Patch TRANSFORMERS_CACHE in sanity check too

training/run_grpo.sh

@@ -46,6 +46,13 @@ python -c "
 import torch, transformers, peft, trl
 print(f'torch={torch.__version__}  transformers={transformers.__version__}')
 print(f'peft={peft.__version__}  trl={trl.__version__}')
+
+# Patch missing constant that llm_blender expects from older transformers
+import transformers.utils.hub as _h
+if not hasattr(_h, 'TRANSFORMERS_CACHE'):
+    from huggingface_hub.constants import HF_HUB_CACHE
+    _h.TRANSFORMERS_CACHE = HF_HUB_CACHE
+
 from trl import GRPOTrainer, GRPOConfig
 print('GRPOTrainer OK')
 "