AM67A: Network version mismatch

Network version - 0x20250630, Expected version - 0x20250429

Hi Andow,

I have ran into this many times.  The key to getting it right is the version of the .wic file you used for the EVM.  It looks like the SD card (or boot image) on your EVM is 10.1 and you are using 11.0 and 11.1 network artifacts.  Please use a 10.1 version to compile the models and this should go away.  Please see the following guide for compatible TIDL versions for each SDK.

SDK Version 11.01.xx.xx

• 11_01_06_00 - Default

These releases have been validated on:

• AM62A - PSDK LINUX 11.01.07.05
• J722S/TDA4AEN/AM67A - PSDK LINUX 11.01.00.03 / PSDK RTOS 11.01.00.04
• J721E/TDA4VM/AM68PA - PSDK LINUX 11.01.00.03 / PSDK RTOS 11.01.00.04
• J721S2/TDA4VL/AM68A - PSDK LINUX 11.01.00.03 / PSDK RTOS 11.01.00.04
• J784S4/TDA4VH/AM69A - PSDK LINUX 11.01.00.03 / PSDK RTOS 11.01.00.04
• AM62 - PSDK LINUX 11.01.05.03

SDK Version 11.00.xx.xx

• 11_01_07_00 - Patch with backward compatibility
• 11_01_05_00 - Patch with backward compatibility
• 11_00_08_00 - Patch with default compatibility
• 11_00_06_00 - Default

These releases have been validated on:

• AM62A - N/A
• J722S/TDA4AEN/AM67A - PSDK LINUX 11.00.00.08 / PSDK RTOS 11.00.00.06
• J721E/TDA4VM/AM68PA - PSDK LINUX 11.00.00.08 / PSDK RTOS 11.00.00.06
• J721S2/TDA4VL/AM68A - PSDK LINUX 11.00.00.08 / PSDK RTOS 11.00.00.06
• J784S4/TDA4VH/AM69A - PSDK LINUX 11.00.00.08 / PSDK RTOS 11.00.00.06
• AM62 - N/A

SDK Version 10.01.xx.xx

• 11_00_07_00 - Patch with backward compatibility
• 10_01_04_00 - Patch with default compatibility
• 10_01_00_02 - Default

These releases have been validated on:

• AM62A - PSDK LINUX 10.01.00.05
• J722S/TDA4AEN/AM67A - PSDK LINUX 10.01.00.04 / PSDK RTOS 10.01.00.04
• J721E/TDA4VM/AM68PA - PSDK LINUX 10.01.00.04 / PSDK RTOS 10.01.00.04
• J721S2/TDA4VL/AM68A - PSDK LINUX 10.01.00.04 / PSDK RTOS 10.01.00.04
• J784S4/TDA4VH/AM69A - PSDK LINUX 10.01.00.05 / PSDK RTOS 10.01.00.04
• AM62 - PSDK LINUX 10.01.00.05 / 10.01.10.04

SDK Version 10.00.xx.xx

• 10_01_03_00 - Patch with backward compatibility
• 10_00_08_00 - Patch with default compatibility
• 10_00_06_00 - Patch with default compatibility
• 10_00_04_00 - Patch with default compatibility
• 10_00_02_00 - Default

These releases have been validated on:

• AM62A - PSDK LINUX 10.00.00.08
• J722S/TDA4AEN/AM67A - PSDK LINUX 10.00.00.08 / PSDK RTOS 10.00.00.05
• J721E/TDA4VM/AM68PA - PSDK LINUX 10.00.00.08 / PSDK RTOS 10.00.00.05
• J721S2/TDA4VL/AM68A - PSDK LINUX 10.00.00.08 / PSDK RTOS 10.00.00.05
• J784S4/TDA4VH/AM69A - PSDK LINUX 10.00.00.08 / PSDK RTOS 10.00.00.05
• AM62 - PSDK LINUX 10.00.07.04

Regards,

Chris

Chris, thank you for your reply.
I’m building my image based on the script for the beagleY-AI AM67A (j722s), and I’m using the rootfs and bootfs from ti-processor-sdk-linux-edgeai-j722s-evm-11_00_00_08-Linux-x86-Install.bin.
Here is how the script works:
https://gitlab.com/beagle-edge-ai/edge-ai-image-builder/-/blob/main/setup-sdk.sh?ref_type=heads

Inside the image, the models work, but when I use download_models.sh from the model zoo, the downloaded models give a version mismatch error.

PYTHONPATH=/usr/lib/python3.12/site-packages/

MODEL_ZOO_PATH=/opt/model_zoo

EDGEAI_SDK_VERSION=11_00_00

EDGEAI_VERSION=11.0

SOC=j722s

Hi Andow,

The model zoo models are probably compiled with the latest, or at least newer, version of TIDL.  Just recompile the model to generate artifacts for your TIDL version and they should work.

Regards,

Chris

I also just rechecked it with version 10.1 model_zoo, and here is the result:

What needs to be done to rebuild the model for the correct TIDL version?
Previously, I used edgeai-tensorlab and edgeai-benchmark. By selecting the required branch version (for example, r11.0) and using run_benchmarks_pc.sh, I would get compiled models, but they didn’t work on my board due to the version mismatch.

Hi Andow,

The easiest way to run is with OSRT.  Here is an example. 

1. cd to edgeai-tidl-tools/examples/osrt_python/ort

Compile:

2. python3 ./onnxrt_ep.py -c -m cl-ort-resnet18-v1

Run inference:

3. python3 ./onnxrt_ep.py -m cl-ort-resnet18-v1

This will download, compile (2), and run (3)  resnet18_opset9.onnx.  You can compile other models adding a configuration to model_configs.py (in edgeai-tidl-tools/examples/osrt_python).  For example, 

"my_model": create_model_config(
task_type="classification",
source=dict(
model_url="",
infer_shape=True,
),
preprocess=dict(
resize=256,
crop=224,
data_layout="NCHW",
resize_with_pad=False,
reverse_channels=False,
),
session=dict(
session_name="onnxrt",
model_path=os.path.join(models_base_path, "my_model.onnx"),
input_mean=[123.675, 116.28, 103.53],
input_scale=[0.017125, 0.017507, 0.017429],
input_optimization=True,
),
postprocess=dict(),
extra_info=dict(num_images=numImages, num_classes=1000),
),

Copy the above text into its own section in model_configs.py.  Change the red sections to you model's name (.onnx) and my_model can be anything.  Just my_model is a key in the dictionary so you will need to remember that when compiling and run.  Place your new models in edgeai-tidl-tools/models/public.  The artifacts will be in edgeai-tidl-tools/model-artifacts/<my_model>/artifacts.  Any images or binaries produces will be in edgeai-tidl-tools/output_binaries and edgeai-tidl-tools/output_images.

Once the configuration is complete, compile and  your model by 

python3 ./onnxrt_ep.py -c -m my_model

python3 ./onnxrt_ep.py  -m my_model

Regards,

Chris

Hi, thanks for the explanation.
I managed to compile the model for different TIDL versions, but the error still remains.
Here is what I get for two different TIDL versions:

rel_11_00 - VX_ZONE_ERROR: [tivxKernelTIDLCreate:964] Network version - 0x20250630, Expected version - 0x20250429
rel_10_01 - VX_ZONE_ERROR: [tivxKernelTIDLCreate:964] Network version - 0x20241120, Expected version - 0x20250429

edgeai-benchmark and modelmaker give me the same result. Chris Tsongas Any ideas?

Hi Andow,

I talked this over with others and I think you have a back ported SD card in your EVM.  Please get the 11.0 wic file installed on an SD card and boot your EVM with it.  

Please get the .wic image from here:

https://www.ti.com/tool/download/PROCESSOR-SDK-LINUX-AM67A/11.00.00.08

and use balena etcher or DD to flash the card.  Then try the 11.0 built artifacts on the EVM.

Regards,

Chris

Hi,
How is the .wic image different from the rootfs from
EDGE_AI_SDK_URL=dr-download.ti.com/.../ti-processor-sdk-linux-edgeai-j722s-evm-11_00_00_08-Linux-x86-Install.bin
SDK_RTOS_URL="">dr-download.ti.com/.../ti-processor-sdk-rtos-j722s-evm-11_00_00_06.tar.gz"
?

I don’t think this works for me because the BeagleY-AI uses slightly different memory addressing than the EVM, so I need to make changes to the device tree.

Hi Andow,

We do not know much about the Beagle.  The basic problem the SD card network version is 0x20250429 and the standard EVM TIDL releases associated with that do not match.   you can try this, try doing a git checkout <version> -b <version> and running "source setup.sh" with that version.  Where version is probably in this range.   

10_01_03_00
10_01_04_00
11_00_06_00
11_00_07_00 (try this last)

I cannot be sure on a Beagle as we only best effort support Beagle device and do not have any locally.

Regards,

Chris

I understand that Beagle is not exactly the same as the EVM, but at the moment our company is very interested in how it happens that the models that come with SDK 11.00.00.08 work, while the ones you build yourself do not.
Does 0x20250429 really not indicate the date 2025.04.29? From this, it follows that the version is definitely not lower than 10_01.
On our project, we’ve already been trying to figure this out for more than three weeks, and it’s starting to get a bit frustrating.

Chris Tsongas said:

git checkout <version> -b <version> and running "source setup.sh"

Which repository are you referring to?

Thank you, that worked! I switched to version 11_00_06_00, and the generated model cl-ort-resnet18-v1 ran successfully on the board. Now I’ll try to port my own model. Thanks again!