292 posts in total. Keep on posting.

2026

• 06-15 EN

  Parallel-Synthesis: Direct KV-Cache Synthesis for Parallel Branches in LLM-Agent Workflows

  Parallel-Synthesis is a plug-and-play framework that lets a synthesizer LLM directly consume the KV caches produced by parallel worker agents, avoiding redundant prefill and reducing time-to-first-token by 2.5–11× while matching or beating text-concatenation-based synthesis on 7 of 9 benchmarks.

• 06-15 中

  Parallel-Synthesis：让 LLM 综合智能体直接消费并行分支的 KV 缓存

  Parallel-Synthesis 提出了一套即插即用框架，让综合智能体的 LLM 直接复用并行 Worker 在解码时产生的 KV 缓存，从根本上消除了多智能体协作中代价高昂的重复预填充过程，在九个基准上的准确率与文本拼接基准持平或更优，首 token 延迟降低 2.5–11 倍。

• 06-14 EN

  GF-DiT: Scheduling GPU Parallelism as a First-Class Resource for Diffusion Transformer Serving

  GF-DiT treats GPU parallelism as a schedulable resource for Diffusion Transformer serving, decomposing requests into reschedulable trajectory tasks and introducing group-free collectives that cut communication-group setup from 778 ms to 60 μs, achieving up to 6.01× throughput gains and 95% latency reduction over static-parallelism baselines.

• 06-14 中

  GF-DiT：把 GPU 并行度当作可调度资源的扩散 Transformer 推理系统

  GF-DiT 把 GPU 并行度从静态部署参数提升为运行时可调度的资源，通过轨迹任务图和无组通信原语（组建开销从 778 ms 降至 60 μs）实现弹性并行推理，吞吐量提升最高 6.01 倍，平均延迟降低最高 95%。

• 06-13 EN

  ForeMoE: Micro-step-level MoE Load Balancing for RL Post-training via Routing Foresight

  ForeMoE exploits the unique structure of RL post-training — where rollout routing decisions are replayed in later stages — to predict and proactively balance MoE expert loads at micro-step granularity, achieving up to 1.45x speedup over state-of-the-art RL training systems on 64 GPUs.

• 06-13 中

  ForeMoE：利用路由预见性实现 RL 后训练中 MoE 微步级负载均衡

  ForeMoE 利用 RL 后训练特有的路由回放结构——rollout 阶段的路由决策在后续阶段被重用——实现对每个梯度微步的 MoE 专家负载精确预测与主动均衡，在 64 张 GPU 上实现最高 1.45× 的端到端加速。

• 06-12 EN

  SliceGPT: Post-Training LLM Compression via Computational Invariance

  SliceGPT exploits an exact structural symmetry in transformers to physically delete rows and columns from every weight matrix, achieving 25% parameter reduction with 99% zero-shot performance on LLAMA2-70B and OPT-66B — no custom hardware kernels required.

• 06-12 中

  SliceGPT 阅读笔记：用计算不变性删除 Transformer 的行与列

  SliceGPT 证明了 Transformer 计算对正交基变换具有精确不变性，并以 PCA 为工具将权重矩阵旋转到方差最集中的方向后直接裁去低方差维度，在 LLAMA2-70B 上以 25% 参数缩减保住 99% 零样本性能，且无需任何自定义 CUDA 算子。

• 06-11 EN

  MegaScale: Engineering 55% MFU at 12,288 GPUs for LLM Training

  MegaScale is ByteDance's full-stack production system for training LLMs at more than 10,000 GPUs, achieving 55.2% Model FLOPs Utilization through co-designed algorithmic optimizations, communication overlapping, and deep observability for fault tolerance.

• 06-11 中

  MegaScale：ByteDance 如何在 12,288 块 GPU 上实现 55% MFU 的大规模 LLM 训练

  MegaScale 是 ByteDance 用于超大规模 LLM 训练的生产系统，通过算法-系统协同设计、通信计算重叠、算子优化和深度可观测性，在 12,288 块 GPU 上实现了 55.2% 的 Model FLOPs Utilization，比 Megatron-LM 提升 1.34 倍。

• 06-10 EN

  KeepKV: Lossless KV Cache Compression via Electoral Votes and ZIP-Merging

  KeepKV introduces Electoral Votes and Zero Inference-Perturbation Merging to achieve single-step lossless KV cache compression, provably fixing the Attention Sag problem that plagues all prior merging methods.

• 06-10 中

  KeepKV：用「选举票」机制和零扰动合并实现无损 KV 缓存压缩

  KeepKV 提出了「选举票」机制和零推理扰动合并（ZIP-Merging），在数学上证明了单步无损 KV 缓存压缩，从根本上解决了所有现有合并方法都存在的「注意力衰落」问题。