Add some sanity checks

aviary/backend/llm/initializers/hf_transformers/base.py

@@ -71,7 +71,7 @@ def _get_model_location_on_disk(self, model_id: str) -> str:
         )
         model_id_or_path = model_id
 
-        if os.path.exists(path):
+        if os.path.exists(path) and os.path.exists(os.path.join(path, "refs", "main")):
             with open(os.path.join(path, "refs", "main"), "r") as f:
                 snapshot_hash = f.read().strip()
             if os.path.exists(

aviary/backend/llm/predictor/continuous_batching_predictor.py

@@ -414,4 +414,5 @@ async def _stream_async(
 
     def check_health(self) -> None:
         super().check_health()
-        self.scheduler.check_health()
+        if self.scheduler:
+            self.scheduler.check_health()

aviary/backend/llm/predictor/predictor.py

@@ -95,6 +95,7 @@ def __init__(self, llm_config: LLMConfig, world_size: int):
         self.local_rank = None
         self.current_device = None
         self.gpu_memory_fraction = 1.0
+        self.pg = None
 
     def init_model(
         self,