[ML]Unsupported ops error while using RKNN.

[yoni13]

I have searched the existing issues, both open and closed, to make sure this is not a duplicate report.

• Yes

The bug

While trying to use XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k on RK3566

E RKNN: [00:39:42.701] Unsupport CPU op: CumSum in this librknnrt.so, please try to register custom op by calling rknn_register_custom_ops or If using rknn, update to the latest toolkit2 and runtime from: https://console.zbox.filez.com/l/I00fc3 (PWD: rknn). If using rknn-llm, update from: https://github.com/airockchip/rknn-llm

We can either dont ship these models and throw error while exporting or add a custom op

Custom OP Example

The OS that Immich Server is running on

Debian GNU/Linux 12 (bookworm) aarch64

Version of Immich Server

related to ML

Version of Immich Mobile App

related to ML

Platform with the issue

• Server
• Web
• Mobile

Your docker-compose.yml content

<default with rknn enabled>

Your .env content

<default with TZ>

Reproduction steps

1.Use RKNN for ML
2.Change Model to XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k
3. Rerun smart search
4. Search something in search bar
...

Relevant log output

E RKNN: [00:39:42.701] Unsupport CPU op: CumSum in this librknnrt.so, please try to register custom op by calling rknn_register_custom_ops or If using rknn, update to the latest toolkit2 and runtime from: https://console.zbox.filez.com/l/I00fc3 (PWD: rknn). If using rknn-llm, update from: https://github.com/airockchip/rknn-llm

Additional information

No response

[yoni13]

I've tried to add a custom op, but i can't verify it now due to my poor internet connections
yoni13@2d7e670
EDIT: Something is wrong, dont use it

[raptr55]

same here with RK3588:

Initializing Immich ML v1.130.2
[03/27/25 08:46:10] INFO     Starting gunicorn 23.0.0                           
[03/27/25 08:46:10] INFO     Listening at: http://[::]:3003 (9)                 
[03/27/25 08:46:10] INFO     Using worker: app.config.CustomUvicornWorker       
[03/27/25 08:46:10] INFO     Booting worker with pid: 10                        
[03/27/25 08:46:13] I        Started server process [10]                        
[03/27/25 08:46:13] I        Waiting for application startup.                   
[03/27/25 08:46:13] I        Created in-memory cache with unloading after 300s  
                             of inactivity.                                     
[03/27/25 08:46:13] I        Initialized request thread pool with 8 threads.    
[03/27/25 08:46:13] I        Application startup complete.                      
[03/27/25 08:49:31] I        Downloading detection model 'buffalo_l'. This may  
                             take a while.                                      

Fetching 12 files:   0%|          | 0/12 [00:00<?, ?it/s]
Fetching 12 files:   8%|▊         | 1/12 [00:00<00:05,  1.99it/s]
Fetching 12 files:  33%|███▎      | 4/12 [00:00<00:01,  5.58it/s]
Fetching 12 files:  67%|██████▋   | 8/12 [00:05<00:03,  1.17it/s]
Fetching 12 files: 100%|██████████| 12/12 [00:05<00:00,  2.01it/s]
[03/27/25 08:49:37] I        Loading detection model 'buffalo_l' to memory      
[03/27/25 08:49:37] I        Loading RKNN model from                            
                             /cache/facial-recognition/buffalo_l/detection/rknpu
                             /rk3588/model.rknn with 3 threads.                 
I RKNN: [08:49:37.613] RKNN Runtime Information, librknnrt version: 2.3.0 (c949ad889d@2024-11-07T11:35:33)
I RKNN: [08:49:37.613] RKNN Driver Information, version: 0.9.8
I RKNN: [08:49:37.613] RKNN Model Information, version: 6, toolkit version: 2.3.0(compiler version: 2.3.0 (c949ad889d@2024-11-07T11:39:30)), target: RKNPU v2, target platform: rk3588, framework name: ONNX, framework layout: NCHW, model inference type: static_shape
W RKNN: [08:49:37.656] query RKNN_QUERY_INPUT_DYNAMIC_RANGE error, rknn model is static shape type, please export rknn with dynamic_shapes
I RKNN: [08:49:37.703] RKNN Runtime Information, librknnrt version: 2.3.0 (c949ad889d@2024-11-07T11:35:33)
I RKNN: [08:49:37.703] RKNN Driver Information, version: 0.9.8
I RKNN: [08:49:37.704] RKNN Model Information, version: 6, toolkit version: 2.3.0(compiler version: 2.3.0 (c949ad889d@2024-11-07T11:39:30)), target: RKNPU v2, target platform: rk3588, framework name: ONNX, framework layout: NCHW, model inference type: static_shape
W RKNN: [08:49:37.740] query RKNN_QUERY_INPUT_DYNAMIC_RANGE error, rknn model is static shape type, please export rknn with dynamic_shapes
I RKNN: [08:49:37.786] RKNN Runtime Information, librknnrt version: 2.3.0 (c949ad889d@2024-11-07T11:35:33)
I RKNN: [08:49:37.786] RKNN Driver Information, version: 0.9.8
I RKNN: [08:49:37.787] RKNN Model Information, version: 6, toolkit version: 2.3.0(compiler version: 2.3.0 (c949ad889d@2024-11-07T11:39:30)), target: RKNPU v2, target platform: rk3588, framework name: ONNX, framework layout: NCHW, model inference type: static_shape
W RKNN: [08:49:37.824] query RKNN_QUERY_INPUT_DYNAMIC_RANGE error, rknn model is static shape type, please export rknn with dynamic_shapes
[03/27/25 08:49:37] I        Loaded RKNN model from                             
                             /cache/facial-recognition/buffalo_l/detection/rknpu
                             /rk3588/model.rknn with 3 threads.                 
warning: det_size is already set in detection model, ignore
[03/27/25 08:49:37] I        Downloading visual model                           
                             'XLM-Roberta-Large-Vit-B-16Plus'. This may take a  
                             while.                                             

Fetching 189 files:   0%|          | 0/189 [00:00<?, ?it/s]
Fetching 189 files:   1%|          | 1/189 [00:00<01:36,  1.95it/s]
Fetching 189 files:   2%|▏         | 4/189 [00:00<00:25,  7.36it/s]
Fetching 189 files:   5%|▍         | 9/189 [00:00<00:15, 11.80it/s]
Fetching 189 files:   6%|▌         | 11/189 [00:01<00:21,  8.18it/s]
Fetching 189 files:   7%|▋         | 13/189 [00:02<00:36,  4.77it/s]
Fetching 189 files:  12%|█▏        | 22/189 [00:02<00:14, 11.16it/s]
Fetching 189 files:  15%|█▍        | 28/189 [00:03<00:16,  9.95it/s]
Fetching 189 files:  18%|█▊        | 34/189 [00:03<00:12, 12.56it/s]
Fetching 189 files:  21%|██        | 40/189 [00:03<00:11, 12.78it/s]
Fetching 189 files:  24%|██▍       | 46/189 [00:05<00:20,  7.07it/s]
Fetching 189 files:  33%|███▎      | 63/189 [00:05<00:08, 15.25it/s]
Fetching 189 files:  36%|███▌      | 68/189 [00:06<00:10, 11.80it/s]
Fetching 189 files:  38%|███▊      | 72/189 [00:06<00:10, 10.99it/s]
Fetching 189 files:  43%|████▎     | 82/189 [00:07<00:10, 10.53it/s]
Fetching 189 files:  47%|████▋     | 88/189 [00:08<00:09, 11.17it/s]
Fetching 189 files:  52%|█████▏    | 99/189 [00:08<00:06, 14.11it/s]
Fetching 189 files:  54%|█████▍    | 102/189 [00:09<00:07, 12.09it/s]
Fetching 189 files:  56%|█████▌    | 106/189 [00:09<00:06, 13.82it/s]
Fetching 189 files:  59%|█████▉    | 112/189 [00:09<00:05, 13.49it/s]
Fetching 189 files:  60%|██████    | 114/189 [00:10<00:06, 12.35it/s]
Fetching 189 files:  63%|██████▎   | 119/189 [00:11<00:08,  8.17it/s]
Fetching 189 files:  69%|██████▉   | 130/189 [00:11<00:04, 13.18it/s]
Fetching 189 files:  70%|██████▉   | 132/189 [00:11<00:04, 11.59it/s]
Fetching 189 files:  72%|███████▏  | 136/189 [00:12<00:04, 12.12it/s]
Fetching 189 files:  73%|███████▎  | 138/189 [00:12<00:04, 11.04it/s]
Fetching 189 files:  78%|███████▊  | 148/189 [00:13<00:03, 12.09it/s]
Fetching 189 files:  78%|███████▊  | 148/189 [00:23<00:03, 12.09it/s]
Fetching 189 files:  79%|███████▉  | 150/189 [00:37<01:03,  1.62s/it]
Fetching 189 files:  80%|███████▉  | 151/189 [01:21<03:16,  5.18s/it]
Fetching 189 files: 100%|██████████| 189/189 [01:21<00:00,  2.32it/s]
[03/27/25 08:50:59] I        Loading visual model                               
                             'XLM-Roberta-Large-Vit-B-16Plus' to memory         
[03/27/25 08:50:59] I        Loading RKNN model from                            
                             /cache/clip/XLM-Roberta-Large-Vit-B-16Plus/visual/r
                             knpu/rk3588/model.rknn with 3 threads.             
I RKNN: [08:50:59.911] RKNN Runtime Information, librknnrt version: 2.3.0 (c949ad889d@2024-11-07T11:35:33)
I RKNN: [08:50:59.911] RKNN Driver Information, version: 0.9.8
I RKNN: [08:50:59.916] RKNN Model Information, version: 6, toolkit version: 2.3.0(compiler version: 2.3.0 (c949ad889d@2024-11-07T11:39:30)), target: RKNPU v2, target platform: rk3588, framework name: ONNX, framework layout: NCHW, model inference type: static_shape
W RKNN: [08:51:00.197] query RKNN_QUERY_INPUT_DYNAMIC_RANGE error, rknn model is static shape type, please export rknn with dynamic_shapes
I RKNN: [08:51:00.355] RKNN Runtime Information, librknnrt version: 2.3.0 (c949ad889d@2024-11-07T11:35:33)
I RKNN: [08:51:00.355] RKNN Driver Information, version: 0.9.8
I RKNN: [08:51:00.355] RKNN Model Information, version: 6, toolkit version: 2.3.0(compiler version: 2.3.0 (c949ad889d@2024-11-07T11:39:30)), target: RKNPU v2, target platform: rk3588, framework name: ONNX, framework layout: NCHW, model inference type: static_shape
W RKNN: [08:51:00.620] query RKNN_QUERY_INPUT_DYNAMIC_RANGE error, rknn model is static shape type, please export rknn with dynamic_shapes
I RKNN: [08:51:00.825] RKNN Runtime Information, librknnrt version: 2.3.0 (c949ad889d@2024-11-07T11:35:33)
I RKNN: [08:51:00.825] RKNN Driver Information, version: 0.9.8
I RKNN: [08:51:00.825] RKNN Model Information, version: 6, toolkit version: 2.3.0(compiler version: 2.3.0 (c949ad889d@2024-11-07T11:39:30)), target: RKNPU v2, target platform: rk3588, framework name: ONNX, framework layout: NCHW, model inference type: static_shape
W RKNN: [08:51:01.198] query RKNN_QUERY_INPUT_DYNAMIC_RANGE error, rknn model is static shape type, please export rknn with dynamic_shapes
[03/27/25 08:51:01] I        Loaded RKNN model from                             
                             /cache/clip/XLM-Roberta-Large-Vit-B-16Plus/visual/r
                             knpu/rk3588/model.rknn with 3 threads.             
[03/27/25 08:51:01] I        Loading recognition model 'buffalo_l' to memory    
[03/27/25 08:51:01] I        Loading RKNN model from                            
                             /cache/facial-recognition/buffalo_l/recognition/rkn
                             pu/rk3588/model.rknn with 3 threads.               
I RKNN: [08:51:01.285] RKNN Runtime Information, librknnrt version: 2.3.0 (c949ad889d@2024-11-07T11:35:33)
I RKNN: [08:51:01.285] RKNN Driver Information, version: 0.9.8
I RKNN: [08:51:01.286] RKNN Model Information, version: 6, toolkit version: 2.3.0(compiler version: 2.3.0 (c949ad889d@2024-11-07T11:39:30)), target: RKNPU v2, target platform: rk3588, framework name: ONNX, framework layout: NCHW, model inference type: static_shape
W RKNN: [08:51:01.371] query RKNN_QUERY_INPUT_DYNAMIC_RANGE error, rknn model is static shape type, please export rknn with dynamic_shapes
I RKNN: [08:51:01.451] RKNN Runtime Information, librknnrt version: 2.3.0 (c949ad889d@2024-11-07T11:35:33)
I RKNN: [08:51:01.451] RKNN Driver Information, version: 0.9.8
I RKNN: [08:51:01.452] RKNN Model Information, version: 6, toolkit version: 2.3.0(compiler version: 2.3.0 (c949ad889d@2024-11-07T11:39:30)), target: RKNPU v2, target platform: rk3588, framework name: ONNX, framework layout: NCHW, model inference type: static_shape
W RKNN: [08:51:01.530] query RKNN_QUERY_INPUT_DYNAMIC_RANGE error, rknn model is static shape type, please export rknn with dynamic_shapes
I RKNN: [08:51:01.606] RKNN Runtime Information, librknnrt version: 2.3.0 (c949ad889d@2024-11-07T11:35:33)
I RKNN: [08:51:01.606] RKNN Driver Information, version: 0.9.8
I RKNN: [08:51:01.606] RKNN Model Information, version: 6, toolkit version: 2.3.0(compiler version: 2.3.0 (c949ad889d@2024-11-07T11:39:30)), target: RKNPU v2, target platform: rk3588, framework name: ONNX, framework layout: NCHW, model inference type: static_shape
W RKNN: [08:51:01.689] query RKNN_QUERY_INPUT_DYNAMIC_RANGE error, rknn model is static shape type, please export rknn with dynamic_shapes
[03/27/25 08:51:01] I        Loaded RKNN model from                             
                             /cache/facial-recognition/buffalo_l/recognition/rkn
                             pu/rk3588/model.rknn with 3 threads.               
[03/27/25 08:54:33] I        Shutting down due to inactivity.                   
[03/27/25 08:54:33] I        Shutting down                                      
[03/27/25 08:54:33] I        Waiting for application shutdown.                  
[03/27/25 08:54:33] I        Application shutdown complete.                     
[03/27/25 08:54:33] I        Finished server process [10]                       
[03/27/25 08:54:33] ERROR    Worker (pid:10) was sent SIGINT!                   
[03/27/25 08:54:34] INFO     Booting worker with pid: 189                       
[03/27/25 08:54:37] I        Started server process [189]                       
[03/27/25 08:54:37] I        Waiting for application startup.                   
[03/27/25 08:54:37] I        Created in-memory cache with unloading after 300s  
                             of inactivity.                                     
[03/27/25 08:54:37] I        Initialized request thread pool with 8 threads.    
[03/27/25 08:54:37] I        Application startup complete.                      
[03/27/25 08:54:38] I        Loading textual model                              
                             'XLM-Roberta-Large-Vit-B-16Plus' to memory         
[03/27/25 08:54:38] I        Loading RKNN model from                            
                             /cache/clip/XLM-Roberta-Large-Vit-B-16Plus/textual/
                             rknpu/rk3588/model.rknn with 3 threads.            
I RKNN: [08:54:39.002] RKNN Runtime Information, librknnrt version: 2.3.0 (c949ad889d@2024-11-07T11:35:33)
I RKNN: [08:54:39.002] RKNN Driver Information, version: 0.9.8
I RKNN: [08:54:40.003] RKNN Model Information, version: 6, toolkit version: 2.3.0(compiler version: 2.3.0 (c949ad889d@2024-11-07T11:39:30)), target: RKNPU v2, target platform: rk3588, framework name: ONNX, framework layout: NCHW, model inference type: static_shape
E RKNN: [08:54:40.986] Unsupport CPU op: CumSum in this librknnrt.so, please try to register custom op by callingrknn_register_custom_ops or If using rknn, update to the latest toolkit2 and runtime from: https://console.zbox.filez.com/l/I00fc3 (PWD: rknn). If using rknn-llm, update from: https://github.com/airockchip/rknn-llm
W RKNN: [08:54:40.987] query RKNN_QUERY_INPUT_DYNAMIC_RANGE error, rknn model is static shape type, please export rknn with dynamic_shapes
I RKNN: [08:54:43.008] RKNN Runtime Information, librknnrt version: 2.3.0 (c949ad889d@2024-11-07T11:35:33)
I RKNN: [08:54:43.008] RKNN Driver Information, version: 0.9.8
I RKNN: [08:54:43.008] RKNN Model Information, version: 6, toolkit version: 2.3.0(compiler version: 2.3.0 (c949ad889d@2024-11-07T11:39:30)), target: RKNPU v2, target platform: rk3588, framework name: ONNX, framework layout: NCHW, model inference type: static_shape
E RKNN: [08:54:44.072] Unsupport CPU op: CumSum in this librknnrt.so, please try to register custom op by callingrknn_register_custom_ops or If using rknn, update to the latest toolkit2 and runtime from: https://console.zbox.filez.com/l/I00fc3 (PWD: rknn). If using rknn-llm, update from: https://github.com/airockchip/rknn-llm
W RKNN: [08:54:44.072] query RKNN_QUERY_INPUT_DYNAMIC_RANGE error, rknn model is static shape type, please export rknn with dynamic_shapes
I RKNN: [08:54:46.369] RKNN Runtime Information, librknnrt version: 2.3.0 (c949ad889d@2024-11-07T11:35:33)
I RKNN: [08:54:46.369] RKNN Driver Information, version: 0.9.8
I RKNN: [08:54:46.371] RKNN Model Information, version: 6, toolkit version: 2.3.0(compiler version: 2.3.0 (c949ad889d@2024-11-07T11:39:30)), target: RKNPU v2, target platform: rk3588, framework name: ONNX, framework layout: NCHW, model inference type: static_shape
E RKNN: [08:54:47.568] Unsupport CPU op: CumSum in this librknnrt.so, please try to register custom op by callingrknn_register_custom_ops or If using rknn, update to the latest toolkit2 and runtime from: https://console.zbox.filez.com/l/I00fc3 (PWD: rknn). If using rknn-llm, update from: https://github.com/airockchip/rknn-llm
W RKNN: [08:54:47.571] query RKNN_QUERY_INPUT_DYNAMIC_RANGE error, rknn model is static shape type, please export rknn with dynamic_shapes
[03/27/25 08:54:47] I        Loaded RKNN model from                             
                             /cache/clip/XLM-Roberta-Large-Vit-B-16Plus/textual/
                             rknpu/rk3588/model.rknn with 3 threads.

[yoni13]

same here

Using XLM and nllb (haven't test this since refactor) might have issue since I forget to check ops after refactor , I should spot that out before releasing it, my fault 😔.

Using other models should work.

[raptr55]

can you suggest a multi language template that would work?

ps you are doing a great job! don't worry bugs happen!

[mertalev]

The SigLIP2 models are very good and should all work (assuming you have enough ram). You can check the docs to see which models are best for your languages. Use this link to check because the numbers were recently updated and aren't in the latest docs yet.

[raptr55]

i have changed the model with "ViT-L-16-SigLIP2-384__webli" and now i receive only warnings:

Initializing Immich ML v1.130.3
[03/28/25 08:37:54] INFO     Starting gunicorn 23.0.0                           
[03/28/25 08:37:54] INFO     Listening at: http://[::]:3003 (9)                 
[03/28/25 08:37:54] INFO     Using worker: app.config.CustomUvicornWorker       
[03/28/25 08:37:54] INFO     Booting worker with pid: 10                        
[03/28/25 08:37:58] I        Started server process [10]                        
[03/28/25 08:37:58] I        Waiting for application startup.                   
[03/28/25 08:37:58] I        Created in-memory cache with unloading after 300s  
                             of inactivity.                                     
[03/28/25 08:37:58] I        Initialized request thread pool with 8 threads.    
[03/28/25 08:37:58] I        Application startup complete.                      
[03/28/25 08:38:17] I        Loading visual model 'ViT-L-16-SigLIP2-384__webli' 
                             to memory                                          
[03/28/25 08:38:17] I        Loading RKNN model from                            
                             /cache/clip/ViT-L-16-SigLIP2-384__webli/visual/rknp
                             u/rk3588/model.rknn with 3 threads.                
I RKNN: [08:38:18.224] RKNN Runtime Information, librknnrt version: 2.3.0 (c949ad889d@2024-11-07T11:35:33)
I RKNN: [08:38:18.224] RKNN Driver Information, version: 0.9.8
I RKNN: [08:38:18.225] RKNN Model Information, version: 6, toolkit version: 2.3.0(compiler version: 2.3.0 (c949ad889d@2024-11-07T11:39:30)), target: RKNPU v2, target platform: rk3588, framework name: ONNX, framework layout: NCHW, model inference type: static_shape
W RKNN: [08:38:18.875] query RKNN_QUERY_INPUT_DYNAMIC_RANGE error, rknn model is static shape type, please export rknn with dynamic_shapes
I RKNN: [08:38:19.200] RKNN Runtime Information, librknnrt version: 2.3.0 (c949ad889d@2024-11-07T11:35:33)
I RKNN: [08:38:19.200] RKNN Driver Information, version: 0.9.8
I RKNN: [08:38:19.201] RKNN Model Information, version: 6, toolkit version: 2.3.0(compiler version: 2.3.0 (c949ad889d@2024-11-07T11:39:30)), target: RKNPU v2, target platform: rk3588, framework name: ONNX, framework layout: NCHW, model inference type: static_shape
W RKNN: [08:38:19.673] query RKNN_QUERY_INPUT_DYNAMIC_RANGE error, rknn model is static shape type, please export rknn with dynamic_shapes
I RKNN: [08:38:19.990] RKNN Runtime Information, librknnrt version: 2.3.0 (c949ad889d@2024-11-07T11:35:33)
I RKNN: [08:38:19.990] RKNN Driver Information, version: 0.9.8
I RKNN: [08:38:19.990] RKNN Model Information, version: 6, toolkit version: 2.3.0(compiler version: 2.3.0 (c949ad889d@2024-11-07T11:39:30)), target: RKNPU v2, target platform: rk3588, framework name: ONNX, framework layout: NCHW, model inference type: static_shape
W RKNN: [08:38:20.581] query RKNN_QUERY_INPUT_DYNAMIC_RANGE error, rknn model is static shape type, please export rknn with dynamic_shapes
[03/28/25 08:38:20] I        Loaded RKNN model from                             
                             /cache/clip/ViT-L-16-SigLIP2-384__webli/visual/rknp
                             u/rk3588/model.rknn with 3 threads.  

[raptr55]

Something strange happens. Every time I search for something my system misses an NVME disk... and the system crashes. I have to physically remove power to the server and restore it so that the system recognizes both disks!

I restored the docker-compose.yaml file so that it would use the CPU for machine learning tasks and as soon as it finishes parsing the whole library I try to do a search to see if the system crashes anyway.

immich is inside an openmediavault docker compose:

Host Name: openmediavault
Version: 7.7.2-1 (Sandworm)
Processor: (board CM3588 by friendlyelec.com) based on RK3588
Kernel: Linux 6.1.99
File system BTRFS:
Total devices 2 FS bytes used 229.81GiB
devid 1 size 953.87GiB used 268.03GiB path /dev/nvme0n1
devid 2 size 953.87GiB used 268.03GiB path /dev/nvme1n1

***** UPDATE *****
now searching in the search bar the problem does not occur.
thus seems to be related only to the new "rknn" service.

[yoni13]

Something strange happens. Every time I search for something my system misses an NVME disk... and the system crashes. I have to physically remove power to the server and restore it so that the system recognizes both disks!

Do you mean your ssd disconnected after using search with RKNN? That's really strange.
I suggest you to log the dmesg somewhere to see if it's related to RKNN or something.
And also open a new issue if needed.

[raptr55]

Something strange happens. Every time I search for something my system misses an NVME disk... and the system crashes. I have to physically remove power to the server and restore it so that the system recognizes both disks!

Do you mean your ssd disconnected after using search with RKNN? That's really strange. I suggest you to log the dmesg somewhere to see if it's related to RKNN or something. And also open a new issue if needed.

yeah... kinda odd.

the portion of dmesg i have is this:

[ 96.907096] BTRFS info (device nvme0n1): first mount of filesystem 8c0d426a-71f5-4764-92a3-b326334aeb92
[ 96.907182] BTRFS info (device nvme0n1): using crc32c (crc32c-generic) checksum algorithm
[ 96.907206] BTRFS info (device nvme0n1): enabling ssd optimizations
[ 96.907215] BTRFS info (device nvme0n1): using free space tree
[ 96.908282] BTRFS error (device nvme0n1): devid 2 uuid 99506759-8f4f-495f-80c4-38cfca213afc is missing
[ 96.908650] BTRFS error (device nvme0n1): failed to read the system array: -2
[ 96.909028] BTRFS error (device nvme0n1): open_ctree failed

log.txt

i'm sorry but i'm not very familiar with in linux systems so i don't know exactly how i can be helpful in identifying the cause. what i can provide is the list of actions to reproduce the problem on my device:

1. enable the hardware acceleration rknn as explained at the link
2. initialize the machine learning job in the web page
3. wait for completion
4. search for something in the search bar

if it is necessary to open a new bug tell me I will do it

[yoni13]

if it is necessary to open a new bug tell me I will do it

I think it needs its own thread

since i am not familiar with that FS, my suggestion is to run examples from rknn model zoo and see if it crashes again.

[yoni13]

The SigLIP2 models are very good and should all work (assuming you have enough ram). You can check the docs to see which models are best for your languages. Use this link to check because the numbers were recently updated and aren't in the latest docs yet.

@mertalev Sorry for inconvenience, can you help me pull down those not working model?

[mertalev]

Is there an easy way to validate the output RKNN model? I can update the exporter.

[yoni13]

Is there an easy way to validate the output RKNN model? I can update the exporter.

my script used to check if No lowering found for XXX is in the output, probably isn't the most ideal way.