GLM-5.3-Flash ormandj runtime v0.4.2 qualification¶
Date: 2026-09-09
Status: retain-baseline/no-promotion; the campaign is concluded.
Result card¶
| Setup | Result |
|---|---|
| Candidate | pinned v0.4.2 source/image, TP=2, 393,216/C1 |
| Matched direct capacity | 4K n12 and 120K n3 improve median TTFT/E2E/decode |
| Qualification | candidate not qualified: strict turnover 58/60 then 57/60; baseline retained |
| Evidence | manifest ยท publication summary |
Why it matters: bounded performance gains are retained without treating them as permission to change the deployed runtime.
Important caveat: strict turnover did not meet its predeclared gate; the candidate route was not admitted.
Fixed comparison boundary¶
The candidate uses image
ghcr.io/ormandj/sglang-glm53-flash-sm120@sha256:877ae29486eb7d385eb2a07921522aa3597f643f9a7c50112751f2aaa14d558a,
from source revision 23570f89. The release tag resolves to 44c026ca, so
the candidate is an exact separately pinned source state, not the tag itself.
The same GLM-5.3-Flash weights, TP=2, 393,216-token/C1 contract, adaptive MTP,
P2P enabled, cuMem disabled, and HiCache disabled remain fixed. This is a
deployment-profile runtime upgrade with PCIe IPC, FP8 head, graph, and kernel
setting deltas; it is not a single-variable image attribution.
Retained evidence and limits¶
Candidate direct preflight and default, enabled, and disabled-thinking controls pass. Baseline refused the synthetic 128K launch-code needle twice at 105,647 actual prompt tokens, while its other functional observations remain retained. Baseline high-context, three-repetition quality, session, and multimodal gates passed. Its summary 60-request canary formatting failed; its strict controlled turnover completed 53/60 with seven count/length failures. All 60 responses arrived without foreign markers; none of those failed populations can support a performance claim.
Candidate startup observed both-rank PCIe IPC and NCCL P2P. It also retained
an IPC-seed autotune warning and four TileLang race warnings. Independent Sol
review refuted the cited races through guarded injective indexes, but that does
not permit a warning-free claim or use of TL_DISABLE_DATA_RACE_CHECK.
Completed direct comparison and human decision¶
The eligible matched capacity cells completed 12/12 at the 4K target and 3/3 at 120K in both arms. Candidate versus baseline changes are respectively -5.19%/-1.31% median TTFT, -4.91%/-1.45% median E2E, and +3.93%/+17.21% median decode. These small populations are descriptive only and establish no p99, service-tail, or universal runtime claim.
The exact 60-request controlled turnover gate completed 58/60 for the
candidate versus 53/60 for the baseline. One pre-recorded unchanged strict
repeat completed 57/60; no loop reruns were used. The candidate's original two
misses were clean stop responses with 33 and 31 rather than 32 words. Its
repeat includes 33-word stop responses at indices 20 and 36 plus a 499-word,
512-token length truncation at index 32. All 60 responses arrived, and every canary passed.
The strict-turnover disposition prevents full candidate qualification despite
comparative performance gains. The user declined a comparative exception and
selected retain-baseline/no-promotion; this concludes the campaign without a
production promotion.
Candidate PowerShell marker
diagnostics are 2/3 because one response lacks literal absolute; review
labels it a false negative with semantic caveats; see the
independent ledger.
The baseline's 3/3 marker responses contain invalid code, so this does not
establish semantic regression.
The user also mentioned other reports. They are user-reported and unverified; they are not corroboration for this finding or its decision. See the dated human decision.
Restoration boundary¶
The exact baseline registry, image, manifest, router, and topology were restored. Direct and routed baseline preflight each passed 2/2; the router was admitting and ready with zero active requests, and the restored unit had zero restarts. P2P remains enabled. The candidate routed gate was intentionally not attempted because strict turnover failed. See restoration.