What's Changed
- fix: proper gpu support for linalg by @polvalente in #113
- Makefile: make MLX_SO overridable by @GenericJam in #111
- fix: handle spaces in env vars by @polvalente in #115
- Fix int32 overflow in elem_count for tensors with >2.1B elements by @edenamram in #116
- Avoid host round-trips for same-device backend copies by @hfiguera in #117
- Support dense Qwen3 generation in EMLXAxon by @hfiguera in #119
- feat: EMLX intermediate representation lowering compiler by @polvalente in #120
- fix: fast rope with more than one head by @polvalente in #122
- feat: lower while and runtime calls by @polvalente in #123
- feat: add fused kv_cache+sdpa by @polvalente in #124
New Contributors
- @GenericJam made their first contribution in #111
- @edenamram made their first contribution in #116
- @hfiguera made their first contribution in #117
Full Changelog: v0.3.0...v0.4.0
Contributors
hfiguera, GenericJam, and 2 other contributors