import subprocess subprocess.run('pip install flash-attn==2.7.0.post2 --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True) import spaces import argparse import os import re import logging from typing import List, Optional, Tuple, Generator from threading import Thread import gradio as gr import PIL.Image import torch import numpy as np from moviepy.editor import VideoFileClip from transformers import AutoModelForCausalLM, TextIteratorStreamer logging.getLogger("httpx").setLevel(logging.WARNING) logging.basicConfig(level=logging.INFO) logger = logging.getLogger(__name__) # --- Global Model Variables --- model = None streamer = None # This should point to the directory containing your SVG file. CUR_DIR = os.path.dirname(os.path.abspath(__file__)) def submit_chat(chatbot, text_input): response = '' chatbot.append([text_input, response]) return chatbot, '' # --- Helper Functions --- latex_delimiters_set = [ { "left": "\\(", "right": "\\)", "display": False }, { "left": "\\begin{equation}", "right": "\\end{equation}", "display": True }, { "left": "\\begin{align}", "right": "\\end{align}", "display": True }, { "left": "\\begin{alignat}", "right": "\\end{alignat}", "display": True }, { "left": "\\begin{gather}", "right": "\\end{gather}", "display": True }, { "left": "\\begin{CD}", "right": "\\end{CD}", "display": True }, { "left": "\\[", "right": "\\]", "display": True } ] def load_video_frames(video_path: Optional[str], n_frames: int = 8) -> Optional[List[PIL.Image.Image]]: """Extracts a specified number of frames from a video file.""" if not video_path: return None try: with VideoFileClip(video_path) as clip: total_frames = int(clip.fps * clip.duration) if total_frames <= 0: return None num_to_extract = min(n_frames, total_frames) indices = np.linspace(0, total_frames - 1, num_to_extract, dtype=int) frames = [PIL.Image.fromarray(clip.get_frame(index / clip.fps)) for index in indices] return frames except Exception as e: print(f"Error processing video {video_path}: {e}") return None def parse_model_output(response_text: str, enable_thinking: bool) -> str: """Formats the model output, separating 'thinking' and 'response' parts if enabled.""" if enable_thinking: # Use a more robust regex to handle nested content and variations think_match = re.search(r"(.*?)", response_text, re.DOTALL) if think_match: thinking_content = think_match.group(1).strip() # Remove the think block from the original text to get the response response_content = re.sub(r".*?", "", response_text, flags=re.DOTALL).strip() return f"**Thinking:**\n```\n{thinking_content}\n```\n\n**Response:**\n{response_content}" else: return response_text # No think tag found, return as is else: # If thinking is disabled, strip the tags just in case the model still generates them return re.sub(r".*?", "", response_text, flags=re.DOTALL).strip() # --- MODIFIED Core Inference Logic (Now with Streaming) --- @spaces.GPU def run_inference( chatbot: List, image_input: Optional[PIL.Image.Image], video_input: Optional[str], do_sample: bool, max_new_tokens: int, enable_thinking: bool, ): """ Runs a single turn of inference and yields the output stream for a gr.Chatbot. This function is now a generator. """ prompt = chatbot[-1][0] if (not image_input and not video_input and not prompt) or not prompt: gr.Warning("A text prompt is required for generation.") # MODIFICATION: Yield the current state and return to avoid errors yield chatbot return # MODIFICATION: Append the new prompt to the existing history # chatbot.append([prompt, ""]) # yield chatbot, "" # Yield the updated chat to show the user's prompt immediately content = [] if image_input: content.append({"type": "image", "image": image_input}) if video_input: frames = load_video_frames(video_input) if frames: content.append({"type": "video", "video": frames}) else: gr.Warning("Failed to process the video file.") chatbot[-1][1] = "Error: Could not process the video file." yield chatbot return content.append({"type": "text", "text": prompt}) messages = [{"role": "user", "content": content}] logger.info(messages) try: if video_input: input_ids, pixel_values, grid_thws = model.preprocess_inputs(messages=messages, add_generation_prompt=True, enable_thinking=enable_thinking, max_pixels=896*896) else: input_ids, pixel_values, grid_thws = model.preprocess_inputs(messages=messages, add_generation_prompt=True, enable_thinking=enable_thinking) except Exception as e: chatbot[-1][1] = f"Error during input preprocessing: {e}" yield chatbot return input_ids = input_ids.to(model.device) if pixel_values is not None: pixel_values = pixel_values.to(model.device, dtype=torch.bfloat16) if grid_thws is not None: grid_thws = grid_thws.to(model.device) gen_kwargs = { "max_new_tokens": max_new_tokens, "do_sample": do_sample, "eos_token_id": model.text_tokenizer.eos_token_id, "pad_token_id": model.text_tokenizer.pad_token_id, "streamer": streamer, "use_cache": True } with torch.inference_mode(): thread = Thread(target=model.generate, kwargs={ "inputs": input_ids, "pixel_values": pixel_values, "grid_thws": grid_thws, **gen_kwargs }) thread.start() # MODIFICATION: Stream output token by token response_text = "" for new_text in streamer: response_text += new_text # Append only the new text chunk to the last response chatbot[-1][1] = response_text yield chatbot # Yield the updated history thread.join() # MODIFICATION: Format the final response once generation is complete formatted_response = parse_model_output(response_text, enable_thinking) chatbot[-1][1] = formatted_response yield chatbot # Yield the final, formatted response logger.info("[OVIS_CONV_START]") [print(f'Q{i}:\n {request}\nA{i}:\n {answer}') for i, (request, answer) in enumerate(chatbot, 1)] # print('New_Q:\n', text_input) # print('New_A:\n', response) logger.info("[OVIS_CONV_END]") def clear_chat(): return [], None, "" # --- UI Helper Functions --- def toggle_media_input(choice: str) -> Tuple: """Switches visibility between Image/Video inputs and their corresponding examples.""" if choice == "Image": return gr.update(visible=True, value=None), gr.update(visible=False, value=None), gr.update(visible=True), gr.update(visible=False) else: # Video return gr.update(visible=False, value=None), gr.update(visible=True, value=None), gr.update(visible=False), gr.update(visible=True) # --- Build Gradio Application --- # @spaces.GPU def build_demo(model_path: str): """Builds the Gradio user interface for the model.""" global model, streamer device = "cuda" print(f"Loading model {model_path} onto device {device}...") model = AutoModelForCausalLM.from_pretrained( model_path, torch_dtype=torch.bfloat16, trust_remote_code=True ).to(device).eval() text_tokenizer = model.text_tokenizer streamer = TextIteratorStreamer(text_tokenizer, skip_prompt=True, skip_special_tokens=True) print("Model loaded successfully.") model_name_display = model_path.split('/')[-1] logo_html = "" logo_svg_path = os.path.join(CUR_DIR, "resource", "logo.svg") if os.path.exists(logo_svg_path): with open(logo_svg_path, "r", encoding="utf-8") as svg_file: svg_content = svg_file.read() font_size = "2.5em" svg_content_styled = re.sub(r'(]*)(>)', rf'\1 height="{font_size}" style="vertical-align: middle; display: inline-block;"\2', svg_content) logo_html = f'{svg_content_styled}' else: logo_html = 'Ovis' print(f"Warning: Logo file not found at {logo_svg_path}. Using text fallback.") html_header = f"""

{logo_html} {model_name_display}

Ovis has been open-sourced on 😊 Huggingface and 🌟 GitHub. If you find Ovis useful, a like❤️ or a star🌟 would be appreciated.
""" prompt_input = gr.Textbox(label="Prompt", placeholder="Enter your text here and press ENTER", lines=1, container=False) with gr.Blocks(theme=gr.themes.Ocean()) as demo: gr.HTML(html_header) gr.Markdown("Note: you might have to increase \"Max New Tokens\" and wait longer to obtain answer when Deep Thinking is enabled.") with gr.Row(): with gr.Column(scale=4): input_type_radio = gr.Radio(choices=["Image", "Video"], value="Image", label="Select Input Type") image_input = gr.Image(label="Image Input", type="pil", visible=True) video_input = gr.Video(label="Video Input", visible=False) with gr.Accordion("Generation Settings", open=True): do_sample = gr.Checkbox(label="Enable Sampling (Do Sample)", value=True) max_new_tokens = gr.Slider(minimum=32, maximum=4096, value=2048, step=32, label="Max New Tokens") enable_thinking = gr.Checkbox(label="Enable Deep Thinking", value=False) with gr.Column(visible=True) as image_examples_col: gr.Examples( examples=[ [os.path.join(CUR_DIR, "examples", "ovis2_math0.jpg"), "Each face of the polyhedron shown is either a triangle or a square. Each square borders 4 triangles, and each triangle borders 3 squares. The polyhedron has 6 squares. How many triangles does it have?\n\nEnd your response with 'Final answer: '."], [os.path.join(CUR_DIR, "examples", "ovis2_math1.jpg"), "A large square touches another two squares, as shown in the picture. The numbers inside the smaller squares indicate their areas. What is the area of the largest square?\n\nEnd your response with 'Final answer: '."], [os.path.join(CUR_DIR, "examples", "ovis2_figure0.png"), "Explain this model."], # [os.path.join(CUR_DIR, "examples", "ovis2_figure1.png"), "Organize the notes about GRPO in the figure."], [os.path.join(CUR_DIR, "examples", "ovis2_multi0.jpg"), "Posso avere un frappuccino e un caffè americano di taglia M? Quanto costa in totale?"], ], inputs=[image_input, prompt_input] ) with gr.Column(visible=False) as video_examples_col: gr.Examples(examples=[[os.path.join(CUR_DIR, "examples", "video_demo.mp4"), "Describe the video."]], inputs=[video_input, prompt_input]) with gr.Column(scale=7): chatbot = gr.Chatbot(label="Ovis", height=600, show_copy_button=True, layout="panel", latex_delimiters=latex_delimiters_set) prompt_input.render() with gr.Row(): generate_btn = gr.Button("Send", variant="primary") clear_btn = gr.Button("Clear", variant="secondary") input_type_radio.change( fn=toggle_media_input, inputs=input_type_radio, outputs=[image_input, video_input, image_examples_col, video_examples_col] ) # MODIFICATION: Update event handlers to use the new function and manage state run_inputs = [chatbot, image_input, video_input, do_sample, max_new_tokens, enable_thinking] # run_outputs = [image_input, prompt_input] generat_click_event = generate_btn.click(submit_chat, [chatbot, prompt_input], [chatbot, prompt_input]).then(run_inference, run_inputs, chatbot) submit_event = prompt_input.submit(submit_chat, [chatbot, prompt_input], [chatbot, prompt_input]).then(run_inference, run_inputs, chatbot) clear_btn.click( fn=lambda: ([], None, None, "", "Image", True, 1024, False), outputs=[chatbot, image_input, video_input, prompt_input, input_type_radio, do_sample, max_new_tokens, enable_thinking] ).then( fn=toggle_media_input, inputs=input_type_radio, outputs=[image_input, video_input, image_examples_col, video_examples_col] ) return demo # --- Main Execution Block --- # def parse_args(): # parser = argparse.ArgumentParser(description="Gradio interface for a single Multimodal Large Language Model.") # parser.add_argument("--model-path", type=str, default='AIDC-AI/Ovis2.5-9B', help="Path to the model checkpoint on Hugging Face Hub or local directory.") # parser.add_argument("--gpu", type=int, default=0, help="GPU index to run the model on.") # parser.add_argument("--port", type=int, default=7860, help="Port to run the Gradio server on.") # parser.add_argument("--server-name", type=str, default="0.0.0.0", help="Server name for the Gradio app.") # return parser.parse_args() # if __name__ == "__main__": # args = parse_args() model_path = 'AIDC-AI/Ovis2.5-9B' demo = build_demo(model_path=model_path) # demo = build_demo(model_path=args.model_path) # demo.launch(server_name=args.server_name, server_port=args.port, share=False, ssl_verify=False, show_error=True) demo.queue().launch()