291 posts in total. Keep on posting.

2026

• 07-10 中

  FlashSVD v1.5：为什么低秩大模型不会自动变快

  FlashSVD v1.5 揭示了 SVD 压缩后的 Transformer 推理依然缓慢的根本原因不是低秩运算本身效率低，而是朴素运行时将每层拆散成数百次细碎内核调用；论文提出密集 KV 解码注意力、打包 MLP 投影与逐层 CUDA 图重放三项机制，在不改变任何压缩检查点的前提下将解码速度提升最高 2.55 倍。

• 07-09 EN

  RATrain: Training-State Lifecycle Scheduling for Dense LLM Training on Bandwidth-Constrained Heterogeneous Supercomputers

  RATrain reframes dense LLM pipeline training as a fine-grained training-state lifecycle scheduling problem, letting a bandwidth-starved, memory-constrained domestic HPC accelerator match 8xA800-class throughput without changing training semantics.

• 07-09 中

  RATrain：面向带宽受限异构超算的稠密大模型训练状态生命周期调度

  RATrain 把稠密大模型的流水线训练重新建模为细粒度的训练状态生命周期调度问题，让一台算力尚可但显存和带宽都很紧张的国产超算加速器，在不改变训练语义的前提下逼近 8xA800 级别的训练吞吐。

• 07-08 EN

  DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation

  DSpark pairs a semi-autoregressive drafter — a parallel backbone plus a lightweight sequential head — with a hardware-aware confidence scheduler that casts verification-length selection as a provably lossless throughput-maximization problem, shipping in DeepSeek-V4 production for a 60-85pct per-user speedup.

• 07-08 中

  DSpark:半自回归生成与置信度调度的投机解码

  DSpark 用「并行主干 + 轻量顺序修正头」的半自回归草稿模型，配合一个把验证长度选择问题严格建模为可证明无损、感知硬件负载的吞吐最大化问题的置信度调度器，已在 DeepSeek-V4 生产环境中实现单用户 60-85% 的生成加速。

• 07-07 EN

  RSPO: Reward-Swap Policy Optimization for Multi-Turn LLM Agents

  RSPO bridges the gap between sparse outcome rewards and dense process rewards in multi-turn LLM agent training by using a reward-swap cyclic framework that expands trajectory diversity without reward misalignment.

• 07-07 中

  RSPO：多轮LLM智能体的奖励交换策略优化

  RSPO 通过奖励交换机制，让稠密过程奖励训练的探索智能体扩展轨迹多样性，再用结果奖励训练最终策略，同时避免奖励错位与奖励黑客问题。

• 07-06 EN

  The Mirage of Optimizing Training Policies: Monotonic Inference Policy Improvement for LLM RL

  MIPU exposes an overlooked objective-level flaw in LLM RL: training-side improvement does not guarantee inference-side improvement under training-inference mismatch, and proposes a two-step framework to fix it.

• 07-06 中

  训练策略的幻觉：为什么LLM强化学习的真正目标是推理策略单调改进

  MIPU揭示了LLM RL训练中被忽视的目标错位问题：在训练-推理不一致的情况下，训练侧的策略改进并不保证推理侧的策略改进，并提出了两步框架来解决这一问题。

• 07-05 EN

  Lynx: Progressive Speculative KV Cache Transfer for Disaggregated LLM Inference

  Lynx challenges the assumption that KV caches must be fully received before decoding begins — splitting them into MSB Anchor and LSB Residual streams to overlap transfer with speculative generation, achieving INT4-level TTFT while matching BF16 accuracy in disaggregated LLM serving.

• 07-05 中

  Lynx 阅读笔记：渐进式推测量化加速解聚合 LLM 推理的 KV 传输

  Lynx 打破「KV 缓存必须完整接收才能开始解码」的假设，将 KV 缓存分成高优先级 MSB Anchor 流和低优先级 LSB Residual 流，Anchor 到达后立即推测解码，Residual 到达后无损验证——在解聚合 LLM 服务中同时实现 INT4 的首 token 延迟和 BF16 的推理精度。

• 07-04 EN

  MosaicKV: Dynamic Two-Dimensional KV Cache Compression for Long-Context LLM Serving — Technical Review

  MosaicKV solves the long-context KV cache bottleneck by applying dynamic per-vector element selection and segment-adaptive strategies across both sequence and channel dimensions, achieving 16x attention speedup and 7.3x throughput gain at only 1.76pct average accuracy loss.