ethos-u-vela crash during int8x8 LSTM model optimization
Hello, I am trying to optimize a custom LSTM model using the Vela compiler. However the Vela compiler crashes without any error messages and result. The actual point of crash: https://gitlab.arm.com/artificial-intelligence/ethos-u/ethos-u-vela/-/blob/main/ethosu/vela/rewrite_graph.py?ref_type=heads#L80. I tried to increase recursion limit based on previous suggestions for other issues inside this repo, no success. Steps to reproduce:
- pip3 install ethos-u-vela (I have version 4.3.0)
- Place "LSTM_TEST_MODEL_600X3_int8x8.tflite" to current working directory
- Used vela parameters:
- accelerator-config ethos-u55-128
- memory-mode Shared_Sram
- optimise Performance
- recursion-limit 10000
- arena_cache_size=4000000
Expected result: Vela compiler creates optimized .tflite file with _vela suffix. Actual result: Python process crashes on Vela compiler step: https://gitlab.arm.com/artificial-intelligence/ethos-u/ethos-u-vela/-/blob/main/ethosu/vela/rewrite_graph.py?ref_type=heads#L80 Input model: LSTM_TEST_MODEL_600X3_int8x8.tflite Note: I tried to reduce model size several times. This one works well: ML_3696_MODEL_100X3_int8x8.tflite
Regards, Dmytro