`spmd_types`

A type system for distributed (SPMD) tensor computations in PyTorch. This package provides two type systems:

Local SPMD types, which allow you to use Megatron-style differentiable collectives in a safe way by tracking whether or not your backward gradients are pending reduction or not.
Global SPMD types, a DTensor-like abstraction for writing code that has the same semantics whether run on a single device or in a distributed fashion, but with explicit communication operations so you are never guessing when a redistribute occurs.

In both cases, the SPMD types makes it possible for you to check that your code computes correct gradients (local SPMD) or gives equivalent results across different parallelizations (global SPMD), without having to actually run a full E2E distributed training run to check for loss matching.

The goal of this package is to provide a flexible type system that can typecheck realistic training code. We have used local SPMD types to typecheck a realistic pretraining codebase, and global SPMD types is actively under construction!

Installation

pip install spmd_types

Quick start

import torch
import torch.distributed as dist
import spmd_types as spmd
import spmd_types.checker
from torch.distributed.device_mesh import init_device_mesh

# Set up a fake process group (no GPUs needed)
dist.init_process_group(backend="fake", rank=0, world_size=8)
mesh = init_device_mesh("cpu", (2, 4), mesh_dim_names=("dp", "tp"))
dp = mesh.get_group("dp")
tp = mesh.get_group("tp")

with spmd.set_current_mesh(mesh), spmd.checker.typecheck():
    x = torch.randn(4)
    spmd.assert_type(x, {dp: spmd.R, tp: spmd.P})       # R on dp, partial on tp
    y = spmd.all_reduce(x, tp, src=spmd.P, dst=spmd.R)  # sum across tp ranks
    spmd.assert_type(y, {dp: spmd.R, tp: spmd.R})       # now replicated everywhere
    z = torch.mul(y, y)                                 # type inference: R * R -> R
    spmd.assert_type(z, {dp: spmd.R, tp: spmd.R})

dist.destroy_process_group()

Documentation

See Local SPMD types for a hands-on guide on porting Megatron-derived training frameworks, including the Megatron-to-spmd_types function mapping table and advice on Invariant vs Replicate.

See Design for the full type system specification, including local vs global SPMD modes, collective signatures with diagrams, forward-backward pairs, expert mode, cross-mesh compatibility, and partition spec redistribute.

License

BSD 3-Clause License. See CONTRIBUTING.md for how to contribute.

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

spmd_types-0.2.5.tar.gz (110.3 kB view details)

Uploaded Aug 24, 2026 Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

The dropdown lists show the available interpreters, ABIs, and platforms. Enable javascript to be able to filter the list of wheel files.

spmd_types-0.2.5-py3-none-any.whl (123.7 kB view details)

Uploaded Aug 24, 2026 Python 3

File details

Details for the file spmd_types-0.2.5.tar.gz.

File metadata

Download URL: spmd_types-0.2.5.tar.gz
Upload date: Aug 24, 2026
Size: 110.3 kB
Tags: Source
Uploaded using Trusted Publishing? Yes
Uploaded via: twine/7.0.0 CPython/3.13.14

File hashes

Hashes for spmd_types-0.2.5.tar.gz
Algorithm	Hash digest
SHA256	`2f104f6384d1b33facd8fa7a631e871c5f05360543508145a8d8435e0e0b03a8`
MD5	`28871356b9ee5f453a94c254ffff5190`
BLAKE2b-256	`f456c39c2e8523c1133fae620465bbdd87289253c23a5b12bd425ffd5a37a3c4`

See more details on using hashes here.

Provenance

The following attestation bundles were made for spmd_types-0.2.5.tar.gz:

Publisher: publish_release.yml on meta-pytorch/spmd_types

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

Statement:
- Statement type: https://in-toto.io/Statement/v1
- Predicate type: https://docs.pypi.org/attestations/publish/v1
- Subject name: spmd_types-0.2.5.tar.gz
- Subject digest: 2f104f6384d1b33facd8fa7a631e871c5f05360543508145a8d8435e0e0b03a8
- Sigstore transparency entry: 2581678100
- Sigstore integration time: Aug 24, 2026
Source repository:
- Permalink: meta-pytorch/spmd_types@6dbe9e23406db055a4bc174ce721897983b49115
- Branch / Tag: refs/heads/main
- Owner: https://github.com/meta-pytorch
- Access: public
Publication detail:
- Token Issuer: https://token.actions.githubusercontent.com
- Runner Environment: github-hosted
- Publication workflow: publish_release.yml@6dbe9e23406db055a4bc174ce721897983b49115
- Trigger Event: workflow_dispatch