This work presents LongStraw, an objective-aware, architecture-aware system for resident-state virtualization, response replay, and distributed-gradient execution that bounds the live training graph by the response suffix while reusing the expensive prompt computation across the complete GRPO group.
Abstract
Long-context RL post-training is constrained by the lifetime of state and gradients, not attention cost alone. In GRPO, one multi-million-token prompt must serve old-policy and reference scoring plus multiple policy responses, while conventional autograd keeps the prompt graph and all response graphs live alongside model weights, caches, and distributed communication buffers. We present LongStraw, an objective-aware, architecture-aware system for resident-state virtualization, response replay, and distributed-gradient execution. Its transaction captures the shared prompt without autograd, retains only the architecture-required state on explicitly owned pages, restores that state for each group member, scores old/reference branches without a graph, replays one policy response at a time with autograd, and accumulates the resulting gradients before one distributed finalization and optimizer step. This schedule bounds the live training graph by the response suffix while reusing the expensive prompt computation across the complete GRPO group. We instantiate this design for two incompatible model structures. Qwen3.6-27B combines 48 recurrent GDN layers with 16 full-attention layers; LongStraw keeps the compact recurrent state and physically CP8-sharded KV pages, composes global attention through cross-rank LSE/output merging, and performs blockwise response replay. GLM-5.2 combines a 78-layer MLA/DSA attention stack with a 256-expert, top-8 MoE tail. Its implementation keeps CP-sharded MLA latent pages and DSA indexer-key pages in CPU memory, stages one layer at a time, reconstructs IndexShare-aware global sparse selection over CP32, and dispatches routed response tokens over EP32. The two paths share one transaction contract while specializing the retained state, replay operator, and collective communication to the architecture...
Prefill-Pressure Adaptive Scheduling (P-PAS), a lightweight policy that dynamically adapts the scheduling budget based on concurrent prefill and decode state, is introduced, maintaining low end-to-end latency across changing load regimes, avoiding the limitations of a fixed MBT.
LazyTrain is proposed, an optimization layer over a layer-streaming executor that formulates checkpoint selection, activation placement, recomputation, and CPU-GPU-NVMe communication overlap as a mixed-integer scheduling problem, then executes the solved policy during training.
Xiao-Jun Wu, Cehao Yang, Honghao Liu et al.· 0 citations
This work proposes Akashic, a low-overhead memory system built around MemAttention, which organizes context into bounded chunks and models semantic relationships across chunks, preserving cross-chunk evidence without repeatedly rewriting the full history.
Yang Liu, ZhaoKai Luo, Huayi Jin et al.· 0 citations
Modern large language model (LLM) inference suffers from severe Time-To-First-Token (TTFT) bottlenecks. Existing prefix KV caching mechanisms are inherently stateless, forcing a trade-off between cross-chunk attention accuracy and online recomputation overhead. To address this issue, we propose Pegasus, a novel stateful prefix KV caching system that aims to achieve full-context attention accuracy while avoiding costly recomputation. To handle the exponential growth of context states under limited memory capacity, Pegasus employs a Recursive Path-Pruning Caching (RPPC) algorithm to selectively cache high-value states based on access frequency, memory footprint, and asymmetric latency benefit. In addition, Pegasus introduces a transition-based KV management mechanism to mitigate cache-miss overhead. By exploiting the sparsity of state-dependent KV variations, it replaces expensive attention recomputation and I/O-intensive tensor reloading with lightweight sparse state transitions. Extensive experiments show that Pegasus improves end-to-end serving throughput by 45.9% on average, reduces TTFT by up to 78.5%, and lowers cache-miss recovery overhead by more than 72%.
Fahao Chen, Peng Li, Dongxiao Yu et al.· Fall Joint Computer Conferen...· 0 citations
LLM serving caches prompt KV state, yet most front ends still re-tokenize the full request on every call. Coding agents pay most: sessions repeatedly submit a long transcript after a small append, which can shift token boundaries near the end of the prior sequence. Across 153,951 calls the median append is ~1.4K characters; only 1.0-3.6% of calls start or rebuild a session, yet those carrymulti-million-character contexts. Fleet prompt-cache hit rate is 94.1%, and as it approaches 0.99, tokenization grows from 10% to 64% of time to first token (TTFT) in component measurements. TokTier is a stateful CPU+GPU tokenization service for this two-mode workload, under one contract: emitted token IDs are always identical to full reference tokenization. For session continuations it re-tokenizes a small window around the append and splices only when a per-request check finds a stable pre-tokenization boundary; failed checks widen the window or fall back to full reference tokenization. For calls without a reusable prefix it runs exact GPT-family regex pre-tokenization and BPE on a GPU. A sampled shadow verifier re-checks live traffic. Across 17 production tokenizer families, differential campaigns cover 1.5x10^10 split checks, a 12.4 TB real-text corpus, and 93,000+ replayed agent steps, with zero divergence. Incremental repair takes 0.5-1.1 ms from 100K to 3M characters, up to 437x faster than HF tokenization and 2.1x faster at 1M characters than the strongest cache-based baseline (Gigatoken) fully prewarmed. GPU tokenization encodes a 1M-character request in 0.87 ms, up to 491x below HF and 23.4x below the fastest published CPU method on the same protocol. With vLLM, median TTFT drops 16-34% and P99 TTFT 23% under recorded bursts. Under a 50 ms P99 objective, a four-core repair pool plus one GPU sustains 1,821 requests/s, where a 16-core stateless front end saturates at 40 requests/s.
Multi-task supervised fine-tuning (SFT) often casts a heterogeneous data mixture as a single optimization problem, even though different tasks may reach their best generalization at different times. msft exposes this mismatch through task-wise roll-out, exclusion, and rollback, but its original formulation materializes the scheduler state as full-model checkpoints, making stage transitions costly to store, restore, and deploy. This paper introduces AuroSFT, a parameter-efficient framework that recasts the carried state of overfitting-aware multi-task SFT as a compact, mergeable adapter state. AuroSFT freezes the pretrained backbone, trains only injected adapters, rolls back adapter checkpoints at task-wise peaks, and continues on the remaining active mixture. At the layer level, each adapter applies an AuroRA-inspired adaptive nonlinear layer to a low-rank weight factor rather than to the sample representation. The resulting update remains linear in the input, rank-bounded, and exactly mergeable into the frozen projection. Under the retained-backbone comparison protocol, AuroSFT achieves 61.36% average accuracy, compared with 59.85% for the corresponding msft reference row, and obtains higher accuracy on all five backbones. Our code is available at the anonymous repository: https://anonymous.4open.science/r/AuroSFT-80D1.