support deepseek's token usage format

[manukall]

Deepseek includes a choice with empty content in their token usage chunk when streaming. otherwise, their API is compatible with OpenAI's. This commit adds support for this chunk format.

Example of a token usage chunk from Deepseek:

data: {
   "id":"339eade8-b499-472e-bc0d-b267b6e4bceb",
   "object":"chat.completion.chunk",
   "created":1743968719,
   "model":"deepseek-chat",
   "system_fingerprint":"fp_3d5141a69a_prod0225",
   "choices":[
      {
         "index":0,
         "delta":{
            "content":""
         },
         "logprobs":null,
         "finish_reason":"stop"
      }
   ],
   "usage":{
      "prompt_tokens":11,
      "completion_tokens":11,
      "total_tokens":22,
      "prompt_tokens_details":{
         "cached_tokens":0
      },
      "prompt_cache_hit_tokens":0,
      "prompt_cache_miss_tokens":11
   }
}

[brainlid]

Hi @manukall! Thanks for this addition! Can you use the sample data to create a test that identifies the behavior as being for Deepseek? That would really help prevent accidental regressions. Then I'd be happy to merge it in. 🙂

[manukall]

@brainlid I think I added the test you've asked for. Makes total sense.

However, looking at the test, I'm not sure we should emit a MessageDelta when the content = "". Would it make more sense to return :skip in those cases?

[brainlid]

Hi @manukall! Please check out the v0.4.0-rc.0 release. There are other breaking changes that I expect are needed for DeepSeek to be supported, but your TokenUsage update would be best applied to that branch as TokenUsage is being stored in the received Message's metadata.usage.