Compilation for Dataflow

Memory Systems in Dataflow Architectures is an IBM-offered Coursera course about memory systems for dataflow architectures used in AI computing. It examines how memory capacity, bandwidth, and data movement affect LLM inference, including how GPUs, TPUs, and other AI accelerators store and move KV caches.

torch.compile is PyTorch's model compilation tool. It inspects a model before execution and converts as much of its computation as possible into an optimized region, allowing compiler transformations such as kernel fusion, operation reordering, and memory planning. Python scalar reads such as `.item()` can cause graph breaks that fragment these regions and force synchronization between Python and the accelerator, reducing available optimization and potentially slowing execution.

Dynamo is the first compilation stage used by PyTorch's torch.compile system. It traces Python execution, captures tensor operations in an FX graph, and records guards—assumptions about the inputs—that determine when the captured graph can be reused.

深入浅出PyTorch is a Chinese-language educational course and repository by Datawhale China covering PyTorch and its use in deep-learning workflows. It progresses from prerequisites, installation, tensors, automatic differentiation, CUDA and cuDNN to data loading, model construction, loss functions, optimization, training, evaluation, visualization, deployment with ONNX, and source-code analysis of common computer-vision, natural-language-processing, audio, and video models. The course is organized into stages for deep-learning foundations, advanced PyTorch operations, and case analysis. Its materials are stored as Markdown files and Jupyter notebooks and include practical exercises such as Fashion-MNIST and fruit-and-vegetable classification, model fine-tuning, mixed-precision training, data augmentation, TensorBoard and Weights & Biases visualization, and ecosystem overviews including torchvision, PyTorchVideo, torchtext, and torchaudio. The repository includes supplementary video lessons, a planned content roadmap, and contribution instructions, and is licensed under CC BY-NC-SA 4.0.

HFlow is an open-source SDK from Hebbian Robotics for multimodal data pipelines in robotics and physical AI. It lets teams implement transformations, quality checks, labels, and enrichments as plain Python functions while HFlow handles orchestration, provenance, storage, versioning, and dataset curation. The pipeline accepts supported MCAP episodes and can import LeRobot Dataset v3 repositories. It runs code in-process for development or generates Airflow 3 DAGs for scheduled execution, producing canonical MCAP episodes, artifacts, provenance records, and a Parquet catalog. DuckDB SQL creates version-pinned manifests for curation, while a queryable catalog records metadata and quality evidence without requiring the underlying recordings to be loaded.