Chinese contentLLVMhammer：通过 Rowhammer 位翻转让 LLM 推理变慢 4-16%
有人用 Rowhammer 硬件攻击搞坏了 Triton 编译，四个主流模型实测慢 4-16%，搞 GPU 推理的都该看看这篇。
#Triton
13 stories · 1 in 7 days · 7 in 30 days
Stories, products and papers tagged "Triton", newest first.
Oct 23
Oct 16
Chinese contentBitSparse：面向 ReLU 稀疏激活的无损压缩格式
训练 LLM 显存吃紧的可以看看：BitSparse 靠 ReLU 稀疏性无损压缩激活缓存，NanoGPT 和 Nemotron-H 上实测省三成显存，几乎不拖慢速度。
Oct 13
AWS boosts MoE throughput 10% with SageMaker kernel optimizations
The techniques target the exact bottlenecks that slow self-hosted MoE inference, and they are documented well enough to reproduce in your own SageMaker deployments. Teams serving large sparse models on NVIDIA H200 instances can apply them directly to cut latency and cost per token. Compared with relying on a hosted model API, this matters most for builders who need control over quantization and serving internals.
Oct 10
Chinese contentCrusoeAI 将赞助 PyTorchCon 北美站并举办 AI Infra Night
PyTorch 官方发的赞助消息，CrusoeAI 会讲 Triton 和 MoE 微调，搞基础设施的可以去现场或报名晚场活动。
Oct 9
KernelAgent Multi-Agent Harness Speeds Up Triton Kernel Optimization
Hand-tuning GPU kernels has been a specialist bottleneck, and embedding hardware-performance signals into a multi-agent loop points toward automated performance work inside normal PyTorch pipelines. Practitioners writing Triton kernels or relying on torch.compile can benchmark the harness against their own workloads. The talk is scheduled for PyTorch Conference North America 2026, with registration open now.
Oct 4
AMD FlyDSL Brings MLIR-Native GEMM Backend to TorchInductor
The comparison of FlyDSL against Triton for GEMM kernels signals that AMD is building its own kernel-generation stack inside PyTorch rather than relying on Triton alone. Developers running torch.compile on AMD Instinct hardware should watch for the benchmark numbers to judge whether FlyDSL backends are ready for production training and inference workloads.
Sep 29
KernelZero-7B beats Claude and DeepSeek on GPU kernel generation
A 7B open model trained with an automatic curriculum now beats frontier LLMs on kernel generation, suggesting specialised small models can outperform general giants on narrow engineering tasks. Builders writing CUDA or Triton kernels can watch for KernelZero's release to automate optimisation of Torch modules. The comparison against DeepSeek-V4-Pro on Triton gives a concrete benchmark for teams weighing frontier APIs against compact task-specific models.
Sep 14
Attention Quantization for Tabular Foundation Models
This paper details a method to make tabular foundation models run faster by optimizing the attention calculation process, which is a key part of these models' performance.
Sep 10
Stencil Computation for AI and HPC
TinyTC outperforms PyTorch/Triton on Intel GPUs for stencil computations, offering better performance for scientific computing applications.
Aug 17
Chinese content华为开源AscendNPU IR：毕昇编译器核心支持Triton与Ascend 950
华为把毕昇编译器的核心AscendNPU IR开源了，支持Triton，能直接在Ascend 950上写算子，比之前好用多了。
Jul 12
Chinese contentNVIDIA 基于块的 GPU 编程教程：从 cuTile 和 Triton 内核到 Flash Attention
手把手教你用 TileGym 在 Colab 上跑 cuTile 和 Triton，从向量加法到 Flash Attention，每个步骤都和 PyTorch 对照，适合想深入 GPU 编程的人。
Jun 15
Chinese contentFlash-KMeans：IO感知型精确K-Means，在GPU上比FAISS快200倍以上
开源Flash-KMeans在H200上比FAISS快200多倍，做精确k-means不近似，适合大规模数据聚类。