Tag
KVCache Meetup Korea 2026, co-hosted by PyTorchKR and XCENA, featured talks from XCENA, SK hynix, FuriosaAI, and LMCache on CXL memory pooling, GPU/HBM roofline analysis, batch scheduling on RNGD, and process-isolated KV Cache serving.
PyTorch announced a live Q&A webinar for the PyTorch 2.15 release, where Meta engineers will walk through new capabilities including the rewritten NCCL backend, NIXL one-sided transports, finer FSDP2/pipelining memory control, and faster NVFP4/FP8 decoding on NVIDIA Blackwell and Apple Silicon.
The PyTorch Foundation project Ray announced a Special Interest Group (SIG) for Ray on Accelerators, uniting hardware vendors, cloud providers, and the Ray community around a vendor-neutral roadmap covering hardware-aware testing, topology-aware scheduling, and Ray Direct Transport.
River API now supports training LLMs with arbitrary custom loss functions written in PyTorch, which are traced locally, sent as a graph, and executed at scale on River's remote GPU cluster with parallelization and memory-efficient streaming.
PyTorch highlights its ambassador meetup at PyTorchCon North America, celebrating community leaders who grow PyTorch Foundation projects like PyTorch, vLLM, DeepSpeed, Ray, Helion, and Safetensors through events, content, and mentoring.
PyTorch Day Japan 2026 on December 10 in Tokyo will feature Masahiro Tanaka of DeepSpeed presenting 'DeepSpeed at the Frontier: Community-Driven Optimizations for State-of-the-Art Training Performance,' covering innovations in distributed AI training performance and scalability.
PyTorch 2.15 发布，包含 3,117 次提交，重写的 NCCL 后端成为默认、新增 NIXL 单向传输、FSDP2 与流水线并行的更细粒度控制，并在 NVIDIA Blackwell（NVFP4）和 Apple Silicon（FP8）上带来更快的 LLM 解码，同时扩展了 AMD、Intel、ROCm 等平台支持。
A minimal, self-contained manual autograd implementation for a transformer-based classifier in PyTorch, designed to build intuition for how autograd works under the hood and how gradients flow through attention blocks.
Attendees at PyTorch Conference North America 2026 in San Jose completed a full day of instructor-led training preparing for the new PyTorch Certified Associate (PTCA) exam, offered via Linux Foundation Education. Positive attendee feedback highlighted hands-on GPU labs and coverage of benchmarking and profiling beyond model training.
A developer shares ApexW, an experimental PyTorch optimizer using compact optimizer states and a Triton-based execution path, claiming ~1.5x faster optimizer steps and halved optimizer-state memory on large-tensor workloads — while openly noting regressions on many-small-tensor models and minimal end-to-end gains.
Practical notes from Session 1 of a GPU kernel engineering series, covering CUDA 13's GPU compute-capability cutoff, the driver/toolkit/PyTorch-runtime split, async GPU timing pitfalls, and fixing Nsight Compute's ERR_NVGPUCTRPERM profiler counter permissions error.
Sebastian Raschka released an 11-hour, eight-video course on building reasoning models from scratch in PyTorch, covering evaluation, inference scaling, reinforcement learning with verifiable rewards, and distillation, serving as a sequel to his earlier Coding LLMs from the Ground Up course.
Hugging Face 发布了一个 torch-neuronx 模型支持查询 Space，可查看哪些 Hub 上的模型能在 AWS Trainium 上直接运行，并提供可复现的运行脚本和 NEFF 缓存以跳过编译。目前仅覆盖 text-generation 与 image-text-to-text 类 Transformers 模型，未来将扩展到 diffusers 和 sentence transformers。
PyTorch Foundation 欢迎其 2026 届 24 名大使，他们来自尼日利亚、布基纳法索、新加坡、印度、罗马尼亚、西班牙、英国和美国等全球社区，旨在推动开源 AI 的教育、指导与贡献。
PyTorch Foundation 于 PyTorch Conference North America 2026 宣布新增九家成员，包括 Platinum 级的 Nebius、Nscale 和 Shopify，以及 Silver 级的 LightSeek Foundation 和 Lablup，进一步巩固其作为开源 AI 中心枢纽的地位。
PyTorch 公布了 2026 年 PyTorch 贡献者奖获奖名单，在 #PyTorchCon 2026 上表彰了在 PyTorch、vLLM、DeepSpeed、Ray、Helion 和 Safetensors 等项目中做出贡献的社区成员，并由 Red Hat 赞助颁奖。
A blog post reveals that a loss function can evaluate to exactly 0.0 in PyTorch while its backward pass executes arbitrary Turing-complete computation, enabling supply-chain attacks where a malicious regularization package silently modifies model weights during training.
TorchTitan 发布重大性能与架构更新，新增对 DeepSeek-V4、Qwen3.5、Kimi K3、Muse Spark Open 等 SOTA 模型的完整 5D 并行支持，并推出统一后训练与智能体工作负载的 TitanRL 框架，FlexShard 与 torch_checkpointing 大幅提升大规模训练效率。
At PyTorchCon 2026, CrusoeAI's Yiwei Song presented Crusoe's LLM post-training platform, including Crusoe Managed RL — a Tinker API-compatible service that lets customers customize open-weights models while Crusoe manages the GPU infrastructure.
Rachad Alao of Cohere speaks at PyTorchCon 2026 about how open science can enable transparent and secure AI foundation model development, advocating for open architectures, independent auditing, and tools like PyTorch, TorchTitan, and vLLM.