SambaTune Reports

Use one of the following reports to understand how to improve model performance:

Host-device breakdown report. Learn to identify if the application is host-bound or device-bound.
Section report. Learn how to read generated data for sections.
Stage latency report. Learn how to read generated latency data for stages in a section.
Stack Tracing report. Trace a PyTorch operator or hyperfunction across different layers in the software stack.
Performance insights report (Beta). Learn about bottlenecks, including RDU versus Host, critical sections, and critital stages.
Hypertuner (Beta). Explore how to fine-tune model configuration by running the model with a combination of the model arguments and sweep arguments.
Check compilation and running logs. Learn about SambaTune’s detailed logs, which you can use for checking the high-level status, debugging errors, and reproducing substeps.