Update handler.py

handler.py

@@ -57,7 +57,7 @@ class EndpointHandler:
             model_id,
             trust_remote_code=True,
             torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
-            attn_implementation="flash_attention_2" if torch.cuda.is_available() else "sdpa",
+            attn_implementation="sdpa",  # Use SDPA - flash_attn not installed in HF containers
             device_map="auto" if torch.cuda.is_available() else None,
             token=hf_token,
         )