2026-11-05

Hybrid language models interleave a few softmax-attention layers with many linear-attention or state-space layers, so every layer keeps a memory of earlier tokens: a key-value cache or a fixed-size recurrent state. Serving them spends more compute and high-bandwidth memory on these memories than the output requires. The prefill runs every layer for every prompt token, and every decode step loads and stores whole recurrent states, which on Qwen3.8-Flash-Next take as many bytes per request as the attention cache of thousands of tokens. We find two properties of trained hybrids that locate this excess. First, writes are easier than reads: writing a deep layer's memory of a token needs little of the deep computation that predicting the next token needs. Trained copies of the deep layers' write paths, which we call emitters, therefore write a prompt's deep memories from the residual stream of a shallower anchor layer, and all prompt tokens but the last run as few as 63% of the layers. Second, recurrent states have a state sink: most of their energy lies in a rank-one term along a fixed direction of each head, and the rest is a low-rank working memory. SEED builds on both: it makes the prefill shallow in depth and the decode state shallow in representation. Each recurrent state is stored as its sink plus low-rank factors, which the Gated DeltaNet, Kimi Delta Attention and Mamba-2 updates extend exactly; we prove that the open-loop error of the periodic compressions back to low rank is bounded independently of their number. SEED keeps the original weights frozen, compresses no attention cache, and runs all layers for every generated token. On six checkpoints of five production hybrids of 30B to over 300B parameters, it is nearly lossless on eleven official exams. Served with SGLang on NVIDIA GB300 GPUs on an agentic-coding trace, it raises the per-GPU throughput of Qwen3.8-Flash-Next 1.2-1.4x.

Speculative decoding accelerates large language model inference with a small drafter, whose per-guess acceptance rate determines the achievable speedup. Feature-level drafters share a design choice: every token, easy or hard, is drafted under a single forward pass. Many of the tokens the target rejects are ones the drafter is itself unsure about yet the target answers decisively. To address this, we introduce Parallel Reconsideration: guided by its own uncertainty, the drafter reconsiders its hesitant guesses with extra passes that run in parallel with the one-pass draft, adding negligible latency. A co-designed training recipe and serving-time drafting algorithm append these reconsidered guesses to the original draft tree as refined candidates. Across models and benchmarks, on average, Parallel Reconsideration improves over the baselines by 5.3% to 10.2% on acceptance length and by 3.5% to 9.2% on speedup.

We present flash-comm, a communication library that achieves efficient transport for LLM decoding by fusing readiness synchronization with data movement. flash-comm introduces self-synchronizing transfers in which each data block carries evidence of its readiness, allowing receivers to directly validate the blocks they need without separate inter-GPU barriers. Based on this principle, a cost-guided protocol selector adapts the readiness encoding to message size and the number of participating GPUs, balancing synchronization overhead against additional data-transfer cost. flash-comm further decouples transport order from reduction order to preserve result consistency across GPUs under asynchronous data arrivals. On a server with eight NVIDIA B300 GPUs, flash-comm reduces all-reduce latency by a median of 28.8% and up to 43.7% compared with custom AR v2 (the state-of-the-art all-reduce implementation). In end-to-end decoding, flash-comm reduces time per output token (TPOT) by 1.7-6.8%. Note that all-reduce accounts for 11.4% of the baseline decoding time, which bounds the latency reduction possible from optimizing this operation alone. The same mechanism also supports standalone all-gather, reducing latency by a median of 23.8% and up to 58.9% across eight B300 GPUs.

Mixture-of-Experts (MoE) models have become prevailing, yet their distributed deployment suffers from severe dynamic load imbalance due to the conflict between static expert placement and inherent routing dynamism. While host DRAM offers capacity to avoid static device storage for load balancing, we identify existing offloading works fundamentally fail in distributed inference due to two overlooked challenges: (1) coupled performance impacts of runtime Host-to-Device transfers due to load dynamism and shared PCIe topology; (2) system-level memory overhead incurred by DRAM usage at scale. In this paper, we propose vExpert that shifts DRAM offloading paradigm from partial residency to full redundancy through virtualized expert storage. We pioneer the first systematic performance model that jointly captures H2D latency, PCIe contention, and load imbalance, building a lightweight allocator to dynamically map physical experts to virtual slots runtime. A novel disaggregated manager is designed to reduce memory overhead in multi-process environments. Evaluated on DeepSeek-V3.1 within an 8-node H100 cluster, vExpert improves load balance ratio by up to 55% and achieves 1.2$\times$ end-to-end speedup.