Unsloth is anotherfinetuning framework but it uses gradient checkpointing to significantly reduce memory requirements at train time.
Feature ‘cvt with.f32.BF16 ‘requires.target sm_80 or higher
As per this ticket, seems to be related to versions of torch and unsloth when run on older GPUs (e.g. T4s on Google Colab).
Solution is to install older versions of unsloth and torch:
pip install --upgrade --force-reinstall --no-cache-dir torch==2.2.0 triton \
--index-url https://download.pytorch.org/whl/cu121
pip install "unsloth[cu121-torch220] @ git+https://github.com/unslothai/unsloth.git"