Nvidia says it can shrink LLM memory 20x without changing model weights

Nvidia's KV Cache Transform Coding (KVTC) compresses LLM key-value cache by 20x without model changes, cutting GPU memory ...

New KV cache compaction technique cuts LLM memory 50x without accuracy loss

MIT researchers developed Attention Matching, a KV cache compaction technique that compresses LLM memory by 50x in seconds — without the hours of GPU training that prior methods required.
i-SCOOP

Mamba 3, a state space model and an alternative to transformers

Mamba 3 is a state space model built for fast inference. Learn what it is, how it works, why it challenges transformers, and ...