Hello,
I am trying to utilize ESM as made available in the HuggingFace Colab Notebook. I do not have a local GPU to use, so have been attempting the only CPU mode. The first issue is in preparing the model and tokenizer where it states " If using GPU, use model.cuda() to transfer the model to GPU." and the code below is:
from transformers import AutoTokenizer, EsmForProteinFolding
To get around this, I simply removed the "model = model.cuda()" and the code ran fine. Next, down the line, I got to the Tokenization step which requires a GPU:
tokenized_input = tokenized_input.cuda()
Obviously when ran the error of "NO GPU" was displayed. So I went ahead and tried to run actual prediction lines:
import torch
with torch.no_grad():
output = model(tokenized_input)**
This displayed the following error:
RuntimeError: "LayerNormKernelImpl" not implemented for 'Half'
So how are you supposed to run this sheet without a GPU? Is there something in the tokenization stage I should alter to provide good inputs for the prediction model?
chAwater
commented
1 year ago
Hello, I am trying to utilize ESM as made available in the HuggingFace Colab Notebook. I do not have a local GPU to use, so have been attempting the only CPU mode. The first issue is in preparing the model and tokenizer where it states " If using GPU, use model.cuda() to transfer the model to GPU." and the code below is:
from transformers import AutoTokenizer, EsmForProteinFolding
tokenizer = AutoTokenizer.from_pretrained("facebook/esmfold_v1") model = EsmForProteinFolding.from_pretrained("facebook/esmfold_v1", low_cpu_mem_usage=True)
model = model.cuda()**
To get around this, I simply removed the "model = model.cuda()" and the code ran fine. Next, down the line, I got to the Tokenization step which requires a GPU:
tokenized_input = tokenized_input.cuda()
Obviously when ran the error of "NO GPU" was displayed. So I went ahead and tried to run actual prediction lines:
import torch
with torch.no_grad(): output = model(tokenized_input)**
This displayed the following error:
RuntimeError: "LayerNormKernelImpl" not implemented for 'Half'
So how are you supposed to run this sheet without a GPU? Is there something in the tokenization stage I should alter to provide good inputs for the prediction model?