Update training.py

training.py

@@ -1,7 +1,8 @@
 # training.py – Memory‑safe: Phi‑3‑mini + Expert Demos + Fast PPO (2 iterations)
 import os
-os.environ["TRITON_DISABLE"] = "1"
-os.environ["TOKENIZERS_PARALLELISM"] = "false"        # Issue #12: prevent OOM from parallel tokenization
+os.environ["TRITON_INTERPRET"] = "1"          # force CPU interpretation, no ptxas
+os.environ["TRITON_CACHE_DIR"] = "/tmp/triton_cache"
+os.environ["TORCHINDUCTOR_CPP_WRAPPER"] = "0"  # stay in Python     # Issue #12: prevent OOM from parallel tokenization
 
 import torch._dynamo
 torch._dynamo.config.disable = True