Skip to content

CHANGELOG.md

Latest commit

 

History

History
56 lines (39 loc) · 2.03 KB

CHANGELOG.md

File metadata and controls

56 lines (39 loc) · 2.03 KB
Raw
title CHANGELOG
type note
permalink mlxtriage/changelog

Changelog

All notable changes to this project will be documented in this file.

v0.2.0 — Reliability Hardening and Traits (2026-03-11)

New Features

  • Added claim_level to JSON output: runtime-qualified when all checks execute and preflight-only when any check is skipped
  • Added reliability metadata fields in JSON output: checks_executed, checks_skipped, and skipped_check_ids
  • Added --strict CLI mode to exit non-zero when any check is skipped
  • Added BF16-aware weight integrity inspection with MLX when available (NaN/Inf/all-zero detection)
  • Added model-intrinsic trait assembly and Tier 0 trait output in JSON
  • Added VLM architecture detection (Check T0.5)
  • Expanded tokenizer diagnostics with thinking-token detection and explicit has_chat_template metadata

Improvements

  • Introduced backend-agnostic check interfaces via a ModelBackend protocol
  • Expanded validation and public documentation for model-family and quantization coverage

v0.1.1 — Packaging and Publication Update (2026-03-03)

Improvements

  • Added trusted publishing workflow for PyPI releases on version tags
  • Updated package release plumbing for PyPI activation
  • Refined public documentation and validation tables for first publication

v0.1.0 — Initial Release

Tier 0 — Sanity Checks (no MLX required)

  • Dtype compatibility audit (BF16/FP16 mismatch detection)
  • Tokenizer and EOS configuration validation
  • Weight file integrity checks (NaN/Inf, all-zero layers)
  • MLX version and known bug cross-reference

Tier 1 — Statistical Smoke Tests (MLX required)

  • Determinism check (10 runs at temp=0)
  • Reference divergence (MLX vs PyTorch/Transformers)
  • Quantization quality gate (perplexity measurement)

Infrastructure

  • Rich terminal output with colored status indicators
  • JSON output format for programmatic use
  • Known bugs database (6 curated entries with MLX issue references)
  • Validated against 15 models across 5 families, 4 quantization levels