MLA packs attention into a single latent KV stream, removing the per-head cache boundary that head-level reuse relies on. RedKnot-MLA processes each immutable document offline at canonical position 0 (keeping certified local-head contributions as MLA-Off); at serving time, query-side RoPE relocation restores request positions while recomputing a small global-head set and protected local token rows (MLA-On), merged before a shared output projection — the packed latents are never split. On DeepSeek-V4-Flash (37 reusable layers, 56/8 local/global split) it reaches 75.29% of the analytical head-row reuse ceiling (78.89% for Pro-0813); hot-artifact TTFT improves 2.02-3.84x, with 78.7-79.5% compute savings per major operator at 256K, with small quality deltas (aggregate F1 +3.24, EM +4.16; one dataset -2.81 F1). A ~2.0x QPS figure is labeled preliminary rather than archival.
RedKnot-MLA: prefix-cache-style reuse finally works for MLA models
MLA packs attention into a single latent KV stream, removing the per-head cache boundary that head-level reuse relies on. RedKnot-MLA processes each immutable document offline at canonical position 0 (keeping certified local-head contributions as MLA-Off); at serving time, query-side RoPE relocation restores request positions while recomputing a small global-head set and protected local token rows (MLA-On), merged before a shared output projection — the packed latents are never split. On DeepSeek-V4-Flash (37 reusable layers, 56/8 local/global split) it reaches 75.29% of the analytical head-row reuse ceiling (78.89% for Pro-0813); hot-artifact TTFT improves 2.02-3.84x, with 78.7-79.5% compute savings per major operator at 256K, with small quality deltas (aggregate F1 +3.24, EM +4.16; one dataset -2.81 F1). A ~2.0x QPS figure is labeled preliminary rather than archival.
Imported from the existing research archive. Migration preserves the recorded judgment and is not a fresh verification.
What this means for you
Standard prefix caching is blind on MLA models (DeepSeek V4/V4.1 serving), which is exactly where RAG and multi-agent workloads reuse documents. RedKnot-MLA brings document-level reuse back to that architecture family — directly relevant to anyone serving DeepSeek models with shared corpora.
Public artifacts and validated results are different.
The recorded fact is the public work and its artifacts. Author-reported outcomes do not establish independent reproduction. Next checks include usable code, task conditions and third-party validation.
Evidence & related material
Technical details & scores
| Arxiv | 2609.07008 (Wed 9 Sep digest, announced 2026-09-10T00:00Z) |
|---|---|
| Mechanism | offline document processing at canonical position 0 (MLA-Off, certified local-head contributions) + query-side RoPE relocation at serving (MLA-On, small global-head recompute + protected local rows); packed latents never split |
| Results | DeepSeek-V4-Flash: 75.29% of analytical head-row reuse ceiling (Pro-0813: 78.89%); hot-artifact TTFT 2.02-3.84x; 78.7-79.5% per-operator compute savings at 256K; F1 +3.24 / EM +4.16 aggregate, one dataset -2.81 F1; ~2.0x QPS preliminary |
| Code | not stated in abstract |