PyTorch
PyTorch is an open-source deep learning framework created by Meta AI that supports dynamic computational graphs and GPU-accelerated computation.
Core Facts
vLLM is an open-source large language model inference engine project built on top of PyTorch.
Triton is a GPU programming language developed by OpenAI.
Hugging Face's Transformers library supports three major frameworks: PyTorch, TensorFlow, and JAX.
Timeline (last 90 days)
pyg-lib is PyG's low-level C++/CUDA extension, providing efficient implementations of KNN, radius search, FPS, and more; its pre-compiled wheels are tightly bound to PyTorch and CUDA versions, and version matching must be ensured during installation.
The tutorial warns: not installing PyTorch and not writing code is equivalent to not having learned AI; the most trap-filled way to learn is only watching videos, only bookmarking tutorials, only reading other people's summaries, and never running through it yourself.
TensorFlow, Scikit-Learn, and PyTorch are currently the three most mainstream frameworks in machine learning engineering practice.
On an Apple Silicon Mac, after moving a PyTorch tensor on the MPS device to the CPU via .to("cpu"), the process's phys_footprint immediately drops to near zero, but the tensor data remains fully in memory.
Compared with del module, the key advantage of migrating a module to the meta device and then combining gc.collect() with torch.mps.empty_cache() to free memory is that freeing can be done without breaking external references, making it suitable for safely unloading stages in a pipeline that are no longer needed without causing dangling reference errors; but a migrated module can no longer perform forward computation unless the parameters are reloaded with .to(device).
In the reproduction experiment (torch 2.14.0, macOS 26.5.2, M5 Pro 48 GB, 4 GiB tensor), right after moving to CPU both phys_footprint and the MPS driver count returned to near zero, and only after the CPU read through the copy page by page did the footprint return to 4.2 GiB, with the results of three runs completely identical.
PyTorch's meta device is a kind of bodiless device: tensors retain only metadata such as shape, dtype, and stride, and allocate no underlying storage; after .to('meta') on all module parameters and buffers, the objects still exist and references remain valid, but the memory usage drops to zero.
ONNX functions like a universal translator for AI models: no matter whether a model was originally trained with PyTorch, TensorFlow, or another framework, it can be exported to the ONNX format and then loaded and executed in a runtime environment that supports ONNX.
Many deep learning frameworks (such as PyTorch) borrowed from scikit-learn's style in their API design, so mastering scikit-learn first makes migration costs very low.
VIDRAFT performed LoRA training on standard Transformer blocks across all 32 chips of the Korean NPU server, achieving 0.9999 consistency with standard PyTorch training.
40 more timeline events
All Facts (20)
vLLM is an open-source large language model inference engine project built on top of PyTorch.
90%VerifiedTriton is a GPU programming language developed by OpenAI.
90%VerifiedHugging Face's Transformers library supports three major frameworks: PyTorch, TensorFlow, and JAX.
90%VerifiedSince its introduction in 2006, CUDA has become the de facto industry standard in the AI/deep learning field, with mainstream frameworks such as PyTorch and TensorFlow using CUDA as their default backend.
80%VerifiedPyTorch uses dynamic computation graphs (Define-by-Run), rebuilding the graph anew with each forward pass.
80%VerifiedPyTorch's Autograd engine builds a directed acyclic graph (DAG) on the fly during tensor operations, recording the order of computations to enable gradient calculation during backpropagation.
80%VerifiedPython deep learning frameworks include TensorFlow and PyTorch
80%VerifiedPyTorch is currently the most popular deep learning framework in both academia and industry.
80%VerifiedPyTorch was open-sourced in 2016 by Meta (formerly Facebook) AI Research.
80%VerifiedTensorFlow 1.x is a classic example of a static computation graph, requiring the entire computation process to be declared and then compiled before execution.
80%VerifiedPyTorch is widely popular in the research community, and its dynamic computation graph mechanism is one of the important reasons.
80%VerifiedTriton Inference Server is an open-source model serving framework from NVIDIA that supports dynamic batching and concurrent model execution, and is compatible with inference backends such as TensorRT, PyTorch, TensorFlow, and ONNX Runtime.
80%VerifiedPyTorch provides dynamic computation graphs and automatic differentiation mechanisms, allowing users to implement backpropagation without doing so manually.
75%VerifiedThe Apple Core ML framework allows developers to convert models trained in PyTorch or TensorFlow to the .mlmodel format for local execution, leveraging parallel computation across the GPU and Neural Engine.
75%VerifiedStable-Baselines3 is rebuilt on PyTorch and provides production-grade implementations of algorithms such as PPO, DQN, and SAC.
75%VerifiedThe machine learning path uses Scikit-learn as an introductory framework, extending further to deep learning frameworks such as TensorFlow and PyTorch.
75%VerifiedPyTorch was transferred to the independent PyTorch Foundation for governance in 2022, and the foundation is affiliated with the Linux Foundation.
75%VerifiedModel fine-tuning typically requires a large investment in annotated data and GPU computing resources.
75%VerifiedTensorFlow was released in 2015 and PyTorch was released in 2016, both using Python as their primary interface.
75%VerifiedPyTorch uses imperative programming (eager execution), where code executes sequentially line by line and tensor operations are evaluated immediately
70%