331 posts in total. Keep on posting.

2026

• 08-08 中

  QEvict 阅读笔记：把 KV Cache 淘汰从「一锤子买卖」变成可逆的三层调度

  QEvict 用全精度/INT2 可恢复量化/永久淘汰三层结构取代传统二元淘汰，让曾经「看起来不重要」的 KV 窗口在注意力回流时能被反量化并重新晋升，而不是永久丢弃。

• 08-07 EN

  SALT: Subspace-Aligned Centroid-Residual Training for Efficient Ultra-LoRA Serving

  SALT pins one shared, subspace-aligned domain centroid in GPU memory and swaps only a rank-1 residual per tenant, recovering high-rank LoRA accuracy while cutting per-adapter memory by up to 16x.

• 08-07 中

  SALT 阅读笔记：子空间对齐质心-残差训练如何让超低秩 LoRA 服务成为可能

  SALT 把一份子空间对齐的质心永久驻留在显存里，每个租户只需要热插拔一个 r≤2 的残差，就能追回接近满秩 LoRA 的精度，同时把单适配器显存开销压缩到原来的 1/16。

• 08-06 EN

  Fine-Grained Compute-Communication Overlap for MoE: Hiding All-to-All Behind GEMM with Tile-Level Signaling

  A producer-consumer GEMM/all-to-all co-design that hides the MoE return communication behind expert compute using per-tile device-resident signals, a remote-owner-aligned row layout, and disjoint SM partitions — up to 2.74x MoE-layer speedup on 4 A100 GPUs without touching NCCL or fusing kernels.

• 08-06 中

  MoE 计算-通信细粒度重叠：用 Tile 级信号把 All-to-All 藏在 GEMM 背后

  一个不做内核融合的 producer-consumer 设计：用 tile 级设备驻留信号、远端归属对齐的行布局、以及互不干扰的 SM 分区，把 MoE 第二次 all-to-all 的通信开销藏进专家计算里，在 4 张 A100 上拿到最高 2.74 倍 MoE 层加速。

• 08-05 EN

  AnchorKV: Anchor-Residual KV Cache Compression

  AnchorKV keeps every single token in the KV cache alive at 20x compression by projecting each token onto its nearest exactly-stored anchor and spending a small residual budget only where it moves the attention output the most.

• 08-05 中

  AnchorKV：锚点-残差表示下的 KV 缓存压缩阅读笔记

  AnchorKV 把每个 token 都投影到离它最近的一个精确保留的『锚点』上，再把剩余的字节预算精准花在最影响注意力输出的残差上，在 20 倍压缩比下一个 token 都不丢，精度还比驱逐类方法高一大截。

• 08-04 EN

  GRPO vs. Dr. GRPO Is Not a Bug-Fix Story: An Impossibility Theorem for Length Bias in Group RL

  A short, sharp impossibility theorem shows that no length-based weighting scheme for group-relative policy optimization can be simultaneously gradient-unbiased and length-invariant — GRPO and Dr. GRPO are two ends of one unavoidable Pareto tradeoff, not a broken algorithm and its fix.

• 08-04 中

  GRPO 和 Dr. GRPO 不是「修复」关系：一个关于长度偏差的不可能性定理

  一篇短小精悍的不可能性定理证明：任何只依赖长度的加权方案都无法同时做到梯度无偏和长度不变——GRPO 与 Dr. GRPO 只是同一条不可避免的 Pareto 权衡曲线上的两个端点，而不是一个算法和它的修复版。

• 08-03 EN

  BLADE: Boundary-Expanded and Layer-Adaptive Dynamic Exit for Efficient LLM Reasoning

  Self-doubt words like 'Wait' are not the only place a reasoning model already has the right answer — BLADE shows that ordinary sentence boundaries hide most of the early-exit opportunity, and picking the right hidden layers matters as much as picking the right checkpoint.

• 08-03 中

  BLADE 阅读笔记：把「早退」这件事从自我怀疑词扩展到普通句子边界

  推理模型早就写出了正确答案，却还要靠「Wait」这种自我怀疑词才敢喊停——BLADE 发现真正的早退机会大多藏在普通句子边界里，选对探针层和选对检查点一样重要。

• 08-02 EN

  FEPLB: Exploiting Copy Engines for Nearly Free MoE Load Balancing in Distributed Training

  MoE routing imbalance wastes 18.6% of GPU time even without an auxiliary loss — FEPLB shows the fix does not need a new communication path, it needs the NVLink Copy Engine hardware that current frameworks leave idle.