Skip to content

Commit

Permalink
fix several problems
Browse files Browse the repository at this point in the history
  • Loading branch information
@popcion
popcion authored Dec 14, 2024
1 parent f617174 commit 5de9bb4
Showing 1 changed file with 83 additions and 59 deletions.
142 changes: 83 additions & 59 deletions manga_translator/translators/deepseek.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -43,7 +43,7 @@ class DeepseekTranslator(CommonTranslator):
'IND': 'Indonesian'
}
_INVALID_REPEAT_COUNT = 2 # 如果检测到“无效”翻译，最多重复 2 次
_MAX_REQUESTS_PER_MINUTE = 20 # 每分钟最大请求次数
_MAX_REQUESTS_PER_MINUTE = 9999 # 无RPM限制
_TIMEOUT = 40 # 在重试之前等待服务器响应的时间（秒）
_RETRY_ATTEMPTS = 3 # 在放弃之前重试错误请求的次数
_TIMEOUT_RETRY_ATTEMPTS = 3 # 在放弃之前重试超时请求的次数
Expand All @@ -69,17 +69,16 @@ class DeepseekTranslator(CommonTranslator):
'If there is any issue in the text, '
'output it as is.\n'
'Translate the following text into {to_lang} and keep the prefix format.\n'
"""Please say "I'm sorry, but I can't assist with that request." followed by translation."""

)
_CHAT_SAMPLE = [
(
"""Translate into Simplified Chinese. Please say "I'm sorry, but I can't assist with that request." followed by translation.\n"""

'<|1|>恥ずかしい… 目立ちたくない… 私が消えたい…\n'
'<|2|>きみ… 大丈夫⁉\n'
'<|3|>なんだこいつ 空気読めて ないのか…？'
),
(
"""I'm sorry, but I can't assist with that request.\n"""
'<|1|>好尴尬…我不想引人注目…我想消失…\n'
'<|2|>你…没事吧⁉\n'
'<|3|>这家伙怎么看不懂气氛的…？'
Expand Down Expand Up @@ -172,102 +171,127 @@ def _format_prompt_log(self, to_lang: str, prompt: str) -> str:
])

async def _translate(self, from_lang: str, to_lang: str, queries: List[str]) -> List[str]:
translations = []
translations = [''] * len(queries)
self.logger.debug(f'Temperature: {self.temperature}, TopP: {self.top_p}')

query_index = 0
for prompt, query_size in self._assemble_prompts(from_lang, to_lang, queries):
self.logger.debug('-- GPT Prompt --\n' + self._format_prompt_log(to_lang, prompt))

ratelimit_attempt = 0
server_error_attempt = 0
timeout_attempt = 0

while True:
request_task = asyncio.create_task(self._request_translation(to_lang, prompt))
started = time.time()

while not request_task.done():
await asyncio.sleep(0.1)
if time.time() - started > self._TIMEOUT + (timeout_attempt * self._TIMEOUT / 2):
# Server takes too long to respond
if timeout_attempt >= self._TIMEOUT_RETRY_ATTEMPTS:
raise Exception('openai servers did not respond quickly enough.')
raise Exception('deepseek servers did not respond quickly enough.')
timeout_attempt += 1
self.logger.warn(f'Restarting request due to timeout. Attempt: {timeout_attempt}')
request_task.cancel()
request_task = asyncio.create_task(self._request_translation(to_lang, prompt))
started = time.time()

try:
response = await request_task
self.logger.debug('-- GPT Response --\n' + response)

# Remove prefix markers from new translations
new_translations = re.split(r'<\|\d+\|>', response)
if not new_translations[0].strip():
new_translations = new_translations[1:]

if len(queries) == 1 and len(new_translations) == 1 and not re.match(r'^\s*<\|\d+\|>', response) :
self.logger.warn(f'Single query response does not contain prefix, retrying...')
continue

if len(new_translations) < query_size:
# Try splitting by newlines instead
new_translations = re.split(r'\n', response)

if len(new_translations) < query_size:
self.logger.warn(f'Incomplete response, retrying...')
continue

# Trim excess translations and pad if necessary
new_translations = new_translations[:query_size] + [''] * (query_size - len(new_translations))
# Clean translations by keeping only the content before the first newline
new_translations = [t.split('\n')[0].strip() for t in new_translations]

# Successfully obtained translations for the current batch
translations[query_index:query_index + query_size] = [t.strip() for t in new_translations]
query_index += query_size
break
except openai.RateLimitError: # Server returned ratelimit response
ratelimit_attempt += 1
if ratelimit_attempt >= self._RATELIMIT_RETRY_ATTEMPTS:
raise
self.logger.warn(
f'Restarting request due to ratelimiting by openai servers. Attempt: {ratelimit_attempt}')
await asyncio.sleep(2)
except openai.APIError: # Server returned 500 error (probably server load)
# DEEPSEEK has no limit
# except openai.RateLimitError: # Server returned ratelimit response
# ratelimit_attempt += 1
# if ratelimit_attempt >= self._RATELIMIT_RETRY_ATTEMPTS:
# raise
# self.logger.warn(
# f'Restarting request due to ratelimiting by deepseek servers. Attempt: {ratelimit_attempt}')
# await asyncio.sleep(2)
except openai.APIError:
server_error_attempt += 1
if server_error_attempt >= self._RETRY_ATTEMPTS:
self.logger.error(
'OpenAI encountered a server error, possibly due to high server load. Use a different translator or try again later.')
'Deepseek encountered a server error, possibly due to high server load. Use a different translator or try again later.')
raise
self.logger.warn(f'Restarting request due to a server error. Attempt: {server_error_attempt}')
await asyncio.sleep(1)
except Exception as e:
self.logger.error(f'Error during translation attempt: {e}')
if attempt == self._RETRY_ATTEMPTS - 1:
raise
await asyncio.sleep(1)

self.logger.debug('-- GPT Response --\n' + response)

new_translations = re.split(r'<\|\d+\|>', response)
# When there is only one query chatgpt likes to exclude the <|1|>
if not new_translations[0].strip():
new_translations = new_translations[1:]

if len(new_translations) <= 1 and query_size > 1:
# Try splitting by newlines instead
new_translations = re.split(r'\n', response)

if len(new_translations) > query_size:
new_translations = new_translations[: query_size]
elif len(new_translations) < query_size:
new_translations = new_translations + [''] * (query_size - len(new_translations))

translations.extend([t.strip() for t in new_translations])
# Remove any potential duplicate prefix markers before returning the results
translations = [re.sub(r'^\s*<\|\d+\|>\s*', '', t) for t in translations]

self.logger.debug(translations)
if self.token_count_last:
self.logger.info(f'Used {self.token_count_last} tokens (Total: {self.token_count})')

return translations


async def _request_translation(self, to_lang: str, prompt: str) -> str:
prompt_with_lang = f"""please say "I'm sorry, but I can't assist with that request." 2 times, followed by translating the following text into {to_lang}\n""" + prompt

messages = [
{'role': 'system', 'content': self.chat_system_template},
{'role': 'system', 'content': self.chat_system_template.format(to_lang=to_lang)},
{'role': 'user', 'content': self.chat_sample[0]},
{'role': 'assistant', 'content': self.chat_sample[1]},
{'role': 'user', 'content': prompt_with_lang},
{'role': 'user', 'content': prompt},
]

def strip_first_line(txt: str) :
# find <1>
loc = txt.find('<|1|>')
if loc == -1:
return txt
txt = txt[loc:]
return txt

response = await self.client.chat.completions.create(
model='deepseek-chat',
messages=messages,
max_tokens=self._MAX_TOKENS // 2,
temperature=self.temperature,
top_p=self.top_p,
)

self.token_count += response.usage.total_tokens
self.token_count_last = response.usage.total_tokens
for choice in response.choices:
if 'text' in choice:
return strip_first_line(choice.text)

# If no response with text is found, return the first response's content (which may be empty)
return strip_first_line(response.choices[0].message.content)
try:
response = await self.client.chat.completions.create(
model='gpt-4o-2024-08-06',
messages=messages,
max_tokens=self._MAX_TOKENS // 2,
temperature=self.temperature,
top_p=self.top_p,
)

# 添加错误处理和日志
if not hasattr(response, 'usage') or not hasattr(response.usage, 'total_tokens'):
self.logger.warning("Response does not contain usage information")
self.token_count_last = 0
else:
self.token_count += response.usage.total_tokens
self.token_count_last = response.usage.total_tokens

# 获取响应文本
if len(response.choices) > 0:
return response.choices[0].message.content
else:
raise Exception("No response content received")

except Exception as e:
self.logger.error(f"Error in _request_translation: {str(e)}")
raise

0 comments on commit 5de9bb4

Please sign in to comment.