[Issue]: Torch rocm not avaiable on gfx1150/gfx1030/gfx906 and windows local build fail

[AmosLewis]

Problem Description

Issue1:
There is no torch package prebuild in https://rocm.nightlies.amd.com/v2/ for gfx1150/gfx1030/gfx906. I need it on Strix/Navi2/Radeon 7 for windows system. @geomin12 is looking at the strix/gfx1150.
Issue2:
At the same time, I tried to manually build pytorch locally follow the instructions: https://github.com/ROCm/TheRock/blob/main/external-builds/pytorch/README.md. But it also failed.

I tried to manually build torch locally today on windows Strix gfx1150, after a few fixes, I find that vs 2022 does not work with pytorch build, it required vs 2019. It might be the root reason why no torch package prebuild in nightlies.

Detailed info https://gist.github.com/AmosLewis/ac3536cc1c4e67abe9a318c35f20efcb?permalink_comment_id=5816129#gis… and https://gist.github.com/AmosLewis/ac3536cc1c4e67abe9a318c35f20efcb?permalink_comment_id=5816463#gis…

Analysis: torch-rocm local build required for visual studio for libomp.dll in build_prod_wheels.py and copy that to torch source code before build. But after I install Microsoft Visual C++ 2015–2019 Redistributable (x64) version 14.28.29914 via VC_redist.x64.exe download from microsoft website, looks like vs2019 does not have this libomp.dll. I also installed VS2022, it has this libomp.dll but pytorch source code side required vs2019 when cmake. VS2022 won't work.

Operating System

Windows10

CPU

AMD Ryzen AI 9 HX 370 w/ Radeon 890M

GPU

AMD Radeon(TM) 890M Graphics

ROCm Version

rocm7

ROCm Component

No response

Steps to Reproduce

python build_prod_wheels.py build `
  --install-rocm --index-url https://rocm.nightlies.amd.com/v2/gfx1150/ `
  --pytorch-dir C:/b/pytorch `
  --pytorch-audio-dir C:/b/audio `
  --pytorch-vision-dir C:/b/vision `
  --output-dir %HOME%/tmp/pyout

With only VS2019 and without Visual Studio 2022 or no VS installed, it will show bug:
`

RuntimeError: VCToolsRedistDir not set, can't copy libomp to torch lib

With visual studios 2022 installed and `$env:VCToolsRedistDir = "C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Redist\MSVC\14.44.35112" enviroment variables setup, it will show:

...
CMake Error at CMakeLists.txt:28 (project):
  Generator

    Visual Studio 16 2019

  could not find any instance of Visual Studio.



-- Configuring incomplete, errors occurred!
Traceback (most recent call last):
  File "C:\Users\chiliu12\Documents\src\TheRock\external-builds\pytorch\build_prod_wheels.py", line 977, in <module>
    main(sys.argv[1:])
  File "C:\Users\chiliu12\Documents\src\TheRock\external-builds\pytorch\build_prod_wheels.py", line 973, in main
    args.func(args)
  File "C:\Users\chiliu12\Documents\src\TheRock\external-builds\pytorch\build_prod_wheels.py", line 442, in do_build
    do_build_pytorch(
  File "C:\Users\chiliu12\Documents\src\TheRock\external-builds\pytorch\build_prod_wheels.py", line 756, in do_build_pytorch
    exec([sys.executable, "setup.py", "bdist_wheel"], cwd=pytorch_dir, env=env)
  File "C:\Users\chiliu12\Documents\src\TheRock\external-builds\pytorch\build_prod_wheels.py", line 171, in exec
    subprocess.check_call(args, cwd=str(cwd), env=full_env)
  File "C:\Program Files\WindowsApps\PythonSoftwareFoundation.Python.3.12_3.12.2800.0_x64__qbz5n2kfra8p0\Lib\subprocess.py", line 413, in check_call
    raise CalledProcessError(retcode, cmd)
subprocess.CalledProcessError: Command '['C:\\Users\\chiliu12\\Documents\\src\\TheRock\\.venv\\Scripts\\python.exe', 'setup.py', 'bdist_wheel']' returned non-zero exit status 1.

(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support

PS C:\WINDOWS\system32> (Get-WmiObject Win32_OperatingSystem).Version
10.0.26100
PS C:\WINDOWS\system32> (Get-WmiObject win32_Processor).Name
AMD Ryzen AI 9 HX 370 w/ Radeon 890M
PS C:\WINDOWS\system32> (Get-WmiObject win32_VideoController).Name
NVIDIA GeForce RTX 4070 Laptop GPU
Microsoft Remote Display Adapter
AMD Radeon(TM) 890M Graphics
PS C:\WINDOWS\system32>

Additional Information

No response

[yuehuang010]

I was able to build 1150 (yesterday).

Did you start the x64 developer prompt? You shouldn't need to set your own environment variables.

[AmosLewis]

I was able to build 1150 (yesterday).

Did you start the x64 developer prompt? You shouldn't need to set your own environment variables.

Are you using VS2019? I can only install part of vs2019 not all the tools.

[yuehuang010]

I used VS2022 as the instructions required. The command prompt might be hard to find; you actually have to browser the start menu folder itself.

[AmosLewis]

I used VS2022 as the instructions required. The command prompt might be hard to find; you actually have to browser the start menu folder itself.

OK, I found it, and use this intead, Rerun all the commands, then got the same x64_Native_Tools_Command_VS2022_Prompt_build_log_gfx1150_windows.md cmake error:

CMake Error:
  Generator

    Visual Studio 16 2019

  could not find any instance of Visual Studio.

[yuehuang010]

In the command line, can you type "where cl.exe". What path does it return? It should look like C:\Program Files\Microsoft Visual Studio\ ... \bin\Hostx64\x64\cl.exe

Can you confirm when you installed VS2022, you checked the "Desktop Development with C++"?

The scripts look right; I built my own rocm so maybe there might be incompatibility between nighties.

[AmosLewis]

In the command line, can you type "where cl.exe". What path does it return? It should look like C:\Program Files\Microsoft Visual Studio\ ... \bin\Hostx64\x64\cl.exe

Can you confirm when you installed VS2022, you checked the "Desktop Development with C++"?

The scripts look right; I built my own rocm so maybe there might be incompatibility between nighties.

Yes, I definitly checked the Desktop Development with C++.

C:\Program Files\Microsoft Visual Studio\2022\Community>where cl.exe
C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.44.35207\bin\Hostx64\x64\cl.exe

I also deleted all related folders(TheRock including .venv & build dir, C:b\pytorch&audio&vision) and do a clean download and build on strix/gfx1150 with vs2022 x64 promt admin mode, still same cmake error for me , here is the detailed log gfx1150_windows_vs2022_prompt_clean_build_log_1023.md

[AmosLewis]

In the command line, can you type "where cl.exe". What path does it return? It should look like C:\Program Files\Microsoft Visual Studio\ ... \bin\Hostx64\x64\cl.exe

Can you confirm when you installed VS2022, you checked the "Desktop Development with C++"?

The scripts look right; I built my own rocm so maybe there might be incompatibility between nighties.

@yuehuang010 And I notice if I just cd in pytorch source code and manually run cmake -G "Visual Studio 17 2022" insead of "Visual Studio 16 2019" in my previous log, pytorch just build success. But when run python build_prod_wheels.py build, it will always choose the "Visual Studio 16 2019" which always fail. Here is my manually build successfully log with cmake -G"Visual Studio 17 2022" ... cmake-vs-17-2022_rocm-torch-build_gfx1150

[patientx]

so anyone actually managed to build 1030 ?

[Alex-Vasile]

I would appreciate torch-rocm prebuilt for gfx906 on linux too

[dezhiAmd]

With visual studio 2022 Microsoft Visual Studio Professional 2022 (64-bit)
Version 17.14.19
Builds are successful