Structured adherence and streaming

• Built out the whole structured-generation stack: JsonAdherence with retries, schema + semantic validators, pluggable grammar wrappers, and a GenerateResult object that carries attempts/violations.
• Added a streaming engine with incremental JSON validation, token callbacks, and stop_on_invalid control; MLX backend now streams tokens while enforcing schemas in real time.
• Exposed high-level helpers (StructuredSpec, generate_structured, generate_many, semantic predicates) plus logging/eval utilities and a JSONL adherence logger.
• Expanded docs and examples (README + examples/structured_generation.py) so users can see structured and streaming workflows end to end, including the new CLI flags.
• Broadened tests: unit coverage for the adherence/streaming engines, a stubbed eval-suite smoke test, and integration demos.
• Polished packaging/release workflow with make publish, release notes, and metadata fixes so PyPI uploads succeed cleanly.

v0.2.0

Rename and compatibility

• Project renamed to mlx-genkit (from mlx-gen-parity).
• Python import mlx_genkit added; continues to re-export the same public API.
• Old import mlx_gen_parity remains supported for compatibility.
• New CLI alias mlxgk-generate; old mlxgp-generate retained.

What else changed

• No functional changes from v0.1.4 besides the rename and packaging metadata.

Install

pip install mlx-genkit

v0.1.4

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.