import spaces import argparse from ast import parse import datetime import json import os import time import hashlib import re import torch import gradio as gr import requests import random from filelock import FileLock from io import BytesIO from PIL import Image, ImageDraw, ImageFont from models import load_image from constants import LOGDIR, DEFAULT_IMAGE_TOKEN from utils import ( build_logger, server_error_msg, violates_moderation, moderation_msg, load_image_from_base64, get_log_filename, ) from threading import Thread import traceback # import torch from conversation import Conversation from transformers import AutoModel, AutoTokenizer, TextIteratorStreamer import subprocess subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True) torch.set_default_device('cuda') logger = build_logger("gradio_web_server", "gradio_web_server.log") headers = {"User-Agent": "Vintern-1B-3.5-Demo Client"} no_change_btn = gr.Button() enable_btn = gr.Button(interactive=True) disable_btn = gr.Button(interactive=False) @spaces.GPU(duration=10) def make_zerogpu_happy(): pass def write2file(path, content): lock = FileLock(f"{path}.lock") with lock: with open(path, "a") as fout: fout.write(content) get_window_url_params = """ function() { const params = new URLSearchParams(window.location.search); url_params = Object.fromEntries(params); console.log(url_params); return url_params; } """ def init_state(state=None): if state is not None: del state return Conversation() def vote_last_response(state, liked, request: gr.Request): conv_data = { "tstamp": round(time.time(), 4), "like": liked, "model": 'Vintern-1B-v3_5', "state": state.dict(), "ip": request.client.host, } write2file(get_log_filename(), json.dumps(conv_data) + "\n") def upvote_last_response(state, request: gr.Request): logger.info(f"upvote. ip: {request.client.host}") vote_last_response(state, True, request) textbox = gr.MultimodalTextbox(value=None, interactive=True) return (textbox,) + (disable_btn,) * 3 def downvote_last_response(state, request: gr.Request): logger.info(f"downvote. ip: {request.client.host}") vote_last_response(state, False, request) textbox = gr.MultimodalTextbox(value=None, interactive=True) return (textbox,) + (disable_btn,) * 3 def vote_selected_response( state, request: gr.Request, data: gr.LikeData ): logger.info( f"Vote: {data.liked}, index: {data.index}, value: {data.value} , ip: {request.client.host}" ) conv_data = { "tstamp": round(time.time(), 4), "like": data.liked, "index": data.index, "model": 'Vintern-1B-v3_5', "state": state.dict(), "ip": request.client.host, } write2file(get_log_filename(), json.dumps(conv_data) + "\n") return def flag_last_response(state, request: gr.Request): logger.info(f"flag. ip: {request.client.host}") vote_last_response(state, "flag", request) textbox = gr.MultimodalTextbox(value=None, interactive=True) return (textbox,) + (disable_btn,) * 3 def regenerate(state, image_process_mode, request: gr.Request): logger.info(f"regenerate. ip: {request.client.host}") # state.messages[-1][-1] = None state.update_message(Conversation.ASSISTANT, content='', image=None, idx=-1) prev_human_msg = state.messages[-2] if type(prev_human_msg[1]) in (tuple, list): prev_human_msg[1] = (*prev_human_msg[1][:2], image_process_mode) state.skip_next = False textbox = gr.MultimodalTextbox(value=None, interactive=True) return (state, state.to_gradio_chatbot(), textbox) + (disable_btn,) * 5 def clear_history(request: gr.Request): logger.info(f"clear_history. ip: {request.client.host}") state = init_state() textbox = gr.MultimodalTextbox(value=None, interactive=True) return (state, state.to_gradio_chatbot(), textbox) + (disable_btn,) * 5 def add_text(state, message, system_prompt, request: gr.Request): if not state: state = init_state() images = message.get("files", []) text = message.get("text", "").strip() # logger.info(f"add_text. ip: {request.client.host}. len: {len(text)}") # import pdb; pdb.set_trace() textbox = gr.MultimodalTextbox(value=None, interactive=False) if len(text) <= 0 and len(images) == 0: state.skip_next = True return (state, state.to_gradio_chatbot(), textbox) + (no_change_btn,) * 5 if args.moderate: flagged = violates_moderation(text) if flagged: state.skip_next = True textbox = gr.MultimodalTextbox( value={"text": moderation_msg}, interactive=True ) return (state, state.to_gradio_chatbot(), textbox) + (no_change_btn,) * 5 images = [Image.open(path).convert("RGB") for path in images] # Init again if send the second image if len(images) > 0 and len(state.get_images(source=state.USER)) > 0: state = init_state(state) # Upload the first image if len(images) > 0 and len(state.get_images(source=state.USER)) == 0: if len(state.messages) == 0: ## In case the first message is an image text = DEFAULT_IMAGE_TOKEN + "\n" + system_prompt + "\n" + text else: ## In case the image is uploaded after some text messages first_user_message = state.messages[0]['content'] state.update_message(Conversation.USER, DEFAULT_IMAGE_TOKEN + "\n" + first_user_message, None, 0) # If the first message is text if len(images) == 0 and len(state.get_images(source=state.USER)) == 0 and len(state.messages) == 0: text = system_prompt + "\n" + text state.set_system_message(system_prompt) state.append_message(Conversation.USER, text, images) state.skip_next = False return (state, state.to_gradio_chatbot(), textbox) + ( disable_btn, ) * 5 model_name = "5CD-AI/Vintern-1B-v3_5" model = AutoModel.from_pretrained( model_name, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True, trust_remote_code=True, ).eval().cuda() tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True, use_fast=False) @spaces.GPU def predict(state, image_path, max_input_tiles=6, temperature=1.0, max_output_tokens=700, top_p=0.7, repetition_penalty=2.5, do_sample=False): # history = state.get_prompt()[:-1] # logger.info(f"==== History ====\n{history}") generation_config = dict(temperature=temperature, max_new_tokens=max_output_tokens, top_p=top_p, do_sample=do_sample, num_beams = 3, repetition_penalty=repetition_penalty) pixel_values = None if image_path is not None: pixel_values = load_image(image_path, max_num=max_input_tiles).to(torch.bfloat16).cuda() if pixel_values is not None: logger.info(f"==== Lenght Pixel values ====\n{len(pixel_values)}") # Check the first user message to see if it is an image index, first_user_message = state.get_user_message(source=state.USER, position='first') if first_user_message is not None and \ DEFAULT_IMAGE_TOKEN not in first_user_message: state.update_message(state.USER, DEFAULT_IMAGE_TOKEN + "\n" + first_user_message, None, index) history = state.get_history() logger.info(f"==== History ====\n{history}") _, message = state.get_user_message(source=state.USER, position='last') response, conv_history = model.chat(tokenizer, pixel_values, message, generation_config, history=history, return_history=True) logger.info(f"==== Conv History ====\n{conv_history}") return response, conv_history def ai_bot( state, temperature, do_sample, top_p, repetition_penalty, max_new_tokens, max_input_tiles, request: gr.Request, ): logger.info(f"ai_bot. ip: {request.client.host}") start_tstamp = time.time() if hasattr(state, "skip_next") and state.skip_next: # This generate call is skipped due to invalid inputs yield ( state, state.to_gradio_chatbot(), gr.MultimodalTextbox(interactive=False), ) + (no_change_btn,) * 5 return if model is None: state.update_message(Conversation.ASSISTANT, server_error_msg) yield ( state, state.to_gradio_chatbot(), gr.MultimodalTextbox(interactive=False), disable_btn, disable_btn, disable_btn, enable_btn, enable_btn, ) return all_images = state.get_images(source=state.USER) all_image_paths = [state.save_image(image) for image in all_images] state.append_message(Conversation.ASSISTANT, state.streaming_placeholder) yield ( state, state.to_gradio_chatbot(), gr.MultimodalTextbox(interactive=False), ) + (disable_btn,) * 5 try: # Stream output logger.info(f"==== Image paths ====\n{all_image_paths}") response, _ = predict(state, all_image_paths[0] if len(all_image_paths) > 0 else None, max_input_tiles, temperature, max_new_tokens, top_p, repetition_penalty, do_sample) # response = "This is a test response" buffer = "" for new_text in response: buffer += new_text state.update_message(Conversation.ASSISTANT, buffer + state.streaming_placeholder, None) yield ( state, state.to_gradio_chatbot(), gr.MultimodalTextbox(interactive=False), ) + (disable_btn,) * 5 except Exception as e: logger.error(f"Error in ai_bot: {e} \n{traceback.format_exc()}") state.update_message(Conversation.ASSISTANT, server_error_msg, None) yield ( state, state.to_gradio_chatbot(), gr.MultimodalTextbox(interactive=True), ) + ( disable_btn, disable_btn, disable_btn, enable_btn, enable_btn, ) return ai_response = state.return_last_message() logger.info(f"==== AI response ====\n{ai_response}") state.end_of_current_turn() yield ( state, state.to_gradio_chatbot(), gr.MultimodalTextbox(interactive=True), ) + (enable_btn,) * 5 finish_tstamp = time.time() logger.info(f"{buffer}") data = { "tstamp": round(finish_tstamp, 4), "like": None, "model": model_name, "start": round(start_tstamp, 4), "finish": round(start_tstamp, 4), "state": state.dict(), "images": all_image_paths, "ip": request.client.host, } write2file(get_log_filename(), json.dumps(data) + "\n") #

Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling

title_html = """

❄️Vintern-1B-v3_5❄️

An Efficient Multimodal Large Language Model for Vietnamese🇻🇳

[📖 Vintern Paper] [🤗 Huggingface]
""" description_html = """

Vintern-1B-v3.5 is the latest in the Vintern series, bringing major improvements over v2 across all benchmarks. This continuous fine-tuning Version enhances Vietnamese capabilities while retaining strong English performance. It excels in OCR, text recognition, and Vietnam-specific document understanding.

""" tos_markdown = """ ### Terms of use By using this service, users are required to agree to the following terms: It only provides limited safety measures and may generate offensive content. It must not be used for any illegal, harmful, violent, racist, or sexual purposes. Please click the "Flag" button if you get any inappropriate answer! We will collect those to keep improving our moderator. For an optimal experience, please use desktop computers for this demo, as mobile devices may compromise its quality. """ # .gradio-container {margin: 5px 10px 0 10px !important}; block_css = """ .gradio-container {margin: 0.1% 1% 0 1% !important; max-width: 98% !important;}; #buttons button { min-width: min(120px,100%); } .gradient-text { font-size: 28px; width: auto; font-weight: bold; background: linear-gradient(45deg, red, orange, yellow, green, blue, indigo, violet); background-clip: text; -webkit-background-clip: text; color: transparent; } .plain-text { font-size: 22px; width: auto; font-weight: bold; } """ js = """ function createWaveAnimation() { const text = document.getElementById('text'); var i = 0; setInterval(function() { const colors = [ 'red, orange, yellow, green, blue, indigo, violet, purple', 'orange, yellow, green, blue, indigo, violet, purple, red', 'yellow, green, blue, indigo, violet, purple, red, orange', 'green, blue, indigo, violet, purple, red, orange, yellow', 'blue, indigo, violet, purple, red, orange, yellow, green', 'indigo, violet, purple, red, orange, yellow, green, blue', 'violet, purple, red, orange, yellow, green, blue, indigo', 'purple, red, orange, yellow, green, blue, indigo, violet', ]; const angle = 45; const colorIndex = i % colors.length; text.style.background = `linear-gradient(${angle}deg, ${colors[colorIndex]})`; text.style.webkitBackgroundClip = 'text'; text.style.backgroundClip = 'text'; text.style.color = 'transparent'; text.style.fontSize = '28px'; text.style.width = 'auto'; text.textContent = 'Vintern-1B'; text.style.fontWeight = 'bold'; i += 1; }, 200); const params = new URLSearchParams(window.location.search); url_params = Object.fromEntries(params); // console.log(url_params); // console.log('hello world...'); // console.log(window.location.search); // console.log('hello world...'); // alert(window.location.search) // alert(url_params); return url_params; } """ def build_demo(): textbox = gr.MultimodalTextbox( interactive=True, file_types=["image", "video"], placeholder="Enter message or upload file...", show_label=False, ) with gr.Blocks( title="❄️ Vintern-1B-v3_5-Demo ❄️", theme="NoCrypt/miku", css=block_css, js=js, ) as demo: state = gr.State() with gr.Row(): with gr.Column(scale=2): gr.HTML(title_html) with gr.Accordion("Settings", open=False) as setting_row: system_prompt = gr.Textbox( value="Bạn là một trợ lý AI đa phương thức hữu ích, hãy trả lời câu hỏi người dùng một cách chi tiết.", label="System Prompt", interactive=True, ) temperature = gr.Slider( minimum=0.0, maximum=1.0, value=1.0, step=0.1, interactive=True, label="Temperature", ) do_sample = gr.Checkbox( label="Sampling", value=False, interactive=True, ) top_p = gr.Slider( minimum=0.0, maximum=1.0, value=0.9, step=0.1, interactive=True, label="Top P", ) repetition_penalty = gr.Slider( minimum=1.0, maximum=3.0, value=2.2, step=0.02, interactive=True, label="Repetition penalty", ) max_output_tokens = gr.Slider( minimum=0, maximum=4096, value=700, step=64, interactive=True, label="Max output tokens", ) max_input_tiles = gr.Slider( minimum=1, maximum=12, value=6, step=1, interactive=True, label="Max input tiles (control the image size)", ) examples = gr.Examples( examples=[ [ { "files": [ "samples/1.jpg", ], "text": "Hãy trích xuất thông tin từ hình ảnh này và trả về kết quả dạng markdown.", } ], [ { "files": [ "samples/2.png", ], "text": "Bạn là một nhà sáng tạo nội dung tài năng. Hãy viết một kịch bản quảng cáo cho sản phẩm này.", } ], [ { "files": [ "samples/3.jpeg", ], "text": "Hãy viết lại một email cho các chủ hộ về nội dung của bảng thông báo.", } ], [ { "files": [ "samples/6.jpeg", ], "text": "Hãy viết trích xuất nội dung của hoá đơn dạng JSON.", } ], ], inputs=[textbox], ) with gr.Column(scale=8): chatbot = gr.Chatbot( elem_id="chatbot", label="Vintern-1B-v3_5-Demo", height=580, show_copy_button=True, show_share_button=True, avatar_images=[ "assets/human.png", "assets/assistant.png", ], bubble_full_width=False, ) with gr.Row(): with gr.Column(scale=8): textbox.render() with gr.Column(scale=1, min_width=50): submit_btn = gr.Button(value="Send", variant="primary") with gr.Row(elem_id="buttons") as button_row: upvote_btn = gr.Button(value="👍 Upvote", interactive=False) downvote_btn = gr.Button(value="👎 Downvote", interactive=False) flag_btn = gr.Button(value="⚠️ Flag", interactive=False) # stop_btn = gr.Button(value="⏹️ Stop Generation", interactive=False) regenerate_btn = gr.Button( value="🔄 Regenerate", interactive=False ) clear_btn = gr.Button(value="🗑️ Clear", interactive=False) with gr.Row(): gr.HTML(description_html) gr.Markdown(tos_markdown) url_params = gr.JSON(visible=False) # Register listeners btn_list = [upvote_btn, downvote_btn, flag_btn, regenerate_btn, clear_btn] upvote_btn.click( upvote_last_response, [state], [textbox, upvote_btn, downvote_btn, flag_btn], ) downvote_btn.click( downvote_last_response, [state], [textbox, upvote_btn, downvote_btn, flag_btn], ) chatbot.like( vote_selected_response, [state], [], ) flag_btn.click( flag_last_response, [state], [textbox, upvote_btn, downvote_btn, flag_btn], ) regenerate_btn.click( regenerate, [state, system_prompt], [state, chatbot, textbox] + btn_list, ).then( ai_bot, [ state, temperature, do_sample, top_p, repetition_penalty, max_output_tokens, max_input_tiles, ], [state, chatbot, textbox] + btn_list, ) clear_btn.click(clear_history, None, [state, chatbot, textbox] + btn_list) textbox.submit( add_text, [state, textbox, system_prompt], [state, chatbot, textbox] + btn_list, ).then( ai_bot, [ state, temperature, do_sample, top_p, repetition_penalty, max_output_tokens, max_input_tiles, ], [state, chatbot, textbox] + btn_list, ) submit_btn.click( add_text, [state, textbox, system_prompt], [state, chatbot, textbox] + btn_list, ).then( ai_bot, [ state, temperature, do_sample, top_p, repetition_penalty, max_output_tokens, max_input_tiles, ], [state, chatbot, textbox] + btn_list, ) return demo if __name__ == "__main__": parser = argparse.ArgumentParser() parser.add_argument("--host", type=str, default="0.0.0.0") parser.add_argument("--port", type=int, default=7860) parser.add_argument("--concurrency-count", type=int, default=10) parser.add_argument("--share", action="store_true") parser.add_argument("--moderate", action="store_true") args = parser.parse_args() logger.info(f"args: {args}") logger.info(args) demo = build_demo() demo.queue(api_open=False).launch( server_name=args.host, server_port=args.port, share=args.share, max_threads=args.concurrency_count, )