Skip to content

Explanations

This section contains discussion about the overall design and implementation of this package.

  • Named Abstractions -- The core abstraction stack: NamedDimension, NamedDimCollection, NamedShape, and NamedLinop.
  • How Named Dimensions Propagate -- How named dimensions are inferred, renamed, resolved, and incremented as operators are constructed, composed, adjointed, and normalized.
  • Configuration Options -- The torchlinops.config flags: reduce_identity_in_normal, cache_adjoint_normal, log_device_transfers, shape_inference, and the config.using() context manager.
  • Why This Package? -- Motivation: the problems with traditional linops and how named dimensions solve them.
  • Design Notes -- Key design decisions and tradeoffs (staticmethods, caching in lists, pickle-ability, shape matching).
  • Copying Linops -- How shallow copy and memory-aware deep copy work, and why they matter.
  • Multi-GPU Execution -- Splitting operators across devices and synchronizing parallel execution: BatchSpec, ToDevice, SyncContext, streams, and events.
  • FAQ -- Frequently asked questions.