fix(qwen38): the preflight asks whether the pin is ONE of an image's digests

A by-digest pull leaves the image carrying every digest it was pulled under, so
reading RepoDigests[0] reported a healthy pinned image as wrong and refused a
window that was ready to run. Membership, not position. Also records the measured
step optimum: 106 tok/s at an 8,192-token step against 76 at 16,384 on a 32,768
calibration window, so the wide step has a ceiling, not a plateau.
This commit is contained in:
Frank committed 2026-10-30 16:30:03 -07:00
1 parent 5c678b804c
commit 44d72b9e78
3 files changed
+24 -9

No files matched your search

+10
View File
@@ -571,6 +571,16 @@ vllm-radiance era (this table opens on 09-26; the era page claims 10-10).
of falling through its health wait. All three exist because on 2026-10-29 the unguarded chain's
rig boot died of an OOM, the health loop expired without failing, and the pusher fed four rounds
of 10M tokens into **restored production** — 86 minutes of live prefill that recorded nothing.
- **A calibration step has an OPTIMUM, and it is not the widest step that boots.** The tap's cost is
expert re-reads per document (a 7,168-token document cut into four 2,048-token steps re-streams
each layer's routed expert union four times), which argues for the widest step — but the measured
curve turns over: at the 32,768 calibration window one 2-way prefill of the same document ran
**106 tok/s at an 8,192-token step and 76 tok/s at 16,384**, with the computed activation arena
growing 3.39 → 4.50 GiB, so the wider step pays more compute than it saves in reads. The
200,000-window `QSA score plane` refusal that made the wide step look impossible is an artifact of
serving the production window: the same 16,384-token step needs a 512 MiB plane at 32,768.
`Q38_CALIB_MNBT` is the only knob — the KL modes keep production's step, because their numbers are
only comparable at one shape.
- **A container that exists, a port that answers, and a `/health` 200 are three different claims —
bind identity before believing any of them.** Measured 2026-10-30: `podman inspect` named the rig
while the port still belonged to the *dying* production process (`health ok waited=0s`, window dead