I Ran 3 vLLM Endpoints on One DGX Spark—Here’s What Actually Fit

Published 04/10/2026

Description

Can one NVIDIA DGX Spark run multiple local AI agent endpoints for Claude Code? In this hands-on test, I use Docker Compose to launch separate vLLM servers on a single DGX Spark, each with its own port, model name, and context-window configuration. The goal is to prepare dedicated explorer, implementer, and reviewer endpoints that a router such as LiteLLM could later use for local sub-agents. Contents: 0:00 - Introduction 0:41 - Description 2:27 - Docker Install 6:17 - vLLM install 11:19 - Docker Compose 22:37 - Tests 27:29 - Conclusion I start by checking Docker and NVIDIA GPU access, then launch Qwen3.8-27B NVFP4 in vLLM and verify that its API responds. From there, I build a Docker Compose setup with a shared Hugging Face download cache and three independently configured endpoints. Sharing downloaded model files avoids downloading the checkpoint for every container, but it does not make three running instances share one model allocation in memory. The memory test changed the design. Two Qwen3.8-27B NVFP4 instances started successfully, but a third 27B instance pushed the DGX Spark to its memory limit and failed to load—even with a shorter context window. I replaced the reviewer with a smaller Qwen3.5-9B NVFP4 model and was able to bring all three endpoints online. I then sent requests to the explorer, implementer, and reviewer at the same time. They responded, but concurrent inference on separate endpoints was much slower than using the DGX Spark as a fast single-request machine. Three containers provide separate endpoints and queues; they do not provide three separate GPUs. This video covers: Running vLLM and Qwen NVFP4 models on DGX Spark Configuring multiple vLLM services with Docker Compose Reusing a local Hugging Face model download cache Setting separate ports and context windows for agent roles Testing memory limits with two 27B models and a smaller reviewer Sending concurrent API requests and inspecting the results This is