Qwen3.8-27B (bf16): the token gate and every speed axis are OWED — the checkpoint loads and runs, and has never been adjudicated or measured

[localai-bot]

Goal

Adjudicate and measure the bf16 arm of Qwen/Qwen3.8-27B @ 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0.
#821 owns the NVFP4 and Q4_K_M
arms of the same checkpoint; neither covers bf16, and the bf16 arm is the one
that already runs.

Why this exists

porting-inventory.md deviation 17
records that this checkpoint "fits GB10 and loads on current main with no code
change", and that "its own token-exact gate is OWED and unrun". Loading is not a
token, and a token is not a throughput number, so two axes were open with no
issue naming them.

Scope

• A binding greedy token gate against the pinned oracle (555967922), with
  every first-divergence position adjudicated against kNearTieMnats = 500.
• Throughput / latency / memory against vLLM's production configuration.
• Records: the owning model-matrix.md row, docs/STATUS.md,
  docs/BENCHMARKS.md under a new key.

Out of scope

The NVFP4 and Q4_K_M arms (#821),
the vision path, and any fix for the tie-break divergence
(#910).

Constraints

dgx.casa (GB10, sm_121a). Every GPU command serialized through $HOME/gpu.lock.
The vLLM arm is never --enforce-eager. Build with
-DVLLM_CPP_CUTLASS_FETCH=ON -DVLLM_CPP_TRITON=ON and the configure-log
fast-path lines asserted, or the number is void.