 Alexander Korolev
|
35cb002a16
qwen3.8: restore bench baseline params (MTP k=4, marlin atomic-add 1) on the pinned image — both were garble red herrings
|
1 месяц назад |
|
Alexander Korolev
|
3d067b1081
fix: image presence check accepts digest refs (docker image ls does not); pin qwen3.8 b12x to the validated nightly-20260913 digest (floating latest 20260925 garbles long contexts, forum 382733 #195-198); min-recipe A/B lane (no MTP, no prefix cache)
|
1 месяц назад |
|
Alexander Korolev
|
42e78480b8
qwen: VLLM_MARLIN_USE_ATOMIC_ADD off (live garble caught co-batched with 70K prefill); day-shift notes
|
1 месяц назад |
|
Alexander Korolev
|
7250e2ad11
qwen: VLLM_MARLIN_USE_ATOMIC_ADD off (experimental atomic-add path, suspected intermittent garble)
|
1 месяц назад |
|
Alexander Korolev
|
faca4a98da
exl3: two-mount staging (repo at /hub + relative-symlink dir at /model), kv_cap 0; qwen served name qwen3.8
|
1 месяц назад |
|
Alexander Korolev
|
ad7069edcd
docker: --entrypoint takes tail[0], rest is argv (no repeat); port convention 30000 in recipes + docs
|
1 месяц назад |
|
Alexander Korolev
|
548fcdfc75
dsv41 4x sglang aiden-1.6 (162/176) + qwen38 int4 1x (165/176) lanes, night shift 09-26
|
1 месяц назад |
