[CI] Disable some tests that are failing in periodic (#150059)

Disabling some tests to restore periodic

nogpu avx512 timeout:
59f14d19ae (38492953496-box)

profiler failure: 7ae0ce6360 (38461255009-box)

test_accelerator failure:
87bfd66c3c (39476723746-box)
origin: 146098

test_overrides failure:
bf752c36da (39484562957-box)
origin: 146098

inductor cpu repro:
bb9c426024 (38447525659-box)

functorch eager transforms:
8f858e226b (39488068620-box)
f2cea01f71 (39555064878)
b5281a4a18 (39599355600)
either 148288 or 148261?

2ec9aceaeb/1

Pull Request resolved: https://github.com/pytorch/pytorch/pull/150059
Approved by: https://github.com/ZainRizvi, https://github.com/atalman, https://github.com/malfet

.github/workflows/periodic.yml

@@ -59,8 +59,9 @@ jobs:
       docker-image-name: pytorch-linux-focal-cuda12.6-cudnn9-py3-gcc11
       test-matrix: |
         { include: [
-          { config: "nogpu_AVX512", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
-          { config: "nogpu_AVX512", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
+          { config: "nogpu_AVX512", shard: 1, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
+          { config: "nogpu_AVX512", shard: 2, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
+          { config: "nogpu_AVX512", shard: 3, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
           { config: "nogpu_NO_AVX2", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
           { config: "nogpu_NO_AVX2", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
           { config: "jit_legacy", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge.nvidia.gpu" },

test/functorch/test_eager_transforms.py

@@ -74,6 +74,7 @@ from torch.testing._internal.common_utils import (
     skipIfRocm,
     skipIfTorchDynamo,
     subtest,
+    TEST_CUDA_MEM_LEAK_CHECK,
     TEST_WITH_TORCHDYNAMO,
     TestCase,
     xfailIfTorchDynamo,
@@ -2865,6 +2866,10 @@ class TestLinearize(TestCase):
         self.assertEqual(actual_jvp, expected_jvp)
 
     @dtypes(torch.float)
+    @unittest.skipIf(
+        TEST_CUDA_MEM_LEAK_CHECK,
+        "Leaking memory, see https://github.com/pytorch/pytorch/pull/150059 for example",
+    )
     def test_linearize_return(self, device, dtype):
         x_p = make_tensor((3, 1), device=device, dtype=dtype)
         x_t = make_tensor((3, 1), device=device, dtype=dtype)
@@ -2879,6 +2884,10 @@ class TestLinearize(TestCase):
         self.assertEqual(actual_jvp, expected_jvp)
 
     @dtypes(torch.float)
+    @unittest.skipIf(
+        TEST_CUDA_MEM_LEAK_CHECK,
+        "Leaking memory, see https://github.com/pytorch/pytorch/pull/150059 for example",
+    )
     def test_linearize_composition_vmap(self, device, dtype):
         x_p = make_tensor((3, 1), device=device, dtype=dtype)
         x_t = make_tensor((3, 3, 1), device=device, dtype=dtype)
@@ -2897,6 +2906,10 @@ class TestLinearize(TestCase):
         self.assertEqual(actual_batched_jvp, expected_batched_jvp)
 
     @dtypes(torch.float)
+    @unittest.skipIf(
+        TEST_CUDA_MEM_LEAK_CHECK,
+        "Leaking memory, see https://github.com/pytorch/pytorch/pull/150059 for example",
+    )
     def test_linearize_composition_grad(self, device, dtype):
         x_p = make_tensor((3,), device=device, dtype=dtype)
         x_t = make_tensor((3,), device=device, dtype=dtype)
@@ -2916,6 +2929,10 @@ class TestLinearize(TestCase):
         self.assertEqual(actual_batched_jvp, expected_batched_jvp)
 
     @dtypes(torch.float)
+    @unittest.skipIf(
+        TEST_CUDA_MEM_LEAK_CHECK,
+        "Leaking memory, see https://github.com/pytorch/pytorch/pull/150059 for example",
+    )
     def test_linearize_nested_input_nested_output(self, device, dtype):
         x_p = make_tensor((3, 1), device=device, dtype=dtype)
         x_t = make_tensor((3, 1), device=device, dtype=dtype)
@@ -5151,6 +5168,10 @@ class TestCompileTransforms(TestCase):
     # torch.compile is not supported on Windows CUDA.
     # Triton only supports GPU with SM70 or later.
     @expectedFailureIf((IS_WINDOWS and TEST_CUDA) or (TEST_CUDA and not SM70OrLater))
+    @unittest.skipIf(
+        TEST_CUDA_MEM_LEAK_CHECK,
+        "Leaking memory, see https://github.com/pytorch/pytorch/pull/150059 for example",
+    )
     def test_compile_vmap_hessian(self, device):
         # The model and inputs are a smaller version
         # of code at benchmark repo:

test/inductor/test_cpu_repro.py

@@ -4149,6 +4149,10 @@ class CPUReproTests(TestCase):
                         "__at_align__ std::array", 0, exactly=True
                     ).run(code)
 
+    @unittest.skipIf(
+        os.getenv("ATEN_CPU_CAPABILITY") == "default",
+        "Failing in periodic nogpu_NO_AVX2, see #150059 for example",
+    )
     def test_group_norm_large_input(self):
         class M(torch.nn.Module):
             def __init__(self) -> None:
@@ -4178,6 +4182,10 @@ class CPUReproTests(TestCase):
                 # check for parallel reduction.
                 self.assertEqual(metrics.parallel_reduction_count, 1)
 
+    @unittest.skipIf(
+        os.getenv("ATEN_CPU_CAPABILITY") == "default",
+        "Failing in periodic nogpu_NO_AVX2, see #150059 for example",
+    )
     def test_group_norm_large_size(self):
         # https://github.com/pytorch/pytorch/issues/141541
         # We are using the chunk size of 4096 for cascade summation,

test/profiler/test_profiler.py

@@ -2169,6 +2169,10 @@ assert KinetoStepTracker.current_step() == initial_step + 2 * niters
     @skipIfTorchDynamo("profiler gets ignored if dynamo activated")
     @unittest.skipIf(not torch.cuda.is_available(), "CUDA is required")
     @unittest.skipIf(not kineto_available(), "Kineto is required")
+    @unittest.skipIf(
+        "RelWithAssert" in torch.__config__.show(),
+        "failing in debug build, see https://github.com/pytorch/pytorch/pull/150059 for example",
+    )
     def test_profile_all_threads(self):
         profiling_started = threading.Event()
         profiling_ended = threading.Event()

test/test_accelerator.py

@@ -10,6 +10,8 @@ from torch.testing._internal.common_utils import NoTest, run_tests, TEST_MPS, Te
 if not torch.accelerator.is_available():
     print("No available accelerator detected, skipping tests", file=sys.stderr)
     TestCase = NoTest  # noqa: F811
+    # Skip because failing when run on cuda build with no GPU, see #150059 for example
+    sys.exit()
 
 TEST_MULTIACCELERATOR = torch.accelerator.device_count() > 1
 

test/test_overrides.py

@@ -1,5 +1,6 @@
 # Owner(s): ["module: __torch_function__"]
 
+import sys
 import torch
 import numpy as np
 import inspect
@@ -9,6 +10,7 @@ import pickle
 import collections
 import unittest
 import contextlib
+import os
 
 from torch.testing._internal.common_utils import TestCase, run_tests, TEST_WITH_CROSSREF, TEST_WITH_TORCHDYNAMO
 from torch.overrides import (
@@ -29,6 +31,14 @@ from torch.utils._pytree import tree_map
 
 Tensor = torch.Tensor
 
+if os.getenv("ATEN_CPU_CAPABILITY") in ("default", "avx2"):
+    # This test is not supported on ARM
+    print(
+        "Skipping due to failing when cuda build runs on non cuda machine, "
+        + "see https://github.com/pytorch/pytorch/pull/150059 for example"
+    )
+    sys.exit()
+
 # The functions below simulate the pure-python torch functions in the
 # torch.functional namespace. We use examples local to this file rather
 # than any of the real examples implemented in Python since in the