Introduction to the Tool

If the model and environment configurations are correct but accuracy issues still occur, you can use the Large Language Model Debug Tool for precise troubleshooting and root cause analysis, improving development efficiency, as shown in Table 1.

**Table 1** **Large language model debug tool**
Tool Name	Description
Bad case analysis tool (msit Bad Case)	Provides both script-based and command-line interaction modes for automated bad case analysis, enabling users to quickly locate faults. For details, see msit Bad Case Analysis Tool.
Data dump tool (msit llm dump)	Dumps intermediate data generated during ATB model inference. The dumped data is used for subsequent accuracy comparison. For details about how to use the dump function in ATB scenarios, see ATB Model Data Dumping. For details about how to use the dump function in PyTorch scenarios, see Accuracy Data Profiling in PyTorch.
LLM accuracy comparison tool (msit llm compare)	Provides one-click accuracy comparison function, enabling rapid whole-network accuracy comparison in inference scenarios. For details, see LLM Accuracy Comparison.
Accuracy pre-check tool (msit llm opcheck)	Provides single-operator accuracy pre-check capabilities for built-in operators in the ATB to check whether operator accuracy meets requirements. For details, see opcheck Single-Operator Accuracy Pre-Check User Guide.

Parent topic: Using MSIT for Troubleshooting