okay after testing DiffusionGemma a bit more internally we genuinely can’t tell if this is the start of something big or if everyone’s just getting distracted by crazy TPS numbers again lol but one thing that stood out REALLY fast for us was how different the H100 vs A100 behavior felt compared to normal transformer inference on some runs the H100s scaled almost exactly how you’d want them to the A100s were still good, but once concurrency started increasing the gap widened way more than w...