From cd6a1fd399121994a6e75bc2c8427050c003023e Mon Sep 17 00:00:00 2001 From: qingxu fu <505030475@qq.com> Date: Fri, 14 Apr 2023 13:52:56 +0800 Subject: [PATCH] =?UTF-8?q?=E5=BD=93=E6=97=A0=E6=B3=95=E6=AD=A3=E5=B8=B8?= =?UTF-8?q?=E5=88=87=E5=89=B2PDF=E6=96=87=E6=A1=A3=E6=97=B6=EF=BC=8C?= =?UTF-8?q?=E5=BC=BA=E5=88=B6=E5=88=87=E5=89=B2?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- crazy_functions/crazy_utils.py | 41 +++++++++++++++++++++++++--------- request_llm/bridge_chatgpt.py | 2 +- 2 files changed, 32 insertions(+), 11 deletions(-) diff --git a/crazy_functions/crazy_utils.py b/crazy_functions/crazy_utils.py index 6806d4b..189d948 100644 --- a/crazy_functions/crazy_utils.py +++ b/crazy_functions/crazy_utils.py @@ -104,7 +104,7 @@ def request_gpt_model_in_new_thread_with_ui_alive( mutable[0] += f"[Local Message] 警告，在执行过程中遭遇问题, Traceback：\n\n{tb_str}\n\n" if retry_op > 0: retry_op -= 1 - mutable[0] += f"[Local Message] 重试中 {retry_times_at_unknown_error-retry_op}/{retry_times_at_unknown_error}：\n\n" + mutable[0] += f"[Local Message] 重试中，请稍等 {retry_times_at_unknown_error-retry_op}/{retry_times_at_unknown_error}：\n\n" if "Rate limit reached" in tb_str: time.sleep(30) time.sleep(5) @@ -312,7 +312,6 @@ def breakdown_txt_to_satisfy_token_limit(txt, get_token_fn, limit): if get_token_fn(prev) < limit: break if cnt == 0: - print('what the fuck ?') raise RuntimeError("存在一行极长的文本！") # print(len(post)) # 列表递归接龙 @@ -325,8 +324,18 @@ def breakdown_txt_to_satisfy_token_limit(txt, get_token_fn, limit): return cut(txt, must_break_at_empty_line=False) +def force_breakdown(txt, limit, get_token_fn): + """ + 当无法用标点、空行分割时，我们用最暴力的方法切割 + """ + for i in reversed(range(len(txt))): + if get_token_fn(txt[:i]) < limit: + return txt[:i], txt[i:] + return "Tiktoken未知错误", "Tiktoken未知错误" + def breakdown_txt_to_satisfy_token_limit_for_pdf(txt, get_token_fn, limit): - def cut(txt_tocut, must_break_at_empty_line): # 递归 + # 递归 + def cut(txt_tocut, must_break_at_empty_line, break_anyway=False): if get_token_fn(txt_tocut) <= limit: return [txt_tocut] else: @@ -338,28 +347,40 @@ def breakdown_txt_to_satisfy_token_limit_for_pdf(txt, get_token_fn, limit): if must_break_at_empty_line: if lines[cnt] != "": continue - print(cnt) prev = "\n".join(lines[:cnt]) post = "\n".join(lines[cnt:]) if get_token_fn(prev) < limit: break if cnt == 0: - # print('what the fuck ? 存在一行极长的文本！') - raise RuntimeError("存在一行极长的文本！") + if break_anyway: + prev, post = force_breakdown(txt_tocut, limit, get_token_fn) + else: + raise RuntimeError(f"存在一行极长的文本！{txt_tocut}") # print(len(post)) # 列表递归接龙 result = [prev] - result.extend(cut(post, must_break_at_empty_line)) + result.extend(cut(post, must_break_at_empty_line, break_anyway=break_anyway)) return result try: + # 第1次尝试，将双空行（\n\n）作为切分点 return cut(txt, must_break_at_empty_line=True) except RuntimeError: try: + # 第2次尝试，将单空行（\n）作为切分点 return cut(txt, must_break_at_empty_line=False) except RuntimeError: - # 这个中文的句号是故意的，作为一个标识而存在 - res = cut(txt.replace('.', '。\n'), must_break_at_empty_line=False) - return [r.replace('。\n', '.') for r in res] + try: + # 第3次尝试，将英文句号（.）作为切分点 + res = cut(txt.replace('.', '。\n'), must_break_at_empty_line=False) # 这个中文的句号是故意的，作为一个标识而存在 + return [r.replace('。\n', '.') for r in res] + except RuntimeError as e: + try: + # 第4次尝试，将中文句号（。）作为切分点 + res = cut(txt.replace('。', '。。\n'), must_break_at_empty_line=False) + return [r.replace('。。\n', '。') for r in res] + except RuntimeError as e: + # 第5次尝试，没办法了，随便切一下敷衍吧 + return cut(txt, must_break_at_empty_line=False, break_anyway=True) diff --git a/request_llm/bridge_chatgpt.py b/request_llm/bridge_chatgpt.py index 1839dcf..e9dfc6b 100644 --- a/request_llm/bridge_chatgpt.py +++ b/request_llm/bridge_chatgpt.py @@ -96,7 +96,7 @@ def predict_no_ui_long_connection(inputs, llm_kwargs, history=[], sys_prompt="", # 看门狗，如果超过期限没有喂狗，则终止 if len(observe_window) >= 2: if (time.time()-observe_window[1]) > watch_dog_patience: - raise RuntimeError("程序终止。") + raise RuntimeError("用户取消了程序。") else: raise RuntimeError("意外Json结构："+delta) if json_data['finish_reason'] == 'length': raise ConnectionAbortedError("正常结束，但显示Token不足，导致输出不完整，请削减单次输入的文本量。")