An independent, reproducible serving lab for running Qwen3.8-27B on one NVIDIA DGX Spark GB10.
The reference stack uses a mixed-NVFP4 target, a DFlash2 probabilistic K7 drafter, and an 8.3 MiB native rank-1 PEFT LoRA. One vLLM process exposes both a clean base model ID and a per-request adapter model ID. Requests without the adapter stay on the clean CUDA-graph path, while native LoRA identity isolates prefix-cache entries.
The repository includes:
- a digest-pinned ARM64 vLLM/DFlash2 Docker build;
- mixed FP8/NVFP4 output-projection to PEFT-LoRA conversion;
- clean base and adapter aliases from one server;
- C1/C8 benchmark harnesses and sanitized JSON results;
- OpenAI Chat, Responses, and Anthropic Messages validation;
- forced-tool and prefix-cache-isolation gates;
- architecture, memory, methodology, and license documentation.
Measured clean-base C1 throughput recovered from 30.32 tok/s under the previous always-on hook to 45.41 tok/s with native LoRA routing. See the full technical article for protocol, limitations, and interpretation.