DLA Model Quantization Issue: Reformatting CopyNode Fallback Error on Orin

Autonomous Vehicles DRIVE AGX Orin DRIVE AGX Orin General
1

Software Version
DRIVE OS 6.0.8.1

Target Operating System
Linu

Hardware Platform
DRIVE AGX Orin Developer Kit (not sure its number)

SDK Manager Version
other

Host Machine Version
native Ubuntu Linux 20.04 Host installed with DRIVE OS Docker Containers

Issue Description
I am working on INT8 quantization for a DLA model on the Orin platform and encountered some issues. Could you please help analyze the possible causes?

Case 1: Input = kDLA_HWC4, Output = kDLA_LINEAR

  • The output tensor is already aligned to 64 bytes.

  • During quantization, the following error is reported:

Default DLA is enabled but layer Reformatting CopyNode for Output Tensor 1 to ... is not supported on DLA and falling back to GPU is not enabled.

  • if I randomly remove a few output tensors from the model, the quantization can succeed without the above error.

  • Question: What could be the reason that removing some outputs bypasses the error?

Case 2: Input = kDLA_HWC4, Output = kCHW32

  • In this case, quantization usually succeeds even without removing any outputs.

  • However, there is still a probability of failure, reporting the same Reformatting CopyNode error.

  • Question: Could this intermittent behavior be due to caching issues in the quantization environment, or possibly interference from other users running quantization/inference tasks at the same time?

image

**
Error String**
Default DLA is enabled but layer Reformatting CopyNode for Output Tensor 1 to … is not supported on DLA and falling back to GPU is not enabled.

Logs

5

Dear @qiuwen ,
There is no update from you for a period, assuming this is not an issue anymore.
Hence, we are closing this topic. If need further support, please open a new one.
Thanks
Could you use verbose flag for more insights ?
Is it possible to share model and repro steps?