Advanced Practical Deep Learning NLP: LLM Architecture and Fine-tuning in Practice

It seems that Google has recently released the latest version of Tensorflow in an incomplete state that has not been completely backtested with the existing source. I have solved the problem by downgrading tensorflow from the current 2.17.0 to the previous stable 2.15.0 version as follows. I have also updated the github source.

# Delete the current tensorflow version

!pip uninstall --y tensorflow

# tensorflow version downgrade

!pip install tensorflow==2.15.0

# Check if downgrade was done

import tensorflow as tf

print(tf.__version__)

# Install package matching the downgraded version of TensorFlow

!apt install --allow-change-held-packages libcudnn8=8.1.0.77-1+cuda11.2

# Install the compatible cuDNN version for CUDA 11.8

!apt install --allow-change-held-packages libcudnn8=8.1.0.77-1+cuda11.2

# Uninstall existing TensorFlow and related packages

!pip uninstall -y -q tensorflow keras tensorflow-estimator tensorflow-text

# Install the required protobuf version

!pip install protobuf~=3.20.3

# Install TensorFlow Datasets

!pip install -q tensorflow_datasets

# Install TensorFlow 2.15.0 and TensorFlow Text 2.15.0

!pip install -q -U tensorflow-text==2.15.0

By doing this, you can complete the training task in Colab within 1 hour based on T4 GPU.