import hashlib import os import re import json import gradio as gr from chat_template import ChatTemplate from llama_cpp_python_streamingllm import StreamingLLM # ========== 让聊天界面的文本框等高 ========== custom_css = r''' #area > div { height: 100%; } #RAG-area { flex-grow: 1; } #RAG-area > label { height: 100%; display: flex; flex-direction: column; } #RAG-area > label > textarea { flex-grow: 1; max-height: 20vh; } #VO-area { flex-grow: 1; } #VO-area > label { height: 100%; display: flex; flex-direction: column; } #VO-area > label > textarea { flex-grow: 1; max-height: 20vh; } #prompt > label > textarea { max-height: 63px; } ''' # ========== 适配 SillyTavern 的模版 ========== def text_format(text: str, _env=None, **env): if _env is not None: for k, v in _env.items(): text = text.replace(r'{{' + k + r'}}', v) for k, v in env.items(): text = text.replace(r'{{' + k + r'}}', v) return text # ========== 哈希函数 ========== def x_hash(x: str): return hashlib.sha1(x.encode('utf-8')).hexdigest() # ========== 读取配置文件 ========== with open('rp_config.json', encoding='utf-8') as f: tmp = f.read() with open('rp_sample_config.json', encoding='utf-8') as f: cfg = json.load(f) cfg['setting_cache_path']['value'] += x_hash(tmp) cfg.update(json.loads(tmp)) # ========== 给引号加粗 ========== reg_q = re.compile(r'“(.+?)”') def chat_display_format(text: str): return reg_q.sub(r' **\g<0>** ', text) # ========== 温度、采样之类的设置 ========== with gr.Blocks() as setting: with gr.Row(): setting_path = gr.Textbox(label="模型路径", max_lines=1, scale=2, **cfg['setting_path']) setting_cache_path = gr.Textbox(label="缓存路径", max_lines=1, scale=2, **cfg['setting_cache_path']) setting_seed = gr.Number(label="随机种子", scale=1, **cfg['setting_seed']) setting_n_gpu_layers = gr.Number(label="n_gpu_layers", scale=1, **cfg['setting_n_gpu_layers']) with gr.Row(): setting_ctx = gr.Number(label="上下文大小(Tokens)", **cfg['setting_ctx']) setting_max_tokens = gr.Number(label="最大响应长度(Tokens)", interactive=True, **cfg['setting_max_tokens']) setting_n_keep = gr.Number(value=10, label="n_keep", interactive=False) setting_n_discard = gr.Number(label="n_discard", interactive=True, **cfg['setting_n_discard']) with gr.Row(): setting_temperature = gr.Number(label="温度", interactive=True, **cfg['setting_temperature']) setting_repeat_penalty = gr.Number(label="重复惩罚", interactive=True, **cfg['setting_repeat_penalty']) setting_frequency_penalty = gr.Number(label="频率惩罚", interactive=True, **cfg['setting_frequency_penalty']) setting_presence_penalty = gr.Number(label="存在惩罚", interactive=True, **cfg['setting_presence_penalty']) setting_repeat_last_n = gr.Number(label="惩罚范围", interactive=True, **cfg['setting_repeat_last_n']) with gr.Row(): setting_top_k = gr.Number(label="Top-K", interactive=True, **cfg['setting_top_k']) setting_top_p = gr.Number(label="Top P", interactive=True, **cfg['setting_top_p']) setting_min_p = gr.Number(label="Min P", interactive=True, **cfg['setting_min_p']) setting_typical_p = gr.Number(label="Typical", interactive=True, **cfg['setting_typical_p']) setting_tfs_z = gr.Number(label="TFS", interactive=True, **cfg['setting_tfs_z']) with gr.Row(): setting_mirostat_mode = gr.Number(label="Mirostat 模式", **cfg['setting_mirostat_mode']) setting_mirostat_eta = gr.Number(label="Mirostat 学习率", interactive=True, **cfg['setting_mirostat_eta']) setting_mirostat_tau = gr.Number(label="Mirostat 目标熵", interactive=True, **cfg['setting_mirostat_tau']) # ========== 下载模型 ========== if os.path.exists(setting_path.value): print(f"The file {setting_path.value} exists.") else: from huggingface_hub import snapshot_download os.mkdir("downloads") os.mkdir("cache") snapshot_download(repo_id='TheBloke/CausalLM-7B-GGUF', local_dir=r'downloads', allow_patterns='causallm_7b.Q5_K_M.gguf') snapshot_download(repo_id='Limour/llama-python-streamingllm-cache', repo_type='dataset', local_dir=r'cache') # ========== 加载模型 ========== model = StreamingLLM(model_path=setting_path.value, seed=setting_seed.value, n_gpu_layers=setting_n_gpu_layers.value, n_ctx=setting_ctx.value) setting_ctx.value = model.n_ctx() # ========== 聊天的模版 默认 chatml ========== chat_template = ChatTemplate(model) # ========== 展示角色卡 ========== with gr.Blocks() as role: with gr.Row(): role_usr = gr.Textbox(label="用户名称", max_lines=1, interactive=False, **cfg['role_usr']) role_char = gr.Textbox(label="角色名称", max_lines=1, interactive=False, **cfg['role_char']) role_char_d = gr.Textbox(lines=10, label="故事描述", **cfg['role_char_d']) role_chat_style = gr.Textbox(lines=10, label="回复示例", **cfg['role_chat_style']) # model.eval_t([1]) # 这个暖机的 bos [1] 删了就不正常了 if os.path.exists(setting_cache_path.value): # ========== 加载角色卡-缓存 ========== tmp = model.load_session(setting_cache_path.value) print(f'load cache from {setting_cache_path.value} {tmp}') tmp = chat_template('system', text_format(role_char_d.value, char=role_char.value, user=role_usr.value)) setting_n_keep.value = len(tmp) tmp = chat_template(role_char.value, text_format(role_chat_style.value, char=role_char.value, user=role_usr.value)) setting_n_keep.value += len(tmp) # ========== 加载角色卡-第一条消息 ========== chatbot = [] for one in cfg["role_char_first"]: one['name'] = text_format(one['name'], char=role_char.value, user=role_usr.value) one['value'] = text_format(one['value'], char=role_char.value, user=role_usr.value) if one['name'] == role_char.value: chatbot.append((None, chat_display_format(one['value']))) print(one) else: # ========== 加载角色卡-角色描述 ========== tmp = chat_template('system', text_format(role_char_d.value, char=role_char.value, user=role_usr.value)) setting_n_keep.value = model.eval_t(tmp) # 此内容永久存在 # ========== 加载角色卡-回复示例 ========== tmp = chat_template(role_char.value, text_format(role_chat_style.value, char=role_char.value, user=role_usr.value)) setting_n_keep.value = model.eval_t(tmp) # 此内容永久存在 # ========== 加载角色卡-第一条消息 ========== chatbot = [] for one in cfg["role_char_first"]: one['name'] = text_format(one['name'], char=role_char.value, user=role_usr.value) one['value'] = text_format(one['value'], char=role_char.value, user=role_usr.value) if one['name'] == role_char.value: chatbot.append((None, chat_display_format(one['value']))) print(one) tmp = chat_template(one['name'], one['value']) model.eval_t(tmp) # 此内容随上下文增加将被丢弃 # ========== 保存角色卡-缓存 ========== with open(setting_cache_path.value, 'wb') as f: pass tmp = model.save_session(setting_cache_path.value) print(f'save cache {tmp}') # ========== 上传缓存 ========== from huggingface_hub import login, CommitScheduler login(token=os.environ.get("HF_TOKEN"), write_permission=True) CommitScheduler(repo_id='Limour/llama-python-streamingllm-cache', repo_type='dataset', folder_path='cache') # ========== 流式输出函数 ========== def btn_submit_com(_n_keep, _n_discard, _temperature, _repeat_penalty, _frequency_penalty, _presence_penalty, _repeat_last_n, _top_k, _top_p, _min_p, _typical_p, _tfs_z, _mirostat_mode, _mirostat_eta, _mirostat_tau, _role, _max_tokens): # ========== 初始化输出模版 ========== t_bot = chat_template(_role) completion_tokens = [] # 有可能多个 tokens 才能构成一个 utf-8 编码的文字 history = '' # ========== 流式输出 ========== for token in model.generate_t( tokens=t_bot, n_keep=_n_keep, n_discard=_n_discard, im_start=chat_template.im_start_token, top_k=_top_k, top_p=_top_p, min_p=_min_p, typical_p=_typical_p, temp=_temperature, repeat_penalty=_repeat_penalty, repeat_last_n=_repeat_last_n, frequency_penalty=_frequency_penalty, presence_penalty=_presence_penalty, tfs_z=_tfs_z, mirostat_mode=_mirostat_mode, mirostat_tau=_mirostat_tau, mirostat_eta=_mirostat_eta, ): if token in chat_template.eos or token == chat_template.nlnl: t_bot.extend(completion_tokens) print('token in eos', token) break completion_tokens.append(token) all_text = model.str_detokenize(completion_tokens) if not all_text: continue t_bot.extend(completion_tokens) history += all_text yield history if token in chat_template.onenl: # ========== 移除末尾的换行符 ========== if t_bot[-2] in chat_template.onenl: model.venv_pop_token() break if t_bot[-2] in chat_template.onerl and t_bot[-3] in chat_template.onenl: model.venv_pop_token() break if history[-2:] == '\n\n': # 各种 'x\n\n' 的token,比如'。\n\n' print('t_bot[-4:]', t_bot[-4:], repr(model.str_detokenize(t_bot[-4:])), repr(model.str_detokenize(t_bot[-1:]))) break if len(t_bot) > _max_tokens: break completion_tokens = [] # ========== 查看末尾的换行符 ========== print('history', repr(history)) # ========== 给 kv_cache 加上输出结束符 ========== model.eval_t(chat_template.im_end_nl, _n_keep, _n_discard) t_bot.extend(chat_template.im_end_nl) # ========== 显示用户消息 ========== def btn_submit_usr(message: str, history): # print('btn_submit_usr', message, history) if history is None: history = [] return "", history + [[message.strip(), '']], gr.update(interactive=False) # ========== 模型流式响应 ========== def btn_submit_bot(history, _n_keep, _n_discard, _temperature, _repeat_penalty, _frequency_penalty, _presence_penalty, _repeat_last_n, _top_k, _top_p, _min_p, _typical_p, _tfs_z, _mirostat_mode, _mirostat_eta, _mirostat_tau, _usr, _char, _rag, _max_tokens): # ========== 需要临时注入的内容 ========== rag_idx = None if len(_rag) > 0: rag_idx = model.venv_create() # 记录 venv_idx t_rag = chat_template('system', _rag) model.eval_t(t_rag, _n_keep, _n_discard) model.venv_create() # 与 t_rag 隔离 # ========== 用户输入 ========== t_msg = history[-1][0] t_msg = chat_template(_usr, t_msg) model.eval_t(t_msg, _n_keep, _n_discard) # ========== 模型输出 ========== _tmp = btn_submit_com(_n_keep, _n_discard, _temperature, _repeat_penalty, _frequency_penalty, _presence_penalty, _repeat_last_n, _top_k, _top_p, _min_p, _typical_p, _tfs_z, _mirostat_mode, _mirostat_eta, _mirostat_tau, _char, _max_tokens) for _h in _tmp: history[-1][1] = _h yield history, str((model.n_tokens, model.venv)) # ========== 输出完毕后格式化输出 ========== history[-1][1] = chat_display_format(history[-1][1]) yield history, str((model.n_tokens, model.venv)) # ========== 及时清理上一次生成的旁白 ========== if vo_idx > 0: print('vo_idx', vo_idx, model.venv) model.venv_remove(vo_idx) print('vo_idx', vo_idx, model.venv) if rag_idx and vo_idx < rag_idx: rag_idx -= 1 # ========== 响应完毕后清除注入的内容 ========== if rag_idx is not None: model.venv_remove(rag_idx) # 销毁对应的 venv model.venv_disband() # 退出隔离环境 yield history, str((model.n_tokens, model.venv)) print('venv_disband', vo_idx, model.venv) # ========== 待实现 ========== def btn_rag_(_rag, _msg): retn = '' return retn vo_idx = 0 # ========== 输出一段旁白 ========== def btn_submit_vo(_n_keep, _n_discard, _temperature, _repeat_penalty, _frequency_penalty, _presence_penalty, _repeat_last_n, _top_k, _top_p, _min_p, _typical_p, _tfs_z, _mirostat_mode, _mirostat_eta, _mirostat_tau, _max_tokens): global vo_idx vo_idx = model.venv_create() # 创建隔离环境 # ========== 模型输出旁白 ========== _tmp = btn_submit_com(_n_keep, _n_discard, _temperature, _repeat_penalty, _frequency_penalty, _presence_penalty, _repeat_last_n, _top_k, _top_p, _min_p, _typical_p, _tfs_z, _mirostat_mode, _mirostat_eta, _mirostat_tau, '旁白', _max_tokens) for _h in _tmp: yield _h, str((model.n_tokens, model.venv)) # ========== 给用户提供默认回复 ========== def btn_submit_suggest(_n_keep, _n_discard, _temperature, _repeat_penalty, _frequency_penalty, _presence_penalty, _repeat_last_n, _top_k, _top_p, _min_p, _typical_p, _tfs_z, _mirostat_mode, _mirostat_eta, _mirostat_tau, _usr, _max_tokens): model.venv_create() # 创建隔离环境 # ========== 模型输出 ========== _tmp = btn_submit_com(_n_keep, _n_discard, _temperature, _repeat_penalty, _frequency_penalty, _presence_penalty, _repeat_last_n, _top_k, _top_p, _min_p, _typical_p, _tfs_z, _mirostat_mode, _mirostat_eta, _mirostat_tau, _usr, _max_tokens) _h = '' for _h in _tmp: yield _h, str((model.n_tokens, model.venv)) model.venv_remove() # 销毁隔离环境 yield _h, str((model.n_tokens, model.venv)) # ========== 聊天页面 ========== with gr.Blocks() as chatting: with gr.Row(equal_height=True): chatbot = gr.Chatbot(height='60vh', scale=2, value=chatbot, avatar_images=(r'assets/user.png', r'assets/chatbot.webp')) with gr.Column(scale=1, elem_id="area"): rag = gr.Textbox(label='RAG', show_copy_button=True, elem_id="RAG-area") vo = gr.Textbox(label='VO', show_copy_button=True, elem_id="VO-area") s_info = gr.Textbox(value=str((model.n_tokens, model.venv)), max_lines=1, label='info', interactive=False) msg = gr.Textbox(label='Prompt', lines=2, max_lines=2, elem_id='prompt', autofocus=True, **cfg['msg']) with gr.Row(): btn_rag = gr.Button("RAG") btn_submit = gr.Button("Submit") btn_retry = gr.Button("Retry") btn_com1 = gr.Button("自定义1") btn_com2 = gr.Button("自定义2") btn_com3 = gr.Button("自定义3") btn_rag.click(fn=btn_rag_, outputs=rag, inputs=[rag, msg]) btn_submit.click( fn=btn_submit_usr, api_name="submit", inputs=[msg, chatbot], outputs=[msg, chatbot, btn_submit] ).then( fn=btn_submit_bot, inputs=[chatbot, setting_n_keep, setting_n_discard, setting_temperature, setting_repeat_penalty, setting_frequency_penalty, setting_presence_penalty, setting_repeat_last_n, setting_top_k, setting_top_p, setting_min_p, setting_typical_p, setting_tfs_z, setting_mirostat_mode, setting_mirostat_eta, setting_mirostat_tau, role_usr, role_char, rag, setting_max_tokens], outputs=[chatbot, s_info] ).then( fn=btn_submit_vo, inputs=[setting_n_keep, setting_n_discard, setting_temperature, setting_repeat_penalty, setting_frequency_penalty, setting_presence_penalty, setting_repeat_last_n, setting_top_k, setting_top_p, setting_min_p, setting_typical_p, setting_tfs_z, setting_mirostat_mode, setting_mirostat_eta, setting_mirostat_tau, setting_max_tokens], outputs=[vo, s_info] ).then( fn=btn_submit_suggest, inputs=[setting_n_keep, setting_n_discard, setting_temperature, setting_repeat_penalty, setting_frequency_penalty, setting_presence_penalty, setting_repeat_last_n, setting_top_k, setting_top_p, setting_min_p, setting_typical_p, setting_tfs_z, setting_mirostat_mode, setting_mirostat_eta, setting_mirostat_tau, role_usr, setting_max_tokens], outputs=[msg, s_info] ).then( fn=lambda: gr.update(interactive=True), outputs=btn_submit ) # ========== 用于调试 ========== btn_com1.click(fn=lambda: model.str_detokenize(model._input_ids), outputs=rag) @btn_com2.click(inputs=setting_cache_path, outputs=s_info) def btn_com2(_cache_path): _tmp = model.load_session(setting_cache_path.value) print(f'load cache from {setting_cache_path.value} {_tmp}') global vo_idx vo_idx = 0 model.venv = [0] return str((model.n_tokens, model.venv)) # ========== 开始运行 ========== demo = gr.TabbedInterface([chatting, setting, role], ["聊天", "设置", '角色'], css=custom_css) gr.close_all() demo.queue().launch(share=False)