Repository navigation
Phase C: MLX backend port for AMICATorchNG (Apple-native GPU) #76
Description
Activity
- added a commit that references this issue
on Jul 8, 2026 Done via PR #79 (squash-merged into the epic branch
feature/issue-74-epic-apple-gpu, commit 4ff3a11).AMICAMLXNG(pyAMICA/mlx_impl/core.py) is a v1 MVP: single-model, generalized-Gaussian, natural gradient. Hybrid design forced by MLX 0.32 (GPU float32 hot path + CPU-stream inv/slogdet hoisted to once/iter + host-side SciPy for the missing lgamma/digamma), carrying the Phase A ufp/y guard. Validated on real EEG: sufficient stats match the NumPy float64 reference to rtol ~1e-4, converged LL matches the PyTorch float32 backend to ~2e-6. Optional dependency (Apple-only), so CI skips the MLX tests (and mlx_impl is omitted from the coverage gate). Newton, the other PDF families, sharing, multi-model, and save/load are fast-follows. Whether MLX beats CPU/MPS is Phase B's question. Phase C of epic #74.- added 2 commits that reference this issue
on Jul 8, 2026
Context
Higher-ceiling Apple-GPU path of the epic. MLX (Apple's array framework) beats PyTorch
MPS by ~2-3x on the same hardware, is Apple-native with a real compiler / lazy graph, and
can fuse the per-block work that PyTorch MPS runs eager and largely unfused. GPU arrays are
float32-only, so this depends on Phase A.
Approach
AMICATorchNGper-block E/M-step to MLX arrays behind a backend switch, keepingthe PyTorch path as the float64 parity / CUDA backend.
float32 on GPU).
Acceptance
(matched LL within tolerance).
Dependencies
Blocked by Phase A (MLX GPU is float32-only). Research:
.context/mps_pathways.mdPathway C.