feat(BACKEND-TENSTORRENT-KEEPQUANT): the 27B fits — packed GDN projections, int8-dot captured decode, per-request logits gather (#3042) - #3183

Merged

lu-zero merged 9 commits into

mudler:mainfrom

lu-zero:row/BACKEND-TT-KEEPQUANT-W3

Sep 13, 2026

Commits on Sep 13, 2026

wip(BACKEND-TENSTORRENT-KEEPQUANT): W3 seed — 27B alloc-trace instrumentation (#3042)
lu-zero
committed
0202f4f
spec(BACKEND-TENSTORRENT-KEEPQUANT): W4d W4 — packed row-reorder keeps the GDN projections block-encoded (#3042)
lu-zero
committed
58c9d9b
spec(BACKEND-TENSTORRENT-KEEPQUANT): W4d W5 — decode-shaped capture; mixed prefill steps run eager (#3042)
lu-zero
committed
5d2b1f4
spec(BACKEND-TENSTORRENT-KEEPQUANT): W4d W6 — int8-dot as the captured keep-quant decode (#3042)
lu-zero
committed
553155c
feat(BACKEND-TENSTORRENT-KEEPQUANT): the 27B fits — packed GDN projections, int8-dot captured decode, per-request logits gather (#3042)
lu-zero
committed
53959be
fix(BACKEND-TENSTORRENT-KEEPQUANT): gate the warm-slot release behind the TT build (#3042)
lu-zero
committed
45ad8a6
record: re-anchor the three line citations the wave shifted (#3042)
lu-zero
committed
192ba68
record: re-anchor LoadGgufSharedEmbedAndHeadBf16 to 1065 (#3042)
lu-zero
committed
8b50c3f
docs: document VT_TT_RELEASE_WARM_ROWS (#3042)
lu-zero
committed
df45fa8