Question
How do we prove decode correctness on the constrained validation host? Weigh: porting the model repo's inference/model.py as a CPU reference (PR drumih#29 precedent: the gated-DeltaNet kernels were validated against a straight-line CPU reference of the mlx-vlm math), random-weight forward equivalence, comparison against a reference run on larger hardware, and real-model greedy identity across optimization changes. Decide the tolerance regime and which tests gate which phase of the port.
Question
How do we prove decode correctness on the constrained validation host? Weigh: porting the model repo's
inference/model.pyas a CPU reference (PR drumih#29 precedent: the gated-DeltaNet kernels were validated against a straight-line CPU reference of the mlx-vlm math), random-weight forward equivalence, comparison against a reference run on larger hardware, and real-model greedy identity across optimization changes. Decide the tolerance regime and which tests gate which phase of the port.