update app

app.py

@@ -5,6 +5,7 @@ import json
 import time
 import asyncio
 from threading import Thread
+from typing import Iterable
 
 import gradio as gr
 import spaces
@@ -15,14 +16,75 @@ import cv2
 
 from transformers import (
     Qwen2_5_VLForConditionalGeneration,
-    AutoModel,
+    Qwen3VLForConditionalGeneration,
     AutoTokenizer,
     AutoProcessor,
     TextIteratorStreamer,
 )
 from transformers.image_utils import load_image
+from gradio.themes import Soft
+from gradio.themes.utils import colors, fonts, sizes
+
+colors.thistle = colors.Color(
+    name="thistle",
+    c50="#F9F5F9", c100="#F0E8F1", c200="#E7DBE8", c300="#DECEE0",
+    c400="#D2BFD8", c500="#D8BFD8", c600="#B59CB7", c700="#927996",
+    c800="#6F5675", c900="#4C3454", c950="#291233",
+)
+
+class ThistleTheme(Soft):
+    def __init__(
+        self,
+        *,
+        primary_hue: colors.Color | str = colors.gray,
+        secondary_hue: colors.Color | str = colors.thistle,
+        neutral_hue: colors.Color | str = colors.slate,
+        text_size: sizes.Size | str = sizes.text_lg,
+        font: fonts.Font | str | Iterable[fonts.Font | str] = (
+            fonts.GoogleFont("Outfit"), "Arial", "sans-serif",
+        ),
+        font_mono: fonts.Font | str | Iterable[fonts.Font | str] = (
+            fonts.GoogleFont("IBM Plex Mono"), "ui-monospace", "monospace",
+        ),
+    ):
+        super().__init__(
+            primary_hue=primary_hue,
+            secondary_hue=secondary_hue,
+            neutral_hue=neutral_hue,
+            text_size=text_size,
+            font=font,
+            font_mono=font_mono,
+        )
+        super().set(
+            background_fill_primary="*primary_50",
+            background_fill_primary_dark="*primary_900",
+            body_background_fill="linear-gradient(135deg, *primary_200, *primary_100)",
+            body_background_fill_dark="linear-gradient(135deg, *primary_900, *primary_800)",
+            button_primary_text_color="black",
+            button_primary_text_color_hover="white",
+            button_primary_background_fill="linear-gradient(90deg, *secondary_400, *secondary_400)",
+            button_primary_background_fill_hover="linear-gradient(90deg, *secondary_600, *secondary_600)",
+            button_primary_background_fill_dark="linear-gradient(90deg, *secondary_600, *secondary_800)",
+            button_primary_background_fill_hover_dark="linear-gradient(90deg, *secondary_500, *secondary_500)",
+            button_secondary_text_color="black",
+            button_secondary_text_color_hover="white",
+            button_secondary_background_fill="linear-gradient(90deg, *primary_300, *primary_300)",
+            button_secondary_background_fill_hover="linear-gradient(90deg, *primary_400, *primary_400)",
+            button_secondary_background_fill_dark="linear-gradient(90deg, *primary_500, *primary_600)",
+            button_secondary_background_fill_hover_dark="linear-gradient(90deg, *primary_500, *primary_500)",
+            slider_color="*secondary_300",
+            slider_color_dark="*secondary_600",
+            block_title_text_weight="600",
+            block_border_width="3px",
+            block_shadow="*shadow_drop_lg",
+            button_primary_shadow="*shadow_drop_lg",
+            button_large_padding="11px",
+            color_accent_soft="*primary_100",
+            block_label_background_fill="*primary_200",
+        )
+
+thistle_theme = ThistleTheme()
 
-# Constants for text generation
 MAX_MAX_NEW_TOKENS = 2048
 DEFAULT_MAX_NEW_TOKENS = 1024
 MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "4096"))
@@ -47,19 +109,19 @@ model_x = Qwen2_5_VLForConditionalGeneration.from_pretrained(
     torch_dtype=torch.float16
 ).to(device).eval()
 
-# Load Qwen2.5-VL-7B-Abliterated-Caption-it
-MODEL_ID_Q = "prithivMLmods/Qwen2.5-VL-7B-Abliterated-Caption-it"
+# Load Qwen3-VL-4B-Instruct
+MODEL_ID_Q = "Qwen/Qwen3-VL-4B-Instruct"
 processor_q = AutoProcessor.from_pretrained(MODEL_ID_Q, trust_remote_code=True)
-model_q = Qwen2_5_VLForConditionalGeneration.from_pretrained(
+model_q = Qwen3VLForConditionalGeneration.from_pretrained(
     MODEL_ID_Q,
     trust_remote_code=True,
     torch_dtype=torch.float16
 ).to(device).eval()
 
-# Load Lumian2-VLR-7B-Thinking
-MODEL_ID_Y = "prithivMLmods/Lumian2-VLR-7B-Thinking"
+# Load Qwen3-VL-8B-Instruct
+MODEL_ID_Y = "Qwen/Qwen3-VL-8B-Instruct"
 processor_y = AutoProcessor.from_pretrained(MODEL_ID_Y, trust_remote_code=True)
-model_y = Qwen2_5_VLForConditionalGeneration.from_pretrained(
+model_y = Qwen3VLForConditionalGeneration.from_pretrained(
     MODEL_ID_Y,
     trust_remote_code=True,
     torch_dtype=torch.float16
@@ -74,7 +136,8 @@ def downsample_video(video_path):
     total_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))
     fps = vidcap.get(cv2.CAP_PROP_FPS)
     frames = []
-    frame_indices = np.linspace(0, total_frames - 1, 10, dtype=int)
+    # Use a maximum of 10 frames to avoid excessive memory usage
+    frame_indices = np.linspace(0, total_frames - 1, min(total_frames, 10), dtype=int)
     for i in frame_indices:
         vidcap.set(cv2.CAP_PROP_POS_FRAMES, i)
         success, image = vidcap.read()
@@ -95,20 +158,15 @@ def generate_image(model_name: str, text: str, image: Image.Image,
                    repetition_penalty: float = 1.2):
     """
     Generates responses using the selected model for image input.
-    Yields raw text and Markdown-formatted text.
     """
     if model_name == "Qwen2.5-VL-7B-Instruct":
-        processor = processor_m
-        model = model_m
+        processor, model = processor_m, model_m
     elif model_name == "Qwen2.5-VL-3B-Instruct":
-        processor = processor_x
-        model = model_x
-    elif model_name == "Qwen2.5-VL-7B-Abliterated-Caption-it":
-        processor = processor_q
-        model = model_q
-    elif model_name == "Lumian2-VLR-7B-Thinking":
-        processor = processor_y
-        model = model_y
+        processor, model = processor_x, model_x
+    elif model_name == "Qwen3-VL-4B-Instruct":
+        processor, model = processor_q, model_q
+    elif model_name == "Qwen3-VL-8B-Instruct":
+        processor, model = processor_y, model_y
     else:
         yield "Invalid model selected.", "Invalid model selected."
         return
@@ -117,21 +175,11 @@ def generate_image(model_name: str, text: str, image: Image.Image,
         yield "Please upload an image.", "Please upload an image."
         return
 
-    messages = [{
-        "role": "user",
-        "content": [
-            {"type": "image", "image": image},
-            {"type": "text", "text": text},
-        ]
-    }]
+    messages = [{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": text}]}]
     prompt_full = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
     inputs = processor(
-        text=[prompt_full],
-        images=[image],
-        return_tensors="pt",
-        padding=True,
-        truncation=False,
-        max_length=MAX_INPUT_TOKEN_LENGTH
+        text=[prompt_full], images=[image], return_tensors="pt", padding=True,
+        truncation=True, max_length=MAX_INPUT_TOKEN_LENGTH
     ).to(device)
     streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)
     generation_kwargs = {**inputs, "streamer": streamer, "max_new_tokens": max_new_tokens}
@@ -152,20 +200,15 @@ def generate_video(model_name: str, text: str, video_path: str,
                    repetition_penalty: float = 1.2):
     """
     Generates responses using the selected model for video input.
-    Yields raw text and Markdown-formatted text.
     """
     if model_name == "Qwen2.5-VL-7B-Instruct":
-        processor = processor_m
-        model = model_m
+        processor, model = processor_m, model_m
     elif model_name == "Qwen2.5-VL-3B-Instruct":
-        processor = processor_x
-        model = model_x
-    elif model_name == "Qwen2.5-VL-7B-Abliterated-Caption-it":
-        processor = processor_q
-        model = model_q
-    elif model_name == "Lumian2-VLR-7B-Thinking":
-        processor = processor_y
-        model = model_y
+        processor, model = processor_x, model_x
+    elif model_name == "Qwen3-VL-4B-Instruct":
+        processor, model = processor_q, model_q
+    elif model_name == "Qwen3-VL-8B-Instruct":
+        processor, model = processor_y, model_y
     else:
         yield "Invalid model selected.", "Invalid model selected."
         return
@@ -174,43 +217,38 @@ def generate_video(model_name: str, text: str, video_path: str,
         yield "Please upload a video.", "Please upload a video."
         return
 
-    frames = downsample_video(video_path)
-    messages = [
-        {"role": "system", "content": [{"type": "text", "text": "You are a helpful assistant."}]},
-        {"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})
-    inputs = processor.apply_chat_template(
-        messages,
-        tokenize=True,
-        add_generation_prompt=True,
-        return_dict=True,
-        return_tensors="pt",
-        truncation=False,
-        max_length=MAX_INPUT_TOKEN_LENGTH
+    frames_with_ts = downsample_video(video_path)
+    if not frames_with_ts:
+        yield "Could not process video.", "Could not process video."
+        return
+
+    messages = [{"role": "user", "content": [{"type": "text", "text": text}]}]
+    images_for_processor = []
+    for frame, timestamp in frames_with_ts:
+        messages[0]["content"].append({"type": "image"})
+        images_for_processor.append(frame)
+
+    prompt_full = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+    inputs = processor(
+        text=[prompt_full], images=images_for_processor, return_tensors="pt", padding=True,
+        truncation=True, max_length=MAX_INPUT_TOKEN_LENGTH
     ).to(device)
     streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)
     generation_kwargs = {
-        **inputs,
-        "streamer": streamer,
-        "max_new_tokens": max_new_tokens,
-        "do_sample": True,
-        "temperature": temperature,
-        "top_p": top_p,
-        "top_k": top_k,
-        "repetition_penalty": repetition_penalty,
+        **inputs, "streamer": streamer, "max_new_tokens": max_new_tokens,
+        "do_sample": True, "temperature": temperature, "top_p": top_p,
+        "top_k": top_k, "repetition_penalty": repetition_penalty,
     }
     thread = Thread(target=model.generate, kwargs=generation_kwargs)
     thread.start()
     buffer = ""
     for new_text in streamer:
         buffer += new_text
+        buffer = buffer.replace("<|im_end|>", "")
         time.sleep(0.01)
         yield buffer, buffer
 
+
 # Define examples for image and video inference
 image_examples = [
     ["Explain the content in detail.", "images/D.jpg"],
@@ -228,42 +266,32 @@ video_examples = [
 ]
 
 css = """
-.submit-btn {
-    background-color: #2980b9 !important;
-    color: white !important;
-}
-.submit-btn:hover {
-    background-color: #3498db !important;
+#main-title h1 {
+    font-size: 2.3em !important;
 }
-.canvas-output {
-    border: 2px solid #4682B4;
-    border-radius: 10px;
-    padding: 20px;
+#output-title h2 {
+    font-size: 2.1em !important;
 }
 """
 
 # Create the Gradio Interface
-with gr.Blocks(css=css, theme="bethecloud/storj_theme") as demo:
-    gr.Markdown("# **[Qwen2.5-VL-Outpost](https://huggingface.co/collections/prithivMLmods/multimodal-implementations-67c9982ea04b39f0608badb0)**")
+with gr.Blocks(css=css, theme=thistle_theme) as demo:
+    gr.Markdown("# **Qwen3-VL-Outpost**", elem_id="main-title")
     with gr.Row():
-        with gr.Column():
+        with gr.Column(scale=2):
             with gr.Tabs():
                 with gr.TabItem("Image Inference"):
                     image_query = gr.Textbox(label="Query Input", placeholder="Enter your query here...")
                     image_upload = gr.Image(type="pil", label="Image", height=290)
-                    image_submit = gr.Button("Submit", elem_classes="submit-btn")
-                    gr.Examples(
-                        examples=image_examples,
-                        inputs=[image_query, image_upload]
-                    )
+                    image_submit = gr.Button("Submit", variant="primary")
+                    gr.Examples(examples=image_examples, inputs=[image_query, image_upload])
+                    
                 with gr.TabItem("Video Inference"):
                     video_query = gr.Textbox(label="Query Input", placeholder="✦︎ Enter your query here...")
                     video_upload = gr.Video(label="Video", height=290)
-                    video_submit = gr.Button("Submit", elem_classes="submit-btn")
-                    gr.Examples(
-                        examples=video_examples,
-                        inputs=[video_query, video_upload]
-                    ) 
+                    video_submit = gr.Button("Submit", variant="primary")
+                    gr.Examples(examples=video_examples, inputs=[video_query, video_upload])
+                    
             with gr.Accordion("Advanced options", open=False):
                 max_new_tokens = gr.Slider(label="Max new tokens", minimum=1, maximum=MAX_MAX_NEW_TOKENS, step=1, value=DEFAULT_MAX_NEW_TOKENS)
                 temperature = gr.Slider(label="Temperature", minimum=0.1, maximum=4.0, step=0.1, value=0.6)
@@ -271,33 +299,18 @@ with gr.Blocks(css=css, theme="bethecloud/storj_theme") as demo:
                 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():
-            with gr.Column(elem_classes="canvas-output"):
-                gr.Markdown("## Output")
-                output = gr.Textbox(label="Raw Output", interactive=False, lines=5, show_copy_button=True)
-            
-                with gr.Accordion("(Result.md)", open=False):
-                    markdown_output = gr.Markdown()
+        with gr.Column(scale=3):
+            gr.Markdown("## Output", elem_id="output-title")
+            output = gr.Textbox(label="Raw Output", interactive=False, lines=14, show_copy_button=True)
+            with gr.Accordion("(Result.md)", open=False):
+                markdown_output = gr.Markdown()
                 
             model_choice = gr.Radio(
-                choices=["Qwen2.5-VL-7B-Instruct", "Qwen2.5-VL-3B-Instruct", "Lumian2-VLR-7B-Thinking", "Qwen2.5-VL-7B-Abliterated-Caption-it"],
+                choices=["Qwen3-VL-4B-Instruct", "Qwen3-VL-8B-Instruct", "Qwen2.5-VL-3B-Instruct", "Qwen2.5-VL-7B-Instruct"],
                 label="Select Model",
-                value="Qwen2.5-VL-7B-Instruct"
-            )
-            gr.Markdown("**Model Info 💻** | [Report Bug](https://huggingface.co/spaces/prithivMLmods/Qwen2.5-VL/discussions)")
-            
-            gr.Markdown(
-                """
-            > [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct): The Qwen2.5-VL-7B-Instruct model is a multimodal AI model developed by Alibaba Cloud that excels at understanding both text and images. It's a Vision-Language Model (VLM) designed to handle various visual understanding tasks, including image understanding, video analysis, and even multilingual support.  
-            >
-            > [Qwen2.5-VL-7B-Abliterated-Caption-it](prithivMLmods/Qwen2.5-VL-7B-Abliterated-Caption-it): Qwen2.5-VL-7B-Abliterated-Caption-it is a fine-tuned version of Qwen2.5-VL-7B-Instruct, optimized for Abliterated Captioning / Uncensored Captioning. This model excels at generating detailed, context-rich, and high-fidelity captions across diverse image categories and variational aspect ratios, offering robust visual understanding without filtering or censorship.  
-                """
+                value="Qwen3-VL-4B-Instruct"
             )
 
-            gr.Markdown("> [Lumian2-VLR-7B-Thinking](https://huggingface.co/prithivMLmods/Lumian2-VLR-7B-Thinking): The Lumian2-VLR-7B-Thinking model is a high-fidelity vision-language reasoning (experimental model) system designed for fine-grained multimodal understanding. Built on Qwen2.5-VL-7B-Instruct, this model enhances image captioning, sampled video reasoning, and document comprehension through explicit grounded reasoning. It produces structured reasoning traces aligned with visual coordinates, enabling explainable multimodal reasoning.")
-        
-            gr.Markdown(">⚠️note: all the models in space are not guaranteed to perform well in video inference use cases.")
-            
     image_submit.click(
         fn=generate_image,
         inputs=[model_choice, image_query, image_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],
@@ -310,4 +323,4 @@ with gr.Blocks(css=css, theme="bethecloud/storj_theme") as demo:
     )
 
 if __name__ == "__main__":
-    demo.queue(max_size=50).launch(share=True, mcp_server=True, ssr_mode=False, show_error=True)
+    demo.queue(max_size=50).launch(mcp_server=True, ssr_mode=False, show_error=True)