Qwen3.8 27B PRO 6000 comprehensive campaign evidence

This sanitized bundle supports the dated finding.

Campaign boundary

• Campaign: 2026-09-04-qwen38-27b-pro6000-possibility
• Measured hardware: two equal RTX PRO 6000 Blackwell Max-Q GPUs under Docker Desktop/WSL2; single TP1, dual independent TP1, and TP2 measured
• Measurement: direct online streaming; natural-completion screening plus a matched 100-request unique-canary sustained-output workload
• Decision: DP2 bounded aggregate-throughput winner; TP2 rejected; no promotion
• Not measured: broad quality, agentic/SWE, multimodal, routed/client, load-balancer/failover, and complete power/energy telemetry

Common campaign artifacts

• artifact-manifest.json — campaign role ledger
• source-registry.json — dated official, community, and local priors
• workload-manifest.json — workload and metric definitions
• run-plan.json — executed stage order and stop rules
• configuration.json — immutable identities and exact runtime/topology settings
• summary.json — machine-readable metrics, gates, limitations, and no-promotion decision
• friction-log.md — runtime, correctness, and restoration friction with dispositions
• restoration.json — exact starting-state restoration
• publication-summary.md — bounded public copy and claim ledger

Headline matched artifacts

• SGLang Inferact TP1: finalist-k12-chunk1k-4k-c8-canary-long256-n100.json
• SGLang Inferact TP2: opt-tp2-k12-chunk1k-4k-c8-canary-long256-n100.json
• DP2 combined: dp2-combined-4k-c16-canary-long256-n100.json, with replica A and replica B
• RadixArk K8: radixark-dflash-k8-4k-c8-canary-long256-n100.json
• kelnei/vLLM MTP2: kelnei-vllm0271-mtp2-4k-c8-canary-long256-n100.json
• kelnei/vLLM no-spec: kelnei-vllm0271-nospec-4k-c8-canary-long256-n100.json
• vLLM MTP counters: kelnei-vllm0271-mtp2-runtime-metrics.json

Each headline artifact contains request-level timings plus mean, p50, p95, p99, confidence interval, and standard deviation for TTFT, effective prefill, decode, TPOT/mean ITL, and E2E. Aggregate throughput is output tokens divided by the concurrent run wall clock.

Functional and correctness artifacts

• Selected Inferact TP1: preflight
• TP2: failed full preflight and failed isolated JSON repeat
• DP2: replica A and replica B
• RadixArk: target-only, K8, and K12/1K
• kelnei/vLLM: MTP2 and no-spec

TP2's performance artifact completed 100/100 canaries, but its strict JSON gate emitted a duplicate object around a literal closing think delimiter and failed again in isolation. It is rejected.

SGLang optimization artifacts

• Draft depth: K4 C1, C4, C8; K12 C1, C4, C8, 32K; K16 C1, C4, C8, 32K
• K12 chunk sweep: 1K C1, C4, C8, 32K; 8K C1, C4, C8, 32K
• Compile: C1, C4, C8, 32K
• Mamba strategy/slots: lazy96 C1, C4, C8, 32K; Mamba40 C1, C4, C8, 32K, 82K/C8
• Finalist: preflight, natural N100, sustained N100, and 82K/C8

The original target-only/DFlash2 K8 matrix remains retained as the baseline; see files prefixed nospec-, nospec-mamba96-, and dflash-k8-.

Topology, checkpoint, and alternate-runtime artifacts

• TP2: files prefixed opt-tp2-k12-chunk1k-
• DP2: files prefixed dp2-
• RadixArk target-only/K8/K12: files prefixed radixark-
• kelnei/vLLM MTP2 and no-spec: files prefixed kelnei-vllm0271-

The 82K/C8 artifacts are negative interactive-latency evidence. They are not warm-prefix results and must not be used to claim responsive long-context C8.

Graphs and restoration

• graph-manifest.json — exact chart selections
• benchmark-graph-data.json — plotted values, source paths, and hashes
• benchmark-matrix.svg — derivative dashboard
• restoration-smoke.json and restoration-router-smoke.json — fresh direct and routed restoration checks

The graph is derivative; raw JSON is authoritative. Publication does not authorize a serve, route, client, or deployment change.