Translated title. Original: Deep Learning • Was ist das und wie funktioniert es? · [mit Video]
Deep Learning is a part of Machine Learning that uses multi-layer neural networks to learn from large data, enabling autonomous task execution after training. Its prominence grows with data and computing power, powering many AI applications like speech recognition, image processing, and self-driving cars.
The document presents TensorFlow retracing warnings from repeated tf.function calls, explains possible causes (looped function creation, varying tensor shapes, Python objects), and offers remedies: move @tf.function outside loops and enable reduce_retracing; includes links to TensorFlow guidance.
Guide explaining llama.cpp as a fast, lightweight local engine with mmap and GGML, wrapped by GUIs; emphasizes debugging wrappers before judging weights, and points to a full tutorial with code and diagrams.
The document describes a PyTorch GitHub release note about refactoring test_torch.py and enabling XPU support, with emphasis on code security and build-related changes. It also references session UI actions and account switching, likely from a placeholder or template in the page.
The change fixes module attribution for triton_op by wrapping backend functions with functools.wraps so the user’s module is recorded, preventing import-time failures in precompile artifacts; behavior of calling the op remains the same.
Adds torch::stable::from_file to the stable ABI, returning a memory-mapped 1-D tensor. Differences from ATen include filename type. Includes tests for full-file mapping and data integrity; linked to the related torchvision helper and PR 198263.
PyTorch fixes Windows wheel startup by preloading cuDNN DLLs from torch/lib when cuDNN isn’t on PATH, emitting warnings on failure rather than crashing, and validating with manual tests on Windows on Arm.
A guideline document for out-of-tree accelerator backends explains how to reuse PyTorch's upstream tests, classify tests by hardware, adapt tests without modifying upstreams, configure for backend differences, and gradually expand coverage, with a bring-up workflow and a reference to PR 189376 and its approval.
wxx is a WebGPU-backed PyTorch-like tensor library with autograd, neural modules, and training utilities; it includes GPT2, ResNet18, and Hyena models, standard training components, and save/load capabilities, with tests available from GitHub.
Translated title. Original: Pytorch erklärt: Was das KI-Framework leistet
The article outlines PyTorch's role in AI workflows, contrasting it with TensorFlow and early CPU-based training. It notes PyTorch's popularity in research, TensorFlow's alignment with Google's infrastructure and TPU chips, and the historical shift toward hardware accelerators for efficient model training and inference.