Merge pull request #407 from allenai/favyen/20251023-model-loader

Add convenience function to load models from Hugging Face.

Author: favyen2

olmoearth_pretrain/model_loader.py

@@ -0,0 +1,68 @@
+"""Load the OlmoEarth models from Hugging Face.
+
+The weights are converted to pth file from distributed checkpoint like this:
+
+    import json
+    from pathlib import Path
+
+    import torch
+
+    from olmo_core.config import Config
+    from olmo_core.distributed.checkpoint import load_model_and_optim_state
+
+    checkpoint_path = Path("/weka/dfive-default/helios/checkpoints/joer/nano_lr0.001_wd0.002/step370000")
+    with (checkpoint_path / "config.json").open() as f:
+        config_dict = json.load(f)
+        model_config = Config.from_dict(config_dict["model"])
+
+    model = model_config.build()
+
+    train_module_dir = checkpoint_path / "model_and_optim"
+    load_model_and_optim_state(str(train_module_dir), model)
+    torch.save(model.state_dict(), "OlmoEarth-v1-Nano.pth")
+"""
+
+import json
+from enum import StrEnum
+
+import torch
+from huggingface_hub import hf_hub_download
+from olmo_core.config import Config
+
+
+class ModelID(StrEnum):
+    """OlmoEarth pre-trained model ID."""
+
+    OLMOEARTH_V1_NANO = "OlmoEarth-v1-Nano"
+    OLMOEARTH_V1_TINY = "OlmoEarth-v1-Tiny"
+    OLMOEARTH_V1_BASE = "OlmoEarth-v1-Base"
+
+
+def load_model(model_id: ModelID, load_weights: bool = True) -> torch.nn.Module:
+    """Initialize and load the weights for the specified model ID.
+
+    The config and weights will be downloaded from Hugging Face.
+
+    Args:
+        model_id: the model ID to load.
+        load_weights: whether to load the weights. Set false to skip downloading the
+            weights from Hugging Face and leave them randomly initialized. Note that
+            the config.json will still be downloaded from Hugging Face.
+    """
+    # We ignore bandit warnings here since we are just downloading config and weights,
+    # not any code.
+    repo_id = f"allenai/{model_id.value}"
+    config_fname = hf_hub_download(repo_id=repo_id, filename="config.json")  # nosec
+    with open(config_fname) as f:
+        config_dict = json.load(f)
+        model_config = Config.from_dict(config_dict["model"])
+
+    model: torch.nn.Module = model_config.build()
+
+    if not load_weights:
+        return model
+
+    pth_fname = hf_hub_download(repo_id=repo_id, filename="weights.pth")  # nosec
+    state_dict = torch.load(pth_fname, map_location="cpu")
+    model.load_state_dict(state_dict)
+    return model