mxdlzg commented 4 months ago

close #1408 close #1386

我已确认该 PR 已自测通过，相关截图如下：（此处放上测试通过的截图，如果不涉及前端改动或从 UI 上无法看出，请放终端启动成功的截图）

理论上早期的版本应该没有返回FinishReason？所以流不会被截断。但是在v0.6.7左右的版本修复stream流问题时这部分被修正了#1382，引入了FinishReason字段。涉及问题#1345、

1328。

但是有个比较尴尬的地方在于，实际上Gemini的API本身就在每个流message中返回STOP（看起来跟早期的Gemini的表现不太一致，不知道为什么会出现这个问题）。

因为是流式返回，没法准确预知哪个是最后一条message，只能暂时去掉FinishReason字段了。

启动：

[SYS] 2024/05/28 - 15:26:32 | One API v0.0.0 started 
[SYS] 2024/05/28 - 15:26:32 | running in debug mode 
[SYS] 2024/05/28 - 15:26:32 | SQL_DSN not set, using SQLite as database 
[SYS] 2024/05/28 - 15:26:32 | database migration started 
[SYS] 2024/05/28 - 15:26:32 | database migrated 
[SYS] 2024/05/28 - 15:26:32 | REDIS_CONN_STRING not set, Redis is not enabled 
[SYS] 2024/05/28 - 15:26:32 | using theme default 
[SYS] 2024/05/28 - 15:26:32 | initializing token encoders 
[SYS] 2024/05/28 - 15:26:33 | token encoders initialized 
[DEBUG] 2024/05/28 - 15:27:06 | 2024052815270615146503779860442 | request body: {
    "model": "gemini-pro",
    "temperature": 0,
    "messages": [
        {
            "role": "user",
            "content": "Why sky is blue?"
        }
    ],
    "stream": true
} 
[SYS] 2024/05/28 - 15:27:06 | failed to get token encoder for model gemini-pro: no encoding for model gemini-pro, using encoder for gpt-3.5-turbo 
[INFO] 2024/05/28 - 15:27:06 | 2024052815270615146503779860442 | user 1 has enough quota 499999999963990, trusted and no need to pre-consume 
[DEBUG] 2024/05/28 - 15:27:06 | 2024052815270615146503779860442 | converted request: 
{"contents":[{"role":"user","parts":[{"text":"Why sky is blue?"}]}],"safety_settings":[{"category":"HARM_CATEGORY_HARASSMENT","threshold":"BLOCK_NONE"},{"category":"HARM_CATEGORY_HATE_SPEECH","threshold":"BLOCK_NONE"},{"category":"HARM_CATEGORY_SEXUALLY_EXPLICIT","threshold":"BLOCK_NONE"},{"category":"HARM_CATEGORY_DANGEROUS_CONTENT","threshold":"BLOCK_NONE"}],"generation_config":{}} 
[GIN] 2024/05/28 - 15:27:12 | 2024052815270615146503779860442 | 200 |  6.303667817s |       127.0.0.1 |    POST /v1/chat/completions
[INFO] 2024/05/28 - 15:27:12 | 2024052815270615146503779860442 | record consume log: userId=1, channelId=1, promptTokens=12, completionTokens=337, modelName=gemini-pro, tokenName=Test, quota=1023, content=模型倍率 1.00，分组倍率 1.00，补全倍率 3.00

manjieqi commented 3 months ago

希望能尽快合并

songquanpeng commented 3 months ago

Thx~

songquanpeng / one-api

fix: Gemini streaming return being truncated by FinishReason. #1477

1328。