Tool Navigation

Ascend has developed dedicated accuracy debugging tools for both large language model (LLM) and traditional small model scenarios. Refer to the following guide to select the appropriate tool and documentation:

• LLM Scenarios
  • Training: Use msprobe, an accuracy debugging tool for LLM training on PyTorch and MindSpore frameworks. It is primarily released on Gitcode; for details, see the msprobe User Guide.
  • Inference: Use the Large Language Model Debugging Tool for LLM inference. It is primarily released on Gitcode; for details, see Large Language Model Debugging Tool.
• Traditional Small Model Scenarios
  • This document focuses on msaccucmp.py, an accuracy comparison tool for traditional small models. It supports comparison before and after ATC model conversion (ONNX, Caffe, and TensorFlow), TensorFlow training comparisons, and comparisons between different offline model versions. In addition, the tool supports expert recommendation output for comparison results, accuracy comparison between .npy files, and single-operator comparison.
  • For traditional small models, you can also use msprobe for training and msit debug for inference scenarios.