deepseek-ai / DeepSeek-V2

DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model
MIT License
3.47k stars 143 forks source link

What's the Prompt and Response length in the Paper? #70

Open JadeRay opened 2 months ago

JadeRay commented 2 months ago

image

The paper has mentioned 50K tps for generation phase and 100K throughput for prefill phase, my question is what's the tested prompt length and repsonse length in avg and P99?