Skip to content

v0.1.4

Choose a tag to compare

View all tags
@strangeloopcanon strangeloopcanon released this 14 Sep 18:42
· 18 commits to main since this release
v0.1.4
8581c3a

Highlights

  • Beam search now applies processors on raw logits (HF parity), then normalizes.
  • Fixed sampling string stop-sequence scope bug (avoids UnboundLocalError at early steps).
  • Training: robust unary/binary tree-map helpers for bf16 compute and microbatch grad accumulation.
  • Soft prompt training falls back when input_embeddings kwarg is unsupported by the model.
  • Docs: README now covers auto-convert loader, limitations, and usage tips.

Commits since v0.1.3

  • Generation parity fixes: beam logits semantics, stop-sequence scope bug, training tree_map robustness, soft-prompt fallback, and README updates
  • chore: bump version to 0.1.4

Parity report

See the attached JSON report generated against Qwen 0.6B for sampling and beam modes.

Assets 3
Loading