laya-mlx 0.4.0

laya-mlx v0.4.0

on Python PyPI

Laya-MLX v0.4.0 selectively ports upstream Laya v0.4.1 runtime features to native MLX inference.

Behavior change

Router() now sends language-undecided inputs to the multilingual checkpoint. Identified English still uses the English checkpoint. Set Router(default="english") to retain the previous fallback behavior.

Changes

• Keep resident-model requests and status reads available during cold checkpoint construction. Deduplicate builds and synchronize targeted unloads per checkpoint.
• Register custom checkpoints with register, unregister, registered, or models=. Source replacement and unregistration cannot resurrect an outdated in-flight model.
• Add predict_tournament to narrow large choice sets without an embedding model. Its probabilities and usage describe the final pass; metadata records finalists and elimination rounds.
• Add scalar and per-option-count min_confidence gates to Agent and Router. Gate results are reported without replacing the original answer or automatically invoking a fallback.
• Use FP32 intermediate computation for CPU parallel-option attention to avoid native FP16 kernel crashes on hosted macOS 26; GPU precision and parameter storage precision are unchanged.
• Support checkpoints trained with option_layout="parallel" throughout MLX inference, compilation, prefix caching, and conversion. Existing checkpoints remain sequential; do not enable parallel layout by editing an old checkpoint's configuration.
• Fix language detection for capitalized acronyms, address fragments, emphasis capitals, and mixed scripts; fix French device-footer removal.
• Release MLX free-buffer caches on unload and eviction while preserving caller-owned and in-flight model references.
• Update English/Chinese documentation, attribution, and CI with separate pinned sequential and parallel-layout upstream references.

Validation

• 231 tests passed on CPU and 231 on GPU before release.
• Three published checkpoints in FP32 and FP16: 378/378 argmax agreements with the pinned sequential upstream reference.
• 100 deterministic finite calls per configuration, 600 total, with zero measured active-memory growth.
• Parallel-layout tests cover all 24 permutations of four options in FP32/FP16, numerical comparison with upstream v0.4.1, compilation, caching, and export/reload.
• Ruff, wheel/sdist strict metadata checks, and installed-wheel smoke tests passed.

Raw checkpoint validation · API and migration details

Install or upgrade:

python -m pip install --upgrade laya-mlx==0.4.0