Repository navigation
2026.11.14.01
Unpack anywhere, then run setup.cmd (or zero-launcher.cmd once a model is there).
New since 2026.11.13.11 (results unchanged; one GPU gives the same logits as before):
- Split models can keep their decoded weights. A model split over several GPUs used to compare the whole model's size with
--budget, so the 27B decoded its weights for every prompt chunk even on 2-4 Pro B60s. Each GPU now counts only its own share:--budget 20over 2 B60s or--budget 10over 4 keeps the 27B's decoded weights on the cards. The default budget is unchanged. (docs/xe2-tuning.md section 13.1) - The prompt's matrix kernel can prefetch its next chunks, opt-in on Arc B-series cards:
ARC_GEMM_XW_PF=1(weights),2(activations) or3(both),ARC_GEMM_XW_PFD=1/2/4chunks ahead..\bench\xw_sweep.ps1 -Model <gguf> -Prefetchtries them all. (section 13.2) - The website shows the newest Pro B60 numbers: the 27B's 4K prompt on 1 / 3 cards, and Flash Next over 1 / 2 / 4 cards.