GLM-5.3-Flash SGLang SM120 benchmark evidence

This directory is the sanitized evidence bundle for the 2026-09-02 local qualification of the pinned community SGLang SM120 recipe. Native benchmark artifacts are authoritative; the narrative finding summarizes them without turning upstream or Discord claims into local measurements.

Campaign boundary

• Campaign: 2026-09-02-glm53-sglang-sm120-qualification
• Candidate: ormandj/GLM-5.3-Flash-W4A16-NVFP4-K32-Experts-FP8-WO@c3cbb9891b67c741bcbf6b176dd7af9265b069db
• Runtime: ormandj/sglang-glm53-flash-sm120 v0.1.1-rc.14 at sha256:0c0637959c3931829f05154087bbefd2c50003fb9b2010200ce0ec82f4d71a53
• Hardware: two RTX PRO 6000 Blackwell Max-Q cards, exclusive TP=2 over PCIe without NVLink, Windows 11/Docker Desktop/WSL2
• Selected recipe: configs/glm53-flash-ormandj-sglang-sm120-tp2-240k-c1-adaptive-mtp-recipe.toml
• Decision: locally verified challenger, no-promotion

This directory preserves the qualification decision as observed. A later operator-approved reserve waiver reclassified the already-passing 393K/C1 profile as verified and authorized its promotion. See the promotion finding and vram-policy-reclassification.json; the original raw artifacts below are unchanged.

Result index

• summary.json — bounded decision and headline metrics
• promotion-publication-summary.md — derivative copy for the later human-approved 393K promotion
• publication-summary.md — compact human-readable result
• source-registry.json and upstream-source-lock.json — date-aware source provenance and stable-versus-rc.14 claim boundary
• configuration-and-identity.json — immutable local candidate identity
• friction-log.md, nccl-cumem-probe.md, and log-audit.md — failures, fix-forward decisions, and bounded final log review
• restoration.json — exact before/after reconciliation; final-restoration-recheck.json — fresh goal-closure health, identity, routed-smoke, GPU, candidate-absence, and shared-memory assertions

Selected 240K/C1 artifacts

• safe240k-preflight-all-low.json — all ten thinking-disabled functional checks
• safe240k-preflight-thinking-enabled.json and safe240k-thinking-control.json — verified bidirectional thinking control
• Capacity r3 at 4K, 120K, and 230K nominal
• safe240k-quality-coding-agent-v2-disabled-r3-visible2048.json — five deterministic coding-agent checks, three repetitions each
• safe240k-multimodal-image-c1.json — deterministic image/OCR corpus
• safe240k-endurance-c1-4k-r60.json — 60-request endurance run
• post-workload-state.json — health, counters, adaptive MTP state, and 3 GiB reserve gate

Controls and rejected envelopes

• The 131K no-speculation and adaptive-MTP controls provide the matched speculation A/B.
• The exact 499,712-token/C4 profile passed functional work and short throughput, but failed the reserve policy; its full499k-unsafe-* artifacts remain retained.
• The 393,216-token/C1 profile passed functional, quality, image, endurance, and deep-context work, but fell to 2,101 MiB free per card after workload; its safe393k-* artifacts retain the physical measurement that was later reclassified under the explicitly recorded model-only reserve waiver.
• Early failed/partial preflights are preserved rather than overwritten.

Real GPU UUIDs, operator paths, credentials, and unsanitized logs remain outside this public evidence bundle. This qualification did not change a route or authorize promotion; the later promotion is a separate dated decision.