|
| 1 | +import gradio as gr |
| 2 | +import copy |
| 3 | +import re |
| 4 | +from threading import Thread |
| 5 | +from transformers import TextIteratorStreamer |
| 6 | +from qwen_vl_utils import process_vision_info |
| 7 | + |
| 8 | + |
| 9 | +def _parse_text(text): |
| 10 | + lines = text.split("\n") |
| 11 | + lines = [line for line in lines if line != ""] |
| 12 | + count = 0 |
| 13 | + for i, line in enumerate(lines): |
| 14 | + if "```" in line: |
| 15 | + count += 1 |
| 16 | + items = line.split("`") |
| 17 | + if count % 2 == 1: |
| 18 | + lines[i] = f'<pre><code class="language-{items[-1]}">' |
| 19 | + else: |
| 20 | + lines[i] = "<br></code></pre>" |
| 21 | + else: |
| 22 | + if i > 0: |
| 23 | + if count % 2 == 1: |
| 24 | + line = line.replace("`", r"\`") |
| 25 | + line = line.replace("<", "<") |
| 26 | + line = line.replace(">", ">") |
| 27 | + line = line.replace(" ", " ") |
| 28 | + line = line.replace("*", "*") |
| 29 | + line = line.replace("_", "_") |
| 30 | + line = line.replace("-", "-") |
| 31 | + line = line.replace(".", ".") |
| 32 | + line = line.replace("!", "!") |
| 33 | + line = line.replace("(", "(") |
| 34 | + line = line.replace(")", ")") |
| 35 | + line = line.replace("$", "$") |
| 36 | + lines[i] = "<br>" + line |
| 37 | + text = "".join(lines) |
| 38 | + return text |
| 39 | + |
| 40 | + |
| 41 | +def _remove_image_special(text): |
| 42 | + text = text.replace("<ref>", "").replace("</ref>", "") |
| 43 | + return re.sub(r"<box>.*?(</box>|$)", "", text) |
| 44 | + |
| 45 | + |
| 46 | +def is_video_file(filename): |
| 47 | + video_extensions = [".mp4", ".avi", ".mkv", ".mov", ".wmv", ".flv", ".webm", ".mpeg"] |
| 48 | + return any(filename.lower().endswith(ext) for ext in video_extensions) |
| 49 | + |
| 50 | + |
| 51 | +def transform_messages(original_messages): |
| 52 | + transformed_messages = [] |
| 53 | + for message in original_messages: |
| 54 | + new_content = [] |
| 55 | + for item in message["content"]: |
| 56 | + if "image" in item: |
| 57 | + new_item = {"type": "image", "image": item["image"]} |
| 58 | + elif "text" in item: |
| 59 | + new_item = {"type": "text", "text": item["text"]} |
| 60 | + elif "video" in item: |
| 61 | + new_item = {"type": "video", "video": item["video"]} |
| 62 | + else: |
| 63 | + continue |
| 64 | + new_content.append(new_item) |
| 65 | + |
| 66 | + new_message = {"role": message["role"], "content": new_content} |
| 67 | + transformed_messages.append(new_message) |
| 68 | + |
| 69 | + return transformed_messages |
| 70 | + |
| 71 | + |
| 72 | +def make_demo(model, processor): |
| 73 | + def call_local_model(model, processor, messages): |
| 74 | + messages = transform_messages(messages) |
| 75 | + |
| 76 | + text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) |
| 77 | + image_inputs, video_inputs = process_vision_info(messages) |
| 78 | + inputs = processor(text=[text], images=image_inputs, videos=video_inputs, padding=True, return_tensors="pt").to(model.device) |
| 79 | + |
| 80 | + tokenizer = processor.tokenizer |
| 81 | + streamer = TextIteratorStreamer(tokenizer, timeout=3600.0, skip_prompt=True, skip_special_tokens=True) |
| 82 | + |
| 83 | + gen_kwargs = {"max_new_tokens": 512, "streamer": streamer, **inputs} |
| 84 | + |
| 85 | + thread = Thread(target=model.generate, kwargs=gen_kwargs) |
| 86 | + thread.start() |
| 87 | + |
| 88 | + generated_text = "" |
| 89 | + for new_text in streamer: |
| 90 | + generated_text += new_text |
| 91 | + yield generated_text |
| 92 | + |
| 93 | + def create_predict_fn(): |
| 94 | + def predict(_chatbot, task_history): |
| 95 | + chat_query = _chatbot[-1][0] |
| 96 | + query = task_history[-1][0] |
| 97 | + if len(chat_query) == 0: |
| 98 | + _chatbot.pop() |
| 99 | + task_history.pop() |
| 100 | + return _chatbot |
| 101 | + print("User: " + _parse_text(query)) |
| 102 | + history_cp = copy.deepcopy(task_history) |
| 103 | + full_response = "" |
| 104 | + messages = [] |
| 105 | + content = [] |
| 106 | + for q, a in history_cp: |
| 107 | + if isinstance(q, (tuple, list)): |
| 108 | + if is_video_file(q[0]): |
| 109 | + content.append({"video": f"file://{q[0]}"}) |
| 110 | + else: |
| 111 | + content.append({"image": f"file://{q[0]}"}) |
| 112 | + else: |
| 113 | + content.append({"text": q}) |
| 114 | + messages.append({"role": "user", "content": content}) |
| 115 | + messages.append({"role": "assistant", "content": [{"text": a}]}) |
| 116 | + content = [] |
| 117 | + messages.pop() |
| 118 | + |
| 119 | + for response in call_local_model(model, processor, messages): |
| 120 | + _chatbot[-1] = (_parse_text(chat_query), _remove_image_special(_parse_text(response))) |
| 121 | + |
| 122 | + yield _chatbot |
| 123 | + full_response = _parse_text(response) |
| 124 | + |
| 125 | + task_history[-1] = (query, full_response) |
| 126 | + print("Qwen-VL-Chat: " + _parse_text(full_response)) |
| 127 | + yield _chatbot |
| 128 | + |
| 129 | + return predict |
| 130 | + |
| 131 | + def create_regenerate_fn(): |
| 132 | + def regenerate(_chatbot, task_history): |
| 133 | + if not task_history: |
| 134 | + return _chatbot |
| 135 | + item = task_history[-1] |
| 136 | + if item[1] is None: |
| 137 | + return _chatbot |
| 138 | + task_history[-1] = (item[0], None) |
| 139 | + chatbot_item = _chatbot.pop(-1) |
| 140 | + if chatbot_item[0] is None: |
| 141 | + _chatbot[-1] = (_chatbot[-1][0], None) |
| 142 | + else: |
| 143 | + _chatbot.append((chatbot_item[0], None)) |
| 144 | + _chatbot_gen = predict(_chatbot, task_history) |
| 145 | + for _chatbot in _chatbot_gen: |
| 146 | + yield _chatbot |
| 147 | + |
| 148 | + return regenerate |
| 149 | + |
| 150 | + predict = create_predict_fn() |
| 151 | + regenerate = create_regenerate_fn() |
| 152 | + |
| 153 | + def add_text(history, task_history, text): |
| 154 | + task_text = text |
| 155 | + history = history if history is not None else [] |
| 156 | + task_history = task_history if task_history is not None else [] |
| 157 | + history = history + [(_parse_text(text), None)] |
| 158 | + task_history = task_history + [(task_text, None)] |
| 159 | + return history, task_history, "" |
| 160 | + |
| 161 | + def add_file(history, task_history, file): |
| 162 | + history = history if history is not None else [] |
| 163 | + task_history = task_history if task_history is not None else [] |
| 164 | + history = history + [((file.name,), None)] |
| 165 | + task_history = task_history + [((file.name,), None)] |
| 166 | + return history, task_history |
| 167 | + |
| 168 | + def reset_user_input(): |
| 169 | + return gr.update(value="") |
| 170 | + |
| 171 | + def reset_state(task_history): |
| 172 | + task_history.clear() |
| 173 | + return [] |
| 174 | + |
| 175 | + with gr.Blocks() as demo: |
| 176 | + gr.Markdown("""<center><font size=8>Qwen2.5-VL OpenVINO demo</center>""") |
| 177 | + |
| 178 | + chatbot = gr.Chatbot(label="Qwen2-VL", elem_classes="control-height", height=500) |
| 179 | + query = gr.Textbox(lines=2, label="Input") |
| 180 | + task_history = gr.State([]) |
| 181 | + |
| 182 | + with gr.Row(): |
| 183 | + addfile_btn = gr.UploadButton("📁 Upload (上传文件)", file_types=["image", "video"]) |
| 184 | + submit_btn = gr.Button("🚀 Submit (发送)") |
| 185 | + regen_btn = gr.Button("🤔️ Regenerate (重试)") |
| 186 | + empty_bin = gr.Button("🧹 Clear History (清除历史)") |
| 187 | + |
| 188 | + submit_btn.click(add_text, [chatbot, task_history, query], [chatbot, task_history]).then( |
| 189 | + predict, [chatbot, task_history], [chatbot], show_progress=True |
| 190 | + ) |
| 191 | + submit_btn.click(reset_user_input, [], [query]) |
| 192 | + empty_bin.click(reset_state, [task_history], [chatbot], show_progress=True) |
| 193 | + regen_btn.click(regenerate, [chatbot, task_history], [chatbot], show_progress=True) |
| 194 | + addfile_btn.upload(add_file, [chatbot, task_history, addfile_btn], [chatbot, task_history], show_progress=True) |
| 195 | + |
| 196 | + gr.Markdown( |
| 197 | + """\ |
| 198 | +<font size=2>Note: This demo is governed by the original license of Qwen2.5-VL. \ |
| 199 | +We strongly advise users not to knowingly generate or allow others to knowingly generate harmful content, \ |
| 200 | +including hate speech, violence, pornography, deception, etc. \ |
| 201 | +(注：本演示受Qwen2-VL的许可协议限制。我们强烈建议，用户不应传播及不应允许他人传播以下内容，\ |
| 202 | +包括但不限于仇恨言论、暴力、色情、欺诈相关的有害信息。)""" |
| 203 | + ) |
| 204 | + |
| 205 | + return demo |
0 commit comments