Update handler.py

handler.py

@@ -36,7 +36,9 @@ class EndpointHandler:
             device_map="auto",
             attn_implementation="sdpa",
         )
-        base.resize_token_embeddings(vocab_size)
+        # mean_resizing=False avoids holding 2x embedding matrix in VRAM during resize.
+        # Safe here because SinLlama adapter contains the correct trained embeddings.
+        base.resize_token_embeddings(vocab_size, mean_resizing=False)
 
         # Merge SinLlama into base so the MCQ adapter sees a plain model (not stacked PeftModel)
         print("Loading and merging SinLlama adapter...")