profiled linum-v2 on an H100 — your denoise loop is clean
I'm building GPU cost attribution tooling, and I needed a real workload to test it against instead of a synthetic benchmark. I picked linum-v2-360p because the weights are public.
Rented an H100 80GB on RunPod, wrapped your generate() with a 100ms NVML sampler, split phases on cuda.synchronize(). 360p / 2s / 50 steps / CFG 10 / seed 16.
What I found:
setup 29.6s 27% GPU-busy 115W (cold weight load)
denoise 34.2s 99% GPU-busy 630W
decode 1.0s 100% GPU-busy 590W
Generation came in at 35.1s, which lines up with your published ~40s, so I trust the rig.
I went in expecting the dual batch-1 CFG passes to leave the GPU idle between kernel launches. They don't. 99% busy at 630W of a 700W cap and the power draw corroborates the utilization, so it isn't NVML rounding up. The only idle is one time cold weight load, which your API already lets you preload away.
So: nothing to sell you. Your denoise loop is saturated. I'm sending this because I found the opposite of what I was looking for, and that seemed worth telling you.
One question I couldn't answer from the outside: does your production path preload the encoder and VAE, or do you scale to zero between requests?
Raw sampler output attached (n=1, cold — caveat stated).
Kevin
github.com/AgentMulder404