MiniMax 渠道 abab6.5 模型未支持，缺失 `max_tokens` 参数情况下，流式/非流式回答被截断。

[MurphyLo]

例行检查

• [x] 我已确认目前没有类似 issue
• [x] 我已确认我已升级到最新版本
• [x] 我已完整查看过项目 README，尤其是常见问题部分
• [x] 我理解并愿意跟进此 issue，协助测试和提供反馈
• [x] 我理解并认可上述内容，并理解项目维护者精力有限，不遵循规则的 issue 可能会被无视或直接关闭

问题描述

如题，使用 v0.6.6-alpha.14 版本，分别运行以下命令：

1. 流式，不指定 max_tokens

   curl https://one-api-host/v1/chat/completions \
   -H "Content-Type: application/json" \
   -H "Authorization: Bearer sk-xxx" \
   -d '{
   "model": "abab6.5-chat",
   "messages": [{"role": "user", "content": "用两句话夸我帅"}],
   "temperature": 0.2,
   "stream": true
   }'

2. 流式，max_tokens=4000

   curl https://one-api-host/v1/chat/completions \
   -H "Content-Type: application/json" \
   -H "Authorization: Bearer sk-xxx" \
   -d '{
   "model": "abab6.5-chat",
   "messages": [{"role": "user", "content": "用两句话夸我帅"}],
   "temperature": 0.2,
   "max_tokens": 4000,
   "stream": true
   }'

3. 非流，不指定 max_tokens

   curl https://one-api-host/v1/chat/completions \
   -H "Content-Type: application/json" \
   -H "Authorization: Bearer sk-xxx" \
   -d '{
   "model": "abab6.5-chat",
   "messages": [{"role": "user", "content": "用两句话夸我帅"}],
   "temperature": 0.2,
   "stream": false
   }'

4. 非流，max_tokens=4000

   curl https://one-api-host/v1/chat/completions \
   -H "Content-Type: application/json" \
   -H "Authorization: Bearer sk-xxx" \
   -d '{
   "model": "abab6.5-chat",
   "messages": [{"role": "user", "content": "用两句话夸我帅"}],
   "temperature": 0.2,
   "max_tokens": 4000,
   "stream": false
   }'

依次有如下响应：

1. 流式，不指定 max_tokens

   
   data: {"id":"xxx","choices":[{"index":0,"delta":{"content":"你","role":"assistant"}}],"created":1714458858,"model":"abab6.5-chat","object":"chat.completion.chunk"}

data: {"id":"xxx","choices":[{"finish_reason":"length","index":0,"delta":{"content":"不仅外表英俊","role":"assistant"}}],"created":1714458858,"model":"abab6.5-chat","object":"chat.completion.chunk"}

data: {"id":"xxx","choices":[{"finish_reason":"length","index":0,"message":{"content":"你不仅外表英俊","role":"assistant"}}],"created":1714458858,"model":"abab6.5-chat","object":"chat.completion","usage":{"total_tokens":76},"base_resp":{"status_code":0,"status_msg":""}}

2. 流式，max_tokens=4000

data: {"id":"xxx","choices":[{"index":0,"delta":{"content":"你","role":"assistant"}}],"created":1714458504,"model":"abab6.5-chat","object":"chat.completion.chunk"}

data: {"id":"xxx","choices":[{"index":0,"delta":{"content":"拥有着令人印象深刻的英俊外表，每一个微笑都像是阳光下的璀璨星辰，让人无法移开视线。你的帅气不仅仅是外表，更是一种由内而外散发","role":"assistant"}}],"created":1714458505,"model":"abab6.5-chat","object":"chat.completion.chunk"}

data: {"id":"xxx","choices":[{"finish_reason":"stop","index":0,"delta":{"content":"的自信和魅力，让人不禁赞叹。","role":"assistant"}}],"created":1714458505,"model":"abab6.5-chat","object":"chat.completion.chunk"}

data: {"id":"xxx","choices":[{"finish_reason":"stop","index":0,"message":{"content":"你拥有着令人印象深刻的英俊外表，每一个微笑都像是阳光下的璀璨星辰，让 人无法移开视线。你的帅气不仅仅是外表，更是一种由内而外散发的自信和魅力，让人不禁赞叹。","role":"assistant"}}],"created":1714458505,"model":"abab6.5-chat","object":"chat.completion","usage":{"total_tokens":117},"base_resp":{"status_code":0,"status_msg":""}}

3. 非流，不指定 `max_tokens`

{"id":"xxx","choices":[{"finish_reason":"length","index":0,"message":{"content":"你不仅外表英俊","role":"assistant"}}],"created":1714458902,"model":"abab6.5-chat","object":"chat.completion","usage":{"total_tokens":76},"base_resp":{"status_code":0,"status_msg":""}}

4. 非流，max_tokens=4000

{"id":"xxx","choices":[{"finish_reason":"stop","index":0,"message":{"content":"你拥有令人瞩目的英俊外表，每一次出现都像是从时尚杂志中走出来的模特。你的帅 气不仅仅是外表，更是一种由内而外散发的独特魅力，让人无法忽视。","role":"assistant"}}],"created":1714458942,"model":"abab6.5-chat","object":"chat.completion","usage":{"total_tokens":111},"base_resp":{"status_code":0,"status_msg":""}}

[QAbot-zh]

补充：除了回答被截断外，新模型的计费也有问题（可能是因为token计算器的不同，minimax 比 gpt 省 token，导致返回响应中的 total_token 小于 gpt3.5 tiktoken 算出来的 prompt_token 导致的）

[songquanpeng]

有点搞，补全是根据total反算出来的，这下有点麻烦

[lem0ke]

有点搞，补全是根据total反算出来的，这下有点麻烦

minimax人员回复可以根据respond返回信息确认token字符数，不知道这个信息是否对你有帮助

[QAbot-zh]

看了下 minimax 输入输出价格一致，所以至少消耗日志上不会出统计差错。如果想避免出现复数token，也可以统计提示和补全的字符数比例然后根据 total 算出一个值来，这样比较兼顾。