Run test_torchinductor_opinfo CPU tests if triton not installed (pyto…

…rch#88934)

These test are not run currently because normal CI workers don't have
triton installed.

Pull Request resolved: pytorch#88934
Approved by: https://github.com/ngimel

test/inductor/test_torchinductor.py

@@ -20,7 +20,6 @@
 from torch.fx.experimental.proxy_tensor import make_fx
 from torch.nn import functional as F
 from torch.testing._internal.common_utils import (
-    IS_FBCODE,
     TEST_WITH_ASAN,
     TEST_WITH_ROCM,
     TestCase as TorchTestCase,
@@ -41,7 +40,7 @@
     from torch._inductor.compile_fx import compile_fx, complex_memory_overlap
     from torch._inductor.ir import IndexingDiv, ModularIndexing
     from torch._inductor.sizevars import SizeVarAllocator
-    from torch._inductor.utils import has_torchvision_roi_align, has_triton, timed
+    from torch._inductor.utils import has_torchvision_roi_align, timed
 
     # This will only pass on pytorch builds newer than roughly 5/15/2022
     assert get_decompositions([torch.ops.aten.trace])
@@ -53,25 +52,10 @@
         sys.exit(0)
     raise unittest.SkipTest("requires sympy/functorch/filelock")
 
-HAS_CPU = False
-try:
-    from subprocess import CalledProcessError
-
-    from torch._inductor.codecache import CppCodeCache
-
-    CppCodeCache.load("")
-    HAS_CPU = not IS_FBCODE
-except (
-    CalledProcessError,
-    OSError,
-    torch._inductor.exc.InvalidCxxCompiler,
-    torch._inductor.exc.CppCompileError,
-):
-    pass
+from torch.testing._internal.inductor_utils import HAS_CPU, HAS_CUDA
 
 aten = torch.ops.aten
 
-HAS_CUDA = has_triton()
 requires_cuda = functools.partial(unittest.skipIf, not HAS_CUDA, "requires cuda")
 
 torch._inductor.config.triton.autotune = False  # too slow

test/inductor/test_torchinductor_opinfo.py

@@ -16,20 +16,22 @@
     onlyNativeDeviceTypes,
     OpDTypes,
     ops,
+    skipCPUIf,
+    skipCUDAIf,
 )
 from torch.testing._internal.common_methods_invocations import op_db
 from torch.testing._internal.common_utils import (
     dtype_abbrs,
     run_tests,
     skipCUDAMemoryLeakCheckIf,
+    skipIfCrossRef,
+    skipIfTorchDynamo,
     suppress_warnings,
-    TEST_WITH_ROCM,
     TestCase,
 )
+from torch.testing._internal.inductor_utils import HAS_CPU, HAS_CUDA
 
 try:
-    from torch._inductor.utils import has_triton
-
     try:
         from .test_torchinductor import check_model, check_model_cuda
     except ImportError:
@@ -120,6 +122,7 @@ def process(device_type):
 
 inductor_skips["cpu"] = {
     "linalg.ldl_solve": {b8, f16, f32, f64, i32, i64},  # segfault
+    "linalg.ldl_factor": {f32, f64},  # flaky
     "__rdiv__": {b8, f16, f32, f64, i32, i64},  # flaky
 }
 
@@ -169,6 +172,8 @@ def process(device_type):
     "argwhere": {b8, f16, f32, f64, i32, i64},
     "bernoulli": {f32, f64},
     "bincount": {i32, i64},
+    "cdouble": {b8, f16, f32, f64, i32, i64},
+    "cfloat": {b8, f16, f32, f64, i32, i64},
     "chalf": {b8, f16, f32, f64, i32, i64},
     "cholesky": {f32, f64},
     "combinations": {b8, f16, f32, f64, i32, i64},
@@ -209,11 +214,10 @@ def process(device_type):
     "linalg.lstsq.grad_oriented": {f32, f64},
     "linalg.matrix_rank": {f32, f64},
     "linalg.matrix_rank.hermitian": {f32, f64},
-    "linalg.lu_solve": {f32, f64},
-    "lu_solve": {f32, f64},
-    "lu_unpack": {f32, f64},
+    "linalg.pinv.singular": {f32, f64},
     "logdet": {f32, f64},
     "masked.norm": {f16},
+    "masked.normalize": {f16},
     "masked_fill": {f16},
     "masked_scatter": {f16, f32, f64},
     "masked_select": {b8, f16, f32, f64, i32, i64},
@@ -225,8 +229,8 @@ def process(device_type):
     "nan_to_num": {f16},
     "nanquantile": {f32, f64},
     "nn.functional.avg_pool1d": {i64},
-    "nn.functional.avg_pool2d": {i64},
-    "nn.functional.adaptive_avg_pool2d": {f16},
+    "nn.functional.avg_pool2d": {i64, f64},
+    "nn.functional.adaptive_avg_pool2d": {f16, f64},
     "nn.functional.ctc_loss": {f32, f64},
     "nn.functional.gaussian_nll_loss": {f32, f64},
     "nn.functional.gelu": {f64},
@@ -243,6 +247,7 @@ def process(device_type):
     "quantile": {f32, f64},
     "rand_like": {f16, f32, f64},
     "randint_like": {f16, f32, f64, i32, i64},
+    "randint": {f16, f32, f64, i32, i64},
     "randn_like": {f16, f32, f64},
     "repeat_interleave": {b8, f16, f32, f64, i32, i64},
     "scatter_add": {f16},
@@ -455,6 +460,10 @@ class TestInductorOpInfo(TestCase):
     @skipCUDAMemoryLeakCheckIf(
         True
     )  # inductor kernels failing this test intermittently
+    @skipCUDAIf(not HAS_CUDA, "Skipped! Triton not found")
+    @skipCPUIf(not HAS_CPU, "Skipped! Supported CPU compiler not found")
+    @skipIfTorchDynamo("Test uses dynamo already")
+    @skipIfCrossRef
     @_ops(op_db[START:END])
     @patch("torch._dynamo.config.raise_on_unsafe_aot_autograd", True)
     def test_comprehensive(self, device, dtype, op):
@@ -599,5 +608,4 @@ def fn(*args, **kwargs):
 instantiate_device_type_tests(TestInductorOpInfo, globals())
 
 if __name__ == "__main__":
-    if has_triton() and not TEST_WITH_ROCM:
-        run_tests()
+    run_tests()

torch/testing/_internal/inductor_utils.py

@@ -0,0 +1,23 @@
+from subprocess import CalledProcessError
+
+from torch._inductor.codecache import CppCodeCache
+from torch._inductor.utils import has_triton
+from torch.testing._internal.common_utils import (
+    IS_FBCODE,
+    TEST_WITH_ROCM,
+)
+import torch
+
+HAS_CPU = False
+try:
+    CppCodeCache.load("")
+    HAS_CPU = not IS_FBCODE
+except (
+    CalledProcessError,
+    OSError,
+    torch._inductor.exc.InvalidCxxCompiler,
+    torch._inductor.exc.CppCompileError,
+):
+    pass
+
+HAS_CUDA = has_triton() and not TEST_WITH_ROCM