Multi-backend refactor + MLX backend MVP; bump to 0.3a1#14

Merged

JulianSlzr merged 29 commits into

mainfrom

mlx

Jun 4, 2026

Commits on Jun 1, 2026

tool/build: Modernize pyproject.toml and add design manuals
JulianSlzr
committed
refactor: Introduce specs/ protocol layer
JulianSlzr
committed
refactor: Align JAX/MLX backends and introduce backend operations
JulianSlzr
committed
refactor: Abstract and implement multi-backend testing utilities
JulianSlzr
committed
docs(sync): add temporary agent guidelines and staging path override
JulianSlzr
committed
refactor(mlx/test_utils): standardize abstract test_utils and backends for specs architecture
JulianSlzr
committed
fix(mlx/test_utils): standardize type idempotency and low-precision tolerances
JulianSlzr
committed
feat(specs/types): add channel_spec abstract property and concrete overrides
JulianSlzr
committed
fix(specs): resolve MRO and duplicate protocol issues in types
JulianSlzr
committed
feat: squash-merge mlx modules (simple through logging) onto rebased foundation
JulianSlzr
committed
fix: resolve specs mixin generic MRO issues and coerce Sequence inputs to mx.array
JulianSlzr
committed
refactor(mlx): Port signal/utils unit tests and fix latency validation
JulianSlzr
committed
chore(release): bump version to 0.3.0rc1
JulianSlzr
committed
refactor(jax): remove dead projection helper to secure checkpoint backward compatibility
JulianSlzr
committed
style(jax): explicitly restate inherited config fields in conditioning and position layers
JulianSlzr
committed
chore(specs): clean up private exports and testonly symbols from package roots
JulianSlzr
committed
refactor(simple): decouple gated unit configs from GatedUnit.Config
JulianSlzr
committed
style(jax): strip config= keyword argument in simple layers
JulianSlzr
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Multi-backend refactor + MLX backend MVP; bump to 0.3a1#14

Multi-backend refactor + MLX backend MVP; bump to 0.3a1#14
JulianSlzr merged 29 commits into
mainfrom
mlx

Commits on Jun 1, 2026

tool/build: Modernize pyproject.toml and add design manuals

refactor: Introduce specs/ protocol layer

refactor: Align JAX/MLX backends and introduce backend operations

refactor: Abstract and implement multi-backend testing utilities

docs(sync): add temporary agent guidelines and staging path override

refactor(mlx/test_utils): standardize abstract test_utils and backends for specs architecture

fix(mlx/test_utils): standardize type idempotency and low-precision tolerances

feat(specs/types): add channel_spec abstract property and concrete overrides

fix(specs): resolve MRO and duplicate protocol issues in types

feat: squash-merge mlx modules (simple through logging) onto rebased foundation

fix: resolve specs mixin generic MRO issues and coerce Sequence inputs to mx.array

refactor(mlx): Port signal/utils unit tests and fix latency validation

chore(release): bump version to 0.3.0rc1

refactor(jax): remove dead projection helper to secure checkpoint backward compatibility

style(jax): explicitly restate inherited config fields in conditioning and position layers

chore(specs): clean up private exports and testonly symbols from package roots

refactor(simple): decouple gated unit configs from GatedUnit.Config

style(jax): strip config= keyword argument in simple layers

Commits on Jun 2, 2026

refactor: unify get_initial_state and Lambda signatures to ChannelSpec

fix(jax): restore explicit_semicausal padding and fix FFT/IRFFT dtypes

feat(mlx): canonicalize dtype conversions and preserve int64/float64

chore: migrate all relative imports to absolute imports for Google3

refactor(mlx): unify duplicate _to_mx_dtype implementations

style(jax): reorder explicit_semicausal in dsp_test.py to minimize diff with main

feat(mlx): add EinsumDense.to_quantized for int8/int4 quantization

feat(mlx): support combined QKV projection quantization in DotProductSelfAttention

Commits on Jun 4, 2026

restore Python 3.12 compatibility

simplify versioning to 0.3a1

refactor(mlx): use mlx_layers backing attribute for Serial combinators