PyTorch
Named by 50 kept items in this workspace.
30 items name it under classification.technologies and 20 under frameworks and tools.
The items that name it 50
Used for quantisation-aware training of the models; df.MLP.from_torch and df.Transformer.from_torch convert them.
Framework named among fast-growing community and independent projects.
Used with Python to implement the reasoning methods from scratch: generation, evaluation, reinforcement learning and distillation.
Machine learning framework that Weco is said to work with.
Framework in which most AI apps are written; it grew up on Nvidia GPUs.
Machine learning framework. Version 2.15 improves the MPS backend for Apple silicon with FP8, faster LLM decoding, more linear algebra and MPS streams. It requires Python 3.11 or later and macOS 15 or later for MPS.
Framework whose code Maxwell replaces with optimised kernels for the particular GPU present.
Baseline: compiled, CUDA-graphed BF16 blocks, FP8 blocks, and an INT4 projection kernel that targets a single query row
Deep learning framework installed (version 2.8.0+cu124) for training and running the models.
Required for GPU inference in reproduction.
Underlying framework in which transformers runs models; the article's local inference work targets keeping the model and generation loop in Python/PyTorch without requiring torch.compile.
Named as an ML community project supporting the benchmark.
Named as an ML community project supporting the AgentX benchmark.
Listed as a prerequisite for the course (deep learning basics)
Provides DTensor, FSDP2, and torch.distributed primitives underlying Halo's distributed stack.
Mentioned as the kind of tool that implements a numerically stable softmax function, used in place of the simplified NumPy example.
Underlying deep learning framework used by Soup's training stack (torch>=2.6.0 required to fix TRL preference trainer import issue)
Used to implement and train the LSTM neural network (TradingLSTM class), including layers, dropout, optimizer, and gradient clipping
Production library used in the Use It step and in the Phase 3 introduction lesson.
Underlying tensor/ML framework required by pocket-tts (2.5+, CPU build recommended on Linux)
In every workspace you can read 59
AI
Security
All technologies in this workspace.