Removed sampling

app.py

@@ -140,7 +140,7 @@ def internvl_inference(image_path: str | None, text_input: str | None = None):
         return "Please upload an image first."
     pixel_values = load_image(image_path, max_num=12).to(torch.bfloat16).cuda()
     prompt = f"<image>\n{text_input}" if text_input else "<image>\n"
-    gen_cfg = dict(max_new_tokens=1024, do_sample=True)
+    gen_cfg = dict(max_new_tokens=1024, do_sample=False)
     return model.chat(tokenizer, pixel_values, prompt, gen_cfg)