PyTorch

Evidence (20)

• 4Hacker News·10/10/2026research
  Build Your Own Decision Model

  The post explains how to turn a language model into a decision model by constraining outputs to a fixed set of options and selecting the highest-probability choice in a single pass. It also shows a simple implementation with Qwen/Qwen3-1.7B and reports accuracy on a CommonsenseQA sample, including a small gain after fine-tuning.

• 7NVIDIA GenAI·10/2/2026product_launch
  NVIDIA adds 64GB DGX Spark for local AI development

  NVIDIA is introducing a 64GB DGX Spark configuration through select partners, aimed at developers, researchers, and enthusiasts building local AI workloads. The system ships with DGX OS and NVIDIA’s AI software stack, supports up to 100-billion-parameter models on device, and can be clustered in pairs to expand memory and performance without additional setup.

• 5Hacker News·9/24/2026research
  Dynamic Abliteration suppresses refusals in Qwen at runtime

  The post describes a proof of concept for dynamically suppressing refusal behavior in the open-weight Qwen3-4B model without changing its weights. It uses multi-layer steering via PyTorch forward hooks to inject signals into residual streams during inference, aiming to preserve the base model while altering responses to sensitive prompts.

• 4Hacker News·9/14/2026open_source
  OpenArch: PyTorch implementations of modern LLM architectures

  OpenArch is an open-source repository that provides PyTorch implementations of modern large language model architectures. It appears aimed at helping developers study, reproduce, and experiment with current LLM design patterns in one place.

• 7The New Stack·9/3/2026research
  Cut GPU inference cold start from 8 minutes to under a minute

  The article breaks down GPU inference startup into six phases and shows that cold-start time can be reduced substantially through configuration and platform changes. It reports that warm-node restarts can drop from 1.5-8 minutes to under 30 seconds, while fresh-node startups can fall from 8-15 minutes to about 5 minutes when using Amazon EKS Auto Mode and related optimizations.

• 5Hacker News·8/15/2026research
  Auto-research with Codex delivers a 232x faster QR kernel

  A Hacker News post describes how the author used Codex in GPU Mode to iteratively optimize a batched square QR decomposition kernel for a contest. The final solution achieved a 232x speedup over the baseline and ranked 12th out of 183 participants, highlighting agent-assisted kernel tuning and auto-research workflows.

• 7The New Stack·8/14/2026security
  Unsigned container images are a growing security risk in AI pipelines

  The article argues that many organizations still do not sign container images, leaving delivery pipelines exposed to tampering, impersonation, and compromised base images. It says the risk is more urgent in AI workflows because model artifacts, datasets, and runtimes are increasingly shipped as OCI artifacts that scanning alone cannot verify for provenance.

• 8The New Stack·8/10/2026breakthrough
  EKS speeds up multi-gigabyte container image pulls

  The article describes how large ML inference container images on Amazon EKS were taking several minutes to pull, delaying pod readiness and leaving GPUs idle. By reworking the image pull pipeline to better use available network, storage, and compute resources, the team reduced multi-gigabyte pulls to seconds; the improvements are now available by default in EKS Auto Mode and were contributed upstream to containerd and the SOCI snapshotter.

• 7Hugging Face Blog·8/10/2026research
  Hugging Face shows cheaper large-scale knowledge distillation for LLMs

  The article presents a method for making knowledge distillation for large language models much cheaper by caching the teacher model’s top-K logits offline and using a fused chunked KL loss. The approach reduces VRAM usage enough to make distillation experiments feasible on a single GPU in some cases, instead of requiring large multi-GPU setups.

• 7Hacker News·8/2/2026research
  Running Kimi K3 on AMD MI355X with Better Performance per Dollar Than B300

  The post reports benchmark results for serving the Kimi K3 model on AMD MI355X GPUs, claiming better performance per dollar than Nvidia B300 in this setup. It also describes a ROCm-side bug in speculative decoding infrastructure and notes that a small PyTorch fix was needed to stabilize the scheduler.

• 5Hacker News·7/26/2026model_release
  Inflect-Micro-v2 Brings Complete Local Voice Models

  Inflect-Micro-v2 introduces complete local text-to-waveform speech models in 3.96M and 9.36M parameter variants. Official releases are available for PyTorch and ONNX Runtime, making the models easier to test and deploy locally.

• 7Hugging Face Blog·7/17/2026framework_update
  NVIDIA and Hugging Face add large-scale fine-tuning for Diffusers models

  NVIDIA NeMo Automodel now integrates with 🤗 Diffusers to support distributed fine-tuning of video and image diffusion models directly from Hugging Face Hub checkpoints, without checkpoint conversion or model rewrites. The collaboration adds production-oriented training features such as memory-efficient sharding, latent caching, multiresolution bucketing, and configurable parallelism across multiple scales.

• 3Hugging Face Blog·7/10/2026research
  Profiling Attention in PyTorch

  This Hugging Face blog post continues a profiling series by examining how attention appears in PyTorch profiler traces. It walks through naive causal attention and shows how discrete operations such as matmul, masking, softmax, and scaling are represented during execution.

• 7The New Stack·7/1/2026product_launch
  Clockwork launches a guarantee to avoid AI training restarts

  Clockwork introduced its YOCO Guarantee, promising that 90% of failures on supported training runs will be recovered without losing progress, rolling back checkpoints, or recomputing work. The offer is built on TorchPass, a fault-tolerance product that can live-migrate GPU training state to a healthy spare or lower-priority node, with recovery ranging from seconds to a few minutes depending on the mode used.

• 6Simon Willison·6/22/2026research
  Porting Moebius image inpainting to the browser with WebGPU

  Simon Willison describes porting the Moebius 0.2B image inpainting model from a PyTorch/NVIDIA CUDA setup to run in the browser. He used Claude Code, ONNX Runtime Web, and WebGPU to build a working demo that lets users upload an image, mask regions, and have the model fill them in.

• 3Hugging Face Blog·6/11/2026research
  Profiling PyTorch Linear Layers and Fused MLPs

  This Hugging Face blog post continues a series on reading PyTorch profiler traces, focusing on replacing a manual matmul-plus-bias pattern with nn.Linear and then stacking layers into an MLP. It explains how PyTorch handles transposition, addmm, and kernel launches, and uses profiling to motivate performance optimizations such as fused MLPs.

• 7Hugging Face Blog·6/8/2026open_source
  OpenEnv expands open governance for agentic RL

  Hugging Face announced that OpenEnv, a library for agentic execution environments, will be coordinated by a broader committee that includes organizations such as Meta-PyTorch, Unsloth, Nvidia, Microsoft, and others. The project is being positioned as a protocol layer for interoperable reinforcement learning environments, with standardized APIs, packaging, and support for HTTP, WebSocket, Docker, and MCP.

• 4Hugging Face Blog·5/29/2026framework_update
  A Beginner’s Guide to torch.profiler in PyTorch

  Hugging Face Blog introduces the first part of a series on profiling in PyTorch, focused on helping beginners read profiler traces and understand how Python calls map to CPU and CUDA activity. The article explains how to use torch.profiler, interpret its tables and traces, and observe what changes when torch.compile is applied.

• 5Hugging Face Blog·5/18/2026framework_update
  PaddleOCR 3.5 adds a Transformers inference backend

  PaddleOCR 3.5 lets supported OCR and document parsing models run with Hugging Face Transformers as an inference backend by setting engine="transformers". The update keeps PaddleOCR managing the OCR/document parsing pipeline while giving developers a more flexible way to integrate models into Transformers-centered workflows.

• 5Hugging Face Blog·5/11/2026research
  AWS building blocks for foundation model training and inference

  This Hugging Face Blog article introduces a layered architecture for foundation model training and inference on AWS, spanning compute, networking, storage, orchestration, and observability. It frames the modern model lifecycle across pre-training, post-training, and inference, and maps common open-source tools to the AWS infrastructure needed to run them at scale.