请增加大模型api调试日志

[shiiuen01]

相同的api参数，在chatbox里可以正常调用，但程序中显示连不上，不知道是什么原因

(epubtranslate) (base) [wenxy@DESKTOP-DABEL8H epubtranslate]$ /home/wenxy/miniconda3/envs/epubtranslate/bin/python /home/wenxy/epubtranslate/main.py
Translating: 0%| | 0/1.0 [00:00<?, ?it/s]
request failed with parsing error, retrying... (1 times)
request failed with parsing error, retrying... (2 times)
request failed with parsing error, retrying... (3 times)
request failed with parsing error, retrying... (4 times)
request failed with parsing error, retrying... (5 times)
request failed with parsing error, retrying... (6 times)
Traceback (most recent call last):
File "/home/wenxy/epubtranslate/main.py", line 17, in
translate(
File "/home/wenxy/miniconda3/envs/epubtranslate/lib/python3.10/site-packages/epub_translator/translator.py", line 30, in translate
_Translator(
File "/home/wenxy/miniconda3/envs/epubtranslate/lib/python3.10/site-packages/epub_translator/translator.py", line 78, in do
self._translate_spine(
File "/home/wenxy/miniconda3/envs/epubtranslate/lib/python3.10/site-packages/epub_translator/translator.py", line 115, in _translate_spine
for translated_text in _translate(
File "/home/wenxy/miniconda3/envs/epubtranslate/lib/python3.10/site-packages/epub_translator/translation/translation.py", line 65, in translate
yield from _sort_translated_texts_by_chunk(
File "/home/wenxy/miniconda3/envs/epubtranslate/lib/python3.10/site-packages/epub_translator/translation/translation.py", line 81, in _sort_translated_texts_by_chunk
for chunk, translated_texts in target:
File "/home/wenxy/miniconda3/envs/epubtranslate/lib/python3.10/site-packages/epub_translator/translation/translation.py", line 63, in _generate_chunks_from_futures
raise err
File "/home/wenxy/miniconda3/envs/epubtranslate/lib/python3.10/site-packages/epub_translator/translation/translation.py", line 58, in _generate_chunks_from_futures
yield future.result()
File "/home/wenxy/miniconda3/envs/epubtranslate/lib/python3.10/concurrent/futures/_base.py", line 451, in result
return self.__get_result()
File "/home/wenxy/miniconda3/envs/epubtranslate/lib/python3.10/concurrent/futures/_base.py", line 403, in __get_result
raise self._exception
File "/home/wenxy/miniconda3/envs/epubtranslate/lib/python3.10/concurrent/futures/thread.py", line 58, in run
result = self.fn(*self.args, **self.kwargs)
File "/home/wenxy/miniconda3/envs/epubtranslate/lib/python3.10/site-packages/epub_translator/translation/translation.py", line 41, in
executor.submit(lambda chunk=chunk: (chunk, _translate_chunk(
File "/home/wenxy/miniconda3/envs/epubtranslate/lib/python3.10/site-packages/epub_translator/translation/translation.py", line 122, in _translate_chunk
for text in _translate_texts(
File "/home/wenxy/miniconda3/envs/epubtranslate/lib/python3.10/site-packages/epub_translator/translation/translation.py", line 180, in _translate_texts
return llm.request_xml(
File "/home/wenxy/miniconda3/envs/epubtranslate/lib/python3.10/site-packages/epub_translator/llm/node.py", line 120, in request_xml
return self._executor.request(
File "/home/wenxy/miniconda3/envs/epubtranslate/lib/python3.10/site-packages/epub_translator/llm/executor.py", line 102, in request
raise last_error
File "/home/wenxy/miniconda3/envs/epubtranslate/lib/python3.10/site-packages/epub_translator/llm/executor.py", line 76, in request
result = parser(response)
File "/home/wenxy/miniconda3/envs/epubtranslate/lib/python3.10/site-packages/epub_translator/llm/node.py", line 117, in parse_response
raise ValueError("No valid XML response found")
ValueError: No valid XML response found

[Moskize91]

这个问题是 LLM 返回了非法结构的 response，可能的原因很多，比如 DeepSeek 可能截断内容等。一般过一段时间重试可能就好了。如果还有问题，方便的话把你运行 translate 的参数和要翻译的原始 epub 文件发出来，我看看能不能重现。

[Moskize91]

epub-translator/scripts/translate_epub.py

Line 17 in f81cd0d

log_dir_path=temp_path / "log",

参考这个文件。这里设置一个文件夹，会把 LLM 调用日志文件写进去。

[pavlozt]

The product generally works, but I only have the reasoning "deepseek-r1-distill-llama-70b" from the available models.
"deepseek-chat" is not available.

Maybe in this case the xml is not formed correctly?

@Moskize91 What do you think?

[Moskize91]

I haven't tested this model, but I have tested it on the another R1 version of DeepSeek. The XML can be generated correctly, but it consumes more tokens and takes more time. @pavlozt

[pavlozt]

Here is a typical file with such an error.

request 2025-09-02 06-30-12 101497.log

Changing the parameters doesn't help much. There is some problem with xml.

Here is the code that I run:

from epub_translator import LLM
import os
from dotenv import load_dotenv

load_dotenv()

llm = LLM(
  key=os.getenv('MODEL_ACCESS_KEY'), # LLM's API key
  url = "https://inference.do-ai.run/v1/", # LLM's base URL
  #model="deepseek-chat", # LLM's model name
  model="deepseek-r1-distill-llama-70b", # LLM's model name
  token_encoding="o200k_base", # Local model for calculating the number of tokens
  temperature=0.2, # Temperature
  retry_times=2, # Retry times. If the request still fails after this number, an error will be reported
  retry_interval_seconds=6.0, # Retry interval (in seconds)
  log_dir_path="./log/",
)


from epub_translator import translate, Language

translate(
  llm=llm, # llm object constructed in the previous step
  source_path="books/book-pythontips-com-en-latest.epub", # Original EPUB file to be translated
  translated_path="books/book-pythontips-rus-latest.epub", # Path to save the translated EPUB
  target_language=Language.RUSSIAN,
  max_chunk_tokens_count=10000,

)

[pavlozt]

It looks like this provider (digitalocean) is cutting off responses? Why might this happen?

[pavlozt]

It seems that this problem is somehow related to reasoning (<think>) #25

I also can’t turn off the thinking mode.

[Moskize91]

@pavlozt I've been perplexed by the "cutting off" issue. I recently heard a rumor from deepSeek: deploying Deep Seek is complex, and the official price is too low. This makes it difficult for third-party vendors to properly implement it, and these issues can lead to various issues with third-party Deep Seek. So why don't third-party vendors address this issue? Because their profit margins are so thin, selling Deep Seek doesn't make much money, so they naturally don't want their engineers to spend the effort to overcome these deployment issues.

So, is it possible to switch to the official Deep Seek? Yes, but it's too slow. In other words, Deep Seek refuses to profit, and you can't pay more to get Deep Seek to provide you with better service alone (this situation will only improve when Deep Seek has a more stable chip). I'm considering making Gemini Flash an optional option in the future.

[pavlozt]

@Moskize91 The market does not always behave like a market.
You can see for yourself what models Digital Ocean provides. There is no Gemini Flash.
Other people may have different reasons for using reasoning models.