Update app.py

app.py

@@ -13,8 +13,12 @@ import numpy as np
 from PIL import Image
 import cv2
 
+from keye_vl_utils import process_vision_info
+
 from transformers import (
     Qwen2_5_VLForConditionalGeneration,
+    AutoModel,
+    AutoTokenizer,
     AutoProcessor,
     TextIteratorStreamer,
 )
@@ -45,6 +49,16 @@ model_x = Qwen2_5_VLForConditionalGeneration.from_pretrained(
     torch_dtype=torch.float16
 ).to(device).eval()
 
+# Load Keye-VL-8B-Preview
+MODEL_ID_K = "Kwai-Keye/Keye-VL-8B-Preview"
+processor_k = AutoModel.from_pretrained(MODEL_ID_K, trust_remote_code=True)
+model_k = Qwen2_5_VLForConditionalGeneration.from_pretrained(
+    MODEL_ID_K,
+    attn_implementation="flash_attention_2",
+    trust_remote_code=True,
+    torch_dtype=torch.float16
+).to(device).eval()
+
 def downsample_video(video_path):
     """
     Downsamples the video to evenly spaced frames.
@@ -82,6 +96,9 @@ def generate_image(model_name: str, text: str, image: Image.Image,
     elif model_name == "Qwen2.5-VL-3B-Instruct":
         processor = processor_x
         model = model_x
+    elif model_name == "Keye-VL-8B-Preview":
+        processor = processor_k
+        model = model_k
     else:
         yield "Invalid model selected."
         return
@@ -132,6 +149,9 @@ def generate_video(model_name: str, text: str, video_path: str,
     elif model_name == "Qwen2.5-VL-3B-Instruct":
         processor = processor_x
         model = model_x
+    elif model_name == "Keye-VL-8B-Preview":
+        processor = processor_k
+        model = model_k
     else:
         yield "Invalid model selected."
         return
@@ -230,7 +250,7 @@ with gr.Blocks(css=css, theme="bethecloud/storj_theme") as demo:
         with gr.Column():
             output = gr.Textbox(label="Output", interactive=False, lines=2, scale=2)
             model_choice = gr.Radio(
-                choices=["Qwen2.5-VL-7B-Instruct", "Qwen2.5-VL-3B-Instruct"],
+                choices=["Qwen2.5-VL-7B-Instruct", "Qwen2.5-VL-3B-Instruct", "Keye-VL-8B-Preview"],
                 label="Select Model",
                 value="Qwen2.5-VL-7B-Instruct"
             )