add custom handler

handler.py

@@ -7,7 +7,7 @@ class EndpointHandler:
         self.model = AutoModel.from_pretrained(
             path, 
             trust_remote_code=True,
-            attn_implementation='sdpa',
+            attn_implementation='sdpa',  # Using sdpa instead of flash_attention_2
             torch_dtype=torch.float16
         )
         self.model = self.model.eval().cuda()

requirements.txt

@@ -3,4 +3,3 @@ torch==2.1.2
 torchvision==0.16.2
 transformers==4.40.0
 sentencepiece==0.1.99
-flash-attn==2.3.6