Open timotheecour4 opened 1 year ago
Lowest I managed to get was 32GB with batch size 1... but that was ages ago
Same, I was hoping there would be some flag configuration that would allow using a GPU (or multiple GPUs) with less memory
Same, I was hoping there would be some flag configuration that would allow using a GPU (or multiple GPUs) with less memory
Since it based on pytorch you might be able to implement some optimization like HF's accelerate or deepspeed for RAM offloading
All of the training scripts specified in the README give errors like the following:
CUDA Version: 11.7 using 8x Tesla V100-SXM2 (with 16GB memory)
reducing
--batch_size 32
didn't help passing--microbatch 1
didn't help