# Owner(s): ["module: primTorch"] import torch import os from enum import Enum from torch.overrides import resolve_name from torch.utils._pytree import tree_map, tree_flatten, tree_unflatten from torch._subclasses.meta_utils import MetaConverter import torch.utils._python_dispatch from torch.testing._internal.common_utils import ( TestCase, skipIfCrossRef, suppress_warnings, TEST_WITH_ASAN, run_tests, skipIfSlowGradcheckEnv, dtype_abbrs ) from torch.testing._internal.common_device_type import ( ops, instantiate_device_type_tests, ) from torch.testing._internal.common_methods_invocations import op_db from torchgen.utils import YamlLoader from torchgen.model import OperatorName import sys import yaml import atexit import re from collections import defaultdict import unittest import warnings import weakref bf16 = torch.bfloat16 f64 = torch.float64 f32 = torch.float32 f16 = torch.float16 c32 = torch.complex32 c64 = torch.complex64 c128 = torch.complex128 i8 = torch.int8 i16 = torch.int16 i32 = torch.int32 i64 = torch.int64 b8 = torch.bool u8 = torch.uint8 @skipIfSlowGradcheckEnv class TestMetaConverter(TestCase): def assertSameVersionCounter(self, m1, m2): # Cannot easily test m1 and m2 have same storage due to # lack of Storage bindings. Use version counter. vc = m1._version self.assertEqual(m2._version, vc) # Doing it this way ensures that we get VC bump even with leaves with torch.no_grad(): m1._base.add_(3) self.assertNotEqual(m1._version, vc) self.assertEqual(m2._version, m1._version) def test_view_of_non_leaf(self): x = torch.randn(4, requires_grad=True) y = x.neg() z1 = y[:] z2 = y[:] to_meta = MetaConverter() m1 = to_meta(z1) m2 = to_meta(z2) self.assertEqual(m1.shape, z1.shape) self.assertTrue(m1._is_view()) self.assertFalse(m1._base.is_leaf) self.assertSameVersionCounter(m1, m2) def test_view_of_leaf(self): x = torch.randn(4, requires_grad=True) z1 = x[:] z2 = x[:] to_meta = MetaConverter() m1 = to_meta(z1) m2 = to_meta(z2) self.assertEqual(m1.shape, z1.shape) self.assertTrue(m1._is_view()) self.assertTrue(m1._base.is_leaf) self.assertSameVersionCounter(m1, m2) def test_leaf(self): x = torch.randn(4, requires_grad=True) to_meta = MetaConverter() m = to_meta(x) self.assertEqual(m.shape, x.shape) self.assertTrue(m.is_leaf) self.assertTrue(m.requires_grad) def test_non_leaf(self): x = torch.randn(4, requires_grad=True) y = x.neg() to_meta = MetaConverter() m = to_meta(y) self.assertEqual(m.shape, y.shape) self.assertFalse(m.is_leaf) self.assertTrue(m.requires_grad) def test_requires_grad_false(self): x = torch.randn(4, requires_grad=False) to_meta = MetaConverter() m = to_meta(x) self.assertEqual(m.shape, x.shape) self.assertFalse(m.requires_grad) # NB: complex stuff is not actually exercised right now because # we have a blanket exclusion for complex conversion def test_view_as_real(self): x = torch.randn(4, dtype=torch.complex64) y = torch.view_as_real(x) m = MetaConverter()(y) self.assertEqual(m.shape, y.shape) self.assertEqual(m.stride(), y.stride()) self.assertEqual(m.dtype, y.dtype) def test_complex_noncontiguous_bug(self): x = torch.randn((2, 2, 4, 9), dtype=torch.complex32)[:, 0, :, :] m = MetaConverter()(x) self.assertEqual(m.shape, x.shape) self.assertEqual(m.stride(), x.stride()) self.assertEqual(m.dtype, x.dtype) def test_view_as_complex(self): x = torch.randn((4, 2), dtype=torch.float32) y = torch.view_as_complex(x) m = MetaConverter()(y) self.assertEqual(m.shape, y.shape) self.assertEqual(m.stride(), y.stride()) self.assertEqual(m.dtype, y.dtype) def test_view_dtype(self): x = torch.randn(4, dtype=torch.float32) y = x.view(dtype=torch.int32) m = MetaConverter()(y) self.assertEqual(m.shape, y.shape) self.assertEqual(m.stride(), y.stride()) self.assertEqual(m.dtype, y.dtype) def test_imag(self): x = torch.randn(4, dtype=torch.complex64) y = x.imag m = MetaConverter()(y) self.assertEqual(m.shape, y.shape) self.assertEqual(m.dtype, y.dtype) self.assertEqual(m.stride(), y.stride()) self.assertEqual(m.storage_offset(), y.storage_offset()) def test_weakref(self): x = torch.randn(4, 4, 4) m = MetaConverter() y = m(x) z = m(x) self.assertIs(y, z) self.assertEqual(len(m.tensor_memo), 1) self.assertEqual(len(m.storage_memo), 1) del x self.assertEqual(len(m.tensor_memo), 0) m.check_for_expired_weak_storages() self.assertEqual(len(m.storage_memo), 0) li = [] for i in range(4): li.append(torch.rand([i])) m(li[-1]) self.assertEqual(len(m.tensor_memo), 4) del li self.assertEqual(len(m.tensor_memo), 0) m.check_for_expired_weak_storages() self.assertEqual(len(m.storage_memo), 0) def test_tensor_outlives_converter(self): m = MetaConverter() ref = weakref.ref(m) x = torch.randn([4, 4]) y = m(x) del m self.assertIs(ref(), None) CHECK_STRIDES = { torch.Tensor.__getitem__, } def should_check_strides(func): if func in CHECK_STRIDES: return True if not isinstance(func, torch._ops.OpOverload): return False # Prims are expected to model strides correctly if func.namespace == "prims": return True # Check if it's a view, by testing if any of the returns have # a non-empty alias set if any(r.alias_info.before_set for r in func._schema.returns if r.alias_info): return True # TODO: check for TensorIterator return False def assert_ref_meta_equal(test_case, func, meta_rs, rs, msg_callable): flat_meta_rs, _ = tree_flatten(meta_rs) flat_rs, _ = tree_flatten(rs) test_case.assertEqual(len(flat_meta_rs), len(flat_rs)) for i, meta_r, r in zip(range(len(flat_rs)), flat_meta_rs, flat_rs): def test_assert(cond, msg): if not cond: raise RuntimeError(f"output {i}: {msg_callable(msg)}") if not isinstance(r, torch.Tensor): continue test_assert(isinstance(meta_r, torch.Tensor), f"but real {i}th result is Tensor") test_assert(meta_r.dtype == r.dtype, f"but real dtype was {r.dtype}") test_assert(meta_r.shape == r.shape, f"but real shape was {r.shape}") # See https://github.com/pytorch/pytorch/issues/78050 if should_check_strides(func): same_strides, _ = torch._prims_common.check_significant_strides(meta_r, r) test_assert(same_strides, f"but real stride was {r.stride()}") test_assert( meta_r.storage_offset() == r.storage_offset(), f"but real storage_offset was {r.storage_offset()}") test_assert(meta_r.requires_grad == r.requires_grad, f"but real requires_grad was {r.requires_grad}") test_assert(meta_r.is_conj() == r.is_conj(), f"but real is_conj was {r.is_conj()}") test_assert(meta_r.is_neg() == r.is_neg(), f"but real is_neg was {r.is_neg()}") # This environment variable controls whether or not we print expected failure # lists at the end of a test suite run. The intended usage looks like this: # # 1. Run `PYTORCH_COLLECT_EXPECT=1 python test/test_meta.py` on a CUDA build # of PyTorch that has LAPACK/MAGMA installed. You can filter `-k test_meta` # or `-k test_dispatch_meta` to only focus on one or another list # 2. Given the printed skip/xfail list, add them to the corresponding lists; # torch.* entries go in meta_function and aten.* entries go in meta_dispatch. # If there are preexisting entries, you need to merge in the entries. # # This is somewhat manual but typically you shouldn't need to do this, unless # you've made a major change (e.g., added a new dtype to PyTorch) and need to # refresh the lists. If you want to do it from scratch, just clear out the # preexisting lists before running. # # WARNING: Python dict literals will silently ignore duplicate keys COLLECT_EXPECT = os.getenv('PYTORCH_COLLECT_EXPECT', '0') == '1' seen_succeeded = {} seen_failed = {} failed_reasons = defaultdict(set) def print_seen(): expected_failures = [] skips = [] def fmt_dtypes(dtypes): r = ', '.join(sorted(dtype_abbrs[d] for d in dtypes)) return '{' + r + '}' for op, failed_dtypes in seen_failed.items(): ops = resolve_name(op) succeeded_dtypes = seen_succeeded.get(op, set()) expected_failures_dtypes = failed_dtypes - succeeded_dtypes skips_dtypes = failed_dtypes & succeeded_dtypes reasons = "" if failed_reasons[op]: reasons = " # " + ", ".join(sorted(failed_reasons[op])) if expected_failures_dtypes: expected_failures.append(f" {ops}: {fmt_dtypes(expected_failures_dtypes)},{reasons}") if skips_dtypes: skips.append(f" {ops}: {fmt_dtypes(skips_dtypes)},") expected_failures.sort() skips.sort() nl = '\n' print(f"""\ expected_failures = {{ {nl.join(expected_failures)} }} skips = {{ {nl.join(skips)} }} """) if COLLECT_EXPECT: atexit.register(print_seen) # Success forces pass; failure forces fail; skip unconditionally skips testing TestExpect = Enum("TestExpect", ("SUCCESS", "XFAILURE", "SKIP")) # unlike print produce strides def verbose_print(e): class Lit: def __init__(self, s): self.s = s def __repr__(self): return self.s def go(t): if isinstance(t, torch.Tensor): return Lit(f"{t} stride={t.stride()}") else: return t return repr(tree_map(go, e)) def run_meta_crossref( test_case, test_expect, func, args, kwargs, *, dtype, device_type, ): to_meta = MetaConverter() do_meta = test_expect is not TestExpect.SKIP if do_meta: try: meta_args = tree_map(to_meta, args) meta_kwargs = tree_map(to_meta, kwargs) except Exception as e: raise RuntimeError( f"failed to convert args to meta; " f"originally (*{args}, **{kwargs})") from e rs = func(*args, **kwargs) # TODO: also handle cases where func raise an exception # For now, only attempt if we managed to convert all tensor types # (if any of them failed, we're in a mixed device situation and # this isn't well supported) if do_meta and to_meta.successful(): # Special cases if func is torch.tensor_split: # Use original indices_or_sections, this argument is data dependent meta_args = (meta_args[0], args[1]) + meta_args[2:] elif func is torch.Tensor.__getitem__: # Ensure boolean tensors use original assert len(args) == 2 flat_args, _ = tree_flatten(args[1]) flat_meta_args, spec = tree_flatten(meta_args[1]) flat_new_args = [] for a, ma in zip(flat_args, flat_meta_args): flat_new_args.append(a if isinstance(a, torch.Tensor) and a.dtype in [torch.int8, torch.bool] else ma) meta_args = (meta_args[0], tree_unflatten(flat_new_args, spec)) elif func is torch.ops.aten.repeat_interleave.Tensor: if kwargs.get("output_size", None) is None: meta_args = args elif func is torch.ops.aten.index.Tensor: # Don't convert boolean tensors to meta as they will have nonzero # called on them indices = [] for meta_index, real_index in zip(meta_args[1], args[1]): if meta_index is not None and meta_index.dtype in [torch.int8, torch.bool]: indices.append(real_index) else: indices.append(meta_index) meta_args = (meta_args[0], indices) if kwargs.get("device", None) is not None: meta_kwargs["device"] = "meta" try: # Suppress warnings, this doesn't matter for test_meta.py # but it does matter if you want to use this decorator # for cross-ref testing, as some tests may be looking at # errors with warnings.catch_warnings(): warnings.simplefilter("ignore") meta_rs = func(*meta_args, **meta_kwargs) except Exception as e: if test_expect is TestExpect.XFAILURE: return rs seen_failed.setdefault(func, set()).add(dtype) if isinstance(e, NotImplementedError): m = RE_NOT_IMPLEMENTED_MSG.search(e.args[0]) if m: failed_reasons[func].add(m.group(1)) if COLLECT_EXPECT: return rs raise RuntimeError(f"""\ failed to run: {resolve_name(func)}( *{verbose_print(meta_args)}, **{verbose_print(meta_kwargs)} )""") from e else: try: delim = ',\n ' assert_ref_meta_equal(test_case, func, meta_rs, rs, lambda msg: f"""\ meta disagrees with real impl: {resolve_name(func)}( {delim.join(map(verbose_print, meta_args))}, {delim.join(k + ": " + verbose_print(v) for k, v in meta_kwargs.items())} ) = ( {verbose_print(meta_rs)} ) {msg} """) except Exception: if test_expect is TestExpect.XFAILURE: return rs seen_failed.setdefault(func, set()).add(dtype) if COLLECT_EXPECT: return rs raise else: seen_succeeded.setdefault(func, set()).add(dtype) if test_expect is TestExpect.XFAILURE and not COLLECT_EXPECT: raise RuntimeError(f"unexpected success {resolve_name(func)}") return rs RE_NOT_IMPLEMENTED_MSG = re.compile(r"Could not run '([^']+)' with arguments ") meta_function_expected_failures = { torch.Tensor.to_sparse : {f64, i32, c128, i64, i16, f16, u8, c64, bf16, b8, i8, f32}, torch.allclose : {f64, f16, c128, c64, bf16, f32}, torch.argwhere : {f64, i32, c128, i64, i16, f16, u8, c64, bf16, b8, i8, f32}, torch.combinations : {f64, i32, c128, i64, i16, f16, u8, c64, bf16, b8, i8, f32}, torch.corrcoef : {f64, i32, c128, i64, i16, u8, c64, bf16, i8, f32}, torch.count_nonzero : {f64, i32, c128, i64, i16, f16, u8, c64, bf16, b8, i8, f32}, torch.cov : {f64, i32, c128, i64, i16, u8, c64, bf16, i8, f32}, torch.functional.istft : {f64, c64, c128, f32}, torch.geqrf : {f64, c64, c128, f32}, torch.linalg.householder_product : {f64, c64, c128, f32}, torch.linalg.solve_triangular : {f64, c64, c128, f32}, torch.masked_select : {f64, i32, c128, i64, i16, f16, u8, c64, bf16, b8, i8, f32}, torch.matrix_exp : {f64, c128, c64, bf16, f32}, torch.nonzero : {f64, i32, c128, i64, i16, c32, f16, u8, c64, bf16, b8, i8, f32}, torch.Tensor.nonzero : {f64, i32, c128, i64, i16, c32, f16, u8, c64, bf16, b8, i8, f32}, torch.ormqr : {f64, c64, c128, f32}, torch.repeat_interleave : {f64, i32, c128, i64, i16, c32, f16, u8, c64, bf16, b8, i8, f32}, torch.take : {f64, i32, c128, i64, i16, f16, u8, c64, bf16, b8, i8, f32}, torch.Tensor.item : {f64, i32, c128, i64, i16, f16, u8, c64, bf16, b8, i8, f32}, torch.bincount : {i32, i64, u8, i16, i8}, torch.bucketize : {f64, i32, i64, f16, u8, i16, bf16, i8, f32}, torch.frexp : {f64, f16, bf16, f32}, torch.functional.unique : {f64, i32, i64, u8, i16, bf16, b8, i8, f32}, torch.functional.unique_consecutive : {f64, i32, i64, u8, i16, bf16, b8, i8, f32}, torch.histc : {f64, bf16, f32}, torch.histogram : {f64, f32}, torch.histogramdd : {f64, f32}, torch.kthvalue : {f64, i32, i64, u8, i16, bf16, i8, f32}, torch.logcumsumexp : {f64, bf16, f32}, torch.median : {f64, i32, i64, u8, i16, bf16, i8, f32}, torch.mode : {f64, i32, i64, f16, u8, i16, bf16, b8, i8, f32}, torch.multinomial : {f64, bf16, f32}, torch.nn.functional.ctc_loss : {f64, f32}, torch.nn.functional.gaussian_nll_loss : {f64, bf16, f32}, torch.nn.functional.max_pool3d : {f64, f32}, torch.nn.functional.max_pool3d_with_indices : {f64, f32}, torch.nn.functional.max_unpool1d : {f64, f32}, torch.nn.functional.max_unpool2d : {f64, f32}, torch.nn.functional.max_unpool3d : {f64, f32}, torch.nn.functional.multi_margin_loss : {f64, f32}, torch.nn.functional.multilabel_margin_loss : {f64, f32}, torch.nn.functional.one_hot : {i64}, torch.nn.functional.pdist : {f64, f32}, torch.nn.functional.rrelu : {f64, bf16, f32}, torch.polar : {f64, f32}, torch.segment_reduce : {f64, f16, bf16, f32}, torch.searchsorted : {f64, i32, i64, f16, u8, i16, bf16, i8, f32}, torch.symeig : {f64, f32, c128, c64}, torch.cholesky : {f64, f32, c128, c64}, torch.cholesky_inverse : {f64, f32, c128, c64}, torch.cholesky_solve : {f64, f32, c128, c64}, torch.linalg.eig : {f64, f32, c128, c64}, torch.linalg.eigvals : {f64, f32, c128, c64}, torch.linalg.lstsq : {f64, f32, c128, c64}, } """ # This is some sample code for how we could dump these dicts into YAML # file for easier reading/writing import yaml print(yaml.dump( {resolve_name(k): [dtype_abbrs[d] for d in v] for k, v in meta_function_expected_failures.items()}, default_flow_style=None)) import sys sys.exit() """ meta_function_skips = { torch.Tensor.__rmatmul__ : {bf16, c128, f64, f32, f16, c64}, torch.Tensor.matmul : {f64, f32, c128, c64}, torch.fft.fft2 : {i8, i64, u8, c128, b8, f64, i16, f32, i32, c64, c32, f16}, torch.fft.fft : {i8, i64, u8, c128, b8, f64, i16, f32, i32, c64, c32, f16}, torch.fft.fftn : {i8, i64, u8, c128, b8, f64, i16, f32, i32, c64, c32, f16}, torch.fft.ifft2 : {i8, i64, u8, c128, b8, f64, i16, f32, i32, c64, c32, f16, c32}, torch.fft.ifft : {c128, c64, c32, f16}, torch.fft.ifftn : {i8, i64, u8, c128, b8, f64, i16, f32, i32, c64, c32, f16}, torch.fft.hfft: {f16}, torch.fft.hfftn: {f16}, torch.fft.hfft2: {f16}, torch.fft.ihfft: {f16}, torch.fft.ihfft2 : {i8, i64, u8, f64, b8, f32, i32, i16, f16, c32, f16}, torch.fft.ihfftn : {i8, i64, u8, f64, b8, f32, i32, i16, c32, f16}, torch.fft.irfft2 : {f16}, torch.fft.irfft : {f16}, torch.fft.irfftn : {f16}, torch.fft.rfft2 : {i8, i64, u8, f64, b8, f32, i32, i16, c32, f16}, torch.fft.rfft : {i8, i64, u8, f64, b8, f32, i32, i16, c32, f16}, torch.fft.rfftn : {i8, i64, u8, f64, b8, f32, i32, i16, c32, f16}, torch.functional.atleast_2d : {bf16, i8, c32, i64, u8, c128, b8, f64, i16, i32, f32, f16, c64}, torch.functional.atleast_3d : {bf16, i8, c32, i64, u8, c128, b8, f64, i16, i32, f32, f16, c64}, torch.functional.cartesian_prod : {bf16, i8, i64, u8, c128, b8, f64, i16, i32, f32, f16, c64}, torch.functional.einsum : {bf16, c128, f64, f32, f16, c64}, torch.functional.stft : {c128, f32, c64, f64}, torch.functional.tensordot : {bf16, i8, i64, u8, c128, f64, i16, f32, i32, c64}, torch.inner : {bf16, i8, i64, u8, c128, f64, i16, f32, i32, c64}, torch.linalg.lu_solve : {c128, c64}, torch.linalg.matrix_norm : {c128, f32, c64, f64}, torch.linalg.matrix_power : {c128, c64}, torch.linalg.matrix_rank : {c128, c64}, torch.linalg.svd : {c128, c64}, torch.matmul : {bf16, c128, f64, f32, f16, c64}, torch.nanquantile : {f64, f32}, torch.narrow : {bf16, i8, i64, u8, c128, b8, f64, i16, i32, f32, f16, c32, c64}, torch.nn.functional.batch_norm : {f64, f32}, torch.nn.functional.binary_cross_entropy : {bf16, f64, f32, f16}, torch.nn.functional.dropout3d : {bf16, f64, f32, f16}, torch.nn.functional.local_response_norm : {bf16, f64, f32, f16}, torch.svd : {c128, c64}, torch.take_along_dim : {bf16, i8, i64, u8, c128, b8, f64, i16, i32, f32, f16, c64}, torch.vstack : {bf16, i8, c32, i64, u8, c128, b8, f64, i16, i32, f32, f16, c64}, torch.aminmax : {i8, i64, u8, f64, b8, f32, i32, i16}, torch.cummax : {bf16, i8, i64, u8, f64, b8, f32, i32, i16}, torch.cummin : {bf16, i8, i64, u8, f64, b8, f32, i32, i16}, torch.diff : {b8}, torch.equal : {bf16, i8, c32, i64, u8, c128, b8, f64, i16, i32, f32, f16, c64}, torch.functional.cdist : {f64, f32}, torch.nanmean : {bf16, f64, f32, f16}, torch.nn.functional.cross_entropy : {bf16, f64, f32}, torch.nn.functional.interpolate : {bf16, f64, f32, u8}, torch.nn.functional.nll_loss : {bf16, f64, f32}, torch.linalg.pinv : {f64, f32}, torch.linalg.cond : {c128, c64, f32, f64}, torch.linalg.vander: {c128, c64, f32, f64, i16, i32, i64, i8, u8}, torch.linalg.vecdot : {bf16, f64, f32, f16}, torch.empty : {bf16, i8, c32, i64, u8, c128, b8, f64, i16, i32, f32, f16, c64}, # This fails for arguments dispatched to grid_sampler_3d, but succeeds # for grid_sampler_2d, so we can't just xfail it torch.nn.functional.grid_sample : {f64, f32}, } meta_function_device_expected_failures = defaultdict(dict) meta_function_device_skips = defaultdict(dict) meta_function_device_expected_failures['cpu'] = { torch.native_batch_norm: {bf16}, torch.native_layer_norm: {bf16}, } meta_function_device_expected_failures['cuda'] = { torch.corrcoef: {bf16, f16}, # aten::_local_scalar_dense torch.cov: {f16}, # aten::_local_scalar_dense torch.functional.unique: {f16}, # aten::_unique2, aten::unique_dim torch.functional.unique_consecutive: {f16}, # aten::unique_consecutive torch.geqrf: {f32, f64}, # aten::geqrf torch.histc: {i16, i32, i64, i8}, # aten::histc, aten::histc.out torch.kthvalue: {f16}, # aten::kthvalue.values torch.linalg.householder_product: {f32, f64}, # aten::linalg_householder_product, aten::linalg_householder_product.out torch.linalg.solve_triangular: {f32, f64}, # aten::linalg_solve_triangular, aten::linalg_solve_triangular.out torch.logcumsumexp: {bf16, f16}, # aten::_logcumsumexp, aten::_logcumsumexp.out torch.matrix_exp: {f16}, # aten::linalg_matrix_exp torch.median: {f16}, # aten::median, aten::median.dim_values torch.multinomial: {f16}, # aten::multinomial, aten::multinomial.out torch.nn.functional.gaussian_nll_loss: {f16}, # aten::_local_scalar_dense torch.nn.functional.max_pool3d: {bf16, f16}, # aten::max_pool3d_with_indices torch.nn.functional.max_pool3d_with_indices: {bf16, f16}, # aten::max_pool3d_with_indices torch.nn.functional.max_unpool1d: {f16}, # aten::max_unpool2d torch.nn.functional.max_unpool2d: {f16}, # aten::max_unpool2d torch.nn.functional.max_unpool3d: {f16}, # aten::max_unpool3d torch.nn.functional.multi_margin_loss: {bf16, f16}, # aten::multi_margin_loss torch.nn.functional.multilabel_margin_loss: {bf16, f16}, # aten::multilabel_margin_loss_forward torch.nn.functional.rrelu: {f16}, # aten::rrelu_with_noise torch.ormqr: {f32, f64}, # aten::ormqr, aten::ormqr.out } meta_function_device_skips['cpu'] = { torch.narrow_copy: {b8, bf16, c128, c32, c64, f16, f32, f64, i16, i32, i64, i8, u8}, torch.native_batch_norm: {f32, f64}, } meta_function_device_skips['cuda'] = { torch.cummax: {f16}, torch.cummin: {f16}, torch.functional.tensordot: {f16}, torch.inner: {f16}, torch.linalg.matrix_power: {f32, f64}, torch.linalg.matrix_rank: {f32, f64}, torch.linalg.svd: {f32, f64}, torch.nn.functional.cross_entropy: {f16}, torch.nn.functional.interpolate: {f16}, torch.nn.functional.nll_loss: {f16}, torch.svd: {f32, f64}, # This fails for arguments dispatched to grid_sampler_3d, but succeeds # for grid_sampler_2d, so we can't just xfail it torch.nn.functional.grid_sample : {f16}, } # This is a __torch_function__ mode that, when enabled, interposes every # Torch API call and runs the operator as normal, and then reruns it # with meta inputs, and then checks that everything about the output agrees. # Most of the logic deals with faithfully replicating the original tensor # as a meta tensor, which is nontrivial because there are a lot of subsystems # that may potentially be exercised. # # That being said, this class is a little overkill for what it is doing in # this test file (since I could have just inlined __torch_function__ on the # OpInfo call, and OpInfos generally have very regular inputs), but it will be # useful for more comprehensive testing e.g., as seen in # https://github.com/pytorch/pytorch/pull/75994 The big benefit is it is # A LOT more efficient that torch dispatch mode (at the cost of less coverage) class MetaCrossRefFunctionMode(torch.overrides.TorchFunctionMode): test_case: TestCase device_type: str dtype: torch.dtype def __init__(self, test_case, *, device, dtype): self.test_case = test_case self.device_type = torch.device(device).type self.dtype = dtype def __torch_function__(self, func, types, args=(), kwargs=None): kwargs = kwargs or {} if torch.jit.is_tracing() or isinstance(func, torch.ScriptMethod): return func(*args, **kwargs) if self.dtype in meta_function_skips.get(func, set()): test_expect = TestExpect.SKIP elif self.dtype in meta_function_device_skips[self.device_type].get(func, set()): test_expect = TestExpect.SKIP elif self.dtype in meta_function_expected_failures.get(func, set()): test_expect = TestExpect.XFAILURE elif self.dtype in meta_function_device_expected_failures[self.device_type].get(func, set()): test_expect = TestExpect.XFAILURE else: test_expect = TestExpect.SUCCESS return run_meta_crossref( self.test_case, test_expect, func, args, kwargs, dtype=self.dtype, device_type=self.device_type ) aten = torch.ops.aten # these always fail meta_dispatch_expected_failures = { aten.allclose.default: {f16, bf16, f32, f64, c64, c128}, # NotImplementedError: 'aten::_local_scalar_dense' aten._fft_c2c.out : {f16, c64, i8, f64, c128, i32, i64, f32, c32, b8, i16, u8}, aten._fft_r2c.out : {f16, i8, f64, i32, i64, f32, b8, i16, u8}, aten.cholesky.default : {c64, c128, f64, f32}, aten.cholesky.out : {c64, c128, f64, f32}, aten.cholesky_inverse.default : {c64, c128, f64, f32}, aten.cholesky_inverse.out : {c64, c128, f64, f32}, aten.cholesky_solve.default : {c64, c128, f64, f32}, aten.cholesky_solve.out : {c64, c128, f64, f32}, aten.count_nonzero.default : {c64, f16, i8, f64, c128, i64, bf16, f32, i32, b8, i16, u8}, aten.count_nonzero.dim_IntList : {c64, f16, i8, f64, c128, i64, bf16, f32, i32, b8, i16, u8}, aten.geqrf.default : {c64, c128, f64, f32}, aten.linalg_eig.default : {c64, c128, f64, f32}, aten.linalg_householder_product.default : {c64, c128, f64, f32}, aten.linalg_householder_product.out : {c64, c128, f64, f32}, aten.linalg_lstsq.default : {c64, c128, f64, f32}, aten.linalg_matrix_exp.default : {c64, bf16, f32, f64, c128}, aten.linalg_solve_triangular.default : {c64, c128, f64, f32}, aten.linalg_solve_triangular.out : {c64, c128, f64, f32}, aten.masked_select.default : {c64, f16, i8, f64, c128, i64, bf16, f32, i32, b8, i16, u8}, aten.masked_select.out : {c64, f16, i8, f64, c128, i64, bf16, f32, i32, b8, i16, u8}, aten.native_group_norm.default : {bf16}, aten.nonzero.default : {c64, f16, i8, f64, c128, i64, bf16, f32, i32, c32, b8, i16, u8}, aten.nonzero.out : {c64, f16, i8, f64, c128, i64, bf16, f32, i32, c32, b8, i16, u8}, aten.ormqr.default : {c64, c128, f64, f32}, aten.ormqr.out : {c64, c128, f64, f32}, aten.polar.out : {f32, f64}, aten.symeig.default : {c64, c128, f64, f32}, aten.take.default : {c64, f16, i8, f64, c128, i64, bf16, f32, i32, b8, i16, u8}, aten.take.out : {c64, f16, i8, f64, c128, i64, bf16, f32, i32, b8, i16, u8}, aten.tensordot.out : {c64, i8, f64, c128, i64, bf16, f32, i32, i16, u8}, aten.to_sparse.default : {c64, f16, i8, f64, c128, i64, bf16, f32, i32, b8, i16, u8}, aten.to_sparse.sparse_dim : {c64, f16, i8, f64, c128, i64, bf16, f32, i32, b8, i16, u8}, aten._ctc_loss.default : {f32, f64}, # Shape of second output depends on data. aten._ctc_loss.Tensor : {f32, f64}, # Shape of second output depends on data. aten._histogramdd_bin_edges.default : {f32, f64}, aten._histogramdd_from_bin_cts.default : {f32, f64}, aten._histogramdd_from_bin_tensors.default : {f32, f64}, aten._local_scalar_dense.default : {c32, c64, f16, i8, f64, c128, i64, bf16, f32, i32, b8, i16, u8}, aten._pdist_forward.default : {f32, f64}, aten._unique2.default : {i8, f64, i64, bf16, f32, i32, b8, i16, u8}, aten.bincount.default : {i64, i8, i32, i16, u8}, aten.bucketize.Tensor : {f16, i8, f64, i64, bf16, f32, i32, i16, u8}, aten.bucketize.Tensor_out : {f16, i8, f64, i64, bf16, f32, i32, i16, u8}, aten.equal.default : {c64, f16, i8, f64, c128, i64, bf16, f32, i32, b8, i16, u8}, aten.frexp.Tensor : {bf16, f32, f16, f64}, aten.grid_sampler_3d.default : {f32, f64}, aten.histc.default : {bf16, f32, f64}, aten.histc.out : {bf16, f32, f64}, aten.histogram.bin_ct : {f32, f64}, aten.histogram.bins_tensor : {f32, f64}, aten.kthvalue.default : {i8, f64, i64, bf16, f32, i32, i16, u8}, aten.logcumsumexp.default : {bf16, f32, f64}, aten.logcumsumexp.out : {bf16, f32, f64}, aten.max_pool3d_with_indices.default : {f32, f64}, aten.max_unpool2d.default : {f32, f64}, aten.max_unpool3d.default : {f32, f64}, aten.median.default : {i8, f64, i64, bf16, f32, i32, i16, u8}, aten.median.dim : {i8, f64, i64, bf16, f32, i32, i16, u8}, aten.mode.default : {f16, i8, f64, i64, bf16, f32, i32, b8, i16, u8}, aten.multi_margin_loss.default : {f32, f64}, aten.multilabel_margin_loss_forward.default : {f32, f64}, aten.multinomial.default : {bf16, f32, f64}, aten.multinomial.out : {bf16, f32, f64}, aten.nll_loss2d_forward.default : {bf16, f32, f64}, aten.polar.default : {f32, f64}, aten.rrelu_with_noise.default : {bf16, f32, f64}, aten.searchsorted.Tensor : {f16, i8, f64, i64, bf16, f32, i32, i16, u8}, aten.searchsorted.Tensor_out : {f16, i8, f64, i64, bf16, f32, i32, i16, u8}, aten.segment_reduce.default : {bf16, f32, f16, f64}, aten.unique_consecutive.default : {i8, f64, i64, bf16, f32, i32, b8, i16, u8}, aten.unique_dim.default : {i8, f64, i64, bf16, f32, i32, b8, i16, u8}, aten.upsample_nearest3d.vec : {bf16, f32, f64, u8}, } # these sometimes pass and sometimes fail meta_dispatch_skips = { aten.index.Tensor: {i64, bf16, f16, u8, b8, f32, i8, f64, i16, i32, c32, c64, c128}, # at::nonzero doesn't have a Meta function aten._to_copy.default: {i64, bf16, f16, u8, b8, f32, i8, f64, i16, i32, c32, c64, c128}, aten.aminmax.default: {i64, u8, b8, f32, i8, f64, i16, i32}, aten.cummax.default: {i64, bf16, u8, b8, f32, i8, f64, i16, i32}, aten.cummin.default: {i64, bf16, u8, b8, f32, i8, f64, i16, i32}, aten.linalg_lu_solve.default: {c32, c64, c128}, aten.linalg_lu_solve.out: {c32, c64, c128}, aten.linalg_pinv.atol_rtol_tensor: {f32, f64}, aten.linalg_pinv.atol_rtol_tensor_out: {f32, f64}, aten.empty.memory_format: {b8, bf16, c128, c64, c32, f16, f32, f64, i16, i32, i64, i8, u8}, } meta_dispatch_device_expected_failures = defaultdict(dict) meta_dispatch_device_skips = defaultdict(dict) meta_dispatch_device_expected_failures['cpu'] = { aten.narrow_copy.out: {b8, bf16, c128, c32, c64, f16, f32, f64, i16, i32, i64, i8, u8}, # aten::narrow_copy.out aten.native_batch_norm.default: {bf16}, aten.native_layer_norm.default: {bf16}, } meta_dispatch_device_expected_failures['cuda'] = { aten._unique2.default: {f16}, # aten::_unique2 aten._use_cudnn_ctc_loss.default: {f32, f64}, # aten::_use_cudnn_ctc_loss aten._use_cudnn_ctc_loss.Tensor: {f32, f64}, # aten::_use_cudnn_ctc_loss.Tensor aten.cudnn_grid_sampler.default: {f16, f32, f64}, # aten::cudnn_grid_sampler aten.geqrf.default: {f32, f64}, # aten::geqrf aten.grid_sampler_3d.default: {f16}, # aten::grid_sampler_3d aten.histc.default: {i16, i32, i64, i8}, # aten::histc aten.histc.out: {i16, i32, i64, i8}, # aten::histc.out aten.kthvalue.default: {f16}, # aten::kthvalue.values aten.linalg_eigvalsh.out: {f32, f64}, # aten::linalg_eigvalsh.out aten.linalg_householder_product.default: {f32, f64}, # aten::linalg_householder_product aten.linalg_householder_product.out: {f32, f64}, # aten::linalg_householder_product.out aten.linalg_matrix_exp.default: {f16}, # aten::linalg_matrix_exp aten.linalg_solve_triangular.default: {f32, f64}, # aten::linalg_solve_triangular aten.linalg_solve_triangular.out: {f32, f64}, # aten::linalg_solve_triangular.out aten.log_sigmoid_forward.default: {bf16, f16, f64, f32}, aten.log_sigmoid_forward.output : {bf16, f16, f64, f32}, # aten::log_sigmoid_forward.output aten.logcumsumexp.default: {bf16, f16}, # aten::_logcumsumexp aten.logcumsumexp.out: {bf16, f16}, # aten::_logcumsumexp.out aten.max_pool3d_with_indices.default: {bf16, f16}, # aten::max_pool3d_with_indices aten.max_unpool2d.default: {f16}, # aten::max_unpool2d aten.max_unpool3d.default: {f16}, # aten::max_unpool3d aten.median.default: {f16}, # aten::median aten.median.dim: {f16}, # aten::median.dim_values aten.multi_margin_loss.default: {bf16, f16}, # aten::multi_margin_loss aten.multilabel_margin_loss_forward.default: {bf16, f16}, # aten::multilabel_margin_loss_forward aten.multinomial.default: {f16}, # aten::multinomial aten.multinomial.out: {f16}, # aten::multinomial.out aten.native_group_norm.default: {bf16, f16}, aten.nll_loss2d_forward.default: {f16}, # aten::nll_loss2d_forward aten.ormqr.default: {f32, f64}, # aten::ormqr aten.ormqr.out: {f32, f64}, # aten::ormqr.out aten.rrelu_with_noise.default: {f16}, # aten::rrelu_with_noise aten.tensordot.out: {f16}, # aten::tensordot.out aten.unique_consecutive.default: {f16}, # aten::unique_consecutive aten.unique_dim.default: {f16}, # aten::unique_dim aten.upsample_nearest3d.vec: {f16}, # aten::upsample_nearest3d.vec } meta_dispatch_device_skips['cpu'] = { aten._embedding_bag_forward_only.default: {f16, f32, f64}, aten.native_batch_norm.default: {f32, f64}, } meta_dispatch_device_skips['cuda'] = { aten._conj.default: {c32, f16}, # file issue aten._linalg_svd.default: {c64, c128}, # aten::linalg_eigvalsh.out aten.cudnn_batch_norm.default: {f32, f64}, aten.log_softmax.int : {c32, c64}, aten.softmax.int : {c32, c64}, aten.softmax.int : {c32, c64}, aten.cummax.default: {f16}, aten.cummin.default: {f16}, # ROCm stuff; technically this should be expected failure but it's # not worth it; these should get unified anyway aten.miopen_batch_norm.default: {f32}, } class MetaCrossRefDispatchMode(torch.utils._python_dispatch.TorchDispatchMode): test_case: TestCase device: torch.device dtype: torch.dtype def __init__(self, test_case, *, device, dtype): self.test_case = test_case # save TLS self.precision = test_case.precision self.rel_tol = test_case.rel_tol self.device_type = torch.device(device).type self.dtype = dtype def __torch_dispatch__(self, func, types, args=(), kwargs=None): kwargs = kwargs or {} self.test_case.precision = self.precision self.test_case.rel_tol = self.rel_tol if self.dtype in meta_dispatch_skips.get(func, set()): test_expect = TestExpect.SKIP elif self.dtype in meta_dispatch_device_skips[self.device_type].get(func, set()): test_expect = TestExpect.SKIP elif self.dtype in meta_dispatch_expected_failures.get(func, set()): test_expect = TestExpect.XFAILURE elif self.dtype in meta_dispatch_device_expected_failures[self.device_type].get(func, set()): test_expect = TestExpect.XFAILURE else: test_expect = TestExpect.SUCCESS return run_meta_crossref( self.test_case, test_expect, func, args, kwargs, dtype=self.dtype, device_type=self.device_type, ) # NB: we're running these tests only on CUDA because there are some # inconsistencies between CUDA and CPU, and running on CUDA makes it easier # to ignore the CPU case when inconsistencies arise. Ideally we deal # with the inconsistencies but this takes time. @skipIfSlowGradcheckEnv class TestMeta(TestCase): @unittest.skipIf(TEST_WITH_ASAN, "Skipped under ASAN") @skipIfCrossRef @suppress_warnings @ops(op_db) def test_meta(self, device, dtype, op): # run the OpInfo sample inputs, cross-referencing them with the # meta implementation and check the results are the same. All # the heavy lifting happens in MetaCrossRefFunctionMode func = op.get_op() samples = op.sample_inputs(device, dtype, requires_grad=False) for sample_input in samples: args = [sample_input.input] + list(sample_input.args) kwargs = sample_input.kwargs with MetaCrossRefFunctionMode(self, dtype=dtype, device=device): expected = func(*args, **kwargs) if isinstance(expected, torch.Tensor) and op.supports_out: func(*args, **kwargs, out=expected) @unittest.skipIf(TEST_WITH_ASAN, "Skipped under ASAN") @skipIfCrossRef @suppress_warnings @ops(op_db) def test_dispatch_meta(self, device, dtype, op): func = op.get_op() samples = op.sample_inputs(device, dtype, requires_grad=False) for sample_input in samples: args = [sample_input.input] + list(sample_input.args) kwargs = sample_input.kwargs with MetaCrossRefDispatchMode.push(self, dtype=dtype, device=device): expected = func(*args, **kwargs) if isinstance(expected, torch.Tensor) and op.supports_out: func(*args, **kwargs, out=expected) def test_empty_quantized(self): r = torch.empty(2 ** 52, device='meta', dtype=torch.qint8) self.assertEqual(r.device.type, 'meta') def test_map_location_deserialize(self): import io t = torch.rand(10) b = io.BytesIO() torch.save(t, b) b.seek(0) r = torch.load(b, map_location=torch.device("meta")) self.assertEqual(r.device.type, 'meta') self.assertEqual(r.shape, t.shape) self.assertEqual(r.dtype, t.dtype) self.assertEqual(r.storage().data_ptr(), 0) instantiate_device_type_tests(TestMeta, globals()) def print_op_str_if_not_supported(op_str): op = OperatorName.parse(op_str) packet = getattr(torch.ops.aten, str(op.name)) overload = getattr(packet, op.overload_name if op.overload_name else "default") if any(overload in d for d in [meta_dispatch_skips, meta_dispatch_device_skips['cuda']]): print(f"{overload} # SKIP") if any(overload in d for d in [meta_dispatch_expected_failures, meta_dispatch_device_expected_failures['cuda']]): print(overload) if __name__ == "__main__": COMPARE_XLA = os.getenv('PYTORCH_COMPARE_XLA', None) if COMPARE_XLA is not None: with open(COMPARE_XLA, "r") as f: d = yaml.load(f, Loader=YamlLoader) ops = d.get("full_codegen", []) + d.get("supported", []) + d.get("autograd", []) for op_str in ops: print_op_str_if_not_supported(op_str) sys.exit(0) COMPARE_TEXT = os.getenv('PYTORCH_COMPARE_TEXT', None) if COMPARE_TEXT is not None: with open(COMPARE_TEXT, "r") as f: for op_str in f: print_op_str_if_not_supported(op_str.strip()) sys.exit(0) run_tests()