Is there a way to manually update the cuda required file in the db runtime? There are some rather annoying bugs still in TF 2.11 that have been fixed ...

Data Engineering

Join discussions on data engineering best practices, architectures, and optimization strategies within the Databricks Community. Exchange insights and solutions with fellow data engineers.

Is there a way to manually update the cuda required file in the db runtime?

There are some rather annoying bugs still in TF 2.11 that have been fixed in TF 2.12.

Sadly the latest DB runtime 13.1 (beta) only supports the older TF 2.11 even tho 2.12 was released a while already before the beta was released.

Nowhere i was able to find if it is possible to manually update the required cuda & cuDNN files.

For anyone curious, in training data augmentation with tf.keras.layers.RandomTranslation & tf.keras.layers.RandomZoom is extremely slow because of a bad implementation for GPU.

As far as i see manually updating only TF to 2.12 doesn't work as this lacks the parity with cuDNN and noting works anymore.

Model trains extremely slow after upgrade to tf 2.9 · Issue #56242 · tensorflow/tensorflow (github.c...