What and Where Information Is Logged

1. NPU Initialization Configurations and Initialization Success

The log shows the NPU initialization configurations, including any update to npu.global_options (for details about the option, see Overview), and the initialization success message. As no argument is passed to the npu.open call, initialization is performed on NPU:0 by default.

2. Preprocessing H2D Thread Started and HDC Channel Created

The log is printed only when you use Dataset as the preprocessing pipeline and Iterator is passed to the function call.

In this log example, the TF Adapter starts the preprocessing H2D thread and creates an HDC channel named AnonymousIterator0. The channel name is the same as the value of shared_name of the Iterator in TF2.

3. Iteration Offload to NPU Detected and Training Started

If iteration offload is enabled, Graph xxx can loop will be determined as true. In this example, the number of training iterations (loop size) to be offloaded to the NPU is set to 32. As the log suggests, 32 asynchronous data transfers are started and 32 training requests are issued.

4. Training Progress

The log indicates the training process.

The following message indicates that 32 asynchronous data transfers are completed successfully. If any transfer error occurs, an error message will be printed.

5. Training Process Exited

After the training is complete, the process will exit, the created HDC data transfer thread will be destroyed, and Graph Engine will be switched off.