unbug on X: "Turns out you can tune Strata for more prefill speed. On Qwen3.8-Flash Q6-KXL setup: +72% prefill on short context, +27% on long context, and VRAM hit rate stays pinned at 100%. The main knobs are --pcie-frac, layer_split, and --trim-stage-weights. The catch: MTP still isn’t supported with parallel request."

unbug

@unbug

Turns out you can tune Strata for more prefill speed. On Qwen3.8-Flash Q6-KXL setup: +72% prefill on short context, +27% on long context, and VRAM hit rate stays pinned at 100%. The main knobs are --pcie-frac, layer_split, and --trim-stage-weights. The catch: MTP still isn’t supported with parallel request.

6:08 PM · Oct 15, 2026·

16K

Views

10

8

113

121