Update app.py

app.py

@@ -19,6 +19,8 @@ from transformers import (
     Qwen2VLForConditionalGeneration,
     AutoProcessor,
     AutoTokenizer,
+    AutoModel, 
+    AutoImageProcessor,
     TextIteratorStreamer,
 )
 
@@ -31,10 +33,11 @@ MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "4096"))
 
 device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
 
-# Load SkyCaptioner-V1
-MODEL_ID_M = "Skywork/SkyCaptioner-V1"
-processor_m = AutoProcessor.from_pretrained(MODEL_ID_M, trust_remote_code=True)
-model_m = Qwen2_5_VLForConditionalGeneration.from_pretrained(
+# Load Llama-3.1-Nemotron-Nano-VL-8B-V1
+MODEL_ID_M = "nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1"
+processor_m = AutoImageProcessor.from_pretrained(MODEL_ID_M, trust_remote_code=True)
+tokenizer_m = AutoTokenizer.from_pretrained(MODEL_ID_M)
+model_m = AutoModel.from_pretrained(
     MODEL_ID_M,
     trust_remote_code=True,
     torch_dtype=torch.float16
@@ -58,16 +61,6 @@ model_k = Qwen2VLForConditionalGeneration.from_pretrained(
     torch_dtype=torch.float16
 ).to(device).eval()
 
-# Load Imgscope-OCR-2B-0527
-MODEL_ID_Y = "prithivMLmods/Imgscope-OCR-2B-0527"
-processor_y = AutoProcessor.from_pretrained(MODEL_ID_Y, trust_remote_code=True)
-model_y = Qwen2VLForConditionalGeneration.from_pretrained(
-    MODEL_ID_Y,
-    trust_remote_code=True,
-    torch_dtype=torch.float16
-).to(device).eval()
-
-
 def downsample_video(video_path):
     """
     Downsamples the video to evenly spaced frames.
@@ -99,7 +92,7 @@ def generate_image(model_name: str, text: str, image: Image.Image,
     """
     Generates responses using the selected model for image input.
     """
-    if model_name == "SkyCaptioner-V1":
+    if model_name == "Llama-3.1-Nemotron-Nano-VL-8B-V1":
         processor = processor_m
         model = model_m
     elif model_name == "SpaceThinker-3B":
@@ -108,9 +101,6 @@ def generate_image(model_name: str, text: str, image: Image.Image,
     elif model_name == "coreOCR-7B-050325-preview":
         processor = processor_k
         model = model_k
-    elif model_name == "Imgscope-OCR-2B-0527":
-        processor = processor_y
-        model = model_y
     else:
         yield "Invalid model selected."
         return
@@ -156,7 +146,7 @@ def generate_video(model_name: str, text: str, video_path: str,
     """
     Generates responses using the selected model for video input.
     """
-    if model_name == "SkyCaptioner-V1":
+    if model_name == "Llama-3.1-Nemotron-Nano-VL-8B-V1":
         processor = processor_m
         model = model_m
     elif model_name == "SpaceThinker-3B":
@@ -165,9 +155,6 @@ def generate_video(model_name: str, text: str, video_path: str,
     elif model_name == "coreOCR-7B-050325-preview":
         processor = processor_k
         model = model_k
-    elif model_name == "Imgscope-OCR-2B-0527":
-        processor = processor_y
-        model = model_y
     else:
         yield "Invalid model selected."
         return
@@ -269,7 +256,7 @@ 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=["SkyCaptioner-V1", "SpaceThinker-3B", "coreOCR-7B-050325-preview", "Imgscope-OCR-2B-0527"],
+                choices=["Llama-3.1-Nemotron-Nano-VL-8B-V1", "SpaceThinker-3B", "coreOCR-7B-050325-preview"],
                 label="Select Model",
                 value="SkyCaptioner-V1"
             )