Skip to content
This repository has been archived by the owner on Apr 26, 2021. It is now read-only.

Latest commit

 

History

History
81 lines (61 loc) · 3.88 KB

README-old.md

File metadata and controls

81 lines (61 loc) · 3.88 KB
Raw

pytorch-world

Stable snapshots of the pytorch ecosystem can be found in releases.

The pytorch ecosystem, from a single nix-user's perspective, as I get around to building it. Any user-submitted packages are welcome : ).

Releases

This repository will tag releases according to the working version of pytorch. There will also a post-fixed snapshot number when new packages are working. All releases will only be tested for python-3.6 unless otherwise noted.

  • unstable:

    • darwin support for dev outputs
    • disabled test phase (for faster builds). Override doCheck to run the test suite.
    • adds a cudaArchList argument that allows users to test on the latest nvidia hardware without waiting for this package.
    • adds .dev output which gives top-level access to generated C/C++ code (libtorch).

  • v1.2.0-1:

    • runs all tests except for utils
    • adds mkl support to magma-2.5.0
    • adds cuda support to openMPI
    • adds openMPI support
    • allows users to build NamedTensors
    • allows users to get access to generated binaries from build process
    • adds darwin support if cudaSupport is disabled.
    • packages: (+probtorch)

  • v1.1.0

    • builds without FBGEMM
    • packages: None

Features

The current features are supported:

Python-3.6.9 support pytorch-v1.1.0 pytorch-v1.2.0 libtorch-v1.7.0
vanilla ✔️ ✔️ ✔️
mkl+mkldnn ✔️ ✔️ ✔️
openMPI ✔️ ✔️ ✔️
FBGEMM ❌ ✔️ ✔️
Full CPU (all above)* ✔️ ✔️ ✔️
Full CPU + NamedTensors + Binaries ❕ ✔️ ✔️
CUDA ✔️ ✔️ ✔️
CUDA+NCCL ✔️ ✔️ ✔️
CUDA+NCCL+openMPI ✔️ ✔️ ✔️
CUDA+NCCL+openMPI+mkl ✔️ ✔️ ✔️
Full CUDA (all above)* ✔️ ✔️ ✔️
Full CUDA + NamedTensors + Binaries ❕ ✔️ ✔️

The ":grey_exclamation:" means that namedtensors and binaries weren't attempted in pytorch-v1.1.0.

The "*" implies that FBGEMM is not included in any of the "Full" builds in pytorch-v1.1.0, but is included for all of pytorch-v1.2.0

Using this in your projects

Suggested use is to pull in pytorch-world as a git submodule under some nix/ folder and refer to probtorch/release.nix for an example of how to depend on this library. Do not reference release.nix files as they don't generate site-packages and only output final python binaries.

Cachix

Binaries can be found at pytorch-world.cachix.org

Currently, you will find binaries for the ./release.nix file which includes pytorch-1.2-related builds. This covers all CPU builds, which are darwin-accessible, as well as cuda-enabled pytorch built with cudatoolkit_10_0, and libtorch built with CUDA-9 and CPU. This was built on NixOS with python-3.6 (pin located at ./pin/), please file issues to help support! You can retrieve any of these with the following:

cachix use pytorch-world
nix-build ./release.nix -A pytorch -A pytorch-mkl -A pytorch-cu