The non-obvious point is that most of the speedup came from configuration, not new hardware — teams may already have equivalent capacity locked behind untuned settings. Practitioners running RL fine-tuning pipelines, especially with DeepSeek models, should benchmark their own step times against these figures before provisioning more GPUs. The 40-GPU versus 56-GPU result is a direct cost argument for adding auto-tuning to training stacks.
论文讲了个省显存的训练技巧：专家参数反向传播时就地更新，微调 OLMoE 这类 MoE 模型显存省约两成，速度还更快。
训练模型调学习率衰减的可以看看，UTA 比固定衰减曲线更聪明，能跟着训练状态走，LLaMA 和 ResNet 上都验证了，直接即插即用。
The speed of an optimizer depends as much on kernel-level linear algebra as on step count, a factor most training guides ignore. Engineers tuning large-scale training can apply the QR, Cholesky and eigendecomposition implementation patterns to cut data movement on GPUs. The talk ties directly to GPU MODE's competition, giving concrete reference implementations to study.
用 DFA 训网络总卡在常量预测器？这篇论文把根因拆到共模误差和 tanh 饱和，还给了减批量均值这种直接可用的修复办法。
This research provides a more efficient way to train large language models, which could lead to faster development of new AI products.
AI Engineer 大会的后训练专场正在直播，讲训练怎么决定模型行为，TypeSafe、Hugging Face、MiniMax 都有人讲。
Cursor 说 MoK 已在数万 GPU 上跑生产，比旧 DeepEP 栈端到端快 1.41 倍，搞大规模训练的可以看。
写了个新优化器CMuon，训练扩散Transformer比AdamW快两倍多，675M模型200轮就跑到FID 1.18，厉害得很。
DeepSeek-V4 Flash RL 在 AMD 显卡上跑通了，训练结果还不错，AIME 分数涨了一截。想试试 AMD 平台搞 RL 训练可以看看这篇。
AWS发了篇实战教程，教你用NVIDIA Blackwell在SageMaker上调优训练，从选精度到调batch size都讲清了，搞大模型训练的人别错过。
杨植麟用460万美元和极致架构打脸烧钱竞赛，做AI代理或准备2026年入场的团队，看完会重新思考资源分配策略——建议存下来周末细看。
