Using finetuned ESM2 model with lm-design

[martinez-zacharya]

Has anyone tried using a finetuned ESM2 model for use with lm-design? I can't seem to produce different sequences when using a finetuned model, they still match the pretrained ESM2 outputs.

The only change I've made to the repo's code is for the load_model function in struct_models.py where I simply drop in my finetuned ESM2 weights. I've confirmed that both the resulting embeddings and attentions differ for the same input sequence when comparing the pretrained vs finetuned ESM2.