v2.9.0

recipe/build.sh

@@ -199,15 +199,17 @@ elif [[ ${cuda_compiler_version} != "None" ]]; then
     export CUDAToolkit_ROOT=${PREFIX}
     case ${target_platform} in
         linux-64)
-            export CUDAToolkit_TARGET_DIR=${PREFIX}/targets/x86_64-linux
+            CUDA_TARGET=x86_64-linux
             ;;
         linux-aarch64)
-            export CUDAToolkit_TARGET_DIR=${PREFIX}/targets/sbsa-linux
+            CUDA_TARGET=sbsa-linux
             ;;
         *)
             echo "unknown CUDA arch, edit build.sh"
             exit 1
     esac
+    export CUDAToolkit_TARGET_DIR=${PREFIX}/targets/${CUDA_TARGET}
+    sed -i -e "s,@CUDA_TARGET@,${CUDA_TARGET}," torch/_inductor/cpp_builder.py
 
     # Compatibility matrix for update: https://en.wikipedia.org/wiki/CUDA#GPUs_supported
     # Warning from pytorch v1.12.1: In the future we will require one to

recipe/patches/0007-Add-conda-prefix-to-inductor-include-paths.patch

@@ -1,7 +1,7 @@
-From b06c89a4b491c3137da0325212fb7481ecabfe19 Mon Sep 17 00:00:00 2001
+From 2ece5dedfbb4783ddfb5985567ce0b3d12076a17 Mon Sep 17 00:00:00 2001
 From: Daniel Petry <[email protected]>
 Date: Tue, 21 Jan 2025 17:45:23 -0600
-Subject: [PATCH 07/13] Add conda prefix to inductor include paths
+Subject: [PATCH 07/15] Add conda prefix to inductor include paths
 
 Currently inductor doesn't look in conda's includes and libs. This results in
 errors when it tries to compile, if system versions are being used of
@@ -13,18 +13,19 @@ end user provides a <filename>_compile_flags.json file.
 Updated to use `sys.prefix` per
 https://github.com/conda-forge/pytorch-cpu-feedstock/issues/424.
 ---
- torch/_inductor/cpp_builder.py | 1 +
- 1 file changed, 1 insertion(+)
+ torch/_inductor/cpp_builder.py | 2 ++
+ 1 file changed, 2 insertions(+)
 
 diff --git a/torch/_inductor/cpp_builder.py b/torch/_inductor/cpp_builder.py
-index e2cb445ed10..120fae77475 100644
+index e2cb445ed10..a8fe5402f66 100644
 --- a/torch/_inductor/cpp_builder.py
 +++ b/torch/_inductor/cpp_builder.py
-@@ -1410,6 +1410,7 @@ def get_cpp_torch_options(
+@@ -1410,6 +1410,8 @@ def get_cpp_torch_options(
          + python_include_dirs
          + torch_include_dirs
          + omp_include_dir_paths
 +        + [sys.prefix + '/include']
++        + [sys.prefix + '/targets/@CUDA_TARGET@/include']
      )
      cflags = sys_libs_cflags + omp_cflags
      ldflags = omp_ldflags