Synthesis history

New items this pass are thin: a second tweet confirming UW SyFi's multi-prize win[^20724] and the public FlashInfer Bench GitHub starter kit[^20725] add minor depth to the contest thread, and Paradigm.xyz's Attention Ke…

vLLM's disaggregated prefilling documentation for both v0.8.5[^20326] and v0.10.2[^20327] confirms the 'experimental' label has not been lifted through the most recent tracked release, directly answering a prior open qu…

Third-party deployment guides from Vultr[^19593] and Spheron[^19594], plus a dedicated vLLM Dynamo integration page[^19595], extend the disaggregation toolchain maturity signal beyond NVIDIA's own Kubernetes documentati…

NVIDIA Dynamo's official Kubernetes documentation for disaggregated communication[^18696] is the most substantive addition this pass: it directly addresses the RDMA KV cache networking obstacle that was previously chara…

The Native Sparse Attention paper (arXiv 2502.11089, ACL 2025) enters the thread as a substantive counter-weight to the sparse-attention-as-stopgap debate: NSA designs hardware-aligned sparse attention from training tim…

Three substantive additions this pass: (1) @superaiwatcher introduces the first explicit counter-narrative in the thread, framing sparse attention as a transitional stopgap before hardware-native linear attention, which…

MIT HAN Lab's Adaptive Drafter paper (arXiv 2511.16665) is now confirmed as ASPLOS'26 with an open-source GitHub repository (mit-han-lab/fastrl) and MIT News coverage; secondary sources characterize the speedup as 2x, s…

Attention-FFN disaggregation moved from a single conference mention to a concrete engineering push this pass: StepFun's StepMesh library, a vLLM RFC, and formal papers on provisioning and hardware challenges all appeare…

MLSys 2026, the ninth annual Conference on Machine Learning and Systems, is underway in Bellevue, Washington (week of May 18–22, 2026)[^7796][^10577]. The conference's inference track is organized around four converging…