Update app.py

app.py

@@ -116,6 +116,7 @@ MODEL_ID_M = "nanonets/Nanonets-OCR-s"
 processor_m = AutoProcessor.from_pretrained(MODEL_ID_M, trust_remote_code=True)
 model_m = Qwen2_5_VLForConditionalGeneration.from_pretrained(
     MODEL_ID_M,
+    attn_implementation="flash_attention_2",
     trust_remote_code=True,
     torch_dtype=torch.float16
 ).to(device).eval()
@@ -130,6 +131,7 @@ processor_g = AutoProcessor.from_pretrained(
 )
 model_g = Qwen2_5_VLForConditionalGeneration.from_pretrained(
     MODEL_ID_G,
+    attn_implementation="flash_attention_2",
     trust_remote_code=True,
     subfolder=SUBFOLDER,
     torch_dtype=torch.float16
@@ -140,6 +142,7 @@ MODEL_ID_L = "scb10x/typhoon-ocr-7b"
 processor_l = AutoProcessor.from_pretrained(MODEL_ID_L, trust_remote_code=True)
 model_l = Qwen2_5_VLForConditionalGeneration.from_pretrained(
     MODEL_ID_L,
+    attn_implementation="flash_attention_2",
     trust_remote_code=True,
     torch_dtype=torch.float16
 ).to(device).eval()
@@ -158,6 +161,7 @@ MODEL_ID_N = "Kwai-Keye/Thyme-RL"
 processor_n = AutoProcessor.from_pretrained(MODEL_ID_N, trust_remote_code=True)
 model_n = Qwen2_5_VLForConditionalGeneration.from_pretrained(
     MODEL_ID_N,
+    attn_implementation="flash_attention_2",
     trust_remote_code=True,
     torch_dtype=torch.float16
 ).to(device).eval()