Re-enable test_compiled_autograd_attribution on s390x

It looks like building python_call.cpp with -O3 triggers
a bug in gcc-14.

As a workaround, ignore offending warning in the code.

GCC bug reference: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115016

.github/workflows/_linux-test.yml

@@ -327,6 +327,7 @@ jobs:
           SCCACHE_REGION: ${{ !contains(matrix.runner, 'b200') && 'us-east-1' || '' }}
           SHM_SIZE: ${{ contains(inputs.build-environment, 'cuda') && '2g' || '1g' }}
           DOCKER_IMAGE: ${{ steps.calculate-docker-image.outputs.docker-image }}
+          DOCKER_IMAGE_S390X: ${{ inputs.docker-image }}
           XLA_CUDA: ${{ contains(inputs.build-environment, 'xla') && '0' || '' }}
           XLA_CLANG_CACHE_S3_BUCKET_NAME: ossci-compiler-clang-cache-circleci-xla
           PYTORCH_TEST_CUDA_MEM_LEAK_CHECK: ${{ matrix.mem_leak_check && '1' || '0' }}
@@ -360,10 +361,12 @@ jobs:
             # if for some reason cleanup action doesn't stop container
             # when job is cancelled
             DOCKER_SHELL_CMD="sleep 12h"
+            USED_IMAGE="${DOCKER_IMAGE_S390X}"
           else
             SHM_OPTS="--shm-size=${SHM_SIZE}"
             JENKINS_USER="--user jenkins"
             DOCKER_SHELL_CMD=
+            USED_IMAGE="${DOCKER_IMAGE}"
           fi
 
           # detached container should get cleaned up by teardown_ec2_linux
@@ -426,7 +429,7 @@ jobs:
             ${JENKINS_USER} \
             -v "${GITHUB_WORKSPACE}:/var/lib/jenkins/workspace" \
             -w /var/lib/jenkins/workspace \
-            "${DOCKER_IMAGE}" \
+            "${USED_IMAGE}" \
             ${DOCKER_SHELL_CMD}
           )
           echo "DOCKER_CONTAINER_ID=${container_name}" >> "${GITHUB_ENV}"

test/dynamo/test_structured_trace.py

@@ -1017,7 +1017,6 @@ def forward(self, x_1: "f32[2][1]cpu"):
         logs = self.buffer.getvalue()
         self.assertTrue(all(event in logs for event in chromium_events))
 
-    @xfailIfS390X
     @requires_tlparse
     @torch._dynamo.config.patch("compiled_autograd", True)
     def test_compiled_autograd_attribution(self):

test/inductor/test_torchinductor.py

@@ -2169,7 +2169,6 @@ class CommonTemplate:
 
     @skipCPUIf(IS_MACOS, "fails on macos")
     @skip_if_halide  # accuracy 4.7% off
-    @xfailIfS390X  # accuracy failure
     def test_multilayer_var_lowp(self):
         def fn(a):
             return torch.var(a)

torch/csrc/distributed/rpc/python_call.cpp

@@ -6,6 +6,12 @@ PythonCall::PythonCall(SerializedPyObj&& serializedPyObj, bool isAsyncExecution)
     : serializedPyObj_(std::move(serializedPyObj)),
       isAsyncExecution_(isAsyncExecution) {}
 
+#if defined(__GNUC__) && __GNUC__ == 14
+/* this warning is falsely triggered with gcc-14 in following function. */
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wfree-nonheap-object"
+#endif
+
 c10::intrusive_ptr<Message> PythonCall::toMessageImpl() && {
   std::vector<char> payload;
   payload.reserve(serializedPyObj_.payload_.length() + 1);
@@ -21,6 +27,10 @@ c10::intrusive_ptr<Message> PythonCall::toMessageImpl() && {
       MessageType::PYTHON_CALL);
 }
 
+#if defined(__GNUC__) && __GNUC__ == 14
+#pragma GCC diagnostic pop
+#endif
+
 std::unique_ptr<PythonCall> PythonCall::fromMessage(const Message& message) {
   TORCH_INTERNAL_ASSERT(
       !message.payload().empty(),