Skip to content

Training Hub v0.9.2

Choose a tag to compare

View all tags
@RobotSail RobotSail released this 13 Jul 16:31
v0.9.2
26068e4
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Compatibility release for RHOAI 3.5 GA (AIPCC index).

Dependency Compatibility

  • AIPCC 3.5 GA dependency compatibility (#117) — Updated pyproject.toml constraints for AIPCC 3.5 GA package availability: raised openpipe-art cap to allow 0.5.18, removed the transformers<=5.3.0 cap from [grpo], and moved verl to a separate [grpo-verl] extra (not on AIPCC index). Also fixed trl 1.x compatibility by renaming the deprecated tokenizer parameter to processing_class in the LoRA SFT trainer, and moved max_seq_length to SFTConfig.max_length to match the trl 1.x API.

Bug Fixes

  • ART 0.5.18 megatron import stub (#118) — openpipe-art 0.5.18 added Qwen3.5 Megatron support with module-level imports from megatron.core and megatron.bridge that crash when megatron isn't installed, even when using the Unsloth (non-Megatron) backend. Added a lightweight stub that injects placeholder modules into sys.modules in the training subprocess so the imports succeed. The stub classes are never instantiated at runtime.

Documentation

  • Add vLLM 0.21+ subprocess hang learning to AGENTS.md (#116)
Assets 6
Loading