2351) (7819f57)

Documentation

  • OOM/timeout playbook and NVIDIA vendor-library dispatch, measured (#1684, #1159) (#2374) (5b62869)
  • website: replace unsourced speed claims with measured results (#2309) (b627f5c)
Assets 10
Loading