misc: minor bug fix & support loading configs with path in the MongoDB

server/app.py

@@ -46,7 +46,8 @@
 
 
 def get_model(dictionary_name: str) -> HookedTransformer:
-	cfg = LanguageModelConfig.from_pretrained_sae(f"{result_dir}/{dictionary_name}")
+	path = client.get_dictionary(dictionary_name, dictionary_series=dictionary_series)['path'] or f"{result_dir}/{dictionary_name}"
+	cfg = LanguageModelConfig.from_pretrained_sae(path)
 	if (cfg.model_name, cfg.model_from_pretrained_path) not in lm_cache:
 		hf_model = AutoModelForCausalLM.from_pretrained(
 			(
@@ -81,8 +82,9 @@ def get_model(dictionary_name: str) -> HookedTransformer:
 
 
 def get_sae(dictionary_name: str) -> SparseAutoEncoder:
+	path = client.get_dictionary(dictionary_name, dictionary_series=dictionary_series)['path'] or f"{result_dir}/{dictionary_name}"
 	if dictionary_name not in sae_cache:
-		sae = SparseAutoEncoder.from_pretrained(f"{result_dir}/{dictionary_name}", device=device)
+		sae = SparseAutoEncoder.from_pretrained(path, device=device)
 		sae.eval()
 		sae_cache[dictionary_name] = sae
 	return sae_cache[dictionary_name]
@@ -398,6 +400,7 @@ def feature_interpretation(
 	custom_interpretation: str | None = None,
 ):
 	model = get_model(dictionary_name)
+	path = client.get_dictionary(dictionary_name, dictionary_series=dictionary_series)['path'] or f"{result_dir}/{dictionary_name}"
 	if type == "custom":
 		interpretation = {
 			"text": custom_interpretation,
@@ -411,8 +414,8 @@ def feature_interpretation(
 	elif type == "auto":
 		cfg = AutoInterpConfig(
 			**{
-				**SAEConfig.from_pretrained(f"{result_dir}/{dictionary_name}").to_dict(),
-				**LanguageModelConfig.from_pretrained_sae(f"{result_dir}/{dictionary_name}").to_dict(),
+				"sae": SAEConfig.from_pretrained(path).to_dict(),
+				"lm": LanguageModelConfig.from_pretrained_sae(path).to_dict(),
 				"openai_api_key": os.environ.get("OPENAI_API_KEY"),
 				"openai_base_url": os.environ.get("OPENAI_BASE_URL"),
 			}
@@ -427,8 +430,8 @@ def feature_interpretation(
 	elif type == "validate":
 		cfg = AutoInterpConfig(
 			**{
-				**SAEConfig.from_pretrained(f"{result_dir}/{dictionary_name}").to_dict(),
-				**LanguageModelConfig.from_pretrained_sae(f"{result_dir}/{dictionary_name}").to_dict(),
+				"sae": SAEConfig.from_pretrained(path).to_dict(),
+				"lm": LanguageModelConfig.from_pretrained_sae(path).to_dict(),
 				"openai_api_key": os.environ.get("OPENAI_API_KEY"),
 				"openai_base_url": os.environ.get("OPENAI_BASE_URL"),
 			}

src/lm_saes/database.py

@@ -87,6 +87,17 @@ def update_feature(self, dictionary_name: str, feature_index: int, feature_data:
     def list_dictionaries(self, dictionary_series: str | None = None):
         return [d['name'] for d in self.dictionary_collection.find({'series': dictionary_series} if dictionary_series is not None else {})]
 
+    def get_dictionary(self, dictionary_name: str, dictionary_series: str | None = None):
+        dictionary = self.dictionary_collection.find_one({'name': dictionary_name, 'series': dictionary_series})
+        if dictionary is None:
+            return None
+        return {
+            'name': dictionary['name'],
+            'n_features': dictionary['n_features'],
+            'series': dictionary['series'],
+            'path': dictionary['path'] if 'path' in dictionary else None
+        }
+
     def get_feature(self, dictionary_name: str, feature_index: int, dictionary_series: str | None = None):
         dictionary = self.dictionary_collection.find_one({'name': dictionary_name, 'series': dictionary_series})
         if dictionary is None:

src/lm_saes/entrypoint.py

@@ -81,6 +81,7 @@ def entrypoint():
     else:
         raise ValueError(f'Unsupported runner: {args.runner}.')
 
-    if dist.is_initialized():
-        dist.destroy_process_group()
+    if tp_size > 1 or ddp_size > 1:
+        if dist.is_initialized():
+            dist.destroy_process_group()