Skip to content

CHANGELOG.md

Latest commit

 

History

History
58 lines (47 loc) · 3.04 KB

CHANGELOG.md

File metadata and controls

58 lines (47 loc) · 3.04 KB
Raw

Changelog

Unreleased

  • Remove the custom project logo and Chinese README/quickstart; keep the README and documentation in English with the banner, demo and architecture visuals.
  • Simplify the documentation home layout and restore the English search pipeline.

0.4.1

  • Restore the README banner, architecture visual and first-use story with a new project logo, English/Chinese entry points and the local demo screenshot.
  • Publish searchable Material documentation with workflow, architecture, compatibility and evidence guides, plus a Chinese quickstart.
  • Keep the 0.4.0 estimator model and original experiment records unchanged; this patch improves presentation and documentation, with installed-package qualification repeated for the new distribution files.

0.4.0

  • Replace divergent TRL/Trainer paths with one package-backed runtime and strict requested/effective configuration; preserve explicit dtype and attention.
  • Preserve native system/multi-turn chat, keep all-token loss by default, add verified assistant masks, reject silent truncation and check causal labels.
  • Execute real updates, stage model/adapter saves, protect output directories and verify reload; report dependency/configuration/access/OOM failures explicitly.
  • Add a torch-free loopback interactive estimator and opt-in redacted evidence export/validation with manual community review.
  • Separate GiB/process peaks/safety/free-device budgets, fix native optimizer and full-gradient accounting, constrain calibration to compatible development data and report fresh same-cohort old/new validation including negative results.
  • Qualify minimum/recommended CPU stacks and native Windows RTX 4080 CUDA paths; CI checks actual training, installed wheels and rendered Python.
  • Publish the same hashed, qualified wheel/sdist through the existing Trusted Publisher; verify official PyPI hashes and fresh public installs.

Migration: regenerate recipes, install their versioned package requirements and remove unknown legacy YAML keys. Training no longer needs TRL/datasets. Default truncation now errors; opt into right truncation deliberately. Regenerate old calibration caches. Confidence remains heuristic; Liger remains experimental.

0.3.0

  • Use Hugging Face safetensors metadata when available for exact parameter counts.
  • Account for all experts in mixture-of-experts model weights and active experts in activation estimates.
  • Keep quantization metadata separate from packed weight memory.
  • Calibrate dynamic memory and allocator overhead without rescaling static weights.
  • Fail clearly when requested QLoRA or 8-bit optimizer dependencies are unavailable.
  • Refuse to overwrite non-empty recipe directories unless --force is passed.
  • Add optional Liger kernels to generated recipes with --liger.
  • Add broader tests, static type checks, formatting checks, and release-tag validation.

0.2.0

  • Add calibrated memory estimates, benchmark reporting, and generated training recipes.
  • Improve activation, attention, and logits memory accounting.
  • Add RTX 4080 benchmark baselines.