llmwish.open(@ciprianveg):Do you need some extra memory on your DGX Spark? I created this repo to help the DGX Spark users that have a spare 10-24 GB GPU at home to squeeze some extra memory out of a single Spark or a Sparks cluster. It moves the spec-decode draft model off your Sparks onto that GPU: the freed GB of memory can be used for extra context, or better quant quality. Supports both TCP and RDMA, shipped as eugr-vllm compatible mods:

2026.09.28 17:33

Do you need some extra memory on your DGX Spark? I created this repo to help the DGX Spark users that have a spare 10-24 GB GPU at home to squeeze some extra memory out of a single Spark or a Sparks cluster. It moves the spec-decode draft model off your Sparks onto that GPU: the freed GB of memory can be used for extra context, or better quant quality. Supports both TCP and RDMA, shipped as eugr-vllm compatible mods: