Update app.py

app.py

@@ -22,11 +22,17 @@ from transformers import (
     TextIteratorStreamer,
 )
 
-# Constants for text generation
-MAX_MAX_NEW_TOKENS = 2048
+# Constants for text generation\ nMAX_MAX_NEW_TOKENS = 2048
 DEFAULT_MAX_NEW_TOKENS = 1024
 MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "4096"))
 
+# Default system prompt for Behemoth-3B-070225-post0.1
+DEFAULT_SYSTEM_PROMPT_BEHEMOTH = (
+    "<|begin_of_text|><|start_header_id|>system<|end_header_id|>  detailed thinking on<|eot_id|>"
+    "<|start_header_id|>user<|end_header_id|>  You are a reasoning model designed to answer complex questions step-by-step, "
+    "Conclude with the solution <|eot_id|><|start_header_id|>assistant<|end_header_id|>"
+)
+
 device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
 
 # Load Behemoth-3B-070225-post0.1
@@ -74,8 +80,8 @@ model_y = Qwen2_5_VLForConditionalGeneration.from_pretrained(
     torch_dtype=torch.float16
 ).to(device).eval()
 
-# Video sampling
-def downsample_video(video_path):
+# Video downsampling helper
+ def downsample_video(video_path):
     """
     Downsamples the video to evenly spaced frames.
     Each frame is returned as a PIL image along with its timestamp.
@@ -98,7 +104,7 @@ def downsample_video(video_path):
 
 @spaces.GPU
 def generate_image(model_name: str, text: str, image: Image.Image,
-                   max_new_tokens: int = 1024,
+                   max_new_tokens: int = DEFAULT_MAX_NEW_TOKENS,
                    temperature: float = 0.6,
                    top_p: float = 0.9,
                    top_k: int = 50,
@@ -106,21 +112,27 @@ def generate_image(model_name: str, text: str, image: Image.Image,
     """
     Generates responses using the selected model for image input.
     """
+    # Model selection
     if model_name == "SkyCaptioner-V1":
         processor = processor_m
         model = model_m
+        system_prompt = None
     elif model_name == "Behemoth-3B-070225-post0.1":
         processor = processor_n
         model = model_n
+        system_prompt = DEFAULT_SYSTEM_PROMPT_BEHEMOTH
     elif model_name == "SpaceThinker-3B":
         processor = processor_z
         model = model_z
+        system_prompt = None
     elif model_name == "coreOCR-7B-050325-preview":
         processor = processor_k
         model = model_k
+        system_prompt = None
     elif model_name == "SpaceOm-3B":
         processor = processor_y
         model = model_y
+        system_prompt = None
     else:
         yield "Invalid model selected."
         return
@@ -129,22 +141,17 @@ def generate_image(model_name: str, text: str, image: Image.Image,
         yield "Please upload an image."
         return
 
-    if model_name == "Behemoth-3B-070225-post0.1":
-        messages = [
-            {"role": "system", "content": [{"type": "text", "text": "detailed thinking on"}]},
-            {"role": "user", "content": [
-                {"type": "image", "image": image},
-                {"type": "text", "text": text},
-            ]}
+    # Build message list
+    messages = []
+    if system_prompt:
+        messages.append({"role": "system", "content": system_prompt})
+    messages.append({
+        "role": "user",
+        "content": [
+            {"type": "image", "image": image},
+            {"type": "text", "text": text},
         ]
-    else:
-        messages = [{
-            "role": "user",
-            "content": [
-                {"type": "image", "image": image},
-                {"type": "text", "text": text},
-            ]
-        }]
+    })
 
     prompt_full = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
     inputs = processor(
@@ -168,7 +175,7 @@ def generate_image(model_name: str, text: str, image: Image.Image,
 
 @spaces.GPU
 def generate_video(model_name: str, text: str, video_path: str,
-                   max_new_tokens: int = 1024,
+                   max_new_tokens: int = DEFAULT_MAX_NEW_TOKENS,
                    temperature: float = 0.6,
                    top_p: float = 0.9,
                    top_k: int = 50,
@@ -176,21 +183,27 @@ def generate_video(model_name: str, text: str, video_path: str,
     """
     Generates responses using the selected model for video input.
     """
+    # Model selection
     if model_name == "SkyCaptioner-V1":
         processor = processor_m
         model = model_m
+        system_prompt = None
     elif model_name == "Behemoth-3B-070225-post0.1":
         processor = processor_n
         model = model_n
+        system_prompt = DEFAULT_SYSTEM_PROMPT_BEHEMOTH
     elif model_name == "SpaceThinker-3B":
         processor = processor_z
         model = model_z
+        system_prompt = None
     elif model_name == "coreOCR-7B-050325-preview":
         processor = processor_k
         model = model_k
+        system_prompt = None
     elif model_name == "SpaceOm-3B":
         processor = processor_y
         model = model_y
+        system_prompt = None
     else:
         yield "Invalid model selected."
         return
@@ -200,21 +213,16 @@ def generate_video(model_name: str, text: str, video_path: str,
         return
 
     frames = downsample_video(video_path)
-    
-    if model_name == "Behemoth-3B-070225-post0.1":
-        system_message = "detailed thinking on"
-    else:
-        system_message = "You are a helpful assistant."
-    
-    messages = [
-        {"role": "system", "content": [{"type": "text", "text": system_message}]},
-        {"role": "user", "content": [{"type": "text", "text": text}]}
-    ]
-    for frame in frames:
-        image, timestamp = frame
-        messages[1]["content"].append({"type": "text", "text": f"Frame {timestamp}:"})
-        messages[1]["content"].append({"type": "image", "image": image})
-    
+    # Build message list
+    messages = []
+    if system_prompt:
+        messages.append({"role": "system", "content": system_prompt})
+    messages.append({"role": "system", "content": [{"type": "text", "text": "You are a helpful assistant."}]})
+    messages.append({"role": "user", "content": [{"type": "text", "text": text}]})
+    for image, timestamp in frames:
+        messages[-1]["content"].append({"type": "text", "text": f"Frame {timestamp}:"})
+        messages[-1]["content"].append({"type": "image", "image": image})
+
     inputs = processor.apply_chat_template(
         messages,
         tokenize=True,
@@ -296,6 +304,7 @@ with gr.Blocks(css=css, theme="bethecloud/storj_theme") as demo:
                 top_p = gr.Slider(label="Top-p (nucleus sampling)", minimum=0.05, maximum=1.0, step=0.05, value=0.9)
                 top_k = gr.Slider(label="Top-k", minimum=1, maximum=1000, step=1, value=50)
                 repetition_penalty = gr.Slider(label="Repetition penalty", minimum=1.0, maximum=2.0, step=0.05, value=1.2)
+                
         with gr.Column():
             output = gr.Textbox(label="Output", interactive=False, lines=2, scale=2)
             model_choice = gr.Radio(
@@ -305,7 +314,7 @@ with gr.Blocks(css=css, theme="bethecloud/storj_theme") as demo:
             )
             
             gr.Markdown("**Model Info 💻** | [Report Bug](https://huggingface.co/spaces/prithivMLmods/VisionScope-R2/discussions)")
-            gr.Markdown("> [SkyCaptioner-V1](https://huggingface.co/Skywork/SkyCaptioner-V1):  structural video captioning model designed to generate high-quality, structural descriptions for video data. It integrates specialized sub-expert models.")
+            gr.Markdown("> [SkyCaptioner-V1](https://huggingface.co/Skywork/SkyCaptioner-V1): structural video captioning model designed to generate high-quality, structural descriptions for video data. It integrates specialized sub-expert models.")
             gr.Markdown("> [SpaceThinker-Qwen2.5VL-3B](https://huggingface.co/remyxai/SpaceThinker-Qwen2.5VL-3B): thinking/reasoning multimodal/vision-language model (VLM) trained to enhance spatial reasoning.")
             gr.Markdown("> [coreOCR-7B-050325-preview](https://huggingface.co/prithivMLmods/coreOCR-7B-050325-preview): model is a fine-tuned version of qwen/qwen2-vl-7b, optimized for document-level optical character recognition (ocr), long-context vision-language understanding.")  
             gr.Markdown("> [SpaceOm](https://huggingface.co/remyxai/SpaceOm): SpaceOm, the reasoning traces in the spacethinker dataset average ~200 thinking tokens, so now included longer reasoning traces in the training data to help the model use more tokens in reasoning.")