# #uma — MDRSS hashtag feed

> Public MDRSS cards tagged #uma.
> Canonical feed: https://mdrss.com/feeds/uma

## Cards (2)

### [G1: CUDA 12/13 ABI Mismatch](https://mdrss.com/llm-engineering/inference-and-quantization/901285/901285.md)

Last verified: 2026-07-14 — refresh when CUDA, PyTorch, or the DGX Spark stack ships a new major version. Use it to give an agent explicit responsibilities, steps and constraints.

Classification: llm-engineering/inference-and-quantization · Feed: llm-engineering · Updated: 2026-08-04T13:54:51.641Z · Version: 1

### [Spark Memory & Thermal Ops](https://mdrss.com/platforms/devops-and-infrastructure/901283/901283.md)

DGX Spark's GB10 chip has one 128GB unified memory (UMA) pool shared by CPU and GPU, and a sustained power ceiling well below its rated figure. Both break discrete-GPU assumptions: headroom isn't what nvidia-smi reports, and a run that starts fast will slow down mid-job with no. Use it to give an agent explicit responsibilities, steps and constraints.

Classification: platforms/devops-and-infrastructure · Feed: platforms · Updated: 2026-08-04T13:54:51.641Z · Version: 1