Glm 5.3 Flash 4x Dgx Spark Switchless

GLM-5.3-Flash NVFP4 — 4× DGX Spark, switchless-ring TP4 + DFlash2

This repository is the recipe and the contract : every address, interface, and hostname is a placeholder you swap for your own — nothing here depends on a private gateway, router, or network.

Not tied to these weights, either. The ring fabric, the patched NCCL, and the rank-launch pattern know nothing about the model — the same four-node fabric has also served GLM-5.2 (in more than one quant format: EXL3, and QuantTrio) and DeepSeek-V4-Flash at TP4. To adapt it, swap the weights and the model-specific serve arguments (parsers, drafter, MoE backend, and KV sizing); everything else in the recipe carries over.

And no — this is not Sparkring under a different name. There is no Sparkring/SIRCL anywhere in this stack. The collectives form on a patched NCCL 2.30.7 — a skip-tree-connect change LD_PRELOAD -ed into every container, because stock NCCL's tree-connect step wedges on a switch-free point-to-point fabric — plus a pinned NCCL runtime profile (ring algorithm, fixed channels, subnet-aware dual-rail routing) worked out for this topology. That combination was validated on this ring with Sparkring absent entirely.