DRIFT
Splits one large language model across a Mac and a separate Windows or Linux PC over an encrypted, framework-neutral network protocol, producing output identical to running the whole model on one machine.
DRIFT runs a single large language model across heterogeneous personal computers, such as a Mac using Apple's MPS backend and a Windows or Linux PC using NVIDIA CUDA, by splitting the model layer by layer and streaming only the hidden state between nodes over a framework-neutral byte protocol built on TCP and msgpack. Because the data plane depends on no ML framework, machines with different GPU vendors can run one model together, and the result is described as bit-for-bit identical to running the whole model on a single machine.
Its decentralization layer streams hidden state peer-to-peer rather than through a central hub, encrypts and authenticates the wire, recovers a dropped node, signs a receipt at every hop that the head verifies against live traffic, and lets nodes discover each other automatically. A model can be split, layer by layer, across as many as 28 to 35 machines depending on the model, though the project describes two to four nodes as the current practical range.