Request for deepseek-v3 benchmark

[UbeCc]

Now serving deepseek-v3 is quite confusing and we cannot check whether we get the correct throughout. Can sglang provide baseline throughout for us users? Thanks!

[ispobock]

Some benchmark results and settings for DeepSeek models can be found in this slides.

Besides, there are two kinds of throughput benchmarking. You can choose for your scenario.

• For serving throughput benchmark, you can use this script. Request rate in the script can be adjusted for different requirement.
• For offline throughput benchmark, you can refer to this script.

[UbeCc]

okay, thanks!