Skip to content

v0.5.24

Latest
Latest

Choose a tag to compare

View all tags
@github-actions github-actions released this 04 Nov 22:54
· 9 commits to main since this release
Immutable release. Only release title and notes can be modified.
v0.5.24
1e8c59b
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Release Highlights

  • Weave 0.54.0 or later is now required. Weave 0.53.9 and earlier fail in weave.init() with AttributeError: 'OTLPSpanExporter' object has no attribute '_session' once opentelemetry-exporter-otlp-proto-http 1.45 is installed. Installing ART now upgrades an older Weave instead of keeping it.
  • KL reference adapters no longer leak into saved checkpoints. With a KL penalty, the Unsloth backend left the reference LoRA loaded, so later checkpoints saved in the same process, including other models' checkpoints, contained it. Saving and loading now drop unused adapters, and load_lora_adapter raises if the file is missing any LoRA weight.
  • Megatron trainer attention runs on SM120 (RTX PRO 6000 Blackwell) through a Triton fallback.
  • Native GDN context parallelism (CP>1) keeps rank transitions in FP32.

What's Changed

  • fix(unsloth): keep KL reference adapters out of saved checkpoints (#1161)
  • fix(deps): require weave>=0.54.0 for OpenTelemetry 1.45 exporters (#1159), with the release smoke test updated in #1162
  • feat(megatron): support SM120 trainer attention fallback (#1156)
  • fix: preserve FP32 precision in native GDN context-parallel rank transitions (#1155)

Full Changelog: v0.5.23...v0.5.24

Assets 5
Loading