Add thread lock to glue model tokenizer call and update demos version.

Fixes race condition when using multiple interpreters in separate threads of LIT server with same model. PiperOrigin-RevId: 417452193
PAIR-code · jameswex · Dec 21, 2021 · Nov 10, 2021 · Nov 10, 2021 · Nov 11, 2021
commit f347fad01a846fcc2ba1bdf96c9a8f7fcc9b1cbf
diff --git a/lit_nlp/examples/models/glue_models.py b/lit_nlp/examples/models/glue_models.py
@@ -3,6 +3,7 @@
 
 import os
 import re
+import threading
 from typing import Optional, Dict, List, Iterable
 
 import attr
@@ -58,6 +59,7 @@ def __init__(self,
                **config_kw):
     self.config = GlueModelConfig(**config_kw)
     self._load_model(model_name_or_path)
+    self._lock = threading.Lock()
 
   def _load_model(self, model_name_or_path):
     """Load model. Can be overridden for testing."""
@@ -76,8 +78,9 @@ def _load_model(self, model_name_or_path):
         config=model_config)
 
   def _get_tokens(self, ex: JsonDict, field_name: str) -> List[str]:
-    return (ex.get("tokens_" + field_name) or
-            self.tokenizer.tokenize(ex[field_name]))
+    with self._lock:
+      return (ex.get("tokens_" + field_name) or
+              self.tokenizer.tokenize(ex[field_name]))
 
   def _preprocess(self, inputs: Iterable[JsonDict]) -> Dict[str, tf.Tensor]:
     # Use pretokenized input if available.