`batch_size` with `auto` defaults to 1 if `No executable batch size found` (#1405)

Fixes https://github.com/EleutherAI/lm-evaluation-harness/issues/1323

lm_eval/models/huggingface.py

@@ -617,7 +617,13 @@ class HFLM(LM):
 
             return batch_size
 
-        batch_size = forward_batch()
+        try:
+            batch_size = forward_batch()
+        except RuntimeError as e:
+            if "No executable batch size found" in str(e):
+                batch_size = 1
+            else:
+                raise
 
         if self.world_size > 1:
             # if multi-GPU, always take minimum over all selected batch sizes