AArch64 xFail for failing tests in inductor/test_cpu_repro (#171095)

robert-hardwick · pytorchmergebot · commit a79095985ccd · 2026-03-09T13:15:40.000Z
This PR enables inductor/test_cpu_repro on AArch64 and marks known failures with xfail and link to issue. Pull Request resolved: #171095 Approved by: https://github.com/jgong5, https://github.com/Skylion007, https://github.com/malfet
diff --git a/.ci/pytorch/test.sh b/.ci/pytorch/test.sh
@@ -1785,7 +1785,7 @@ test_linux_aarch64() {
        inductor/test_split_cat_fx_passes inductor/test_compile inductor/test_torchinductor \
        inductor/test_torchinductor_codegen_dynamic_shapes inductor/test_torchinductor_dynamic_shapes inductor/test_memory \
        inductor/test_triton_cpu_backend inductor/test_triton_extension_backend inductor/test_mkldnn_pattern_matcher inductor/test_cpu_cpp_wrapper \
-       inductor/test_cpu_select_algorithm \
+       inductor/test_cpu_select_algorithm inductor/test_cpu_repro \
        --shard "$SHARD_NUMBER" "$NUM_TEST_SHARDS" --verbose
 }
 
diff --git a/test/inductor/test_cpu_repro.py b/test/inductor/test_cpu_repro.py
@@ -32,13 +32,17 @@
 from torch.nn import functional as F
 from torch.testing._internal.common_utils import (
     instantiate_parametrized_tests,
+    IS_ARM64,
+    IS_CPU_CAPABILITY_SVE256,
+    IS_CPU_EXT_SVE_SUPPORTED,
     IS_FBCODE,
     IS_MACOS,
     MI200_ARCH,
     parametrize,
     skipIfRocmArch,
     slowTest,
     TEST_MKL,
+    xfailIf,
     xfailIfS390X,
 )
 from torch.utils._python_dispatch import TorchDispatchMode
@@ -3469,6 +3473,8 @@ def fn(x):
                 3,
             )
 
+    @xfailIf(IS_ARM64 and not IS_CPU_CAPABILITY_SVE256)
+    # see https://github.com/pytorch/pytorch/issues/142231
     @config.patch({"fx_graph_cache": False, "fx_graph_remote_cache": False})
     def test_two_local_buffers_in_outer_loop_fusion(self):
         def fn(x):
@@ -3807,6 +3813,8 @@ def forward(self, x):
                     self.common(m, (x,))
                     check_metrics_vec_kernel_count(8)
 
+    @xfailIf(IS_ARM64 and IS_CPU_CAPABILITY_SVE256)
+    # see https://github.com/pytorch/pytorch/issues/169958
     @requires_vectorization
     @config.patch("cpp.enable_tiling_heuristics", False)
     def test_transpose_copy(self):
@@ -4054,6 +4062,8 @@ def fn(a, b):
         self.common(fn, (x, y))
         check_metrics_vec_kernel_count(2)
 
+    @xfailIf(IS_ARM64 and IS_CPU_CAPABILITY_SVE256)
+    # see https://github.com/pytorch/pytorch/issues/170877
     def test_transpose_mxn_16_16_bf16_fp16(self):
         def fn(a, b):
             c = a * b
@@ -4698,6 +4708,7 @@ def fn(x, y):
         y = torch.randint(0, 255, (3, 3), dtype=torch.uint8)
         self.common(fn, (x, y))
 
+    @xfailIf(IS_ARM64)  # see https://github.com/pytorch/pytorch/issues/168972
     def test_float32_to_uint8(self):
         # https://github.com/pytorch/pytorch/issues/156788
         @torch.compile
@@ -5393,6 +5404,9 @@ def forward(self, x):
         x = torch.randn(1, 4, 2, 2)
         self.common(fn, (x,))
 
+    @xfailIf(
+        IS_ARM64 and IS_CPU_EXT_SVE_SUPPORTED
+    )  # see https://github.com/pytorch/pytorch/issues/142134
     @parametrize("is_inference", (True, False))
     def test_disabled_amp(self, is_inference):
         class M(torch.nn.Module):
@@ -6096,6 +6110,7 @@ def fn(x):
 
         torch.compile(fn)(torch.randn(2, 2))
 
+    @xfailIf(IS_ARM64)  # https://github.com/pytorch/pytorch/issues/176285
     @skipIfRocmArch(MI200_ARCH)
     @unittest.skipIf(not torch.backends.mkldnn.is_available(), "MKLDNN is not enabled")
     @requires_vectorization
diff --git a/torch/testing/_internal/common_utils.py b/torch/testing/_internal/common_utils.py
@@ -1401,22 +1401,24 @@ def printErrors(self) -> None:
     else:
         unittest.main(argv=argv)
 
+def cpu_supports_feature(feature):
+    if sys.platform != 'linux':
+        return False
+    with open("/proc/cpuinfo") as f:
+        flags = {t for k, v in (line.split(":", 1) for line in f if ":" in line)
+                 if k.strip().lower() in ("flags", "features") for t in v.split()}
+    return feature.lower() in flags
+
 IS_LINUX = sys.platform == "linux"
 IS_WINDOWS = sys.platform == "win32"
 IS_MACOS = sys.platform == "darwin"
 IS_PPC = platform.machine() == "ppc64le"
 IS_X86 = platform.machine() in ('x86_64', 'i386')
 IS_ARM64 = platform.machine() in ('arm64', 'aarch64')
 IS_S390X = platform.machine() == "s390x"
-
-def is_avx512_vnni_supported():
-    if sys.platform != 'linux':
-        return False
-    with open("/proc/cpuinfo", encoding="ascii") as f:
-        lines = f.read()
-    return "vnni" in lines
-
-IS_AVX512_VNNI_SUPPORTED = is_avx512_vnni_supported()
+IS_AVX512_VNNI_SUPPORTED = cpu_supports_feature("vnni")
+IS_CPU_EXT_SVE_SUPPORTED = cpu_supports_feature("sve")
+IS_CPU_CAPABILITY_SVE256 = torch._C._get_cpu_capability() == "SVE256"
 
 if IS_WINDOWS:
     @contextmanager

Original file line number	Diff line number	Diff line change
`@@ -1785,7 +1785,7 @@ test_linux_aarch64() {`
`1785`	`1785`	`inductor/test_split_cat_fx_passes inductor/test_compile inductor/test_torchinductor \`
`1786`	`1786`	`inductor/test_torchinductor_codegen_dynamic_shapes inductor/test_torchinductor_dynamic_shapes inductor/test_memory \`
`1787`	`1787`	`inductor/test_triton_cpu_backend inductor/test_triton_extension_backend inductor/test_mkldnn_pattern_matcher inductor/test_cpu_cpp_wrapper \`
`1788`		`- inductor/test_cpu_select_algorithm \`
	`1788`	`+ inductor/test_cpu_select_algorithm inductor/test_cpu_repro \`
`1789`	`1789`	`--shard "$SHARD_NUMBER" "$NUM_TEST_SHARDS" --verbose`
`1790`	`1790`	`}`
`1791`	`1791`