SparkRing's profiles run three models created by Local Inference Lab, Inc. and licensed under the Local Inference Lab License, Version 1.0: GLM-5.3-Flash-NVFP4-MXFP8-CSF-QAD (glm53-flash-tp2, glm53-flash-tp4, glm53-flash-tp2-dcp2), GLM-5.3-NVFP4-CSF (glm53-nvfp4-csf-tp6, glm53-nvfp4-csf-tp8) and Qwen3.8-Flash-Next-NVFP4-MXFP8-CSF-QAD (qwen38-flash-next-tp2). https://local-inference-lab.ai/

SparkRing serves large language models on two to eight NVIDIA DGX Sparks cabled directly over ConnectX-7, with no switch. One command installs a model on all of them, or several models on separate groups of them.

1. Cable two to eight Sparks directly: a pair from p0 to p0, or a line or ring from each Spark's p0 to the next Spark's p1 (requirements).
2. Pick a --profile from the table below, or build the commands in the Install Builder.
3. On the Spark connected to your network, run:

curl -fsSL https://fujitsupolycom.github.io/sparkring/install.sh | bash -s -- --profile PROFILE

This installs the release the Install Builder describes; the pinned install command installs main or another commit.

The installer asks before it changes anything: --plan previews and --yes approves, and running it again updates or switches models. --on 0-3 places a model on part of the Sparks. Node A restarts a model that stops serving; --no-auto-recover turns that off.

More: documentation · commands · status dashboard · Docker Compose files · pinned install command

SIRCL, the Switchless Inference RDMA Collective Layer, carries the model's collectives between Sparks, including Sparks reached through another Spark's ConnectX-7 relay (architecture). net-sircl brings it to any NCCL application as a network plugin (research-only).

glm53-flash-tp2-dcp2 splits the KV cache across the pair (decode-context parallelism 2). It, glm53-nvfp4-tp8 and swift15-qwen38-flash-next-tp4 are research-only; the other rows are implemented. The six-Spark GLM-5.3 profile glm53-nvfp4-csf-tp6 runs on positions 0-5 of an eight-Spark ring (research-only). More profiles and options: profile catalog.

Output tok/s, measured with llm-inference-bench 0.7.6 at temperature 1 on release 2026.10.2's image, one start per profile installed without research options; decode and prefill at 32K context. The default image, release 2026.10.5's, has no speed record of its own, nor has its parent, release 2026.10.4's. Records: every cell's raw source, rings of four, ring of eight, paths of four.

• Benchmarks and test results
• Architecture
• Container images
• Contributing and repository layout
• Community discussions

SparkRing builds on Local Inference Lab (GitHub · Hugging Face · Discord): the B12X kernel library, the Qwen3.8-Flash-Next and GLM-5.3-Flash Spark NVFP4 checkpoints, llm-inference-bench and their vLLM fork.

SparkRing includes software derived from RoCEnante (b12x/comm/roce, Local Inference Lab), written by Jason (original-el8) in local-inference-lab/b12x PR 295, with dual-rail striping and graph-collective work by Luke Alonso (lukealonso), licensed under Apache-2.0; b12x now lives in FlashInfer at flashinfer/experimental/b12x (provenance).

SparkRing's NCCL patches follow Joseph Rose's nccl-spark-switchless. Thanks to Eugene Rakhmatulin for the eugr/spark-vllm-b12x base image, and to the vLLM, NVIDIA NCCL and ExLlamaV3 contributors. The third-party notices give origins and licenses.

SparkRing code is Apache-2.0. Model weights and bundled components keep their own terms; check the model cards and third-party notices.