async generate_content is very slow

[ascillitoe]

The async performance of the new SDK still seems to be much worse than the old SDK (with transport='grpc_asyncio').

We can do 1000 text classifications in ~5s with the old client, but this consistently takes over 30s with the new client.

Is this a know issue? Or are there certain settings that must be configured with the new client? (e.g. we found the transport option was very important with the previous client).

Environment details

• Programming language: Python
• OS: Ubuntu 22.04.5 LTS
• Language runtime version: 3.10.12
• Package version: 1.7.0

Steps to reproduce

1. Run N basic text completions async with the new google.genai.client.aio.models.generate_content
2. Compare with the old google.generativeai generate_content_async

[andrew-stelmach]

Compare the runtimes before and after the call to generate_content. LLM response speeds are stochastic as every provider changes how many gpus are being used, etc

[yinghsienwu]

Note that google.genai uses REST transport. In general, gRPC is faster than REST. We'll calibrate the runtime performance and see how to improve. Thanks for raising this.

[ascillitoe]

Hi @yinghsienwu, thanks for the response. Are there no plans to add support for gRPC AsyncIO like the old SDK? This seems like a pretty big regression?

[hugbubby]

Same problem here. It's an extremely slow interface. We're getting 100% cpu usage with only a handful of concurrent requests, which is practically unusable for our purposes.

[yinghsienwu]

I compared Vertex SDK (google-cloud-aiplatform) (#1 default grpc_asyncio, and #2 rest_asyncio transport) with #3 google-genai SDK v1.9 (rest, httpx) and #4 aiohttp prototype for async generateContent requests (100, 500, 1000 async requests).

1. Vertex SDK, rest_asyncio and grpc_asyncio transport perform similarly (within the std dev). gRPC should not be the key to better runtime performance.
2. Currently google-genai SDK (httpx) runtime is ~6X of grpc_asyncio runtime when sending 1000 async requests. (same as the observation above (async generate_content is very slow #557 (comment)).
3. If we want to improve runtime performance, using aiohttp in google-genai SDK AsyncClient implementation may achieve similar performance as Vertex SDK's rest_asyncio.

We'll try to put it into our roadmap.

[ascillitoe]

Thanks for confirming @yinghsienwu! Do you think this situation might be improved by the time the old sdk reaches end of life on Aug 31st?

[yinghsienwu]

I think likely to be available in Q2 2025. I'll attach a PR here.