Add Qwen-2.5-VL-3B

app.py

@@ -20,6 +20,16 @@ from transformers import (
 # Local imports
 from qwen_vl_utils import process_vision_info
 
+# Set device agnostic code
+if torch.cuda.is_available():
+    device = "cuda"
+elif (torch.backends.mps.is_available()) and (torch.backends.mps.is_built()):
+    device = "mps"
+else:
+    device = "cpu"
+
+print(f"[INFO] Using device: {device}")
+
 
 def array_to_image_path(image_array):
     if image_array is None:
@@ -41,13 +51,21 @@ def array_to_image_path(image_array):
     
 models = {
     "Qwen/Qwen2.5-VL-7B-Instruct": Qwen2_5_VLForConditionalGeneration.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct", 
+                                                                                      trust_remote_code=True, 
+                                                                                      torch_dtype="auto",
+                                                                                      device_map="auto").eval(),
+
+    "Qwen/Qwen2.5-VL-3B-Instruct": Qwen2_5_VLForConditionalGeneration.from_pretrained("Qwen/Qwen2.5-VL-3B-Instruct", 
                                                                                       trust_remote_code=True, 
                                                                                       torch_dtype="auto",
                                                                                       device_map="auto").eval()
+
 }
 
 processors = {
-    "Qwen/Qwen2.5-VL-7B-Instruct": AutoProcessor.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct", trust_remote_code=True)
+    "Qwen/Qwen2.5-VL-7B-Instruct": AutoProcessor.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct", trust_remote_code=True),
+    "Qwen/Qwen2.5-VL-3B-Instruct": AutoProcessor.from_pretrained("Qwen/Qwen2.5-VL-3B-Instruct", trust_remote_code=True)
+
 }
 
 DESCRIPTION = "[Qwen2.5-VL Demo](https://huggingface.co/collections/Qwen/qwen25-vl-6795ffac22b334a837c0f9a5)"
@@ -94,7 +112,7 @@ def run_example(image, text_input=None, model_id=None):
         padding=True,
         return_tensors="pt",
     )
-    inputs = inputs.to("cuda")
+    inputs = inputs.to(device)
     
     # Inference: Generation of the output
     generated_ids = model.generate(**inputs, max_new_tokens=1024)
@@ -127,7 +145,7 @@ with gr.Blocks(css=css) as demo:
                 model_selector = gr.Dropdown(choices=list(models.keys()), 
                                              label="Model", 
                                              value="Qwen/Qwen2.5-VL-7B-Instruct")
-                text_input = gr.Textbox(label="Question")
+                text_input = gr.Textbox(label="Text Prompt")
                 submit_btn = gr.Button(value="Submit")
             with gr.Column():
                 output_text = gr.Textbox(label="Output Text")