Windows running issue: broken bundled slangc.exe binary and missing compiler location

[sleeplessai]

Traceback (most recent call last):
File "D:\slang-torch-main\examples\inline-mlp-example\mlp_image_fit.py", line 8, in
from image_model import RenderImage
File "D:\slang-torch-main\examples\inline-mlp-example\image_model.py", line 19, in
m = slangtorch.loadModule('image-model.slang',
File "C:\miniconda3\envs\inr\lib\site-packages\slangtorch\slangtorch.py", line 685, in loadModule
rawModule = _loadModule(fileName, moduleName, buildDir, options, sourceDir=outputFolder, verbose=verbose, includePaths=includePaths, dryRun=False, skipNinjaCheck=skipNinjaCheck, extraCudaFlags=extraCudaFlags, extraSlangFlags=extraSlangFlags)
File "C:\miniconda3\envs\inr\lib\site-packages\slangtorch\slangtorch.py", line 563, in _loadModule
resultCpp, metadataCpp = compileSlang(metadata.get("cpp", None), fileName, "torch-binding", options, cppOutName, verbose, includePaths=includePaths, dryRun=dryRun, extraSlangFlags=extraSlangFlags)
File "C:\miniconda3\envs\inr\lib\site-packages\slangtorch\slangtorch.py", line 313, in compileSlang
return True, (_compileSlang(metadata, fileName, targetMode, options, outputFile, includePaths, verbose, extraSlangFlags) if not dryRun else None)
File "C:\miniconda3\envs\inr\lib\site-packages\slangtorch\slangtorch.py", line 342, in _compileSlang
raise RuntimeError(f"Compilation failed with error {result.returncode}")
RuntimeError: Compilation failed with error 3221226505

Try runing examples on Windows with native pytorch package.
My pytorch version is 2.1.2cu118 with nvcc compiler 11.8.
Another success try is on Linux with pytorch 2.6.0cu124 with nvcc compiler 12.4.
How to fix this error? Does slangtorch only support for Linux or latest nvcc compiler?

[bmillsNV]

We currently support 12.4. Can you update your cuda version on Windows and give it a try?

[sleeplessai]

We currently support 12.4. Can you update your cuda version on Windows and give it a try?

Sure, I've already tested the with torch 2.6.0cu12.4 with same version nvcc.
The bad news is the error code remains.
The good one is the issue has been solved by my fixing.

There are two bugs that need to be fixed.

1. The bundled slangc.exe binary seems broken.

When calling slangtorch.loadModule, it uses bundled slangc binary for the compiling slang sources.
I checked this bundled slangc.exe version but gets nothing with argument --version.
To fix this, I replaced slangc.exe and slang.dll in site-packages\slangtorch\bin with a well-worked pair confirmed in my other project.
This slangc.exe build is from VulkanSDK 1.4.309.0 and its version is 2025.6.1.
I've uploaded these two files for your reproduction work.
slangc_vksdk.zip

2. Add -prerelease argument for vswhere.exe to get compiler path from VS preview branch.

The installation of Visual Studio 2022 on my PC and the lab workstation shares the same version of cl compilers (old 14.30, latest 14.44), but comes from different branches (Preview and Community). I didn't notice this difference before.
After fixing the slangc issue, a new error report shows that slangtorch cannot correctly locate the compiler of VS preview branch.

I added a '-prerelease' argument for the line code
https://github.com/shader-slang/slang-torch/blob/main/slangtorch/slangtorch.py#L96
to be
result = subprocess.run([vswhere_path, '-latest', '-products', '*', '-property', 'installationPath', '-prerelease'], stdout=subprocess.PIPE, stderr=subprocess.PIPE)

After these two fixes, the examples may work well on Windows now.
I would like to update the code later and make a new pull request.

PR is here.
#32