fix: install real windows cuda cccl headers

Add cuda-cccl_win-64 and normalize nv/target into the CUDA include tree so Windows PyTorch extension builds can resolve cuda_fp16.h.

Author: DrHepa

.github/workflows/build-native-windows-wheels.yml

@@ -46,7 +46,7 @@ jobs:
         shell: pwsh
         run: |
           $cudaEnv = Join-Path $env:RUNNER_TEMP 'cuda-build'
-          conda create -y -p $cudaEnv -c nvidia/label/cuda-12.8.1 -c nvidia -c conda-forge cuda-nvcc cuda-cudart-dev cuda-cccl libcublas-dev libcusparse-dev libcusolver-dev libcurand-dev libcufft-dev
+          conda create -y -p $cudaEnv -c nvidia/label/cuda-12.8.1 -c nvidia -c conda-forge cuda-nvcc cuda-cudart-dev cuda-cccl cuda-cccl_win-64 libcublas-dev libcusparse-dev libcusolver-dev libcurand-dev libcufft-dev
           $cudaRoot = Join-Path $cudaEnv 'Library'
           $nvcc = Join-Path $cudaRoot 'bin\nvcc.exe'
           if (-not (Test-Path $nvcc)) {

native-wheels/README.md

@@ -129,4 +129,4 @@ Run it from GitHub Actions or with `gh` after the workflow has been committed an
 gh workflow run build-native-windows-wheels.yml -f release_tag=native-wheels-torch270-cu128-v1 -f upload_release=true
 ```
 
-The workflow builds `cp311` and `cp312` wheels on `windows-2022`, installs CUDA Toolkit 12.8 build components plus CUDA dev libraries (`cuBLAS`, `cuSPARSE`, `cuSOLVER`, `cuRAND`, `cuFFT`) into a temporary Conda prefix from the NVIDIA Conda channel, runs the smoke test, and optionally uploads the wheels plus license files to the release tag.
+The workflow builds `cp311` and `cp312` wheels on `windows-2022`, installs CUDA Toolkit 12.8 build components plus CUDA dev libraries (`cuBLAS`, `cuSPARSE`, `cuSOLVER`, `cuRAND`, `cuFFT`, and `cuda-cccl_win-64`) into a temporary Conda prefix from the NVIDIA Conda channel, runs the smoke test, and optionally uploads the wheels plus license files to the release tag.

native-wheels/scripts/build-diff-gaussian.ps1

@@ -51,6 +51,42 @@ function Ensure-Directory {
     New-Item -ItemType Directory -Force -Path $PathValue | Out-Null
 }
 
+function Ensure-CudaCcclHeaders {
+    param([Parameter(Mandatory = $true)][string]$CudaRoot)
+
+    $cudaInclude = Join-Path $CudaRoot 'include'
+    $cudaNvInclude = Join-Path $cudaInclude 'nv'
+    $cudaNvTarget = Join-Path $cudaNvInclude 'target'
+    if (Test-Path $cudaNvTarget) {
+        Write-Host "[native-wheels] CUDA CCCL nv/target header already available at: $cudaNvTarget"
+        return
+    }
+
+    $candidateNvIncludes = @(
+        (Join-Path $CudaRoot 'include\cccl\nv'),
+        (Join-Path $CudaRoot 'targets\x86_64-windows\include\nv'),
+        (Join-Path $CudaRoot 'targets\x86_64-win32\include\nv')
+    )
+    $sourceNvInclude = $candidateNvIncludes | Where-Object { Test-Path (Join-Path $_ 'target') } | Select-Object -First 1
+
+    if (-not $sourceNvInclude) {
+        $foundNvTarget = Get-ChildItem -Path $CudaRoot -Recurse -File -Filter target -ErrorAction SilentlyContinue |
+            Where-Object { $_.FullName -match '[\\/]nv[\\/]target$' } |
+            Select-Object -First 1
+        if ($foundNvTarget) {
+            $sourceNvInclude = Split-Path $foundNvTarget.FullName -Parent
+        }
+    }
+
+    if (-not $sourceNvInclude) {
+        throw "Could not locate CUDA CCCL nv/target header under $CudaRoot. Install the NVIDIA Conda cuda-cccl_win-64 package or provide a full CUDA Toolkit layout."
+    }
+
+    Ensure-Directory -PathValue $cudaNvInclude
+    Copy-Item -Path (Join-Path $sourceNvInclude '*') -Destination $cudaNvInclude -Recurse -Force
+    Write-Host "[native-wheels] Mirrored CUDA CCCL nv headers from $sourceNvInclude to: $cudaNvInclude"
+}
+
 $resolvedOutDir = [System.IO.Path]::GetFullPath($OutDir)
 $resolvedWorkDir = [System.IO.Path]::GetFullPath($WorkDir)
 $venvDir = Join-Path $resolvedWorkDir 'venv'
@@ -79,15 +115,8 @@ $ccclInclude = Join-Path $CudaRoot 'include\cccl'
 if (Test-Path $ccclInclude) {
     $env:INCLUDE = "$ccclInclude;$env:INCLUDE"
     Write-Host "[native-wheels] Added CUDA CCCL include path: $ccclInclude"
-
-    $cudaInclude = Join-Path $CudaRoot 'include'
-    $ccclNvInclude = Join-Path $ccclInclude 'nv'
-    $cudaNvInclude = Join-Path $cudaInclude 'nv'
-    if ((Test-Path $ccclNvInclude) -and -not (Test-Path $cudaNvInclude)) {
-        Copy-Item -Path $ccclNvInclude -Destination $cudaNvInclude -Recurse
-        Write-Host "[native-wheels] Mirrored CUDA CCCL nv headers to: $cudaNvInclude"
-    }
 }
+Ensure-CudaCcclHeaders -CudaRoot $CudaRoot
 
 Invoke-Python -Arguments @('-m', 'venv', $venvDir)
 $venvPython = Join-Path $venvDir 'Scripts\python.exe'

native-wheels/scripts/build-nvdiffrast.ps1

@@ -50,6 +50,42 @@ function Ensure-Directory {
     New-Item -ItemType Directory -Force -Path $PathValue | Out-Null
 }
 
+function Ensure-CudaCcclHeaders {
+    param([Parameter(Mandatory = $true)][string]$CudaRoot)
+
+    $cudaInclude = Join-Path $CudaRoot 'include'
+    $cudaNvInclude = Join-Path $cudaInclude 'nv'
+    $cudaNvTarget = Join-Path $cudaNvInclude 'target'
+    if (Test-Path $cudaNvTarget) {
+        Write-Host "[native-wheels] CUDA CCCL nv/target header already available at: $cudaNvTarget"
+        return
+    }
+
+    $candidateNvIncludes = @(
+        (Join-Path $CudaRoot 'include\cccl\nv'),
+        (Join-Path $CudaRoot 'targets\x86_64-windows\include\nv'),
+        (Join-Path $CudaRoot 'targets\x86_64-win32\include\nv')
+    )
+    $sourceNvInclude = $candidateNvIncludes | Where-Object { Test-Path (Join-Path $_ 'target') } | Select-Object -First 1
+
+    if (-not $sourceNvInclude) {
+        $foundNvTarget = Get-ChildItem -Path $CudaRoot -Recurse -File -Filter target -ErrorAction SilentlyContinue |
+            Where-Object { $_.FullName -match '[\\/]nv[\\/]target$' } |
+            Select-Object -First 1
+        if ($foundNvTarget) {
+            $sourceNvInclude = Split-Path $foundNvTarget.FullName -Parent
+        }
+    }
+
+    if (-not $sourceNvInclude) {
+        throw "Could not locate CUDA CCCL nv/target header under $CudaRoot. Install the NVIDIA Conda cuda-cccl_win-64 package or provide a full CUDA Toolkit layout."
+    }
+
+    Ensure-Directory -PathValue $cudaNvInclude
+    Copy-Item -Path (Join-Path $sourceNvInclude '*') -Destination $cudaNvInclude -Recurse -Force
+    Write-Host "[native-wheels] Mirrored CUDA CCCL nv headers from $sourceNvInclude to: $cudaNvInclude"
+}
+
 $resolvedOutDir = [System.IO.Path]::GetFullPath($OutDir)
 $resolvedWorkDir = [System.IO.Path]::GetFullPath($WorkDir)
 $venvDir = Join-Path $resolvedWorkDir 'venv'
@@ -77,15 +113,8 @@ $ccclInclude = Join-Path $CudaRoot 'include\cccl'
 if (Test-Path $ccclInclude) {
     $env:INCLUDE = "$ccclInclude;$env:INCLUDE"
     Write-Host "[native-wheels] Added CUDA CCCL include path: $ccclInclude"
-
-    $cudaInclude = Join-Path $CudaRoot 'include'
-    $ccclNvInclude = Join-Path $ccclInclude 'nv'
-    $cudaNvInclude = Join-Path $cudaInclude 'nv'
-    if ((Test-Path $ccclNvInclude) -and -not (Test-Path $cudaNvInclude)) {
-        Copy-Item -Path $ccclNvInclude -Destination $cudaNvInclude -Recurse
-        Write-Host "[native-wheels] Mirrored CUDA CCCL nv headers to: $cudaNvInclude"
-    }
 }
+Ensure-CudaCcclHeaders -CudaRoot $CudaRoot
 
 Invoke-Python -Arguments @('-m', 'venv', $venvDir)
 $venvPython = Join-Path $venvDir 'Scripts\python.exe'

validate_text_only_setup.py

@@ -156,7 +156,7 @@ def validate_native_wheels_tooling() -> None:
     require("$ErrorActionPreference = 'Stop'" in build_nvdiffrast, "build-nvdiffrast.ps1 must stop on errors")
     require("Invoke-Expression" not in build_nvdiffrast, "build-nvdiffrast.ps1 must not construct commands via Invoke-Expression")
     require("include\\cccl" in build_nvdiffrast and "$env:INCLUDE" in build_nvdiffrast, "build-nvdiffrast.ps1 must add Conda CUDA CCCL headers to INCLUDE")
-    require("Mirrored CUDA CCCL nv headers" in build_nvdiffrast, "build-nvdiffrast.ps1 must mirror Conda CCCL nv headers into CUDA include")
+    require("Ensure-CudaCcclHeaders" in build_nvdiffrast and "cuda-cccl_win-64" in build_nvdiffrast, "build-nvdiffrast.ps1 must normalize Conda CCCL nv headers into CUDA include")
     require("pip wheel" in build_nvdiffrast and "--no-build-isolation" in build_nvdiffrast, "build-nvdiffrast.ps1 must build wheels via pip wheel --no-build-isolation")
     require("https://github.com/NVlabs/nvdiffrast.git" in build_nvdiffrast and "v0.4.0" in build_nvdiffrast, "build-nvdiffrast.ps1 must pin nvdiffrast source")
 
@@ -165,7 +165,7 @@ def validate_native_wheels_tooling() -> None:
     require("$ErrorActionPreference = 'Stop'" in build_diff, "build-diff-gaussian.ps1 must stop on errors")
     require("Invoke-Expression" not in build_diff, "build-diff-gaussian.ps1 must not construct commands via Invoke-Expression")
     require("include\\cccl" in build_diff and "$env:INCLUDE" in build_diff, "build-diff-gaussian.ps1 must add Conda CUDA CCCL headers to INCLUDE")
-    require("Mirrored CUDA CCCL nv headers" in build_diff, "build-diff-gaussian.ps1 must mirror Conda CCCL nv headers into CUDA include")
+    require("Ensure-CudaCcclHeaders" in build_diff and "cuda-cccl_win-64" in build_diff, "build-diff-gaussian.ps1 must normalize Conda CCCL nv headers into CUDA include")
     require("pip wheel" in build_diff and "--no-build-isolation" in build_diff, "build-diff-gaussian.ps1 must build wheels via pip wheel --no-build-isolation")
     require("https://github.com/autonomousvision/mip-splatting.git" in build_diff and "dda02ab5ecf45d6edb8c540d9bb65c7e451345a9" in build_diff, "build-diff-gaussian.ps1 must pin mip-splatting source")
     require("submodules/diff-gaussian-rasterization" in build_diff, "build-diff-gaussian.ps1 must build the diff-gaussian subdirectory")
@@ -184,6 +184,7 @@ def validate_native_wheels_tooling() -> None:
     require("windows-2022" in workflow, "native wheel workflow must build on Windows")
     require("conda-incubator/setup-miniconda@v4" in workflow and "cuda-12.8.1" in workflow, "native wheel workflow must install CUDA Toolkit 12.8 packages from NVIDIA Conda")
     require("conda create -y -p $cudaEnv" in workflow, "native wheel workflow must install CUDA packages into an explicit prefix")
+    require("cuda-cccl_win-64" in workflow, "native wheel workflow must install the real Windows CCCL header package")
     require("libcusparse-dev" in workflow and "libcublas-dev" in workflow, "native wheel workflow must install CUDA dev headers required by PyTorch")
     require("CUDA_HOME=$cudaRoot" in workflow and "CUDA_PATH=$cudaRoot" in workflow, "native wheel workflow must export CUDA_HOME/CUDA_PATH")
     require("ilammy/msvc-dev-cmd@v1" in workflow, "native wheel workflow must prepare MSVC developer shell")