Update app.py

app.py

@@ -1,11 +1,5 @@
 import gradio as gr
-from transformers import (
-    AutoProcessor,
-    Qwen2_5_VLForConditionalGeneration,
-    TextIteratorStreamer,
-    AutoModelForCausalLM,
-    AutoTokenizer,
-)
+from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration, TextIteratorStreamer
 from transformers.image_utils import load_image
 from threading import Thread
 import time
@@ -15,9 +9,6 @@ import cv2
 import numpy as np
 from PIL import Image
 
-# -----------------------
-# Progress Bar Helper
-# -----------------------
 def progress_bar_html(label: str) -> str:
     """
     Returns an HTML snippet for a thin progress bar with a label.
@@ -38,9 +29,6 @@ def progress_bar_html(label: str) -> str:
 </style>
     '''
 
-# -----------------------
-# Video Processing Helper
-# -----------------------
 def downsample_video(video_path):
     """
     Downsamples the video to 10 evenly spaced frames.
@@ -66,60 +54,45 @@ def downsample_video(video_path):
     vidcap.release()
     return frames
 
-# -----------------------
-# Qwen2.5-VL Model (Multimodal)
-# -----------------------
-MODEL_ID_VL = "Qwen/Qwen2.5-VL-7B-Instruct"  # Alternatively: "Qwen/Qwen2.5-VL-3B-Instruct"
-processor = AutoProcessor.from_pretrained(MODEL_ID_VL, trust_remote_code=True)
-vl_model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
-    MODEL_ID_VL,
+MODEL_ID = "Qwen/Qwen2.5-VL-7B-Instruct"  # Alternatively: "Qwen/Qwen2.5-VL-3B-Instruct" 
+processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)
+model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
+    MODEL_ID,
     trust_remote_code=True,
     torch_dtype=torch.bfloat16
 ).to("cuda").eval()
 
-# -----------------------
-# Text Generation Setup (DeepHermes)
-# -----------------------
-TG_MODEL_ID = "prithivMLmods/DeepHermes-3-Llama-3-3B-Preview-abliterated"
-tg_tokenizer = AutoTokenizer.from_pretrained(TG_MODEL_ID)
-tg_model = AutoModelForCausalLM.from_pretrained(
-    TG_MODEL_ID,
-    device_map="auto",
-    torch_dtype=torch.bfloat16,
-)
-tg_model.eval()
-
-# -----------------------
-# Main Inference Function
-# -----------------------
 @spaces.GPU
 def model_inference(input_dict, history):
     text = input_dict["text"]
     files = input_dict["files"]
 
-    # Video inference branch
     if text.strip().lower().startswith("@video-infer"):
+        # Remove the tag from the query.
         text = text[len("@video-infer"):].strip()
         if not files:
-            yield gr.Error("Please upload a video file along with your @video-infer query.")
+            gr.Error("Please upload a video file along with your @video-infer query.")
             return
+        # Assume the first file is a video.
         video_path = files[0]
         frames = downsample_video(video_path)
         if not frames:
-            yield gr.Error("Could not process video.")
+            gr.Error("Could not process video.")
             return
-        # Build messages starting with the text prompt and then add each frame with its timestamp.
+        # Build messages: start with the text prompt.
         messages = [
             {
                 "role": "user",
                 "content": [{"type": "text", "text": text}]
             }
         ]
+        # Append each frame with a timestamp label.
         for image, timestamp in frames:
             messages[0]["content"].append({"type": "text", "text": f"Frame {timestamp}:"})
             messages[0]["content"].append({"type": "image", "image": image})
-        # Collect images from the frames.
+        # Collect only the images from the frames.
         video_images = [image for image, _ in frames]
+        # Prepare the prompt.
         prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
         inputs = processor(
             text=[prompt],
@@ -127,9 +100,10 @@ def model_inference(input_dict, history):
             return_tensors="pt",
             padding=True,
         ).to("cuda")
+        # Set up streaming generation.
         streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)
         generation_kwargs = dict(inputs, streamer=streamer, max_new_tokens=1024)
-        thread = Thread(target=vl_model.generate, kwargs=generation_kwargs)
+        thread = Thread(target=model.generate, kwargs=generation_kwargs)
         thread.start()
         buffer = ""
         yield progress_bar_html("Processing video with Qwen2.5VL Model")
@@ -139,82 +113,52 @@ def model_inference(input_dict, history):
             yield buffer
         return
 
-    # Multimodal branch if images are provided (non-video)
-    if files:
-        # If more than one file is provided, load them as images.
-        if len(files) > 1:
-            images = [load_image(image) for image in files]
-        elif len(files) == 1:
-            images = [load_image(files[0])]
-        else:
-            images = []
+    if len(files) > 1:
+        images = [load_image(image) for image in files]
+    elif len(files) == 1:
+        images = [load_image(files[0])]
+    else:
+        images = []
 
-        if text == "":
-            yield gr.Error("Please input a text query along with the image(s).")
-            return
-
-        messages = [
-            {
-                "role": "user",
-                "content": [
-                    *[{"type": "image", "image": image} for image in images],
-                    {"type": "text", "text": text},
-                ],
-            }
-        ]
-        prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
-        inputs = processor(
-            text=[prompt],
-            images=images,
-            return_tensors="pt",
-            padding=True,
-        ).to("cuda")
-        streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)
-        generation_kwargs = dict(inputs, streamer=streamer, max_new_tokens=1024)
-        thread = Thread(target=vl_model.generate, kwargs=generation_kwargs)
-        thread.start()
-        buffer = ""
-        yield progress_bar_html("Processing with Qwen2.5VL Model")
-        for new_text in streamer:
-            buffer += new_text
-            time.sleep(0.01)
-            yield buffer
+    if text == "" and not images:
+        gr.Error("Please input a query and optionally image(s).")
         return
-
-    # Text-only branch using DeepHermes text generation.
-    if text.strip() == "":
-        yield gr.Error("Please input a query.")
+    if text == "" and images:
+        gr.Error("Please input a text query along with the image(s).")
         return
 
-    input_ids = tg_tokenizer(text, return_tensors="pt").to(tg_model.device)
-    streamer = TextIteratorStreamer(tg_tokenizer, skip_prompt=True, skip_special_tokens=True)
-    generation_kwargs = {
-        "input_ids": input_ids,
-        "streamer": streamer,
-        "max_new_tokens": 2048,
-        "do_sample": True,
-        "top_p": 0.9,
-        "top_k": 50,
-        "temperature": 0.6,
-        "repetition_penalty": 1.2,
-    }
-    thread = Thread(target=tg_model.generate, kwargs=generation_kwargs)
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                *[{"type": "image", "image": image} for image in images],
+                {"type": "text", "text": text},
+            ],
+        }
+    ]
+    prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+    inputs = processor(
+        text=[prompt],
+        images=images if images else None,
+        return_tensors="pt",
+        padding=True,
+    ).to("cuda")
+    streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)
+    generation_kwargs = dict(inputs, streamer=streamer, max_new_tokens=1024)
+    thread = Thread(target=model.generate, kwargs=generation_kwargs)
     thread.start()
     buffer = ""
-    yield progress_bar_html("Processing text with DeepHermes Model")
+    yield progress_bar_html("Processing with Qwen2.5VL Model")
     for new_text in streamer:
         buffer += new_text
         time.sleep(0.01)
         yield buffer
 
-# -----------------------
-# Gradio Chat Interface
-# -----------------------
 examples = [
     [{"text": "Describe the Image?", "files": ["example_images/document.jpg"]}],
-    [{"text": "Tell me a story about a brave knight."}],
     [{"text": "@video-infer Explain the content of the Advertisement", "files": ["example_images/videoplayback.mp4"]}],
     [{"text": "@video-infer Explain the content of the video in detail", "files": ["example_images/breakfast.mp4"]}],
+    [{"text": "@video-infer Explain the content of the video.", "files": ["example_images/sky.mp4"]}],
 ]
 
 demo = gr.ChatInterface(
@@ -228,5 +172,4 @@ demo = gr.ChatInterface(
     cache_examples=False,
 )
 
-if __name__ == "__main__":
-    demo.launch(debug=True)
+demo.launch(debug=True)