Default Branch

03654935b8 · qserve: dump every thread stack when a start-up phase stalls (#133) · Updated 2026-10-30 01:04:05 -09:00

Branches

5a616c3271 · CUDA attention: batch the TurboQuant verify rows instead of looping them · Updated 2026-10-28 20:33:36 -09:00    yindo

2
40

aea8b958fb · TurboQuant decode rotation: matrix-vector kernel for up to 8 rows (3.6 us instead of 20-29 us) · Updated 2026-10-28 11:24:55 -09:00    yindo

3
33

62b9dd4843 · exl3_dec gemv_r: each row chunk arrives on its own counter slot, replacing the R <= rpb gate (kamel 154d0b3c8) · Updated 2026-10-28 09:25:40 -09:00    yindo

3
29

4180816294 · EAGLE-3 drafter on the MTP draft path, --draft dflash:/eagle:/mtp in the Qwen server, generic draft loader · Updated 2026-10-28 01:21:09 -09:00    yindo

3
2

186dbde1e7 · qualify: search extra venvs and eval suites under QUALIFY_WORK_ROOT instead of a fixed home layout · Updated 2026-10-27 05:51:13 -09:00    yindo

13
33