[zfountas]

Hi MacsHeadroom, first author here. Thanks for the great questions about compute/memory trade-offs.

The quick take: To give you an example of processing speed, with a 7B model on an NVIDIA V100, EM-LLM processes (or generates) about 326 tokens/sec with a 51.2K context window (which is quite competitive for these old GPUs).

More broadly, EM-LLM is designed to make ultra-long contexts (memory-prohibitive for standard O(n^2) attention) computationally tractable. The Appendix C of our paper https://openreview.net/pdf?id=BI2int5SAC details how: significantly better attention scaling, efficient O(nm) memory formation, and large KV cache management via CPU/disk offloading. While there's a slight per-chunk overhead compared to the simplest retrieval methods initially, the crucial part is our ability to handle sequences at scales infeasible for full-context models. For instance, we're successfully using 8B models with 10M token contexts on a single GPU without prohibitive delays.

Regarding RAG in particular, EM-LLM often shows significant gains on tasks needing deep understanding of a single, long, coherent context. A key reason is that EM-LLM allows each layer to retrieve and integrate relevant information from different "episodes" of the context independently, offering more nuance than a typical single RAG step, for similar overall resource use.