A video demonstrates an observability stack built to tune local LLM inference on the NVIDIA DGX Spark, using vLLM, Prometheus, and Grafana. The stack provides real-time visibility into token throughput, TTFT, KV cache utilization, prefix caching, concurrency, speculative decoding acceptance, system utilization, and estimated local-vs-cloud inference cost. The hardware includes the NVIDIA DGX Spark GB10 Grace Blackwell Superchip with 128GB unified memory. The inference setup uses vLLM with Qwen3.6-35B-A3B NVFP4, a 262,144-token context, FP8 KV cache, prefix caching, chunked prefill, and MTP speculative decoding. An example vLLM configuration specifies parameters such as `--gpu-memory-utilization 0.85`, `--max-model-len 262144`, `--max-num-seqs 4`, and `--kv-cache-memory-bytes 17179869184`. The video notes that NVIDIA officially documents vLLM as a DGX Spark inference path, including Qwen3.6 support, and that vLLM's own DGX Spark documentation calls out Prometheus telemetry, batching, KV-cache behavior, and throughput tuning. The dashboard and repository are intended to make this information visible in one place for users running local LLMs, Qwen, coding agents, or other agentic workloads on the DGX Spark.