DeepSeek v4 Flash 0731 on H100 node

Hi,

I've been playing with recent DeepSeek Flash updated model and it seems excellent. Like talking to Claude 4.8.

I'm currently deploying it on VLLm and I'm in progres of tuning performance.

I seem to hit the point where I trade decode with prefill by juggling batching knobs.

I'm pretty sure the issue now lies with scheduling and expert routing, but with no absolute certainty. Tensor core saturation never exceeded 30%.

Large, uncached, prefills often hold compute creating queue pressure of smaller prompts.

I got prefix caching enabled.

Maybe you guys have some pointers on how to extract the most out of it for 8xH100.