{"categories":[{"label":"Artificial Intelligence","url":"https://skillfed.io/packages/category/scientific-engineering-artificial-intelligence/9"}],"enrichment":{"capability":"Provides type annotations and runtime checking for distributed tensor computations, tracking data layout and synchronization state across SPMD parallelism patterns.","skillfed_tags":["distributed-training","type-checking"],"use_cases":["Porting training frameworks to verify gradients are correctly reduced across ranks without E2E training runs.","Typechecking distributed code to catch mismatches between expected and actual tensor layouts before deployment.","Writing distributed logic that is semantically identical whether run on a single device or across multiple ranks.","Verifying that collective operations are applied to tensors in the correct synchronization state.","Debugging gradient flow by asserting tensor state at each computation step."],"what_it_does":"spmd_types is a type system for writing distributed tensor computations with explicit, verifiable semantics. It provides two complementary abstractions: local SPMD types track whether gradients are pending reduction or fully synchronized, enabling safe use of Megatron-style differentiable collectives; global SPMD types offer a DTensor-like interface where code has identical semantics whether run on a single device or distributed, but with explicit communication operations so redistributes are never implicit.\n\nThe package lets you verify correctness of distributed training logic\u2014checking gradient computation and parallelization equivalence\u2014without running full end-to-end distributed training. It includes runtime type checking via a checker module, allowing you to annotate tensors with their layout and synchronization state and assert those invariants hold at each step.","worth_installing":"Yes, if you are building or maintaining distributed training code. The package is actively maintained, has no external dependencies, and offers a concrete way to catch synchronization bugs before expensive distributed runs. It is in beta (global SPMD types are under construction), so expect the API to evolve; use it for new projects where you can adapt to changes."},"id":"spmd-types","links":{"html":"https://skillfed.io/packages/spmd-types","md":"https://skillfed.io/packages/spmd-types.md","pypi":"https://pypi.org/project/spmd-types/"},"maintenance":{"status":"active"},"meta":{"latest_release":"2026-08-06","license_spdx":null,"license_treatment":"permissive","name":"spmd-types","python_support":"supports_current","summary":"A type system for distributed (SPMD) tensor computations in PyTorch"},"popularity":{"monthly_downloads":111736,"position":12403,"tier":"top_15000"},"security":{"n_vulnerabilities":0},"version":"0.2.3"}