Hi all, I wanted to share SparkRing, an experimental inference stack I’ve been building for running GLM-5.2 across four directly connected DGX Sparks - without an Ethernet switch in the inference path. This started as a response to “Your Mikrotik 804 is backordered until September…” So here we ar…

Got this project to the point it needed a new thread. Much improved repo, profiles, etc. 2x, 4x, and (coming soon) 6x ring profiles. Switchless 2x 4x 6x GB10 Clusters Serving GLM, DeepSeek, and Qwen

We need more supply chain issues to drive great innovation like this. I will give this a try once I have enough cables next week (supposedly).

Excellent! Yep, as they say, "necessity is the mother of invention.” Yes, I NEED GLM 5.2 lol. EDIT: Repo is being updated to improve the setup workflow. Tis a mess right now, so good luck to anyone who tries piecing it together at the moment. Will be cleaned up asap.

You are pushing me towards the edge of buying 2 more sparks to make it 4, lol The switch was a great excuse for me not to pursue more in getting more sparks, and now you are ruining it 😭

I would purchase those split DACs day 1. That would take a whole lot of headache out of this. Assuming the sparks can take advantage of them, how bandwidth, latency, etc are affected. Fun experiments for sure! Yeah, I went model specific to start, but hope to ultimately shift towards generic model …

Indeed, the journey is part of the fun. I hope you manage to get your approach model agnostic, the more options the merrier for the end users.

Got lost in the weeds trying a different model, very little done to clean the mess of a repo, but this model is performing just as well in initial testing. I’m going to A/B some stuff on it and then try to pin down a good baseline config that people’s clankers should be able to get stood up. (I PROM…

@FujitsuPolycom Spent all day trying to set it up today. Really appreciate you taking the time sharing this and I got really excited testing it. I finally managed to get it up and running thanks to Claude. I’m also waiting for my Mikrotik to arrive(got 8 nodes sitting here), and I’m blown away you m…

Amazing! Thank you for the detailed report and putting in the effort to slog through that mess. I know that wasn’t easy. The numbers and reports from Claude are super useful! The good/bad news? Most of those issues I have worked out now or are simply not present in the latest quant I’ve switched to…

Ok, the repo should have an easier path to getting this model up, with these specs: Model: madeby561/GLM-5.2-MXFP8-NVFP4-NF3-Hybrid TP4 DCP4 AMTP2-4 (adaptive MTP 2-4 tokens) KV format: NVFP4 MLA + FP8 RoPE + per-token scaling Serving: max 8 sequences, 4096 batched tokens Weight mix: 64 NVFP4 +…

GLM 5.2 on 4x GB10s - Switchless Ring - 22 tps Coding

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

mashie July 29, 2026, 4:41pm 5

If my DAC idea works out then 4-nodes without a switch will be a reality without model specific workarounds.

Topic		Replies	Views
GLM-5.2 on a 4× GB10 cluster: ~22 tok/s decode, 256K ctx, Recipe DGX Spark / GB10 llama , deepseek	200	13243	August 9, 2026
Switchless 2x 4x 6x GB10 Clusters Serving GLM, DeepSeek, and Qwen DGX Spark / GB10 Projects cuda , kernel , performance , deepseek	2	505	September 3, 2026
GLM-5.2 on a 3× GB10 cluster: ~16,13 tok/s decode, 215K ctx FP8 TP=3 + VISION! DGX Spark / GB10 Projects	7	1261	July 26, 2026
Fitting a high-quality REAP-less GLM-5.2 onto 4x DGX Spark DGX Spark / GB10 deepseek	0	2079	June 29, 2026
GLM-5.3-Flash NVFP4 on 3× DGX Spark — TP=3, 512K context, 35 tok/s DGX Spark / GB10 llama	7	1446	August 28, 2026
How to run GLM 4.7 on dual DGX Sparks with vLLM / mods support in spark-vllm-docker DGX Spark / GB10	27	4806	January 2, 2026
Followup: Mystery Solved: 4x Spark, GLM-5.2-nfp4, 24tp/s, 128k ctx, no REAP DGX Spark / GB10	34	2436	July 8, 2026
Running GLM-4.7-FP8 (355B MoE) on 4x DGX Spark with SGLang + EAGLE Speculative Decoding DGX Spark / GB10 Projects	38	2913	June 24, 2026
Why 273 GB/s? Less Is More, Until It Isn’t DGX Spark / GB10	67	3391	March 27, 2026
6x Spark setup DGX Spark / GB10	112	12600	April 25, 2026

GLM 5.2 on 4x GB10s - Switchless Ring - 22 tps Coding

Related topics