A Nim frontend for pytorch, aiming to be mostly auto-generated and internally using ATen.

NimTorch

Master Release
Build Status Build Status

Pytorch - Py + Nim

A Nim frontend for pytorch, aiming to be mostly auto-generated and internally using ATen.

Because Nim compiles to C++, this is not a wrapper or binding library. It generates 1-to-1 native ATen code.

The only requirement from pytorch is ATen's core tensor library. Because of this, nimtorch is extremely versatile and can compile on any kind of device.

Current status

Early stage

  • Automatically generated, from Declarations.yaml, the full ATen API
  • Cuda support ( add -d:cuda when compiling with nim )
  • WASM support ( add -d:wasm when compiling with nim )
  • Automatically generated, from derivatives.yaml, gradient procs
  • Autograd
  • Add missing derivatives
  • More high level pytorch API (Module, Models etc)
  • ...

The final aim is to be as compatible as possible with the pytorch API.

Why

Ease of use of the python language while keeping fully bare metal native C++ performance

Python code

# GRUCell
gi = x.matmul(w_input.t()) + b_input
gh = hidden.matmul(w_recur.t()) + b_recur
i_r, i_i, i_n = gi.chunk(3, 1)
h_r, h_i, h_n = gh.chunk(3, 1)
resetgate = (i_r + h_r).sigmoid()
inputgate = torch.sigmoid(i_i + h_i)
newgate = (i_n + resetgate * h_n).tanh()
hy = newgate + inputgate * (hidden - newgate)

Nim code

# GRUCell
let
  gi = x.matmul(w_input.t()) + b_input
  gh = hidden.matmul(w_recur.t()) + b_recur
  (i_r, i_i, i_nn) = gi.chunk(3, 1)
  (h_r, h_i, h_n)  = gh.chunk(3, 1)
  resetgate = (i_r + h_r).sigmoid()
  inputgate = torch.sigmoid(i_i + h_i)
  newgate = (i_nn + resetgate * h_n).tanh()
  hy = newgate + inputgate * (hidden - newgate)

Getting started

Requirements

Linux: A recent distribution on par with ubuntu 18.04 in terms of libc and basic libraries, gcc compiler

macOS: We compile with 10.13 min version flags but might work even on lower versions, XCode for the compilers

Windows: Windows 10, Visual Studio Runtime 2017 and Visual Studio 2017 (any edition)

WASM: Latest Emscripten compiler and tools

Super easy, using conda

Linux, macOS and Windows

conda create -n nimtorch -c fragcolor nimtorch (add cuda10.0 for cuda 10 linux only or add wasm for wasm version)

source activate nimtorch or on windows: conda activate nimtorch

This will install: nim and ATen binaries, fragments and nimtorch all in one command, nothing else needed.

Make sure you use a recent version of conda and have a compiler installed in your system, on windows you have to add --cc:vcc and be on a developer prompt.

Make sure your system is recent (ubuntu 18.04 reference / macOS High Sierra / Windows 10) and you have cuda 9.2 installed (if you need cuda, linux only, more cuda versions coming, please open a issue if you need a specific version).

Test with with something like:

nim cpp -o:test -r $ATEN/dist/pkgs/nimtorch-\#head/tests/test_xor.nim

or on windows... (because dlls need to be side by side)

nim cpp -o:%ATEN%/lib/test.exe -r %ATEN%/dist/pkgs/nimtorch-#head/tests/test_xor.nim

Semi manual way

Linux, macOS and Windows

Check what version of ATen/PyTorch we need in conda/nimtorch/meta.yaml - should be something like aten ==2018.10.10.1089

Note the version as you will need it in the next step

conda create -n aten -c fragcolor aten={version}

or

WASM

conda create -n aten -c fragcolor aten={version} wasm

or Cuda 10.0 (linux only)

conda create -n aten -c fragcolor aten={version} cuda10.0

activate aten environment

source activate aten or on windows: conda activate aten

  1. Make sure you have a recent Nim and Nimble version in your path
  1. clone the release branch git clone -b release https://github.com/fragcolor-xyz/nimtorch.git
  2. cd nimtorch
  3. nimble develop

finally

run self test nim cpp -o:test -r torch.nim (use -o:%ATEN%/lib/test.exe instead on windows because of dll location)

in the case of WASM:

run self test nim cpp -d:wasm -o:test.js torch.nim && node test.js (needs node.js)

Manual way without requiring conda

Build ATEN

pip2 install pyyaml typing
git clone -b fragcolor-devel https://github.com/fragcolor-xyz/pytorch.git
cd pytorch
git reset --hard <commit hash> # from torch/commit.txt
git submodule update --init
mkdir build && cd build
cmake -DCMAKE_BUILD_TYPE=Release -DUSE_CUDA=OFF -DBUILD_ATEN_ONLY=ON -DCMAKE_INSTALL_PREFIX=`pwd`/output ../
make -j4
make install

# also copy derivatives if we want to run generator.nim or nimble test
# notice generator.nim might need python3 and pyyaml
cp ../tools/autograd/derivatives.yaml `pwd`/output/share/

Test the build

cd 
   
    
ATEN=
    
      nim cpp -r -f -o:/tmp/z01 torch.nim # for eg: ATEN=pathto/pytorch/build/output/

    
   

Notes

  • We suggest setting OMP_WAIT_POLICY environment variable to PASSIVE when running on CPU.
Owner
Giovanni Petrantoni
Founder and CEO @ Fragcolor
Giovanni Petrantoni
Comments
  • OSX: `nim cpp -r torch.nim` fails

    OSX: `nim cpp -r torch.nim` fails

    after building ATEN via https://github.com/fragcolor-xyz/nimtorch/issues/5#issuecomment-427937945:

    ATEN=/tmp/d11/Users/timothee/git_clone/nim/pytorch/built/output/lib nim cpp -r torch.nim
    error: unknown type name 'constexpr'
    
    ATEN=/tmp/d11/Users/timothee/git_clone/nim/pytorch/built/output nim cpp -r --passC:-std=c++11 torch.nim
    /Users/timothee/.cache/nim/torch_d/torch_tensors.cpp:206:14: error: no matching constructor for initialization of 'at::IntList' (aka 'ArrayRef<long long>')
            at::IntList temp(((long*) ((&self[((NI) 0)]))), selfLen_0);
                        ^    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    /tmp/d11/Users/timothee/git_clone/nim/pytorch/built/output/include/ATen/core/ArrayRef.h:67:13: note: candidate constructor not viable: no known conversion from 'long *' to 'const long long *' for 1st argument
      constexpr ArrayRef(const T* data, size_t length)
    
    
  • OSX: building ATEN via `docker build -t docker_aten_native .` fails

    OSX: building ATEN via `docker build -t docker_aten_native .` fails

    Download ATen binaries or build it (instructions under)

    => no OSX option

    ATen build instructions

    cd docker && cd docker-aten-native

    is that a typo?

    is my best bet to try to follow instructions from https://github.com/pytorch/pytorch or https://github.com/pytorch/pytorch/tree/master/aten ?

  • SIGSEGV when running examples

    SIGSEGV when running examples

    Using the latest nimtorch, I'm getting a SIGSEGV when trying to compile one of the examples (test_xor).

    I'm not sure if this is a compiler bug or a problem with Aten?

    Dockerfile:

    FROM continuumio/miniconda
    
    RUN conda install -y -c fragcolor nimtorch
    
    ADD test.nim .
    
    docker build -t nimtorch_conda .
    docker run --rm nimtorch_conda nim c test.nim
    #Hint: used config file '/opt/conda/config/nim.cfg' [Conf]
    #Hint: used config file '/opt/conda/config/config.nims' [Conf]
    #Hint: system [Processing]
    #Hint: widestrs [Processing]
    #Hint: io [Processing]
    #Hint: test [Processing]
    #Hint: torch [Processing]
    #Hint: macros [Processing]
    #Hint: cpp [Processing]
    #Hint: nimline [Processing]
    #Hint: tables [Processing]
    #Hint: hashes [Processing]
    #Hint: strutils [Processing]
    #Hint: parseutils [Processing]
    #Hint: math [Processing]
    #Hint: bitops [Processing]
    #Hint: algorithm [Processing]
    #Hint: unicode [Processing]
    #Hint: os [Processing]
    #Hint: pathnorm [Processing]
    #Hint: osseps [Processing]
    #Hint: posix [Processing]
    #Hint: times [Processing]
    #Hint: options [Processing]
    #Hint: typetraits [Processing]
    #Hint: torch_cpp [Processing]
    #Hint: tensors [Processing]
    #Hint: sequtils [Processing]
    #Hint: sets [Processing]
    #Hint: strformat [Processing]
    #Hint: tensor_ops [Processing]
    #Hint: autograd_macro [Processing]
    #Hint: autograd_backward [Processing]
    #Hint: nn [Processing]
    #Hint: modules [Processing]
    #Hint: init [Processing]
    #Hint: python_helpers [Processing]
    #Hint: functional [Processing]
    #SIGSEGV: Illegal storage access. (Attempt to read from nil?)
    

    test.nim:

    import torch
    import torch/[nn, optim]
    
    let inputs = torch.tensor([
      [0.0, 0.0],
      [0.0, 1.0],
      [1.0, 0.0],
      [1.0, 1.0],
    ])
    
    let targets = torch.tensor([
      [0.0],
      [1.0],
      [1.0],
      [0.0],
    ])
    
    let
      fc1 = nn.Linear(2, 4)
      fc2 = nn.Linear(4, 1)
      loss_fn = nn.MSELoss()
      optimizer = optim.SGD(fc1.parameters & fc2.parameters , lr = 0.01, momentum = 0.1)
    
    set_num_threads(1)
    
    when defined gperftools:
      discard ProfilerStart("test_xor.log")
    
    for i in 0 ..< 50000:
      optimizer.zero_grad()
    
      let predictions = inputs.fc1.relu.fc2.sigmoid
    
      let loss = loss_fn(predictions, targets)
      loss.backward()
      optimizer.step()
    
      if i mod 5000 == 0:
        print(loss)
    
    when defined gperftools:
      ProfilerStop()
    
  • can't install with nimble

    can't install with nimble

    $ nim --version Nim Compiler Version 0.18.1 [Windows: amd64] Compiled at 2018-08-18 Copyright (c) 2006-2018 by Andreas Rumpf

    git hash: b5171f57ef00bffb12387d7daf3487c5e07645f9 active boot switches: -d:release

    $ nimble install nimtorch Prompt: nimtorch not found in any local packages.json, check internet for updated packages? [y/N] y Answer: Downloading Official package list Success Package list downloaded. Tip: 3 messages have been suppressed, use --verbose to show them. Error: Package not found.

  • Cannot compile nimtorch tests on Windows 10

    Cannot compile nimtorch tests on Windows 10

    C:/Program Files/mingw-w64/x86_64-8.1.0-posix-seh-rt_v6-rev0/mingw64/bin/../lib/gcc/x86_64-w64-mingw32/8.1.0/../../../../x86_64-w64-mingw32/bin/ld.exe: cannot find -lC:\Users\vsagar200\work\bin\nim-0.19.0_x64\ATen-windows10-cpu\lib\ATen_cpu.lib
    C:/Program Files/mingw-w64/x86_64-8.1.0-posix-seh-rt_v6-rev0/mingw64/bin/../lib/gcc/x86_64-w64-mingw32/8.1.0/../../../../x86_64-w64-mingw32/bin/ld.exe: cannot find -lC:\Users\vsagar200\work\bin\nim-0.19.0_x64\ATen-windows10-cpu\lib\cpuinfo.lib
    collect2.exe: error: ld returned 1 exit status
    Error: execution of an external program failed: 'g++.exe   -o C:\Users\vsagar200\work\soft\nimtorch\tests\test_xor.exe  C:\Users\vsagar200\nimcache\test_xor_d\torch_test_xor.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\stdlib_system.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\torch_torch.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\fragments_cpp.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\stdlib_macros.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\stdlib_tables.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\stdlib_hashes.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\stdlib_math.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\stdlib_strutils.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\stdlib_parseutils.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\stdlib_bitops.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\stdlib_algorithm.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\stdlib_unicode.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\stdlib_ospaths.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\stdlib_winlean.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\stdlib_dynlib.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\torch_torch_cpp.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\stdlib_os.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\stdlib_times.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\stdlib_options.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\stdlib_typetraits.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\stdlib_strformat.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\torch_tensors.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\stdlib_sequtils.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\stdlib_sets.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\torch_tensor_ops.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\torch_autograd_macro.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\torch_autograd_backward.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\torch_nn.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\torch_init.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\torch_python_helpers.cpp.o C:\Users\vsagar200\nimcache\test_xor_d\torch_optim.cpp.o  -LC:\Users\vsagar200\work\bin\nim-0.19.0_x64\ATen-windows10-cpu\lib -LC:\Users\vsagar200\work\bin\nim-0.19.0_x64\ATen-windows10-cpu\lib64 -lC:\Users\vsagar200\work\bin\nim-0.19.0_x64\ATen-windows10-cpu\lib\ATen_cpu.lib -lC:\Users\vsagar200\work\bin\nim-0.19.0_x64\ATen-windows10-cpu\lib\cpuinfo.lib '
    

    Although the ATEN path is correctly set. Tried on nim 0.18.1 and 0.19.0 the .lib files present at the same location the error is saying it cannot find.

  • install Aten with nimtorch

    install Aten with nimtorch

    thanks for the library. are there any plans to have Aten distributed with nim-torch? This will make it easier to build command-line tools that depend on it.

  • [TODO] [offtopic] discussion regarding nimble limitation (from comments in #6)

    [TODO] [offtopic] discussion regarding nimble limitation (from comments in #6)

    moved to here the discussion started here https://github.com/fragcolor-xyz/nimtorch/issues/6#issuecomment-430511494 to keep each topic separate

    @sinkingsugar

    Nimble has a major flaw, which is it applies all the packages it has on every project you have by default. That's my major concern that easy can create a mess if not considered.

    I will give you an example exactly with nimtorch: Have nimtorch installed as nimble package Also work on the repository Did not run nimble develop Rename a file in the repository Forget to update the module import myrenamedfile into import newlocation/myrenamedfile Build will succeed yet will use import myrenamedfile from nimble this time..

    but it all works if you run nimble develop, right? I think that's expected and I don't see a limitation here (in the sense of making it impossible to do certain things). if you really feel something is ill-designed with nimble it really should be a bug report in https://github.com/nim-lang/nimble/issues/ otherwise it will never get fixed (if there's anything to fix).

    That being said, one possibility would be (and that's doable, not a fundamental flaw IMO): if you call nimble build inside a local package foo, nimble could remove from search path an installed package named foo

  • Add a Gitter chat badge to README.md

    Add a Gitter chat badge to README.md

    fragcolor-xyz/nimtorch now has a Chat Room on Gitter

    @sinkingsugar has just created a chat room. You can visit it here: https://gitter.im/nimtorch/Lobby.

    This pull-request adds this badge to your README.md:

    Gitter

    If my aim is a little off, please let me know.

    Happy chatting.

    PS: Click here if you would prefer not to receive automatic pull-requests from Gitter in future.

  • Is this project still active

    Is this project still active

    Hi,

    I think a working binding to pytorch from nim could be very valuable to support the use of nim in data science. This project seems to be inactive for a long time now and it is not using the current version of pytorch - any chance that it will be updated?

  • Question. Import model trained under a Python / Torch library.

    Question. Import model trained under a Python / Torch library.

    Would it be possible or even advisable to import a pth or pkl, which was trained using FastAI, into NimTorch, for the purpose of exposing in a backend written in Nim (for efficiency and speed)?

  • Add

    Add "install fragments" to the non-conda installation doc

    add nimble install fragments to this part of the readme As a beginner I wasted an hour trying to figure out what did I do wrong, turns out it was just a package error

  • Error: expression 'step(optimizer)' has no type (or is ambiguous)

    Error: expression 'step(optimizer)' has no type (or is ambiguous)

    I am trying to use nimtorch (I am new to pytorch as well). I am struggling to run a first example.

    I have installed in Windows 10 like this:

    conda create -n aten -c fragcolor aten=2019.02.16.2841
    nimble install fragments
    nimble install [email protected]#head
    

    And then I tried to compile the code from here:

    import torch
    import torch/[nn, optim]
    
    let
      inputs = torch.tensor([[0.0, 0.0], [0.0, 1.0], [1.0, 0.0], [1.0, 1.0]])
      targets = torch.tensor([[0.0], [1.0], [1.0], [0.0]])
    
    let
      fc1 = nn.Linear(2, 4)
      fc2 = nn.Linear(4, 1)
      loss_fn = nn.MSELoss()
      optimizer = optim.SGD(fc1.parameters & fc2.parameters, lr = 0.01, momentum = 0.1)
    
    for i in 0 ..< 50000:
      optimizer.zero_grad()
    
      var predictions = fc1(inputs).relu()
      predictions = fc2(predictions).sigmoid()
    
      let loss = loss_fn(predictions, targets)
      loss.backward()
      discard optimizer.step()
    
      if i mod 5000 == 0:
        print(loss)
    

    I compile by doing:

    c:> conda activate aten
    c.> nim cpp ex01
    ....
    C:\Users\mantielero\Documents\src\torch\ex01.nim(22, 25) Error: expression 'step(optimizer)' has no type (or is ambiguous)
    

    which is the line:

    discard optimizer.step()
    

    What am I doing wrong?

  • Nimble installation fails

    Nimble installation fails

    Its been a while since the last commit, pls dont tell me nimtorch is dead, its wonderful, and i want to start using it, and libraries like this would pump up nim's popularity, are you leaving it for a while but planning on taking it back or just completely abandoned?

Code for all the Advent of Code'21 challenges mostly written in python

Advent of Code 21 Code for all the Advent of Code'21 challenges mostly written in python. They are not necessarily the best or fastest solutions but j

May 26, 2022
Repo for my Tensorflow/Keras CV experiments. Mostly revolving around the Danbooru20xx dataset

SW-CV-ModelZoo Repo for my Tensorflow/Keras CV experiments. Mostly revolving around the Danbooru20xx dataset Framework: TF/Keras 2.7 Training SQLite D

Nov 6, 2022
PyTorch - Python + Nim
PyTorch - Python + Nim

Master Release Pytorch - Py + Nim A Nim frontend for pytorch, aiming to be mostly auto-generated and internally using ATen. Because Nim compiles to C+

Dec 3, 2022
Rename Images with Auto Generated Neural Image Captions

Recaption Images with Generated Neural Image Caption Example Usage: Commandline: Recaption all images from folder /home/feng/Downloads/images to folde

May 1, 2022
Small utility to demangle Nim symbols in callgrind files
Small utility to demangle Nim symbols in callgrind files

nim_callgrind A small utility to demangle Nim symbols from callgrind files. Usage Run your (Nim) program with something like this: valgrind --tool=cal

Feb 15, 2022
Justmagic - Use a function as a method with this mystic script, like in Nim

justmagic Use a function as a method with this mystic script, like in Nim. Just

Oct 8, 2022
A bunch of random PyTorch models using PyTorch's C++ frontend
A bunch of random PyTorch models using PyTorch's C++ frontend

PyTorch Deep Learning Models using the C++ frontend Gettting started Clone the repo 1. https://github.com/mrdvince/pytorchcpp 2. cd fashionmnist or

Jul 13, 2021
An Inverse Kinematics library aiming performance and modularity
An Inverse Kinematics library aiming performance and modularity

IKPy Demo Live demos of what IKPy can do (click on the image below to see the video): Also, a presentation of IKPy: Presentation. Features With IKPy,

Nov 24, 2022
Aiming at the common training datsets split, spectrum preprocessing, wavelength select and calibration models algorithm involved in the spectral analysis process
Aiming at the common training datsets split, spectrum preprocessing, wavelength select and calibration models algorithm involved in the spectral analysis process

Aiming at the common training datsets split, spectrum preprocessing, wavelength select and calibration models algorithm involved in the spectral analysis process, a complete algorithm library is established, which is named opensa (openspectrum analysis).

Nov 24, 2022
FAST Aiming at the problems of cumbersome steps and slow download speed of GNSS data
FAST Aiming at the problems of cumbersome steps and slow download speed of GNSS data

FAST Aiming at the problems of cumbersome steps and slow download speed of GNSS data, a relatively complete set of integrated multi-source data download terminal software fast is developed. The software contains most of the data sources required in the process of GNSS scientific research and learning. The way of parallel download greatly improves the efficiency of download.

Sep 20, 2022
Project Aquarium is a SUSE-sponsored open source project aiming at becoming an easy to use, rock solid storage appliance based on Ceph.

Project Aquarium Project Aquarium is a SUSE-sponsored open source project aiming at becoming an easy to use, rock solid storage appliance based on Cep

Jul 21, 2022
An integration of several popular automatic augmentation methods, including OHL (Online Hyper-Parameter Learning for Auto-Augmentation Strategy) and AWS (Improving Auto Augment via Augmentation Wise Weight Sharing) by Sensetime Research.

An integration of several popular automatic augmentation methods, including OHL (Online Hyper-Parameter Learning for Auto-Augmentation Strategy) and AWS (Improving Auto Augment via Augmentation Wise Weight Sharing) by Sensetime Research.

Nov 18, 2022
Official Pytorch implementation of paper "Reverse Engineering of Generative Models: Inferring Model Hyperparameters from Generated Images"
Official Pytorch implementation of paper

Reverse_Engineering_GMs Official Pytorch implementation of paper "Reverse Engineering of Generative Models: Inferring Model Hyperparameters from Gener

Nov 27, 2022
Fast and robust clustering of point clouds generated with a Velodyne sensor.
Fast and robust clustering of point clouds generated with a Velodyne sensor.

Depth Clustering This is a fast and robust algorithm to segment point clouds taken with Velodyne sensor into objects. It works with all available Velo

Dec 1, 2022
[ICLR 2021] Rank the Episodes: A Simple Approach for Exploration in Procedurally-Generated Environments.
[ICLR 2021] Rank the Episodes: A Simple Approach for Exploration in Procedurally-Generated Environments.

[ICLR 2021] RAPID: A Simple Approach for Exploration in Reinforcement Learning This is the Tensorflow implementation of ICLR 2021 paper Rank the Episo

Nov 21, 2022
BARTScore: Evaluating Generated Text as Text Generation
BARTScore: Evaluating Generated Text as Text Generation

This is the Repo for the paper: BARTScore: Evaluating Generated Text as Text Generation Updates 2021.06.28 Release online evaluation Demo 2021.06.25 R

Nov 29, 2022
Tracing Versus Freehand for Evaluating Computer-Generated Drawings (SIGGRAPH 2021)
Tracing Versus Freehand for Evaluating Computer-Generated Drawings (SIGGRAPH 2021)

Tracing Versus Freehand for Evaluating Computer-Generated Drawings (SIGGRAPH 2021) Zeyu Wang, Sherry Qiu, Nicole Feng, Holly Rushmeier, Leonard McMill

Nov 10, 2022