You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Decisions serving for qwen27 (Ornith-1.5-122B): score/decide passes, JointHead scoring, a free decision slot, GPU class profiling #1190
Kept up to date as each pull request is opened. The whole family stacks on #1125 (the qwen27 Metal family); every branch carries its parents' commits until they merge — review only each PR's own commits, named in its body.
TF_Q27_PROFILE logs the round's class GPU times; replies carry gpu_seconds (in rework)
The qwen27 Metal lane takes decision work beside text generation: an HTTP thread enqueues a score or decide task and the lane host's own engine thread runs the pass — first only at idle points (#1189), then beside live traffic on a free slot (#1192). One resident model serves both streams; there is no scheduler beside the engine.
#1191 adds a small trained side head (--joint-head) that scores each option on the post-final-norm hidden row at its own final-token position instead of label logits. The last member wires the existing gpu_profile.Trace per-class timers up on the 27B backend behind TF_Q27_PROFILE, off by default.
Thank you, @grantoverton, for the decisions series. Our reviews on #1189 to #1215 (26 October) asked for the stack to build on main's decision queue, stay opt-in, and converge with the shared decision path in #1361, and we haven't heard back, so we're closing this tracker with the stack. A reworked PR against current main is very welcome, and you can reopen this issue with it.
Activity
ashhart commentedon Oct 27, 2026
Thank you, @grantoverton, for the decisions series. Our reviews on #1189 to #1215 (26 October) asked for the stack to build on main's decision queue, stay opt-in, and converge with the shared decision path in #1361, and we haven't heard back, so we're closing this tracker with the stack. A reworked PR against current main is very welcome, and you can reopen this issue with it.