Skip to content

RELEASE_NOTES.md

Latest commit

 

History

History
43 lines (36 loc) · 2.03 KB

RELEASE_NOTES.md

File metadata and controls

43 lines (36 loc) · 2.03 KB
Raw

Unreleased

0.1.0 — 2026-07-23

First public release.

Bundles ship a generation config

  • compile_hf_bundle and save_hf_bundle now write generation_config.json (pad token aliased to eos), so generate() and pipeline() run without an explicit pad_token_id.

Breaking: Hugging Face compilation is Phi-3-first

  • compile_to_hf and compile_hf_bundle now default to the typed CompileProfile.PHI3 contract: SwiGLU, biasless projections, RMSNorm, and a stock Phi3ForCausalLM bundle with no remote code.
  • The former custom classes were renamed to TorchwrightCustomConfig, TorchwrightCustomModel, and TorchwrightCustomForCausalLM; their modules are now configuration_torchwright_custom.py and modeling_torchwright_custom.py.
  • The custom tokenizer is now TorchwrightCustomTokenizer in tokenization_torchwright_custom.py.
  • Custom model generation requires the explicit architecture="custom" profile and trust_remote_code=True when loading a published bundle.
  • The ONNX-to-HF converter, save_bundle(onnx_path, ...), and read_vocab were removed. Recompile source graphs directly to HF.
  • The obsolete calculator_v2 compatibility name was removed; use examples.calculator_simple.
  • HF bundle builders now stage and validate a complete bundle before replacing the destination. A failed compile leaves an existing published bundle untouched, and successful recompilation removes stale files.
  • Token compilation now requires exactly one Embedding reachable from the output and verifies that it is the same node supplied to the HF API.
  • Schedule metadata distinguishes the emitted schedule and its certified optimality from the independent CP-SAT attempt that proposed a candidate.
  • compile_hf_bundle returns an immutable HFBundleReport containing the published output directory, layer count, and typed selected-schedule provenance.

There are intentionally no compatibility aliases. Pin an older release to load historical Python APIs, or recompile the source graph with the current compiler.