On JetPack 7.2 it is impossible to build or run any TensorRT engine on the DLA. Every attempt (both trtexec --useDLACore=0 --allowGPUFallback and via the API) fails with an error along the lines of:
Cannot create DLA engine, 0 not available
(DLACore 0 not available)
What indicates this is a runtime problem rather than a hardware/driver issue:
The hardware and kernel driver are present: dmesg shows nvdla0/nvdla1, the nodes /dev/nvhost-ctrl-nvdla0 and /dev/nvhost-ctrl-nvdla1 exist, and the user belongs to the video group.
What is missing, however, is the submit node / a complete DLA user-space runtime — TensorRT cannot initialize the DLA core despite the visible HW.
Expected behavior:
Building and executing an engine with --useDLACore=0 should work (with GPU fallback for unsupported layers), as on previous JetPack releases.
Regression compared to JetPack 6.2:
On the same class of hardware (AGX Orin) under JetPack 6.2 / TensorRT 10.3, the DLA runtime works correctly — the same trtexec --useDLACore=0 --allowGPUFallback builds and runs the engine without error. The problem only appears on JP7.2/R39.2.
Question: Is the DLA runtime for R39.2 incomplete/not yet shipped, and is a fix or a missing package planned (e.g. the submit node / DLA user-space libraries)?
JetPack 7.2 use SBSA unified driver and causes the DLA support to not be available by default.
We are working on adding DLA support in the future release but not have a concrete timeline currently.
Sorry, only for clarification.
Referring to Orin-NX DLA issue solution, … I searched on release notes some sentence declaring it solved, but I could not find it.
For example in link here below I cannot find details about DLA.
Sorry again, but for Orin NX we tried
-) installation of JP-7.2.1
-) installation of new Cuda-13.3.1/13.4
-) TRT-11…
but DLA engine cannot be generated …
so for example using below simple command
trtexec --onnx=./resnet34_peoplenet_int8_static_b1.onnx --useDLACore=0 --allowGPUFallback --adjustForDLA --saveEngine=./resnet34_peoplenet_int8_dla0_b1.engine 2>&1 | tee ./dla_trt11_final_cuda.log
we get a lot of below errors :
[09/30/2026-19:00:29] [W] [TRT] DLA Compilation failed in one of the configurations due to error code 8
[09/30/2026-19:00:29] [W] [TRT] {ForeignNode[block_2d_conv_1/convolution + Reshape__261:0 + ONNXTRT_Broadcast_25 + block_2d_conv_1/BiasAdd + block_2d_bn_1/batchnorm/mul__86 + block_2d_bn_1/batchnorm/mul_1 + block_2d_bn_1/batchnorm/sub__87 + block_2d_bn_1/batchnorm/add_1]} cannot be compiled by DLA, falling back to GPU.
[09/30/2026-19:00:29] [E] Error[2]: Error Code: 2: DLA Compile exception
could it be that for Orin-NX DLA support is working in progress ?
I found below 2 docs/url saying that DLA support is on-going.
Are those links contents correctly interpreted ?
Let’s know
thanks for support
M.
here details of links
https://docs.nvidia.com/deeplearning/tensorrt/latest/inference-library/work-with-dla.html
DLA is not supported in TensorRT 11.3.0 or in TensorRT 11.3.1 for DriveOS. TensorRT 10.7 was the last release that supported DLA. DLA was also not supported in TensorRT 11.0, 11.1, or 11.2.
https://docs.nvidia.com/metropolis/deepstream/dev-guide/text/DS_plugin_gst-nvinfer.html
enable-dla: Indicates whether to use the DLA engine for inferencing. Note: DLA is supported only on NVIDIA® Jetson AGX Orin™ and NVIDIA® Jetson Orin NX™. Currently work in progress.
use-dla-core: DLA core to be used. Note: Supported only on Jetson AGX Orin and Jetson Orin NX. Currently work in progress.
It’s first supported in the TensorRT standalone package release.
The future JetPack release will include TensorRT that supports DLA.
(Will share details after confirmation)
Based on your log, it’s more like a compilation issue.
Could you test DLA with a ResNet50 model first?
Install modelopt and update ResNet50 with ir_version=10: