Procedure

Model tuning can be divided into two steps: performance collection and performance analysis.

1. Performance logs collection: You are advised to analyze profiling logs first. For details about how to use the tool, see Model Tuning Performance Collection Tools.
   

   • When collecting performance data for the first time, you are advised to collect only L1 data and do not enable the stack (that is, with_stack is set to False). You can set warmup to 1 and active to 2 to collect data of two steps.
   • If competing product analysis is involved, you are advised to collect performance counters of competing products under the same conditions (including but not limited to the number of collected steps, hyperparameter settings, and used data).
2. Performance bottleneck analysis: After profiling logs are collected, use the performance analysis tool to analyze bottlenecks.
   1. The msprof-analyze tool performs preliminary analysis, identifies performance issues at a fine-grained level, and provides clear direction for further in-depth analysis. For details, see Quick Analysis for Model Tuning (msprof-analyze CLI).
   2. The MindStudio Insight tool identifies bottlenecks and analyzes root causes. For details, see In-depth Analysis for Model Tuning (MindStudio Insight).