Update app.py

app.py

@@ -17,10 +17,15 @@ from transformers import (
     Qwen2_5_VLForConditionalGeneration,
     Qwen2VLForConditionalGeneration,
     AutoProcessor,
+    AutoTokenizer,
     TextIteratorStreamer,
 )
 from transformers.image_utils import load_image
 
+import subprocess
+subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
+
+
 # Constants for text generation
 MAX_MAX_NEW_TOKENS = 2048
 DEFAULT_MAX_NEW_TOKENS = 1024
@@ -46,8 +51,8 @@ model_x = Qwen2VLForConditionalGeneration.from_pretrained(
     torch_dtype=torch.float16
 ).to(device).eval()
 
-# Load Relaxed
-MODEL_ID_Z = "Qwen/Qwen2.5-VL-3B-Instruct"
+# Load Space Thinker
+MODEL_ID_Z = "remyxai/SpaceThinker-Qwen2.5VL-3B"
 processor_z = AutoProcessor.from_pretrained(MODEL_ID_Z, trust_remote_code=True)
 model_z = Qwen2_5_VLForConditionalGeneration.from_pretrained(
     MODEL_ID_Z,
@@ -55,12 +60,13 @@ model_z = Qwen2_5_VLForConditionalGeneration.from_pretrained(
     torch_dtype=torch.float16
 ).to(device).eval()
 
-# Load ImageScope
-MODEL_ID_T = "prithivMLmods/Imgscope-OCR-2B-0527"
-processor_t = AutoProcessor.from_pretrained(MODEL_ID_X, trust_remote_code=True)
-model_t = Qwen2VLForConditionalGeneration.from_pretrained(
+# Load moondream
+MODEL_ID_T = "moondream/moondream-2b-2025-04-14-4bit"
+processor_t = AutoTokenizer.from_pretrained(MODEL_ID_X, trust_remote_code=True)
+model_t = AutoModelForCausalLM.from_pretrained(
     MODEL_ID_T,
     trust_remote_code=True,
+    attn_implementation="flash_attention_2",
     torch_dtype=torch.float16
 ).to(device).eval()
 
@@ -103,10 +109,10 @@ def generate_image(model_name: str, text: str, image: Image.Image,
     elif model_name == "coreOCR-7B-050325-preview":
         processor = processor_x
         model = model_x
-    elif model_name == "Qwen2.5-VL-3B":
+    elif model_name == "SpaceThinker-Qwen2.5VL-3B":
         processor = processor_z
         model = model_z
-    elif model_name == "Imgscope-OCR-2B":
+    elif model_name == "moondream-2b-2025-04-14-4bit":
         processor = processor_t
         model = model_t
     else:
@@ -159,10 +165,10 @@ def generate_video(model_name: str, text: str, video_path: str,
     elif model_name == "coreOCR-7B-050325-preview":
         processor = processor_x
         model = model_x
-    elif model_name == "Qwen2.5-VL-3B":
+    elif model_name == "SpaceThinker-Qwen2.5VL-3B":
         processor = processor_z
         model = model_z
-    elif model_name == "Imgscope-OCR-2B":
+    elif model_name == "moondream-2b-2025-04-14-4bit":
         processor = processor_t
         model = model_t
     else:
@@ -262,7 +268,7 @@ 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=["coreOCR-7B-050325-preview", "typhoon-ocr-7b", "Qwen2.5-VL-3B", "Imgscope-OCR-2B"],
+                choices=["coreOCR-7B-050325-preview", "typhoon-ocr-7b", "SpaceThinker-Qwen2.5VL-3B", "moondream-2b-2025-04-14-4bit"],
                 label="Select Model",
                 value="coreOCR-7B-050325-preview"
             )