Throughput evaluation

[vladislavkruglikov]

Hello

Thanks for your work. In paper you report increasing throughput. However I tried to measure it myself and I did not find any improvement. In fact your method showed slow down on H100. Can you provide script that reproduces throughput improvements that you have reported in your paper?

[Mingyuyang-1]

Hello @vladislavkruglikov. Thank you for your interest in our work. As for the throughput, we included a sample script at Zebra-Llama/benchmark and we have verified it on AMD MI300/325. Could you give it a try? We didn't verify it on H100 but it should also work.

[vladislavkruglikov]

Hello @Mingyuyang-1

Sorry not to mention that I was talking specifically about X-EcoMLA and not Zebra-Llama

[Mingyuyang-1]

Hi @vladislavkruglikov. Thanks for the clarification. We have added the throughput evaluation code for X-EcoMLA on MI300 (X-EcoMLA/speed.sh). We haven't tested it on H100 but it should work as well.