What problem does it solve? When a PyPTO operator produces incorrect numerical results on Ascend NPU, developers struggle to pinpoint which computation step diverges from the golden reference. This Skill provides a systematic checkpoint-based comparison workflow that saves intermediate tensors from both kernel and golden implementations, then compares them file-by-file to locate the first failing operation. ## Core Features & Use Cases - Checkpoint Insertion: Uses pypto.pass_verify_save() in the kernel and torch.save() in the golden function to dump intermediate results, with conditional saving (cond=(idx == batch_size - 1)) for loop scenarios. - Automated Comparison Tool: The compare_accuracy.py script auto-detects output directories, matches jit .data files with golden .pt files, applies dtype-specific tolerances (FP16, FP32, BF16, FP8, INT types), and reports mismatch rates with bisection suggestions. - Precision Visualization: The plot_accuracy.py script parses verification logs and plots rtol/atol trends across checkpoints as PNG charts. - Use Case: While debugging a fused BatchMatmulMaxSum operator, insert checkpoints after the matmul, max-reduce, and sum-reduce stages, run the test, then run the comparison tool to discover the max-reduce stage is the first divergence point. ## Quick Start Ask the AI to insert pass_verify_save checkpoints into your PyPTO kernel and matching torch.save calls into the golden function, then run the compare_accuracy.py script to locate the first precision mismatch.
