windows: CUDA serving build — cuda-guard + cmake/llvm runtime paths (rescued from #2056)

tools/scripts/generated/manifest.windows.ps1

@@ -13,8 +13,8 @@ $script:ContinuumManifest = [ordered]@{
   'airc-firewall' = @{ order = 27; tier = 0; flags = @('grid'); applies = 'has-airc'; accept = 'netsh advfirewall firewall show rule name="airc daemon inbound (continuum grid)"'; source = @{ type = 'command'; run = 'New-NetFirewallRule -DisplayName ''airc daemon inbound (continuum grid)'' -Direction Inbound -Action Allow -Profile Any' } }
   'manifest-gen' = @{ order = 28; tier = 3; flags = @('dev'); accept = 'cargo run -q -p manifest-gen -- --check'; source = @{ type = 'command'; run = 'cargo run -q -p manifest-gen' } }
   'msvc' = @{ order = 30; tier = 3; flags = @('dev'); accept = 'vswhere -latest -requires Microsoft.VisualStudio.Component.VC.Tools.x86.x64 -property installationPath'; source = @{ type = 'winget'; id = 'Microsoft.VisualStudio.2022.BuildTools'; override = '--wait --quiet --add Microsoft.VisualStudio.Workload.VCTools --includeRecommended' } }
-  'cmake' = @{ order = 40; tier = 3; flags = @('dev'); accept = 'cmake --version'; source = @{ type = 'archive'; url = 'https://github.com/Kitware/CMake/releases/download/v3.30.5/cmake-3.30.5-windows-x86_64.zip'; version = '3.30.5'; sha256 = '5ab6e1faf20256ee4f04886597e8b6c3b1bd1297b58a68a58511af013710004b'; extract = 'strip-top-dir' } }
-  'llvm-libclang' = @{ order = 50; tier = 3; flags = @('dev'); accept = 'test-path ~/.continuum/tools/llvm/bin/libclang.dll'; source = @{ type = 'archive'; url = 'https://github.com/llvm/llvm-project/releases/download/llvmorg-18.1.8/clang+llvm-18.1.8-x86_64-pc-windows-msvc.tar.xz'; version = '18.1.8'; sha256 = '22c5907db053026cc2a8ff96d21c0f642a90d24d66c23c6d28ee7b1d572b82e8'; extract = 'members:*/bin/libclang.dll,*/lib/clang/*' } }
+  'cmake' = @{ order = 40; tier = 3; flags = @('dev'); accept = 'cmake --version'; source = @{ type = 'archive'; url = 'https://github.com/Kitware/CMake/releases/download/v3.30.5/cmake-3.30.5-windows-x86_64.zip'; version = '3.30.5'; sha256 = '5ab6e1faf20256ee4f04886597e8b6c3b1bd1297b58a68a58511af013710004b'; extract = 'strip-top-dir' }; runtime_path = @('~/.continuum/tools/cmake/bin') }
+  'llvm-libclang' = @{ order = 50; tier = 3; flags = @('dev'); accept = 'test-path ~/.continuum/tools/llvm/bin/libclang.dll'; source = @{ type = 'archive'; url = 'https://github.com/llvm/llvm-project/releases/download/llvmorg-18.1.8/clang+llvm-18.1.8-x86_64-pc-windows-msvc.tar.xz'; version = '18.1.8'; sha256 = '22c5907db053026cc2a8ff96d21c0f642a90d24d66c23c6d28ee7b1d572b82e8'; extract = 'members:*/bin/libclang.dll,*/lib/clang/*' }; runtime_path = @('~/.continuum/tools/llvm/bin') }
   'cuda' = @{ order = 60; tier = 3; flags = @('dev'); applies = 'has-nvidia'; accept = 'nvcc --version >= 12.8'; source = @{ type = 'redist'; version = '12.9.1'; manifest = 'https://developer.download.nvidia.com/compute/cuda/redist/redistrib_12.9.1.json'; components = @('cuda_nvcc', 'cuda_cudart', 'libcublas', 'libcurand', 'cuda_nvrtc', 'cuda_cccl') }; runtime_path = @('~/.continuum/cuda-*/Library/bin') }
   'build-core' = @{ order = 90; tier = 3; flags = @('dev'); accept = 'continuum-core-server.exe boots past the GPU-detection gate on the target device'; build = @{ features = 'cuda,load-dynamic-ort'; profile = 'release'; crt = 'static'; cmake_generator = 'Visual Studio 17 2022'; cuda_arch = '120'; msvc_host = 'vs2022' } }
   'run' = @{ order = 100; tier = 3; accept = 'continuum-core-server binary present + serves TCP 9100' }

tools/scripts/generated/manifest.windows.sh

@@ -27,4 +27,4 @@ declare -A MOD_ARGS=()
 declare -A MOD_RUN=( ['gh-auth']='gh auth login --hostname github.com --git-protocol https --web' ['airc-firewall']='New-NetFirewallRule -DisplayName '\''airc daemon inbound (continuum grid)'\'' -Direction Inbound -Action Allow -Profile Any' ['manifest-gen']='cargo run -q -p manifest-gen' )
 declare -A MOD_BUILD_FEATURES=( ['build-core']='cuda,load-dynamic-ort' )
 declare -A MOD_BUILD_PROFILE=( ['build-core']='release' )
-declare -A MOD_RUNTIME_PATH=( ['cuda']='~/.continuum/cuda-*/Library/bin' )
+declare -A MOD_RUNTIME_PATH=( ['cmake']='~/.continuum/tools/cmake/bin' ['llvm-libclang']='~/.continuum/tools/llvm/bin' ['cuda']='~/.continuum/cuda-*/Library/bin' )

tools/scripts/install-manifest.toml

@@ -159,6 +159,12 @@ formula = "cmake"
 [module.sources.linux]
 type = "apt"              # run-verify pending on a linux node
 package = "cmake"
+[module.runtime_path]
+# The Windows archive installs to ~/.continuum/tools/cmake; its bin/ MUST be on
+# PATH at BUILD time — the llama crate's build.rs shells out to `cmake` to compile
+# vendored llama.cpp, and nvcc-free directml builds still need it. brew/apt already
+# put cmake on PATH, so this is windows-only. [[windows-build-env-drift]]
+windows = ["~/.continuum/tools/cmake/bin"]
 
 [[module]]
 id = "llvm-libclang"
@@ -183,6 +189,11 @@ accept_macos = "test -f /Library/Developer/CommandLineTools/usr/lib/libclang.dyl
 type = "apt"              # run-verify pending on a linux node
 package = "libclang-dev"
 accept_linux = "test -f /usr/lib/llvm-18/lib/libclang.so.1 || ldconfig -p | grep -q libclang"
+[module.runtime_path]
+# The Windows archive installs libclang.dll to ~/.continuum/tools/llvm/bin; on PATH
+# so bindgen (llama crate) can load it at build time. mac/linux resolve libclang via
+# system paths. [[windows-build-env-drift]]
+windows = ["~/.continuum/tools/llvm/bin"]
 
 [[module]]
 id = "cuda"

tools/scripts/shared/cargo-features.sh

@@ -59,7 +59,13 @@ case "$(uname -s)" in
     # top if Nvidia is present so ORT picks CUDA first (faster) +
     # DirectML stays as a co-listed EP for non-CUDA-supported ops.
     CARGO_GPU_FEATURES="--features directml"
-    if command -v nvidia-smi &>/dev/null; then
+    # candle-cuda's affine.cu compiles via nvcc, which needs the MSVC host
+    # compiler cl.exe on PATH (an active vcvars env). Only add cuda when cl.exe
+    # is actually reachable; otherwise nvcc fatals "Cannot find compiler
+    # 'cl.exe'" and the ENTIRE core build dies. directml needs no kernel
+    # compilation, so it stays as the universal Windows GPU EP and the build
+    # degrades gracefully instead of hard-failing. [[windows-build-env-drift]]
+    if command -v nvidia-smi &>/dev/null && command -v cl.exe &>/dev/null; then
       CARGO_GPU_FEATURES="--features cuda,directml"
     fi
     ;;