8 ms·
The 80 GB are for training with a batch size of 32 times 2048 tokens each. Since the model has only about 560M parameters, you could probably run it on CPU, if
by yorwba 1y ago
The 80 GB are for training with a batch size of 32 times 2048 tokens each. Since the model has only about 560M parameters, you could probably run it on CPU, if a bit slow.