sae-lens 6.6.1__tar.gz → 6.6.3__tar.gz

{sae_lens-6.6.1 → sae_lens-6.6.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: sae-lens
-Version: 6.6.1
+Version: 6.6.3
 Summary: Training and Analyzing Sparse Autoencoders (SAEs)
 License: MIT
 Keywords: deep-learning,sparse-autoencoders,mechanistic-interpretability,PyTorch

{sae_lens-6.6.1 → sae_lens-6.6.3}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "sae-lens"
-version = "6.6.1"
+version = "6.6.3"
 description = "Training and Analyzing Sparse Autoencoders (SAEs)"
 authors = ["Joseph Bloom"]
 readme = "README.md"

{sae_lens-6.6.1 → sae_lens-6.6.3}/sae_lens/__init__.py

@@ -1,5 +1,5 @@
 # ruff: noqa: E402
-__version__ = "6.6.1"
+__version__ = "6.6.3"
 
 import logging
 

{sae_lens-6.6.1 → sae_lens-6.6.3}/sae_lens/loading/pretrained_sae_loaders.py

@@ -1245,14 +1245,31 @@ def get_mwhanna_transcoder_config_from_hf(
     # Extract layer from folder name
     layer = int(folder_name.replace(".safetensors", "").split("_")[-1])
 
-    wandb_config_path = hf_hub_download(
-        repo_id, "wanb-config.yaml", force_download=force_download
-    )
-    base_config_path = hf_hub_download(
-        repo_id, "config.yaml", force_download=force_download
-    )
-    with open(base_config_path) as f:
-        base_cfg_info: dict[str, Any] = yaml.safe_load(f)
+    try:
+        # mwhanna transcoders sometimes have a typo in the config file name, so check for both
+        wandb_config_path = hf_hub_download(
+            repo_id, "wanb-config.yaml", force_download=force_download
+        )
+    except EntryNotFoundError:
+        wandb_config_path = hf_hub_download(
+            repo_id, "wandb-config.yaml", force_download=force_download
+        )
+    try:
+        base_config_path = hf_hub_download(
+            repo_id, "config.yaml", force_download=force_download
+        )
+        with open(base_config_path) as f:
+            base_cfg_info: dict[str, Any] = yaml.safe_load(f)
+    except EntryNotFoundError:
+        # the 14b transcoders don't have a config file for some reason, so just pull the model name from the repo_id
+        qwen_3_size_match = re.search(r"qwen3-(\d+(?:\.\d+)?)b", repo_id)
+        if not qwen_3_size_match:
+            raise ValueError(f"Could not extract Qwen3 size from repo_id: {repo_id}")
+        qwen_3_size = qwen_3_size_match.group(1)
+        base_cfg_info = {
+            "model_name": f"Qwen/Qwen3-{qwen_3_size}B",
+        }
+
     with open(wandb_config_path) as f:
         wandb_cfg_info: dict[str, Any] = yaml.safe_load(f)