python3Packages.torch: fix on aarch64-linux (#439489)

      url = "https://github.com/pytorch/pytorch/commit/231c72240d80091f099c95e326d3600cba866eee.patch";

      hash = "sha256-BBCjxzz2TUkx4nXRyRILA82kMwyb/4+C3eOtYqf5dhk=";

    })

    # Fixes GCC-14 compatibility on ARM

    # Adapted from https://github.com/pytorch/pytorch/pull/157867

    # TODO: remove at the next release

    ./gcc-14-arm-compat.path

  ]

  ++ lib.optionals cudaSupport [

    ./fix-cmake-cuda-toolkit.patch

diff --git a/aten/src/ATen/cpu/vec/sve/vec_bfloat16.h b/aten/src/ATen/cpu/vec/sve/vec_bfloat16.h

index 7f05c2ad166..1632b595c4c 100644

--- a/aten/src/ATen/cpu/vec/sve/vec_bfloat16.h

+++ b/aten/src/ATen/cpu/vec/sve/vec_bfloat16.h

@@ -220,8 +220,12 @@ class Vectorized<BFloat16> {

   Vectorized<BFloat16> le(const Vectorized<BFloat16>& other) const;

 };

-inline std::tuple<Vectorized<float>, Vectorized<float>> convert_bfloat16_float(

-    const Vectorized<c10::BFloat16>& a) {

+#if defined(__GNUC__) && __GNUC__ == 14

+// Workaround for gcc-14.2.0 ICE during RTL pass: vregs when compiling for SVE

+__attribute__((optimize("no-tree-vectorize")))

+#endif

+inline std::tuple<Vectorized<float>, Vectorized<float>>

+convert_bfloat16_float(const Vectorized<c10::BFloat16>& a) {

   static_assert(

       Vectorized<c10::BFloat16>::size() == 2 * Vectorized<float>::size());

   auto zero = svreinterpret_bf16_f32(svdup_n_f32(0.0f));

diff --git a/aten/src/ATen/native/cpu/Activation.cpp b/aten/src/ATen/native/cpu/Activation.cpp

index 52d5383e60f..00c9f4eb253 100644

--- a/aten/src/ATen/native/cpu/Activation.cpp

+++ b/aten/src/ATen/native/cpu/Activation.cpp

@@ -26,6 +26,10 @@ namespace at::native {

 namespace {

+#if defined(__GNUC__) && __GNUC__ == 14 && defined(__aarch64__) && !defined(__ARM_FEATURE_SVE)

+// Workaround for gcc-14.2.0 ICE during RTL pass: expand when compiling for NEON

+__attribute__((optimize("no-tree-vectorize")))

+#endif

 static void log_sigmoid_cpu_kernel(TensorBase &output, TensorBase &buffer, const TensorBase &input) {

   if (at::isReducedFloatingType(input.scalar_type())) {

     AT_DISPATCH_REDUCED_FLOATING_TYPES(input.scalar_type(), "log_sigmoid_cpu", [&]() {

diff --git a/aten/src/ATen/native/cpu/Unfold2d.cpp b/aten/src/ATen/native/cpu/Unfold2d.cpp

index 8ef0741e77a..8c94decfff0 100644

--- a/aten/src/ATen/native/cpu/Unfold2d.cpp

+++ b/aten/src/ATen/native/cpu/Unfold2d.cpp

@@ -169,6 +169,10 @@ static void unfolded2d_acc_channels_last(

 /* note: due to write issues, this one cannot be parallelized as well as

  * unfolded2d_copy */

+#if defined(__GNUC__) && __GNUC__ == 14 && defined(__ARM_FEATURE_SVE) && !defined(__ARM_FEATURE_BF16)

+// Workaround for gcc-14.2.0 ICE during RTL pass: vregs when compiling for SVE without BF16

+__attribute__((optimize("no-tree-vectorize")))

+#endif

 void unfolded2d_acc_kernel(

     ScalarType dtype,

     void *finput_data,