How do you think output TPS depends on language of prompt?

[psydok]

Models are mostly trained in English. Will it affect performance if prompts are written, for example, in Ukrainian or French?

I don't have a suitable dataset to check it yet. Therefore, it is interesting to know if you have conducted similar experiments?

[psydok]

I tried to set up custom dataset for specific language with 1 prompt for load generation. It seems that metrics are about same, although even little better than in English.
Has anyone else made such a comparison? Did you do the same?

P.S. To set up test, I used following parameter: -e GUIDELLM_DATA="source=data:/results/uk.txt.gz ,prompt_tokens=4096,output_tokens=512"