Upload folder using huggingface_hub

README.md

@@ -1,10 +1,10 @@
 ---
 library_name: transformers
 tags:
-- biology
-- protein-structure
-- esmfold2
-- multimodal-protein-model
+  - biology
+  - protein-structure
+  - esmfold2
+  - multimodal-protein-model
 ---
 
 # FastPLMs ESMFold2

__init__.py

@@ -1,12 +1,4 @@
-import importlib
-import sys
-
 from .configuration_esmfold2 import ESMFold2Config
 from .modeling_esmfold2 import ESMFold2Model
 
-
-def ensure_vendored_esm() -> None:
-    sys.modules["esm"] = importlib.import_module(f"{__name__}.esm")
-
-
-__all__ = ["ESMFold2Config", "ESMFold2Model", "ensure_vendored_esm"]
+__all__ = ["ESMFold2Config", "ESMFold2Model"]

modeling_esmfold2.py

@@ -7,13 +7,12 @@ Quickstart::
     model = ESMFold2Model.from_pretrained("biohub/ESMFold2").cuda().eval()
     open("ubq.pdb", "w").write(model.infer_protein_as_pdb("MQIFVKTLTGKT..."))
 
-For multi-chain / ligand / MSA inputs see ``ESMFold2InputBuilder`` in the
-companion ``esm`` package.
+For multi-chain, ligand, and MSA inputs, use ``model.input_types`` together
+with ``model.fold(...)`` or ``model.prepare_structure_input(...)``.
 """
 
 import importlib
 import math
-import sys
 from contextlib import contextmanager
 from pathlib import Path
 from typing import Any, cast
@@ -102,13 +101,6 @@ _NONPOLYMER_ID = 4
 _DEFAULT_CHUNK_SIZE = 64
 
 
-def _ensure_vendored_esm_alias() -> None:
-    package = __package__
-    assert package is not None
-    vendored_esm = importlib.import_module(f"{package}.esm")
-    sys.modules["esm"] = vendored_esm
-
-
 class PairTransition(nn.Module):
     """LayerNorm + SwiGLU feed-forward residual block on the pair representation."""
 
@@ -678,15 +670,10 @@ class ESMFold2Model(PreTrainedModel):
                 pretrained_model_name_or_path, **kwargs
             )
             if config.type == "experimental":
-                experimental_module = importlib.import_module(
-                    f"{__package__}.modeling_esmfold2_experimental"
-                )
-                return experimental_module.ESMFold2ExperimentalModel.from_pretrained(
-                    pretrained_model_name_or_path,
-                    *args,
-                    config=config,
-                    load_esmc=load_esmc,
-                    **kwargs,
+                raise ValueError(
+                    "FastPLMs ESMFold2 supports the released ESMFold2 and "
+                    "ESMFold2-Fast checkpoints. Experimental ESMFold2 configs "
+                    "are not part of the self-contained AutoModel package."
                 )
             kwargs["config"] = config
         # Pop the precision knob before forwarding to the HF loader.