Serialize tokenizer first load; one-shot degradation warning

96 worker threads racing transformers 5.x lazy imports on the FIRST
_get_tokenizer call raised ImportError and degraded that whole first
batch to the char approximation (the old single-threaded path never
raced). First load now holds a threading.Lock; the transformers
'>model_max_length' logging is silenced inside truncation (counting a
2M-token doc before trimming it is the point), and the per-sample
degradation print becomes a one-shot warning.

Co-Authored-By: Claude <noreply@anthropic.com>

evalharness/model/truncation.py

@@ -9,20 +9,33 @@ Requires a tokenizer (transformers) at tokenizer_path or auto from the model.
 """
 
 import os
+import threading
 from functools import lru_cache
 from typing import Optional
 
 DEFAULT_TRUNCATION_TOKENS = 32768 * 4  # 131072, mirrors evalside run.py
 
 
+_TOK_LOCK = threading.Lock()
+
+
 @lru_cache(maxsize=4)
 def _get_tokenizer(tokenizer_path: str):
     if not tokenizer_path or not os.path.exists(tokenizer_path):
         raise FileNotFoundError(
             f'tokenizer not found at {tokenizer_path!r} -- token-level truncation '
             'needs a local tokenizer dir (e.g. /data1/models/DeepSeek-V4-Flash-INT8)')
+    # serialize the FIRST load: 96 worker threads racing transformers 5.x's
+    # lazy imports raised ImportError and silently degraded batches to the
+    # char approximation; after one success lru_cache serves the rest
+    with _TOK_LOCK:
         from transformers import AutoTokenizer
 
+        # the '> model_max_length' warnings are EXPECTED here -- counting a
+        # 2M-token doc before trimming it is the whole point of truncation
+        import logging
+
+        logging.getLogger('transformers').setLevel(logging.ERROR)
         return AutoTokenizer.from_pretrained(tokenizer_path, trust_remote_code=True)