Downloading the model for offline use

[daniloimparato]

Hello, everyone.

This is my first time with ESM/Huggingface models in general. I want to load ESM3 on a local server with no internet access. I have tried some general model loading/saving instructions, but no luck so far [1, 2, 3]. I have even tried copying my local cache to the server (~/.cache/huggingface/hub/models--EvolutionaryScale--esm3-sm-open-v1/). It seems ESM3 works a bit... differently? Would appreciate some guidance.

In other words, how can I change this line from the README to work with a local dir?

model: ESM3InferenceClient = ESM3.from_pretrained("esm3-open").to("cuda")

I will keep banging my head at the keyboard, but figured it wouldn't hurt asking this question here as well.

[PraljakReps]

One thing you can do is download model and save it on your server with internet server first...

from esm.models.esm3 import ESM3
import torch

client = ESM3.from_pretrained('esm3_sm_open_v1') # download model on local cache
torch.save( client, './esm3_sm_open_v1_full.pth')" # save model onto hard drive

Then on your server with no internet connection, do the following:

import torch

client = torch.load('./esm3_sm_open_v1_full.pth') # load model from hard drive

#######
# TEST #
#######

from esm.models.esm3 import ESM3
from esm.sdk.api import ESMProtein, SamplingConfig
from esm.utils.constants.models import ESM3_OPEN_SMALL

# Peptidase S1A, chymotrypsin family: https://www.ebi.ac.uk/interpro/structure/PDB/1utn/
protein = ESMProtein(
    sequence=(
        "FIFLALLGAAVAFPVDDDDKIVGGYTCGANTVPYQVSLNSGYHFCGGSLINSQWVVSAAHCYKSGIQVRLGEDNINVVEG"
        "NEQFISASKSIVHPSYNSNTLNNDIMLIKLKSAASLNSRVASISLPTSCASAGTQCLISGWGNTKSSGTSYPDVLKCLKAP"
        "ILSDSSCKSAYPGQITSNMFCAGYLEGGKDSCQGDSGGPVVCSGKLQGIVSWGSGCAQKNKPGVYTKVCNYVSWIKQTIASN"
    )
)
protein_tensor = client.encode(protein)

output = client.forward_and_sample(
    protein_tensor, SamplingConfig(return_per_residue_embeddings=True)
)
print(output.per_residue_embedding.shape)

[daniloimparato]

Life saver! Thank you very much, @PraljakReps. I can't try it right now, but no reason to think it wouldn't work.