Running MiniMax 2.5 on 2× DGX Spark (Multi-node, vLLM)

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10
1

Hi,

I’m trying to deploy MiniMax 2.5 on a setup of 2× DGX Spark nodes, connected in a cluster (stacked through a switch).

I started with vLLM + Ray, following this guide:

NVIDIA Multi-Spark vLLM Guide

Current setup:

  • 2 DGX Spark nodes connected via switch

  • Ray cluster initialized (head + worker)

  • vLLM installed and running

  • Inter-node communication seems OK

Goal:

  • Run MiniMax 2.5 distributed across both nodes

  • Efficiently utilize all GPUs (tensor / pipeline parallelism)

Questions:

  1. What is the correct way to launch MiniMax 2.5 with vLLM in a multi-node setup?

  2. Is Ray the recommended approach here, or are there better alternatives for this scale?
    (e.g., native distributed backends, MPI, or other orchestration methods)

  3. How should tensor vs. pipeline parallelism be configured across nodes?

  4. Are there specific NCCL / networking optimizations required for DGX Spark over a switch?

  5. Any known limitations or best practices for running large models like MiniMax 2.5 in this environment?

I’m also open to approaches other than Ray if they are more stable or performant for multi-node inference.

If anyone has experience running large models across 2 DGX Spark nodes, I’d really appreciate guidance, recommended architecture, or example configs.

Thanks!

2

I highly recommend to use the community docker version of eugr instead.

For expected performances for two nodes head over to the famous Spark Arena

where you will also find a recipe ready to use with the community docker image.

A switch for just to nodes will lead to higher latencies (AFAIR).

Networking documentation:

The AWQ 4bit quant still seems to be the fastest variant (see rest of the leaderboard).