GUI Description
Function
The RL tab page displays the pipeline of each phase in the RL process, enabling developers to fully understand the running status, identify issues, and improve RL performance.
GUI Display
- The Task Trace Timeline area is displayed only when control flow data collected using mstx is imported. For details about the mstx collection mode, see "Sampling and Parsing msprof_tx" in "Ascend PyTorch Profiler" of "Profiling Instructions."
- When importing profile data of Volcano Engine Reinforcement Learning for LLMs (VerL) and MindSpeed, keep their files in different folders. Do not mix them together during import.
The RL tab page consists of the parameter configuration area (area 1) and task trace timeline (area 2), as shown in Figure 1.
- Area 1: In the parameter configuration area, the Framework and Algorithm of the imported data are automatically identified and displayed. If the data of more than 16 cards is imported, the data may not be completely displayed on the RL tab page. You can click Refresh to parse all data and refresh the task trace timeline.
- Area 2: The Task Trace Timeline area displays the execution time of each task on each card. The horizontal coordinate is the timeline, and the vertical coordinate is the rank ID of each device. Different colors represent different tasks. Forward and backward micro batch marks are available in blue blocks, helping locate fine-grained performance issues during training.
You can zoom in or out and move the timeline by dragging the sliders at the right and bottom of the timeline, or by holding down Ctrl and scrolling the mouse wheel.
Parent topic: Reinforcement Learning (RL)