Closed guidomeijer closed 2 years ago
Hey there I had the same problem today, also using a v3-8 TPU vm instance . Maybe not the most satisfying solution, but I did solve this by passing the --cpu flag. My finetuned model completed in 6 min when using the cpu.
Not entirely sure why cpu is not default, but afaik there is not any particular reason to not do the conversion on GPU/TPU
I compressed the weights of a GPT-J-6B model using the
slim_model
script and am now trying to convert them into pycharm hugging face format. I'm running this on a v3-8 TPU and so far everything worked fine. However, when running theto_hf_weights
script I get an out of memory error which I find quite odd since the weights are already compressed and I'm running on a big TPU.jax==0.2.12 jaxlib==0.1.67 (I tried 0.1.68 but getting errors) Any help would be greatly appreciated :)