Search papers, labs, and topics across Lattice.
1
0
2
End-to-end learned subspace projections shatter prior compression limits, achieving 70% weight reduction in Llama-2-7B with near-dense performance and drastically shrinking KV cache memory.
