Research topic
activeOne layer, many personal devices.
Cross-device tensor parallelism
Shard attention and FFN matmuls across a user’s own phones and laptops so they compute the same layer together — pooling FLOPs and weight memory past what any single device can hold.
Why it matters
Personal fleets already outnumber any one machine’s capacity. Edge tensor parallelism is communication-bound and device-skewed; the interesting question is when the aggregate still beats running alone.
We’re studying placement across heterogeneous personal devices — not a datacenter rack — so latency, battery, and link quality are first-class constraints.
Open questions
How do you partition layers when devices differ in FLOPs, memory, and interconnect by an order of magnitude?
When does the communication cost erase the pooled-compute win?