[Kernel] Upload a MoE config file for Mixtral8x7B 8GPU on AMD_Instinct_MI300X_OAM machine (fp16)

[Jacob0226]

This is just a tuned MoE config file which includes boarder examples.
The Jira ticket GPUAI-2067 adopts this config for better MoE performance.

[shajrawi]

@divakar-amd can you please review? see details here https://ontrack-internal.amd.com/browse/GPUAI-2067

@Jacob0226 I cannot access the documentation wiki in JIRA - can you please explain how the tuning results were gathered?

[Jacob0226]

@divakar-amd can you please review? see details here https://ontrack-internal.amd.com/browse/GPUAI-2067

@Jacob0226 I cannot access the documentation wiki in JIRA - can you please explain how the tuning results were gathered?

I share the ticket to both of you.
I use the script benchmarks/kernels/benchmark_mixtral_moe_rocm.py to get the runtime on various bs values.
For the Mixtral8x7B, it needs more bs value and this new MoE config basically cover bs value from 1 to 32768. The original one in the vllm repo has a max number of bs 4000.

[Jacob0226]

The jira ticket may contain too much info you don't need here.
When I benchmark Mixtral with vLLM, I found I need more bs config in MoE, so I use benchmark_mixtral_moe_rocm.py to do the tunings.