MLA with rotary embeddings at half latent rank keeps validation loss nearly unchanged while cutting KV-cache memory by roughly half on small language models.
Fu, Stefano Ermon, Atri Rudra, and Christopher Ré
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Latent Multi-Head Attention for Small Language Models
MLA with rotary embeddings at half latent rank keeps validation loss nearly unchanged while cutting KV-cache memory by roughly half on small language models.