The October 10 review checked both GB10 machines directly, inspected the running containers and gateway selectors, and read retained October 5 and 9 speed reports. The pair is still a DGX Spark plus an ASUS GX10, with one shared coder deployment. No lab service was stopped, switched, or benchmarked during this review.
The running coder
Both nodes run Qwen3.8-Flash-Next NVFP4 on vLLM 0.30.0, with tensor parallelism 2, expert parallelism, and three native MTP draft positions. The head serves requests; the second rank runs headless. The multiprocessing executor and RoCE interfaces are configured explicitly. This inspection did not measure link bandwidth or prove the selected transport's performance.
The inspected settings are 262,144 context, eight sequences, 8,192 batched prefill tokens, and 0.74 GPU memory utilization. Chunked prefill is enabled, KV dtype is set to auto, and the Mamba SSM cache override is bfloat16. Inductor compilation is disabled; the launch requests FULL_DECODE_ONLY CUDA graphs and disables breakable graphs.
The coder gateway caps requested output at 32,768. Its proxy leaves thinking unset when the client omits it, allowing the model default to stand. The reasoning parser is qwen3 and the tool parser is qwen3_coder. A selector description saying “thinking on” is less precise than that actual policy.
Eight slots do not certify eight full windows
The October 10 startup log reports 20.95 GiB available KV cache memory on the head worker, a 1,014,349-token cache pool, and 3.87× maximum concurrency at 262,144 tokens per request. That is the engine's startup capacity calculation, not a successful retrieval or concurrent full-window test.
Eight slots describe admission and scheduling. Requests with shorter contexts can share the pool; eight full 262K sessions were not qualified. The earlier July Coder-Next retrieval result stays attached to its original model.
System memory also deserves its own observation. Samples taken at different moments during the inspection showed about 10.8 GiB available and 8.5 GiB swap used on the coordinator, and 15.8 GiB available and 2.7 GiB swap used on the worker. Swap occupancy alone does not establish active paging or a leak. These are dated samples, not a continuous headroom guarantee.
A fixed hour of ordinary traffic
From 20:00 to 21:00 CDT on October 10—01:00 to 02:00 UTC on October 11—the head log contained 319 throughput samples, of which 310 had nonzero generation. Their median engine-wide generation rate was 47.9 tok/s, with a 103.4 tok/s peak sample. This is aggregate mixed traffic, not the speed of one request or a controlled benchmark.
At the sampling instants there were at most four running requests and one waiting request. KV usage peaked at 14.6%. The logged prefix-cache hit-rate gauge moved from 74.5% to 76.8%; it is a gauge of the engine's counters, not a newly isolated cache-hit measurement for that hour.
The reviewed head log had no ERROR, Traceback, OutOfMemory, or preempted lines in this window. That narrow observation does not establish application correctness, kernel health, or reliability at long context. No new quality suite or full-context soak was run.
Startup and warning lines are part of the configuration
On the retained October 10 boot, the API argument log appeared at 17:22:08.885 CDT and application startup completed at 17:36:38.062 CDT: about 869.18 seconds, or 14½ minutes. The head logged 453.45 seconds for the target weights, 88.63 seconds for another weight-loading stage, and 178 seconds for engine initialization. Those durations do not constitute an additive breakdown of every startup step.
A persistent coder avoids paying that startup cost for every idle gap. Container uptime, an early health response, and completed model readiness are different events.
Warnings also help distinguish requested settings from effective behavior. The head warned about unsupported collective optimizations, experimental MTP/cache behavior, disabled Inductor compilation, and the SSM dtype override. It also called VLLM_MTP_DRAFT_VOCAB an unknown environment variable. A read-only search of the installed model sources found no consumer of that variable. The reduced-draft-vocabulary overlay's effect is therefore unverified on this running image; upstream overlay speed claims are not credited to this lab.
The October runtime probes
Retained short-prompt runs used the same two prompts and one wave each at 1, 4, and 8 concurrent requests. All three runs completed 26 of 26 wave requests plus one prefill probe, with no recorded HTTP errors. The client requested temperature zero, medium reasoning effort, streaming usage, and at most 320 output positions for the waves.
- October 5, Qwen NVFP4 / vLLM: single-request code 67.25 tok/s, prose 44.79 tok/s; the roughly 6K-input probe reached first output in 1.851 s.
- October 5, GLM-5.3-Flash EXL3 / TensorFold: code 97.59 tok/s, prose 43.14 tok/s; roughly 6K-input first output in 8.062 s.
- October 9, Qwen NVFP4 / TensorFold: code 80.09 tok/s, prose 50.26 tok/s; roughly 6K-input first output in 2.108 s.
The benchmark page retains the 4- and 8-request results. The harness's “aggregate” field is the sum of individual request decode rates, rather than total wave output divided by wave wall time. The prefill-rate field is input positions divided by time to first output, including gateway and queueing overhead; it is not a pure kernel measurement. First output can include reasoning.
These were single runs on different dates, with different tokenization and completion lengths. Effective thinking policy, warm/cache state, and runtime revisions were not independently frozen. The generated asserts were not executed and prose was not graded. These probes show workload-dependent timing; they do not promote a new champion or establish coding quality.
Follow the alias all the way to the launcher
The coordinator still labels its Studio agent target as Nemotron Lightning. The Studio's active YAML and managed launcher now map the lightning/agent alias to SuperQwen3.8-27b-abliterated, MLX 4-bit, with thinking disabled and one decode/prompt request at a time. The supervisor retains a 2 GB prompt-cache budget and checks weight residency after the health endpoint comes up.
September's Lightning HumanEval and speed numbers describe the earlier Nemotron artifact. They are not scores for SuperQwen. Gemma, the local Qwen fallback, and persistent embeddings keep their distinct lifecycle policies.
The manual coder-switch script stops the competing two-node backend, waits for the selected server to become healthy, and then moves coder and vision together. Its inspected alternatives are Qwen/vLLM and GLM/TensorFold. The historical Qwen/TensorFold speed probe is not the active coder. A configured alternate proxy is not evidence that its server is loaded.
The practical checklist is in the dual-GB10 operations guide, with the current placement and dated measurements. The October 2 audit remains the earlier hardware and routing record.