Tag
Google 与 Meta 联合发布开源项目 TorchTPU，为 Google TPU 提供原生 PyTorch 后端，支持真正的 eager 执行与 torch.compile 集成，并在 Llama 3 405B 和 Qwen 3 MoE 上实现近线性扩展。
PyTorch's Accelerator Integration Working Group published a technical recap of H1 2026 progress, including the Cross-Repository CI Relay (CRCR), refactored device-agnostic test suites, maturation of the OpenReg reference backend, distributed and torch.compile support, and the new Additional Compute Platforms portal to standardize vendor-neutral hardware enablement.
vLLM is introducing hardware-agnostic layers to balance high performance on cutting-edge hardware with portability across different accelerators, addressing compatibility issues with torch.compile.
The PyTorch Conference North America features engineering teams presenting major breakthroughs in torch.compile, custom kernel authoring, and scaling efficiency, with sessions led by experts from various organizations.
PyTorch 2.14 introduces significant updates including NVGEMM for CUTLASS kernels in Inductor, a new nccl2 backend for distributed computing, first-class fault tolerance, and native linear algebra on Apple Silicon.
A blog post explaining PyTorch FX graphs, which are an intermediate representation used throughout the PyTorch 2.0 compile ecosystem. It covers the core objects Graph, Node, and GraphModule, and how to understand and work with them.
A technique to accelerate matrix multiplication when M and N are small but K is large, as encountered in MoE routers and small-batch decoding, by decomposing K and running partial GEMMs in parallel. The approach beats PyTorch Inductor on most shapes using a custom Triton kernel.
This post explains the Decompose-K technique for accelerating skinny large-K matrix multiplications by splitting the K dimension into chunks, running batched matmuls, and summing partials. It provides a PyTorch implementation and benchmarks showing significant speedups over standard torch.compile for bad-shaped matmuls.
The author explains operator fusion as a key mechanism behind torch.compile's speedups, and provides a minimal 500-line Python implementation and notebook as an educational tool.
The article explains the torch.compile stack in PyTorch, detailing steps from API to Dynamo, FX graph, ATen ops, and Torch Inductor for JIT compilation.
Published a custom kernel to further optimize LTX-2.3 from Lightricks, achieving 1.52x speedup on GB10, building upon previous torch.compile and cuDNN attention optimizations.
Announces Part 2 of a profiling tutorial covering linear layer tracing, gemm epilogues, MLP tracing, and comparisons of torch compile vs Liger kernels, with a link to the full content.
Sayak Paul describes a project to profile and optimize Diffusers pipelines using torch.compile, and announces a tutorial series by Ari G. on the topic.