PyTorch's CUDA Tile: a Python DSL for tiling kernels onto Tensor Cores
PyTorch · x · 2026-11-10
PyTorch introduces CUDA Tile, a Python-based DSL with tile-centric primitives that automatically targets Tensor Cores and other advanced GPU features. It can take straightforward PyTorch code computing multiple terms separately and fuse it into one tiled kernel keeping matrices in registers. The post includes examples, docs and sample code.
More from Infra
- vLLM-Omni benchmarks: 2048px image generation in 0.63s on 4x B300 — vllm_project · 2026-11-10
- ReCaVSR one-step streaming 4x video super-resolution ported to ComfyUI — LatentSpacer · 2026-11-10
- Hugging Face livestream today: model quantization with llama.cpp and vLLM — ben_burtenshaw · 2026-11-10
- Running Startlux 4B 9B Instant Q40 on a cluster of 4 Raspberry Pi 5 boards — bartlomiej__ · 2026-11-10
- How banned Nvidia chips actually reach China: stolen trucks, fake paperwork, hair-dryer sticker swaps — Caneki97 · 2026-11-10
- Tencent reportedly picks Bittensor subnet Chutes as launch partner for flagship model Hy4 — markjeffrey · 2026-11-10