import gradio as gr import json import copy import pandas as pd from vocab import tokenizer_factory from character_util import iter_vocab from utils.log_util import logger from functools import lru_cache default_user_input = """\ Replace this text in the input field to see how tokenization works. Buenos días! 华为发布Mate60手机。 ラグビーワールドカップ2023フランス""" # default_tokenizer_name_1 = "Meta/llama3" default_tokenizer_name_1 = "gradientai/Llama-3-8B-Instruct-Gradient-1048k" default_tokenizer_name_2 = "openai/gpt-4" @lru_cache def _tokenize( text: str, tokenizer_name: str, color_num: int = 5, add_special_token: bool = False ): logger.info("param=" + json.dumps({"text": text, "tokenizer_type": tokenizer_name}, ensure_ascii=False)) pos_tokens = [] tokenizer = tokenizer_factory.get_tokenizer(tokenizer_name) if add_special_token: encoding = tokenizer.encode(text, add_special_tokens=True) else: encoding = tokenizer.encode(text, add_special_tokens=False) table = [] for idx, token_id in enumerate(encoding): decode_text = tokenizer.decode([token_id]) # 特殊字符解码后会统一变成 �，对应 "\ufffd" pos_tokens.extend([(decode_text, str(idx % color_num))]) # token "Byte": # 这是 utf-8编码吧？ token = tokenizer.convert_ids_to_tokens([token_id], skip_special_tokens=False)[0] if isinstance(token, bytes): try: token_str = token.decode("utf-8") except: token_str = token.decode("utf-8", errors="ignore") logger.error(f"{idx}: decode_error: " + json.dumps( # gpt_35_turbo 经常有token会decode error，这里用来记录一下 {"tokenizer_type": tokenizer_name, "token": str(token), "token_str": token_str}, ensure_ascii=False)) token_bytes = token # json_dumps = json.dumps(token_str) elif isinstance(token, str): token_str = token token_bytes = bytes(token_str, "utf-8") # json_dumps = json.dumps(token_str) else: logger.error(f"{idx}: wrong type for token {token_id} {type(token)} " + json.dumps( {"text": text, "tokenizer_type": tokenizer_name}, ensure_ascii=False)) token_str = token token_bytes = token # continue # ⭐ # TODO: gpt3.5_turbo错误： 只有id和text是对的，token和 utf8都是错的。说明 convert_ids_to_tokens 出错了。 table.append( {"TokenID": token_id, "Token": token_str, # utf-8解码后的字符串，为什么有些是 <0xE7>，表示什么？比如llama "Text": decode_text, # # "Bytes": token_bytes, # bytes类型在gradio前端页面被解码成字符串，比如 b'\xe4\xb8\xad' 仍然显示成 "中"。因此 str(token_bytes) "UTF8 Bytes": str(token_bytes), # "Unicode": json_dumps # unicode, 如果是ascii码，就直接显示。如果不是ascii码，就显示unicode } ) table_df = pd.DataFrame(table) logger.info(f"tokenizer_type={tokenizer_name}, Tokens={table[:4]}") return pos_tokens, len(encoding), table_df def tokenize( text: str, tokenizer_name: str, color_num: int = 5, add_special_token: bool = False ): """ tokenize wrapper As gr.Update would be overwritten after passing to frontend, we apply lru_cache in _tokenize. """ pos_tokens, num_tokens, table_df = _tokenize(text, tokenizer_name, color_num, add_special_token) return gr.update(value=pos_tokens, label=f"Tokens: {num_tokens}"), table_df def tokenize_pair(text, tokenizer_type_1, tokenizer_type_2): """ input_text.change """ pos_tokens_1, table_df_1 = tokenize(text, tokenizer_type_1) pos_tokens_2, table_df_2 = tokenize(text, tokenizer_type_2) return pos_tokens_1, table_df_1, pos_tokens_2, table_df_2 @lru_cache def basic_count(tokenizer_name): stats = iter_vocab(tokenizer_name) return stats['vocab_size'], f'{stats["organization"]}' # return tokenizer.vocab_size, f'{stats["中文汉字数"]["中文单字"]}/{stats["中文汉字数"]["中文多字"]}' # def get_compress_rate(tokenizer_name, all_corpus, unit): # tokenizer = tokenizer_factory.get_tokenizer(tokenizer_name) # compress_rate_stats = tokenize_corpus(tokenizer, all_corpus) # compress_rate = unit_convertor(compress_rate_stats, unit) # return compress_rate @lru_cache def get_overlap_token_size(tokenizer_name_1, tokenizer_name_2): tokenizer1 = tokenizer_factory.get_tokenizer(tokenizer_name_1) tokenizer2 = tokenizer_factory.get_tokenizer(tokenizer_name_2) vocab_set_1 = tokenizer1.get_vocab().keys() vocab_set_2 = tokenizer2.get_vocab().keys() token1 = next(iter(vocab_set_1)) token2 = next(iter(vocab_set_2)) if type(token1) != type(token2): # bytes str if isinstance(token1, str): vocab_set_1 = set([token.encode("utf-8") for token in vocab_set_1]) if isinstance(token2, str): vocab_set_2 = set([token.encode("utf-8") for token in vocab_set_2]) overlap_tokens = vocab_set_1 & vocab_set_2 overlap_token_size = len(overlap_tokens) logger.info( f"{overlap_token_size} OverlapTokens of {tokenizer_name_1} {tokenizer_name_2}: {list(overlap_tokens)[:10]}") return overlap_token_size, overlap_token_size def on_load(url_params, request: gr.Request): """ onLoad """ text = None tokenizer_type_1 = None tokenizer_type_2 = None try: url_params = json.loads(url_params) except: url_params = {} if request: logger.info(str(request.headers)) client_ip = request.client.host # local_ip = socket.gethostbyname(socket.gethostbyname("")) # headers = request.kwargs['headers'] # if headers and 'x-forwarded-for' in headers: # x_forwarded_for = headers['x-forwarded-for'] # client_ip = x_forwarded_for.split(' ')[0] if x_forwarded_for else "" # if "referer" in request.headers: # not work for huggingface-space # url_params = parse_qs(urlparse(request.headers["referer"]).query) # url_params = {k: v[0] for k, v in url_params.items() if len(v) > 0} tokenizer_type_1 = url_params.get("tokenizer1", default_tokenizer_name_1) tokenizer_type_2 = url_params.get("tokenizer2", default_tokenizer_name_2) text = url_params.get("text", default_user_input) logger.info(f"client_ip: {client_ip}; params: {url_params}") return text, tokenizer_type_1, tokenizer_type_2 # def compress_rate_unit_change(unit): # return gr.update(label=f"Compress Rate: {unit}"), gr.update(label=f"Compress Rate: {unit}"), def test_coding(): bytes1 = b'\xe4\xb8\xad' print(bytes1) # b'\xe4\xb8\xad' if __name__ == "__main__": print(get_overlap_token_size("gpt-35-turbo", "gpt-4")) # print(basic_count("internlm_chat_7b"))