Results of llama-batched-bench skewed with MoE models?

[enh6]

It seems llama-batched-bench just uses token id 0 as inputs for a batch. For MoE models, same input tokens in a batch will activate the same experts. While in reality different inputs in a batch may activate different experts. The performance should be different?
Maybe we can use random tokens like llama-bench?

[ggerganov]

Yes, this is a mistake and should be fixed. Would you like to submit a PR?

[enh6]

Yeah, I can submit a PR later.