Pipeline Optimization

Enablement

You can set the following environment variables to enable this feature, which is applicable to network scenarios with severe host bound problems.

export TASK_QUEUE_ENABLE=2

Working Principles

The task_queue operator delivery queue supports three configuration levels (the add operator is used as an example). You can configure the queue as required. For details, see "TASK_QUEUE_ENABLE" in Ascend Extension for PyTorch Environment Variable Reference.

Precautions

• When ASCEND_LAUNCH_BLOCKING is set to 1, the task_queue operator queue is forcibly closed. In this case, the TASK_QUEUE_ENABLE configuration becomes invalid. For details about the ASCEND_LAUNCH_BLOCKING configuration, see "ASCEND_LAUNCH_BLOCKING" in Ascend Extension for PyTorch Environment Variable Reference.

• When TASK_QUEUE_ENABLE is set to 2, the peak NPU memory usage may increase due to the increase of concurrent memory access.