fix load pretrained legacy with state dict change

sae_lens/training/sparse_autoencoder.py

@@ -312,7 +312,13 @@ def load_from_pretrained_legacy(cls, path: str):
 
         # Create an instance of the class using the loaded configuration
         instance = cls(cfg=state_dict["cfg"])
-        instance.load_state_dict(state_dict["state_dict"], strict=False)
+        new_state_dict = instance.state_dict()
+        if "scaling_factor" not in state_dict["state_dict"]:
+            assert isinstance(instance.cfg.d_sae, int)
+            state_dict["state_dict"]["scaling_factor"] = torch.ones(
+                instance.cfg.d_sae, dtype=instance.cfg.dtype, device=instance.cfg.device
+            )
+        instance.load_state_dict(new_state_dict, strict=True)
 
         return instance
 

tests/unit/training/test_sparse_autoencoder.py

@@ -112,6 +112,43 @@ def test_SparseAutoencoder_save_and_load_from_pretrained(tmp_path: Path) -> None
         )
 
 
+def test_SparseAutoencoder_save_and_load_from_pretrained_lacks_scaling_factor(
+    tmp_path: Path,
+) -> None:
+    cfg = build_sae_cfg(device="cpu")
+    model_path = str(tmp_path)
+    sparse_autoencoder = SparseAutoencoder(cfg)
+    sparse_autoencoder_state_dict = sparse_autoencoder.state_dict()
+    # sometimes old state dicts will be missing the scaling factor
+    del sparse_autoencoder_state_dict["scaling_factor"]  # = torch.tensor(0.0)
+    sparse_autoencoder.save_model(model_path)
+
+    assert os.path.exists(model_path)
+
+    sparse_autoencoder_loaded = SparseAutoencoder.load_from_pretrained(model_path)
+    sparse_autoencoder_loaded.cfg.verbose = True
+    sparse_autoencoder_loaded.cfg.checkpoint_path = cfg.checkpoint_path
+    sparse_autoencoder_loaded.cfg.device = "cpu"  # might autoload onto mps
+    sparse_autoencoder_loaded = sparse_autoencoder_loaded.to("cpu")
+    sparse_autoencoder_loaded_state_dict = sparse_autoencoder_loaded.state_dict()
+    # check cfg matches the original
+    assert sparse_autoencoder_loaded.cfg == cfg
+
+    # check state_dict matches the original
+    for key in sparse_autoencoder.state_dict().keys():
+        if key == "scaling_factor":
+            assert isinstance(cfg.d_sae, int)
+            assert torch.allclose(
+                torch.ones(cfg.d_sae, dtype=cfg.dtype, device=cfg.device),
+                sparse_autoencoder_loaded_state_dict[key],
+            )
+        else:
+            assert torch.allclose(
+                sparse_autoencoder_state_dict[key],
+                sparse_autoencoder_loaded_state_dict[key],
+            )
+
+
 def test_sparse_autoencoder_forward(sparse_autoencoder: SparseAutoencoder):
     batch_size = 32
     d_in = sparse_autoencoder.d_in