debug(mlx): root-cause Mac batched-decode bug to MLX core quantized/rope kernel (B>1, L=1) — full ablation - #134

Merged

cursor[bot] merged 5 commits into

mainfrom

AgentMemory/mlx-batched-decode-fix-2815

Jun 14, 2026

Commits on Jun 14, 2026

diag(mlx): per-layer batched-vs-serialized hidden diff to localize gemma-4 batch>1 decode divergence

cursoragent
and
FluffyAIcode
committed
235c8ac
diag(mlx): also capture per-layer INPUT to split embedding vs attention (layer-0 row>0 divergence)

cursoragent
and
FluffyAIcode
committed
b83ca66
diag(mlx): --kakeya-cache layer-diff (concat cache) to isolate in-place cache write vs SDPA

cursoragent
and
FluffyAIcode
committed
6b61c12
fix(mlx): --manual-sdpa candidate fix (manual batched matmul SDPA, works around mx.fast.scaled_dot_product_attention batch>1+GQA bug)

cursoragent
and
FluffyAIcode
committed
e04f579
debug(mlx): full root-cause chain — batch>1 decode bug is MLX core quantized/rope kernel at L=1 B>1 (ruled out rotation/cache/shared-kv/embed/mask/SDPA); not Python-patchable, upstream MLX

cursoragent
and
FluffyAIcode
committed
c20e0e7