What problem does it solve? Porting generative image models from PyTorch/diffusers to MLX often produces subtle numerical mismatches that are hard to locate. This Skill provides a repeatable export-then-compare workflow to prove numerical parity or isolate the exact layer where outputs diverge. ## Core Features & Use Cases - Export-then-compare workflow: Dump deterministic tensors from the PyTorch reference, load them in MLX, and compare with explicit rtol/atol thresholds. - RNG parity handling: Export exact initial latents from the reference and inject them into the MLX denoising loop, avoiding false mismatches from differing random generators. - Divergence checklist: Covers common causes such as NCHW vs NHWC layouts, broadcasting errors, dtype casting, RoPE details, and scheduler math. - Use Case: When an MLX-ported transformer produces different images than diffusers, export per-block hidden states from both sides, binary-search the forward path, and pinpoint the first diverging checkpoint. ## Quick Start Ask the AI to debug why the MLX model output diverges from the diffusers reference by exporting and comparing intermediate tensors.
