OSSForks/pytorch
1
0
Fork 0
mirror of https://github.com/zebrajr/pytorch.git synced 2025-12-06 12:20:52 +01:00
Code Issues Actions 38 Packages Projects Releases Wiki Activity
39d1c91c33
pytorch/test/dynamo/test_repros.py
Ryan Guo 39d1c91c33 [dynamo] Restrict support for out= variants of torch operators (#140202) 
There has been a series of attempts to provide support for resizing in
torch operators like `torch.sigmoid(x, out=y)`, i.e., `y` would have a
different shape before and after this expression. Prior to this patch,
we have some checks to graph break if the shape changed.

This patch extends
1. extends the existing check and graph break for any shape change, not
   just for `TensorVariable` with source field.
2. removes an old code path which was introduced to address the shape
   change, but became obselete in that regard because we added extra
   checks to graph break upon shape change. Moreover, this old code path
   is unsound, it tries to replace references to the old
   `TensorVariable` the new one returned by `wrap_fx_proxy`, but it only
   does the replacement in `symbolic_locals`, which breaks when cells
   are involved. In general the old `TensorVariable` could be _anywhere_,
   think the `replace_all` we had for immutable VTs.

Pull Request resolved: https://github.com/pytorch/pytorch/pull/140202
Approved by: https://github.com/jansel
ghstack dependencies: #140035, #140036, #140149, #140150, #140151, #140201
2024-11-13 03:14:23 +00:00

6349 lines
202 KiB
Python
Raw Blame History

"""
PYTEST_DONT_REWRITE (prevents pytest from rewriting assertions, which interferes
with test_rewrite_assert_with_msg and test_rewrite_assert_without_msg)
"""
# Owner(s): ["module: dynamo"]
import collections
import contextlib
import copy
import dataclasses
import functools
import gc
import importlib
import inspect
import itertools
import os
import random
import unittest
import warnings
import weakref
from abc import ABC
from collections import namedtuple
from copy import deepcopy
from enum import Enum, IntEnum
from functools import wraps
from typing import Any, Dict, Iterator, List, Literal, Tuple, TypedDict
from unittest import mock
import numpy as np
import torch
import torch._dynamo.test_case
import torch._dynamo.testing
import torch._dynamo.utils
import torch._functorch.config
import torch.library
import torch.utils._pytree as pytree
from torch import nn
from torch._dynamo.debug_utils import same_two_models
from torch._dynamo.testing import CompileCounter, rand_strided, same, skipIfPy312
from torch._inductor.utils import fresh_inductor_cache
from torch.nn import functional as F
from torch.testing._internal.common_cuda import PLATFORM_SUPPORTS_FLASH_ATTENTION
from torch.testing._internal.common_utils import (
disable_translation_validation_if_dynamic_shapes,
instantiate_parametrized_tests,
parametrize,
skipIfWindows,
TEST_WITH_ROCM,
)
from torch.testing._internal.two_tensor import TwoTensor
_orig_module_call = torch.nn.Module.__call__
# Custom operator that only supports CPU and Meta
lib = torch.library.Library("test_sample", "DEF") # noqa: TOR901
lib.define("foo(Tensor self) -> Tensor")
lib.impl("foo", torch.sin, "CPU")
requires_cuda = unittest.skipUnless(torch.cuda.is_available(), "requires cuda")
_GLOBAL_CPU_TENSOR = torch.randn(3)
HAS_MSGSPEC = importlib.util.find_spec("msgspec")
if HAS_MSGSPEC:
import msgspec
HAS_OMEGACONG = importlib.util.find_spec("omegaconf")
if HAS_OMEGACONG:
from omegaconf import OmegaConf
def exists(val):
return val is not None
def maybe(fn):
@wraps(fn)
def inner(x, *args, **kwargs):
if not exists(x):
return x
return fn(x, *args, **kwargs)
return inner
def is_fx_tracing_test() -> bool:
"""
Copied from the hpc trainer codebase
"""
return torch.nn.Module.__call__ is not _orig_module_call
def has_detectron2():
try:
from detectron2.layers.mask_ops import _paste_masks_tensor_shape
return _paste_masks_tensor_shape is not None
except ImportError:
return False
def _do_paste_mask(masks, boxes, img_h: int, img_w: int, skip_empty: bool = True):
# from detectron2 mask_ops.py
device = masks.device
if skip_empty and not torch.jit.is_scripting():
x0_int, y0_int = torch.clamp(boxes.min(dim=0).values.floor()[:2] - 1, min=0).to(
dtype=torch.int32
)
x1_int = torch.clamp(boxes[:, 2].max().ceil() + 1, max=img_w).to(
dtype=torch.int32
)
y1_int = torch.clamp(boxes[:, 3].max().ceil() + 1, max=img_h).to(
dtype=torch.int32
)
else:
x0_int, y0_int = 0, 0
x1_int, y1_int = img_w, img_h
x0, y0, x1, y1 = torch.split(boxes, 1, dim=1) # each is Nx1
N = masks.shape[0]
img_y = torch.arange(y0_int, y1_int, device=device, dtype=torch.float32) + 0.5
img_x = torch.arange(x0_int, x1_int, device=device, dtype=torch.float32) + 0.5
img_y = (img_y - y0) / (y1 - y0) * 2 - 1
img_x = (img_x - x0) / (x1 - x0) * 2 - 1
# img_x, img_y have shapes (N, w), (N, h)
gx = img_x[:, None, :].expand(N, img_y.size(1), img_x.size(1))
gy = img_y[:, :, None].expand(N, img_y.size(1), img_x.size(1))
grid = torch.stack([gx, gy], dim=3)
if not torch.jit.is_scripting():
if not masks.dtype.is_floating_point:
masks = masks.float()
img_masks = F.grid_sample(masks, grid.to(masks.dtype), align_corners=False)
if skip_empty and not torch.jit.is_scripting():
return img_masks[:, 0], (slice(y0_int, y1_int), slice(x0_int, x1_int))
else:
return img_masks[:, 0], ()
def global_fn(x):
return torch.sin(x)
def cat(tensors, dim=0):
# from detectron2 wrappers.py
assert isinstance(tensors, (list, tuple))
if len(tensors) == 1:
return tensors[0]
return torch.cat(tensors, dim)
def shapes_to_tensor(x, device=None):
# from detectron2 wrappers.py
if torch.jit.is_scripting():
return torch.as_tensor(x, device=device)
if torch.jit.is_tracing():
assert all(
isinstance(t, torch.Tensor) for t in x
), "Shape should be tensor during tracing!"
# as_tensor should not be used in tracing because it records a constant
ret = torch.stack(x)
if ret.device != device: # avoid recording a hard-coded device if not necessary
ret = ret.to(device=device)
return ret
return torch.as_tensor(x, device=device)
fw_graph = [None]
bw_graph = [None]
def aot_graph_capture_backend(gm, args):
from functorch.compile import min_cut_rematerialization_partition
from torch._functorch.aot_autograd import aot_module_simplified
def fw_compiler(gm, _):
fw_graph[0] = gm
return gm
def bw_compiler(gm, _):
bw_graph[0] = gm
return gm
return aot_module_simplified(
gm,
args,
fw_compiler,
bw_compiler,
partition_fn=min_cut_rematerialization_partition,
keep_inference_input_mutations=True,
)
class Boxes:
# from detectron2 poolers.py
def __init__(self, tensor: torch.Tensor):
"""
Args:
tensor (Tensor[float]): a Nx4 matrix. Each row is (x1, y1, x2, y2).
"""
device = (
tensor.device if isinstance(tensor, torch.Tensor) else torch.device("cpu")
)
tensor = torch.as_tensor(tensor, dtype=torch.float32, device=device)
if tensor.numel() == 0:
# Use reshape, so we don't end up creating a new tensor that does not depend on
# the inputs (and consequently confuses jit)
tensor = tensor.reshape((-1, 4)).to(dtype=torch.float32, device=device)
assert tensor.dim() == 2 and tensor.size(-1) == 4, tensor.size()
self.tensor = tensor
def __len__(self) -> int:
return self.tensor.shape[0]
@property
def device(self):
return self.tensor.device
def convert_boxes_to_pooler_format(box_lists):
# from detectron2 structures.py
boxes = torch.cat([x.tensor for x in box_lists], dim=0)
# __len__ returns Tensor in tracing.
sizes = shapes_to_tensor([x.__len__() for x in box_lists], device=boxes.device)
indices = torch.repeat_interleave(
torch.arange(len(box_lists), dtype=boxes.dtype, device=boxes.device), sizes
)
return cat([indices[:, None], boxes], dim=1)
ReformerBackwardOutput = namedtuple(
"ReformerBackwardOutput",
["attn_output", "hidden_states", "grad_attn_output", "grad_hidden_states"],
)
ReformerEncoderOutput = namedtuple(
"ReformerEncoderOutput",
["hidden_states", "all_hidden_states", "all_attentions", "past_buckets_states"],
)
class _ReversibleFunction(torch.autograd.Function):
# taken from modeling_reformer.py in huggingface
@staticmethod
def forward(
ctx,
hidden_states,
layers,
attention_mask,
head_mask,
num_hashes,
all_hidden_states,
all_attentions,
past_buckets_states,
use_cache,
orig_sequence_length,
output_hidden_states,
output_attentions,
):
all_buckets = ()
# split duplicated tensor
hidden_states, attn_output = torch.chunk(hidden_states, 2, dim=-1)
for layer_id, (layer, layer_head_mask) in enumerate(zip(layers, head_mask)):
if output_hidden_states is True:
all_hidden_states.append(hidden_states)
attn_output = layer(attn_output)
all_buckets = all_buckets + (attn_output,)
# Add last layer
if output_hidden_states is True:
all_hidden_states.append(hidden_states)
# attach params to ctx for backward
ctx.save_for_backward(attn_output.detach(), hidden_states.detach())
ctx.layers = layers
ctx.all_buckets = all_buckets
ctx.head_mask = head_mask
ctx.attention_mask = attention_mask
# Concatenate 2 RevNet outputs
return torch.cat([attn_output, hidden_states], dim=-1)
@staticmethod
def backward(ctx, grad_hidden_states):
grad_attn_output, grad_hidden_states = torch.chunk(
grad_hidden_states, 2, dim=-1
)
# free memory
del grad_attn_output
# num of return vars has to match num of forward() args
# return gradient for hidden_states arg and None for other args
return (
grad_hidden_states,
None,
None,
None,
None,
None,
None,
None,
None,
None,
None,
None,
)
class ReformerEncoder(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.dropout = 0.5
self.layer_norm = torch.nn.LayerNorm(512, eps=1.0e-12)
self.layers = [torch.nn.Linear(256, 256)]
def forward(
self,
hidden_states,
attention_mask=None,
head_mask=[None] * 6,
num_hashes=None,
use_cache=False,
orig_sequence_length=64,
output_hidden_states=False,
output_attentions=False,
):
# hidden_states and attention lists to be filled if wished
all_hidden_states = []
all_attentions = []
past_buckets_states = [((None), (None)) for i in range(len(self.layers))]
# concat same tensor for reversible ResNet
hidden_states = torch.cat([hidden_states, hidden_states], dim=-1)
hidden_states = _ReversibleFunction.apply(
hidden_states,
self.layers,
attention_mask,
head_mask,
num_hashes,
all_hidden_states,
all_attentions,
past_buckets_states,
use_cache,
orig_sequence_length,
output_hidden_states,
output_attentions,
)
# Apply layer norm to concatenated hidden states
hidden_states = self.layer_norm(hidden_states)
# Apply dropout
hidden_states = torch.nn.functional.dropout(
hidden_states, p=self.dropout, training=self.training
)
return ReformerEncoderOutput(
hidden_states=hidden_states,
all_hidden_states=all_hidden_states,
all_attentions=all_attentions,
past_buckets_states=past_buckets_states,
)
class ListConfig:
class ValueNode:
def __init__(self, value):
self.value = value
def _dereference_node(self):
return self
def _is_missing(self):
return False
def _value(self):
return self.value
# Based on an example from omegaconfig.listconfig
class ListIterator(Iterator[Any]):
def __init__(self, lst: Any, resolve: bool) -> None:
self.resolve = resolve
self.iterator = iter(lst.__dict__["_content"])
self.index = 0
def __next__(self) -> Any:
x = next(self.iterator)
if self.resolve:
x = x._dereference_node()
if x._is_missing():
raise AssertionError
self.index = self.index + 1
if isinstance(x, ListConfig.ValueNode):
return x._value()
raise AssertionError
def __iter__(self):
return self._iter_ex(True)
def _iter_ex(self, resolve: bool) -> Iterator[Any]:
try:
return ListConfig.ListIterator(self, resolve)
except Exception:
raise AssertionError from None
def __init__(self) -> None:
self._content = [
ListConfig.ValueNode(1),
ListConfig.ValueNode(3),
ListConfig.ValueNode(torch.tensor([7.0])),
]
def longformer_chunk(hidden_states, window_overlap=256):
"""convert into overlapping chunks. Chunk size = 2w, overlap size = w"""
# non-overlapping chunks of size = 2w
hidden_states = hidden_states.view(
hidden_states.size(0),
hidden_states.size(1) // (window_overlap * 2),
window_overlap * 2,
hidden_states.size(2),
)
# use `as_strided` to make the chunks overlap with an overlap size = window_overlap
chunk_size = list(hidden_states.size())
chunk_size[1] = chunk_size[1] * 2 - 1
chunk_stride = list(hidden_states.stride())
chunk_stride[1] = chunk_stride[1] // 2
return hidden_states.as_strided(size=chunk_size, stride=chunk_stride)
class PartialT5(torch.nn.Module):
# Highly simplified T5Attention prefix
def __init__(self) -> None:
super().__init__()
self.q = torch.nn.Linear(512, 512)
self.k = torch.nn.Linear(512, 512)
self.v = torch.nn.Linear(512, 512)
def forward(
self,
hidden_states,
key_value_states=None,
past_key_value=None,
query_length=None,
):
batch_size, seq_length = hidden_states.shape[:2]
real_seq_length = seq_length
if past_key_value is not None:
assert (
len(past_key_value) == 2
), f"past_key_value should have 2 past states: keys and values. Got {len(past_key_value)} past states"
real_seq_length += (
past_key_value[0].shape[2] if query_length is None else query_length
)
def shape(states):
"""projection"""
return states.view(batch_size, -1, 8, 64).transpose(1, 2)
def project(hidden_states, proj_layer, key_value_states, past_key_value):
"""projects hidden states correctly to key/query states"""
if key_value_states is None:
# self-attn
# (batch_size, n_heads, seq_length, dim_per_head)
hidden_states = shape(proj_layer(hidden_states))
elif past_key_value is None:
# cross-attn
# (batch_size, n_heads, seq_length, dim_per_head)
hidden_states = shape(proj_layer(key_value_states))
if past_key_value is not None:
if key_value_states is None:
# self-attn
# (batch_size, n_heads, key_length, dim_per_head)
hidden_states = torch.cat([past_key_value, hidden_states], dim=2)
else:
# cross-attn
hidden_states = past_key_value
return hidden_states
# get query states
query_states = shape(
self.q(hidden_states)
) # (batch_size, n_heads, seq_length, dim_per_head)
# get key/value states
key_states = project(
hidden_states,
self.k,
key_value_states,
past_key_value[0] if past_key_value is not None else None,
)
value_states = project(
hidden_states,
self.v,
key_value_states,
past_key_value[1] if past_key_value is not None else None,
)
# compute scores
scores = torch.matmul(query_states, key_states.transpose(3, 2))
# (truncated here )
return scores, value_states
class ChunkReformerFeedForward(torch.nn.Module):
# simplified from HF modeling_reformer.py
def __init__(self) -> None:
super().__init__()
self.layer_norm = torch.nn.LayerNorm(256, eps=1e-12)
self.dense = torch.nn.Linear(256, 256)
self.output = torch.nn.Linear(256, 256)
def forward(self, attention_output):
return apply_chunking_to_forward(
self.forward_chunk,
attention_output + 1,
)
def forward_chunk(self, hidden_states):
hidden_states = self.layer_norm(hidden_states)
hidden_states = self.dense(hidden_states)
return self.output(hidden_states)
def apply_chunking_to_forward(forward_fn, *input_tensors):
# simplified from HF model_utils.py
assert len(input_tensors) > 0
tensor_shape = input_tensors[0].shape[1]
assert all(input_tensor.shape[1] == tensor_shape for input_tensor in input_tensors)
num_args_in_forward_chunk_fn = len(inspect.signature(forward_fn).parameters)
if num_args_in_forward_chunk_fn != len(input_tensors):
raise ValueError
return forward_fn(*input_tensors)
def _validate_model_kwargs(fn, model_kwargs):
# simplified from transformers.generation.utils._validate_model_kwargs
unused_model_args = []
model_args = set(inspect.signature(fn).parameters)
for key, value in model_kwargs.items():
if value is not None and key not in model_args:
unused_model_args.append(key)
if unused_model_args:
raise ValueError(
f"The following `model_kwargs` are not used by the model: {unused_model_args} (note: typos in the"
" generate arguments will also show up in this list)"
)
class FakeMamlInner(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.linear = torch.nn.Linear(784, 5)
def forward(self, x, ignored=None, bn_training=False):
return self.linear(x.view(x.shape[0], -1))
class PartialMaml(torch.nn.Module):
# Highly simplified version of maml.meta.Meta.finetuning
def __init__(self) -> None:
super().__init__()
self.net = FakeMamlInner()
self.update_step_test = 10
self.update_lr = 0.4
def forward(self, x_spt, y_spt, x_qry, y_qry):
querysz = x_qry.size(0)
corrects = [0 for _ in range(self.update_step_test + 1)]
# in order to not ruin the state of running_mean/variance and bn_weight/bias
# we finetuning on the copied model instead of self.net
net = deepcopy(self.net)
# 1. run the i-th task and compute loss for k=0
logits = net(x_spt)
loss = F.cross_entropy(logits, y_spt)
grad = torch.autograd.grad(loss, net.parameters())
fast_weights = [
p[1] - self.update_lr * p[0] for p in zip(grad, net.parameters())
]
# this is the loss and accuracy before first update
with torch.no_grad():
# [setsz, nway]
logits_q = net(x_qry, net.parameters(), bn_training=True)
# [setsz]
pred_q = F.softmax(logits_q, dim=1).argmax(dim=1)
# scalar
correct = torch.eq(pred_q, y_qry).sum().item()
corrects[0] = corrects[0] + correct
# this is the loss and accuracy after the first update
with torch.no_grad():
# [setsz, nway]
logits_q = net(x_qry, fast_weights, bn_training=True)
# [setsz]
pred_q = F.softmax(logits_q, dim=1).argmax(dim=1)
# scalar
correct = torch.eq(pred_q, y_qry).sum().item()
corrects[1] = corrects[1] + correct
del net
accs = torch.tensor(corrects) / querysz
return accs
def softmax_backward_data(parent, grad_output, output, dim, self):
from torch import _softmax_backward_data
return _softmax_backward_data(grad_output, output, parent.dim, self.dtype)
class XSoftmax(torch.autograd.Function):
# transformers.models.deberta.modeling_deberta.XSoftmax
@staticmethod
def forward(self, input, mask, dim):
self.dim = dim
rmask = ~(mask.to(torch.bool))
output = input.masked_fill(rmask, torch.tensor(torch.finfo(input.dtype).min))
output = torch.softmax(output, self.dim)
output.masked_fill_(rmask, 0)
self.save_for_backward(output, rmask)
return output
@staticmethod
def backward(self, grad_output):
(output, rmask) = self.saved_tensors
inputGrad = softmax_backward_data(self, grad_output, output, self.dim, output)
return inputGrad, None, None
class ModelOutput(collections.OrderedDict):
"""based on file_utils.py in HuggingFace"""
def __getitem__(self, k):
if isinstance(k, str):
inner_dict = dict(self.items())
return inner_dict[k]
else:
return self.to_tuple()[k]
def __setattr__(self, name, value):
if name in self.keys() and value is not None:
# Don't call self.__setitem__ to avoid recursion errors
super().__setitem__(name, value)
super().__setattr__(name, value)
def __setitem__(self, key, value):
# Will raise a KeyException if needed
super().__setitem__(key, value)
# Don't call self.__setattr__ to avoid recursion errors
super().__setattr__(key, value)
def to_tuple(self):
return tuple(self[k] for k in self.keys())
def create_rand_mask_from_inputs(
from_blocked_mask,
to_blocked_mask,
rand_attn,
num_attention_heads,
num_rand_blocks,
batch_size,
from_seq_length,
from_block_size,
):
"""taken from HF modeling_big_bird.py"""
num_windows = from_seq_length // from_block_size - 2
rand_mask = torch.stack(
[p1[i1.flatten()] for p1, i1 in zip(to_blocked_mask, rand_attn)]
)
rand_mask = rand_mask.view(
batch_size, num_attention_heads, num_windows, num_rand_blocks * from_block_size
)
rand_mask = torch.einsum("blq,bhlk->bhlqk", from_blocked_mask[:, 1:-1], rand_mask)
return rand_mask
class SequentialAppendList(torch.nn.Sequential):
"""from timm/models/vovnet.py"""
def forward(self, x: torch.Tensor, concat_list: List[torch.Tensor]) -> torch.Tensor:
for i, module in enumerate(self):
if i == 0:
concat_list.append(module(x))
else:
concat_list.append(module(concat_list[-1]))
x = torch.cat(concat_list, dim=1)
return x, concat_list
class BatchNormAct2d(torch.nn.BatchNorm2d):
"""Taken from timm"""
def __init__(
self,
num_features,
eps=1e-5,
momentum=0.1,
affine=True,
track_running_stats=True,
act_layer=torch.nn.ReLU,
inplace=True,
):
super().__init__(
num_features,
eps=eps,
momentum=momentum,
affine=affine,
track_running_stats=track_running_stats,
)
self.act = act_layer(inplace=inplace)
@torch.jit.ignore
def _forward_python(self, x):
return super().forward(x)
def forward(self, x):
if torch.jit.is_scripting():
x = self._forward_jit(x)
else:
x = self._forward_python(x)
x = self.act(x)
return x
def get_parameter_dtype(parameter):
"""from huggingface model_utils.py"""
try:
return next(parameter.parameters()).dtype
except StopIteration:
# For nn.DataParallel compatibility in PyTorch 1.5
def find_tensor_attributes(module):
tuples = [(k, v) for k, v in module.__dict__.items() if torch.is_tensor(v)]
return tuples
gen = parameter._named_members(get_members_fn=find_tensor_attributes)
first_tuple = next(gen)
return first_tuple[1].dtype
class DummyConfig:
attn_layers = ["local", "lsh", "local", "lsh", "local", "lsh"]
lsh_attn_chunk_length = 64
local_attn_chunk_length = 64
def _get_min_chunk_len(config):
"""from hf_Reformer"""
attn_types = config.attn_layers
attn_types_set = set(attn_types)
if len(attn_types_set) == 1 and attn_types[0] == "lsh":
return config.lsh_attn_chunk_length
elif len(attn_types_set) == 1 and attn_types[0] == "local":
return config.local_attn_chunk_length
elif len(attn_types_set) == 2 and attn_types_set == {"lsh", "local"}:
return min(config.lsh_attn_chunk_length, config.local_attn_chunk_length)
else:
raise NotImplementedError(
f"Only attn layer types 'lsh' and 'local' exist, but `config.attn_layers`: {config.attn_layers}. Select "
"attn layer types from ['lsh', 'local'] only."
)
def _stable_argsort(vector, dim):
"""from hf_Reformer"""
# this function scales the vector so that torch.argsort is stable.
# torch.argsort is not stable on its own
scale_offset = torch.arange(vector.shape[dim], device=vector.device).view(1, 1, -1)
scale_offset = scale_offset.expand(vector.shape)
scaled_vector = vector.shape[dim] * vector + (scale_offset % vector.shape[dim])
return torch.argsort(scaled_vector, dim=dim)
def _get_sorted_bucket_idx_and_undo_sorted_bucket_idx(buckets):
"""from hf_Reformer"""
# no gradients are needed
with torch.no_grad():
# hash-based sort
sorted_bucket_idx = _stable_argsort(buckets, dim=-1)
# create simple indices to scatter to, to have undo sort
indices = (
torch.arange(sorted_bucket_idx.shape[-1], device=buckets.device)
.view(1, 1, -1)
.expand(sorted_bucket_idx.shape)
)
# get undo sort
undo_sorted_bucket_idx = sorted_bucket_idx.new(*sorted_bucket_idx.size())
undo_sorted_bucket_idx.scatter_(-1, sorted_bucket_idx, indices)
return sorted_bucket_idx, undo_sorted_bucket_idx
class CustomList1(list):
def __call__(self, x):
for processor in self:
x = processor(x)
return x
def clear(self):
pass # this prevents RestrictedListSubclassVariable from kicking in
class CustomList2(list):
def __call__(self, x):
for processor in self:
x = processor(x)
return x
def length_times_10(self):
return len(self) * 10
def append_twice(self, x):
self.extend([x, x])
def _merge_criteria_processor_list(default_list, custom_list):
# simplified transformers/generation/utils.py
if len(custom_list) == 0:
return default_list
for default in default_list:
for custom in custom_list:
if type(custom) is type(default):
raise ValueError
default_list.extend(custom_list)
return default_list
class FeedForwardLayer(nn.Module):
def __init__(self, d_model, dim_feedforward, activation, dropout) -> None:
super().__init__()
self.linear1 = nn.Linear(d_model, dim_feedforward)
self.activation = activation
self.dropout1 = nn.Dropout(dropout)
self.linear2 = nn.Linear(dim_feedforward, d_model)
self.dropout2 = nn.Dropout(dropout)
def forward(self, x):
return self.dropout2(
self.linear2(self.dropout1(self.activation(self.linear1(x))))
)
class TransformerEncoderLayer(nn.Module):
def __init__(
self,
d_model,
nhead,
dim_feedforward=2048,
dropout=0.1,
activation=nn.ReLU(),
layer_norm_eps=1e-5,
):
super().__init__()
self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout)
self.norm1 = nn.LayerNorm(d_model, eps=layer_norm_eps)
self.norm2 = nn.LayerNorm(d_model, eps=layer_norm_eps)
self.dropout = nn.Dropout(dropout)
self.ff_block = FeedForwardLayer(d_model, dim_feedforward, activation, dropout)
def forward(self, src, src_mask=None, src_key_padding_mask=None):
x = src
x = self.norm1(x + self._sa_block(x, src_mask, src_key_padding_mask))
x = self.norm2(x + self._ff_block(x))
return x
# self-attention block
def _sa_block(self, x, attn_mask, key_padding_mask):
x = self.self_attn(
x,
x,
x,
attn_mask=attn_mask,
key_padding_mask=key_padding_mask,
need_weights=False,
)[0]
return self.dropout(x)
# feed forward block
def _ff_block(self, x):
return self.ff_block(x)
class MockModule(torch.nn.Module):
def inner_fn(self, left, right):
return tuple(left) == tuple(right)
def fn(self, tensor):
if type(tensor) is int:
return False
torch.add(tensor, tensor)
return self.inner_fn(tensor.shape, (1, 2, 3))
class IncByOne:
def __init__(self, x):
self.x = x + 1
class IncByTwo:
def __init__(self, x):
self.x = x + 2
class ReproTests(torch._dynamo.test_case.TestCase):
def test_do_paste_mask(self):
torch._dynamo.utils.counters.clear()
cnt = torch._dynamo.testing.CompileCounter()
opt__do_paste_mask = torch.compile(_do_paste_mask, backend=cnt)
opt__do_paste_mask(
torch.randn(1, 1, 28, 28),
torch.tensor([[0.0, 1, 2, 4]]) * 1,
427,
640,
True,
)
opt__do_paste_mask(
torch.randn(1, 1, 28, 28),
torch.tensor([[0.0, 1, 2, 4]]) * 2,
427,
640,
True,
)
opt__do_paste_mask(
torch.randn(1, 1, 28, 28),
torch.tensor([[0.0, 1, 2, 4]]) * 3,
612,
612,
True,
)
opt__do_paste_mask(
torch.randn(1, 1, 28, 28),
torch.tensor([[0.0, 1, 2, 4]]) * 4,
612,
612,
True,
)
opt__do_paste_mask(
torch.randn(1, 1, 28, 28),
torch.tensor([[0.0, 1, 2, 4]]) * 2,
427,
640,
False,
)
# (dynamic shapes, static shapes)
self.assertIn(cnt.frame_count, (5, 7))
self.assertIn(cnt.op_count, (92, 106, 119))
def test_convert_boxes_to_pooler_format(self):
boxes1 = [
Boxes(torch.arange(0, 8).reshape((2, 4))),
Boxes(torch.arange(8, 16).reshape((2, 4))),
]
boxes2 = [
Boxes(torch.arange(16, 20).reshape((1, 4))),
Boxes(torch.arange(20, 24).reshape((1, 4))),
]
correct1 = convert_boxes_to_pooler_format(boxes1)
correct2 = convert_boxes_to_pooler_format(boxes2)
fn = convert_boxes_to_pooler_format
cnt = torch._dynamo.testing.CompileCounter()
opt_fn = torch._dynamo.optimize(cnt)(fn)
self.assertTrue(same(opt_fn(boxes1), correct1))
self.assertTrue(same(opt_fn(boxes2), correct2))
# repeat_interleave is a dynamic shape operator we do not execute/
# In the future, we could reduce the frame_count down to 1
# by guarding on the exact values of `Tensor repeats` arg
if torch._dynamo.config.assume_static_by_default:
self.assertExpectedInline(cnt.frame_count, """4""")
self.assertExpectedInline(cnt.op_count, """10""")
else:
self.assertExpectedInline(cnt.frame_count, """4""")
self.assertExpectedInline(cnt.op_count, """14""")
def test_boxes_len(self):
def fn(boxes):
return len(boxes) + boxes.__len__() + boxes.tensor
boxes1 = Boxes(torch.arange(0, 8).reshape((2, 4)))
cnt = torch._dynamo.testing.CompileCounter()
opt_fn = torch._dynamo.optimize_assert(cnt)(fn)
self.assertTrue(same(opt_fn(boxes1), boxes1.tensor + 4.0))
if torch._dynamo.config.assume_static_by_default:
self.assertExpectedInline(cnt.frame_count, """1""")
self.assertExpectedInline(cnt.op_count, """1""")
else:
self.assertExpectedInline(cnt.frame_count, """1""")
self.assertExpectedInline(cnt.op_count, """2""")
def _reformer(self, nopython):
input = torch.randn([1, 64, 256])
model = ReformerEncoder()
torch.manual_seed(1337)
correct = copy.deepcopy(model)(input)
cnt = torch._dynamo.testing.CompileCounter()
torch.manual_seed(1337)
opt_model = torch._dynamo.optimize(cnt, nopython=nopython)(model)
self.assertTrue(same(opt_model(input), correct))
return cnt
@requires_cuda
def test_sub_alpha_scalar_repro(self):
@torch.compile(backend="aot_eager")
def f(x):
return x.sub(1, alpha=2)
f(torch.ones(2, device="cuda", dtype=torch.float64))
# https://github.com/pytorch/pytorch/issues/113010
def test_out_overload_non_contiguous(self):
def f(x, y):
return torch.abs(x, out=y.T)
f_compiled = torch.compile(f, backend="aot_eager")
x_ref = torch.arange(4, dtype=torch.float32).reshape(2, 2)
y_ref = torch.arange(4, dtype=torch.float32).reshape(2, 2)
x_test = torch.arange(4, dtype=torch.float32).reshape(2, 2)
y_test = torch.arange(4, dtype=torch.float32).reshape(2, 2)
out_ref = f(x_ref, y_ref)
out_test = f_compiled(x_test, y_test)
self.assertEqual(out_ref, out_test)
self.assertEqual(y_ref, y_test)
# https://github.com/pytorch/pytorch/issues/109053
def test_view_dtype_overload(self):
def f(x):
return x.view(torch.int32)
f_compiled = torch.compile(f, backend="aot_eager")
x1 = torch.ones(4, requires_grad=True)
out_ref = f(x1)
out_test = f_compiled(x1)
self.assertEqual(out_ref, out_test)
x2 = torch.ones(4, requires_grad=False)
out_ref = f(x2)
out_test = f_compiled(x2)
self.assertEqual(out_ref, out_test)
# https://github.com/pytorch/pytorch/issues/90552
def test_intermediate_leaf_requires_grad(self):
def f(x):
leaf = torch.ones(2, requires_grad=True)
return leaf, leaf * 2
f_compiled = torch.compile(f, backend="aot_eager")
x = torch.arange(4, dtype=torch.float32).reshape(2, 2)
leaf, out = f(x)
leaf_test, out_test = f_compiled(x)
out.sum().backward()
out_test.sum().backward()
self.assertEqual(leaf.grad, leaf_test.grad)
# https://github.com/pytorch/pytorch/issues/113263
def test_unpack_hooks_dont_run_during_tracing(self):
def f(x, y):
return x * y
f_compiled = torch.compile(f, backend="aot_eager")
pack_count = 0
unpack_count = 0
def pack_hook(x):
nonlocal pack_count
pack_count += 1
return x
# unpack hook shouldn't run during compilation, while we trace the forward
def unpack_hook(x):
nonlocal unpack_count
unpack_count += 1
return x
x = torch.ones(4, requires_grad=True)
y = torch.ones(4, requires_grad=False)
with torch.autograd.graph.saved_tensors_hooks(pack_hook, unpack_hook):
out_test = f_compiled(x, y)
self.assertEqual(pack_count, 1)
self.assertEqual(unpack_count, 0)
out_test.sum().backward()
self.assertEqual(pack_count, 1)
self.assertEqual(unpack_count, 1)
# https://github.com/pytorch/pytorch/issues/113263
def test_unpack_hooks_can_be_disabled(self):
def f(x, y):
return x * y
f_compiled = torch.compile(f, backend="aot_eager")
x = torch.ones(4, requires_grad=True)
y = torch.ones(4, requires_grad=False)
with torch.autograd.graph.disable_saved_tensors_hooks("hooks are disabled"):
out_test = f_compiled(x, y)
out_test.sum().backward()
# https://github.com/pytorch/pytorch/issues/113263
def test_disabling_unpack_hooks_within_compiled_region(self):
def g(z):
with torch.autograd.graph.disable_saved_tensors_hooks("hooks are disabled"):
return z + 5
def f(x, y):
z = x * y
return g(z)
f_compiled = torch.compile(f, backend="aot_eager")
x = torch.ones(4, requires_grad=True)
y = torch.ones(4, requires_grad=False)
out_test = f_compiled(x, y)
out_test.sum().backward()
# See https://github.com/pytorch/pytorch/issues/97745
def test_gan_repro_trying_to_backward_through_the_graph_a_second_time(self):
def f(a, b):
c = torch.ones(2, 2)
d = torch.ones(2, 2)
e = torch.matmul(a, c)
g_loss = torch.abs(e - d).mean()
g_loss.backward()
fake_d_pred = torch.matmul(b, e.detach())
d_loss = fake_d_pred.mean()
d_loss.backward()
a_ref = torch.randn(2, 2, requires_grad=True)
b_ref = torch.randn(2, 2, requires_grad=True)
out_ref = f(a_ref, b_ref)
a_test = a_ref.clone().detach().requires_grad_(True)
b_test = b_ref.clone().detach().requires_grad_(True)
out_test = torch.compile(f, backend="aot_eager")(a_test, b_test)
self.assertEqual(out_ref, out_test)
self.assertEqual(a_ref.grad, a_test.grad)
self.assertEqual(b_ref.grad, b_test.grad)
# https://github.com/pytorch/pytorch/issues/111603
def test_tuple_enum_as_key_dict(self):
class MyEnum(Enum):
A = "a"
class SomeModel(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.linear = torch.nn.Linear(1, 1)
def forward(self, x) -> torch.Tensor:
return self.linear(x[MyEnum.A])
x = {MyEnum.A: torch.rand(8, 1)}
model_pytorch = SomeModel()
model = torch.compile(model_pytorch)
# Executing twice works
model(x)
y = model(x)
self.assertEqual(y, model_pytorch(x))
def test_embedding_backward_broadcasting_decomp(self):
def f(grad_output, indices):
num_weights = 10
padding_idx = 1
scale_grad_by_freq = True
return torch.ops.aten.embedding_dense_backward(
grad_output, indices, num_weights, padding_idx, scale_grad_by_freq
)
f_compiled = torch.compile(f, backend="aot_eager")
grad_output = torch.ones(2, 4, 3, dtype=torch.float16)
indices = torch.ones(2, 4, dtype=torch.int64)
out_ref = f(grad_output, indices)
out_test = f_compiled(grad_output, indices)
self.assertEqual(out_ref, out_test)
def test_reformer_eval(self):
with torch.no_grad():
cnt = self._reformer(nopython=True)
self.assertEqual(cnt.frame_count, 1)
self.assertEqual(cnt.op_count, 11)
def test_reformer_train(self):
with torch.enable_grad():
cnt = self._reformer(nopython=False)
expected_op_count = (
"""11""" if torch._dynamo.config.inline_inbuilt_nn_modules else """5"""
)
self.assertExpectedInline(cnt.frame_count, """1""")
self.assertExpectedInline(cnt.op_count, expected_op_count)
@disable_translation_validation_if_dynamic_shapes
def test_longformer_chunk(self):
input1 = torch.randn([1, 4096, 1])
input2 = torch.randn([12, 4096, 64])
correct1 = longformer_chunk(input1)
correct2 = longformer_chunk(input2)
fn = longformer_chunk
cnt = torch._dynamo.testing.CompileCounter()
opt_fn = torch._dynamo.optimize_assert(cnt)(fn)
self.assertTrue(same(opt_fn(input1), correct1))
self.assertTrue(same(opt_fn(input2), correct2))
self.assertTrue(same(opt_fn(input1), correct1))
self.assertTrue(same(opt_fn(input2), correct2))
if torch._dynamo.config.assume_static_by_default:
if torch._dynamo.config.automatic_dynamic_shapes:
self.assertExpectedInline(cnt.frame_count, """2""")
self.assertExpectedInline(cnt.op_count, """8""")
else:
self.assertExpectedInline(cnt.frame_count, """2""")
self.assertExpectedInline(cnt.op_count, """4""")
else:
self.assertExpectedInline(cnt.frame_count, """2""")
self.assertExpectedInline(cnt.op_count, """19""")
def test_hf_t5_forward(self):
input = torch.randn([1, 2048, 512])
model = PartialT5()
correct = model(input)
cnt = torch._dynamo.testing.CompileCounter()
opt_model = torch._dynamo.optimize_assert(cnt)(model)
self.assertTrue(same(opt_model(input), correct))
if torch._dynamo.config.assume_static_by_default:
self.assertExpectedInline(cnt.frame_count, """1""")
self.assertExpectedInline(cnt.op_count, """11""")
else:
self.assertExpectedInline(cnt.frame_count, """1""")
self.assertExpectedInline(cnt.op_count, """11""")
def test_module_in_skipfiles(self):
model = nn.Linear(10, 10)
cnt = torch._dynamo.testing.CompileCounter()
torch.compile(model, backend=cnt, fullgraph=True)(torch.randn([5, 10]))
self.assertEqual(cnt.frame_count, 1)
self.assertEqual(cnt.op_count, 1)
def test_function_in_skipfiles(self):
cnt = torch._dynamo.testing.CompileCounter()
torch.compile(torch.sin, backend=cnt, fullgraph=True)(torch.randn([5, 10]))
self.assertEqual(cnt.frame_count, 1)
self.assertEqual(cnt.op_count, 1)
def test_slicing_dynamic_shape(self):
def fn(y):
x = torch.ones(8)
idx = y[0]
out = x[idx:]
return (out + 3) * 5
counter = torch._dynamo.testing.CompileCounter()
opt_fn = torch._dynamo.optimize(counter)(fn)
out = opt_fn(torch.ones(10, dtype=torch.long))
# idx should be 1 -> slicing off [1:] of 8 elem tensor
self.assertEqual(list(out.shape), [7])
self.assertEqual(counter.op_count, 2)
self.assertEqual(counter.frame_count, 1)
self.assertEqual(list(opt_fn(torch.tensor([4])).shape), [4])
def test_slicing_dynamic_shape_setitem(self):
def fn(input_lengths: torch.Tensor, new_ones_1):
getitem_13 = input_lengths[3]
new_ones_1[(3, slice(getitem_13, None, None))] = 0
setitem_13 = new_ones_1
return (setitem_13,)
x = torch.randn(10).to(dtype=torch.int64)
y = torch.randn(10, 204)
ref = fn(x, y)
opt_fn = torch._dynamo.optimize("aot_eager")(fn)
res = opt_fn(x, y)
self.assertTrue(same(ref, res))
@torch._dynamo.config.patch(error_on_recompile=True)
@torch.fx.experimental._config.patch(use_duck_shape=False)
def test_dynamic_shape_disable_duck_size(self):
class TestModel(nn.Module):
def __init__(
self,
):
super().__init__()
def forward(self, x: torch.Tensor, val: int) -> torch.Tensor:
return x + val
main_model = TestModel().to(memory_format=torch.channels_last)
opt_model = torch.compile(main_model, backend="eager", dynamic=True)
x1 = torch.rand(2, 5, 10, 10).to(memory_format=torch.channels_last)
x2 = torch.rand(2, 5, 4, 8).to(memory_format=torch.channels_last)
o1_ref = main_model(x1, 4)
o1 = opt_model(x1, 4)
o2_ref = main_model(x2, 20)
o2 = opt_model(x2, 20)
def test_chunk_reformer_ff(self):
input = torch.randn([1, 4096, 256])
model = ChunkReformerFeedForward()
correct = model(input)
cnt = torch._dynamo.testing.CompileCounter()
opt_model = torch._dynamo.optimize_assert(cnt)(model)
self.assertTrue(same(opt_model(input), correct))
self.assertEqual(cnt.frame_count, 1)
self.assertLessEqual(cnt.op_count, 10)
# see: https://github.com/pytorch/pytorch/issues/80067
# NB: When you remove the expectedFailure, don't forget to
# uncomment/adjust the assertEqual below
@unittest.expectedFailure
@torch._dynamo.config.patch(
fake_tensor_propagation=True, capture_scalar_outputs=True
)
def test_maml_item_capture(self):
a = torch.randn(5, 1, 28, 28)
b = torch.zeros(5, dtype=torch.int64)
c = torch.randn(75, 1, 28, 28)
d = torch.zeros(75, dtype=torch.int64)
model = PartialMaml()
correct = model(a, b, c, d)
cnt = torch._dynamo.testing.CompileCounter()
opt_model = torch._dynamo.optimize(cnt)(model)
for _ in range(10):
self.assertTrue(same(opt_model(a, b, c, d), correct))
# if torch._dynamo.config.assume_static_by_default:
# self.assertExpectedInline(cnt.frame_count, """2""")
# else:
# self.assertExpectedInline(cnt.frame_count, """3""")
# TODO(jansel): figure out why op count depends on imports
self.assertIn(cnt.op_count, (36, 35, 34, 29, 28, 27))
# see: https://github.com/pytorch/pytorch/issues/80067
@torch._dynamo.config.patch(capture_scalar_outputs=False)
def test_maml_no_item_capture(self):
a = torch.randn(5, 1, 28, 28)
b = torch.zeros(5, dtype=torch.int64)
c = torch.randn(75, 1, 28, 28)
d = torch.zeros(75, dtype=torch.int64)
model = PartialMaml()
correct = model(a, b, c, d)
cnt = torch._dynamo.testing.CompileCounter()
opt_model = torch._dynamo.optimize(cnt)(model)
for _ in range(10):
self.assertTrue(same(opt_model(a, b, c, d), correct))
if torch._dynamo.config.assume_static_by_default:
self.assertExpectedInline(cnt.frame_count, """4""")
else:
self.assertExpectedInline(cnt.frame_count, """5""")
def test_hf_model_output(self):
ex = ModelOutput(a=torch.randn(10), b=torch.randn(10), c=torch.randn(10))
def fn1(x):
return x["a"] + 1
def fn2(x):
return x.a + 1
def fn3(x):
return x.to_tuple()[0] + 1
def fn4(x):
return x[0] + 1
cnt = torch._dynamo.testing.CompileCounter()
for fn in (fn1, fn2, fn3, fn4):
cnt.clear()
opt_fn = torch._dynamo.optimize_assert(cnt)(fn)
self.assertTrue(same(opt_fn(ex), ex.a + 1))
self.assertEqual(cnt.frame_count, 1)
self.assertEqual(cnt.op_count, 1)
@disable_translation_validation_if_dynamic_shapes
def test_create_rand_mask_from_inputs(self):
args = [
torch.randn([1, 64, 64]),
torch.randn([1, 64, 64]),
torch.zeros([1, 12, 62, 3], dtype=torch.int64),
12,
3,
1,
4096,
64,
]
correct = create_rand_mask_from_inputs(*args)
fn = create_rand_mask_from_inputs
cnt = torch._dynamo.testing.CompileCounter()
opt_fn = torch._dynamo.optimize_assert(cnt)(fn)
self.assertTrue(same(opt_fn(*args), correct))
if torch._dynamo.config.assume_static_by_default:
self.assertExpectedInline(cnt.frame_count, """1""")
self.assertExpectedInline(cnt.op_count, """8""")
else:
self.assertExpectedInline(cnt.frame_count, """1""")
self.assertExpectedInline(cnt.op_count, """11""")
def test_rng_state(self):
def fn():
state = torch.get_rng_state()
before = torch.rand(1000)
torch.set_rng_state(state)
after = torch.rand(1000)
return before, after
cnt = torch._dynamo.testing.CompileCounter()
opt_fn = torch._dynamo.optimize(cnt)(fn)
before, after = opt_fn()
self.assertTrue(same(before, after))
self.assertEqual(cnt.frame_count, 2)
self.assertEqual(cnt.op_count, 2) # rand, rand
try:
graph, _ = torch._dynamo.export(fn)()
# See https://github.com/pytorch/pytorch/pull/87490
self.fail("unexpected export success")
except torch._dynamo.exc.Unsupported:
pass
def test_threading_local(self):
import threading
foo = threading.local()
foo.x = torch.rand(1)
def f(x):
return torch.cat([x, foo.x])
cnt = torch._dynamo.testing.CompileCounter()
opt_f = torch._dynamo.optimize(cnt, nopython=True)(f)
inp = torch.ones(1)
out = f(inp)
opt_out = opt_f(inp)
self.assertEqual(opt_out, out)
self.assertEqual(cnt.frame_count, 1)
def test_seq_append_list(self):
x = torch.randn(4, 10)
model = SequentialAppendList(
torch.nn.Linear(10, 10),
torch.nn.ReLU(),
torch.nn.Linear(10, 10),
torch.nn.ReLU(),
)
# this one is tricky because it mutates the list provided as an input
l1 = [x]
l2 = [x]
correct, _ = model(x, l1)
cnt = torch._dynamo.testing.CompileCounter()
opt_model = torch._dynamo.optimize_assert(cnt)(model)
result, l3 = opt_model(x, l2)
self.assertTrue(same(result, correct))
self.assertTrue(same(l1, l2))
self.assertIs(l2, l3)
self.assertEqual(cnt.frame_count, 1)
self.assertEqual(cnt.op_count, 5)
def test_batch_norm_act(self):
a = torch.randn(5, 1, 28, 28)
model = BatchNormAct2d(1).eval()
correct = model(a)
cnt = torch._dynamo.testing.CompileCounter()
if not torch._dynamo.config.specialize_int:
# _local_scalar_dense causes graph break w 0-dim tensor
opt_model = torch._dynamo.optimize(cnt)(model)
self.assertTrue(same(opt_model(a), correct))
return
opt_model = torch._dynamo.optimize_assert(cnt)(model)
self.assertTrue(same(opt_model(a), correct))
self.assertEqual(cnt.frame_count, 1)
self.assertEqual(cnt.op_count, 2)
def test_get_parameter_dtype(self):
model = SequentialAppendList(
torch.nn.Linear(10, 10),
torch.nn.ReLU(),
)
def fn(model, x):
return x + torch.randn(10, dtype=get_parameter_dtype(model))
cnt = torch._dynamo.testing.CompileCounter()
opt_fn = torch._dynamo.optimize_assert(cnt)(fn)
self.assertEqual(opt_fn(model, torch.randn(10)).dtype, torch.float32)
self.assertEqual(cnt.frame_count, 1)
self.assertEqual(cnt.op_count, 2)
def test_nn_parameter(self):
def test_fn():
a = torch.nn.Parameter(torch.randn(5, 5))
# Checks that TensorVariable stores the type information correctly
self.assertTrue(isinstance(a, torch.nn.Parameter))
return a
cnt = torch._dynamo.testing.CompileCounter()
opt_test_fn = torch._dynamo.optimize(cnt)(test_fn)
out = opt_test_fn()
self.assertTrue(isinstance(out, torch.nn.Parameter))
def test_Size(self):
def test_fn():
a = torch.randn(4)
x = torch.Size([1, 2, 3])
# Checks that SizeVariable return torch.Size object
assert isinstance(x, torch.Size)
# Causes graph breaks and checks reconstruction of SizeVariable
# object
self.assertIsInstance(x, torch.Size)
return a
cnt = torch._dynamo.testing.CompileCounter()
opt_test_fn = torch._dynamo.optimize(cnt)(test_fn)
opt_test_fn()
# See https://github.com/pytorch/pytorch/issues/100067
def test_copy_weird_strides(self):
# This test requires inductor's copy() decomp to preserve strides properly.
def test_fn(a):
b = torch.zeros(48, 4, 256, 513)
b[:, 0, 1:256, 1:256] = a
c = b.view(4, 12, 1024, 513)
d = c.transpose(2, 1)
d.add_(1)
return d
sh, st, dt, dev, rg = (
(48, 255, 255),
(787968, 513, 1),
torch.float16,
"cpu",
True,
)
a = rand_strided(sh, st, dt, dev).requires_grad_(rg)
compiled_f = torch.compile(test_fn, backend="aot_eager_decomp_partition")
out1 = test_fn(a)
out2 = compiled_f(a)
self.assertEqual(out1, out2)
def test_indexing_with_list(self):
def test_fn():
def run_test(tensor, *idx):
npt = tensor.numpy()
assert npt[idx].shape == tensor[idx].shape
x = torch.arange(0, 10)
cases = [
[None, None],
[1, None],
]
for case in cases:
run_test(x, *case)
return torch.randn(4)
cnt = torch._dynamo.testing.CompileCounter()
opt_test_fn = torch._dynamo.optimize(cnt)(test_fn)
opt_test_fn()
def test_reformer_min_chunk_len(self):
def fn(cfg):
t = torch.empty(10)
t.fill_(_get_min_chunk_len(cfg))
return t[0]
cfg = DummyConfig()
cnt = torch._dynamo.testing.CompileCounter()
opt_fn = torch._dynamo.optimize_assert(cnt)(fn)
self.assertEqual(opt_fn(cfg), 64)
# With unspec int, maximum computation is preserved
self.assertExpectedInline(cnt.frame_count, """1""")
self.assertExpectedInline(cnt.op_count, """3""")
def test_reformer_sorting(self):
x = torch.zeros([1, 12, 4096], dtype=torch.int64)
correct = _get_sorted_bucket_idx_and_undo_sorted_bucket_idx(x)
fn = _get_sorted_bucket_idx_and_undo_sorted_bucket_idx
cnt = torch._dynamo.testing.CompileCounter()
opt_fn = torch._dynamo.optimize_assert(cnt)(fn)
self.assertTrue(same(opt_fn(x), correct))
if torch._dynamo.config.assume_static_by_default:
self.assertExpectedInline(cnt.frame_count, """1""")
self.assertExpectedInline(cnt.op_count, """14""")
else:
self.assertExpectedInline(cnt.frame_count, """1""")
self.assertExpectedInline(cnt.op_count, """16""")
def test_recursive_map(self):
# https://github.com/pytorch/torchdynamo/issues/132
def _recursive_map(struct, batch_dim=0):
for k, v in struct.items():
if v is not None:
if isinstance(v, dict):
_recursive_map(v)
else:
struct[k] = v
def toy_example(a, b, v):
x = a / (torch.abs(a) + 1)
if v is not None:
_recursive_map(v)
return x * b
cnt = torch._dynamo.testing.CompileCounter()
opt_toy_example = torch._dynamo.optimize(cnt)(toy_example)
opt_toy_example(
torch.randn(10),
torch.randn(10),
{"layer0": {"memory_keys": torch.randn(10)}},
)
self.assertEqual(cnt.frame_count, 1)
self.assertEqual(cnt.op_count, 4)
def test_issue114171(self):
device = torch.device("cpu")
def fcnn(in_dim, out_dim, hidden_dim, activation=torch.nn.GELU):
layers = [
torch.nn.Linear(in_dim, hidden_dim, device=device),
activation(),
torch.nn.Linear(hidden_dim, out_dim, device=device),
]
return torch.nn.Sequential(*layers)
class testmodel(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.interaction_networks = torch.nn.ModuleList(
[fcnn(262, 1174, 400) for _ in range(4)]
)
def interact(self, x, cycle):
return self.interaction_networks[cycle](x)
model = testmodel()
forward_aot = torch.compile(
model.interact, fullgraph=True, dynamic=True, backend="eager"
)
x = torch.rand([111, 262], device=device)
y2 = forward_aot(x, 2) # previously failed
def test_issue175(self):
n_heads = 2
d_model = 64
model = TransformerEncoderLayer(d_model, n_heads)
inp = torch.randn(1, d_model)
cnt = torch._dynamo.testing.CompileCounter()
opt_model = torch._dynamo.optimize(cnt, nopython=True)(model)
opt_model(inp)
opt_model(inp)
self.assertEqual(cnt.frame_count, 1)
self.assertEqual(12, cnt.op_count)
def test_exec_import(self):
def fn1():
exec("import math")
def fn2():
try:
math.sqrt(4)
return False
except NameError:
return True
def fn3():
fn1()
return fn2()
self.assertTrue(fn3())
opt_fn3 = torch._dynamo.optimize("eager")(fn3)
self.assertTrue(opt_fn3())
def test_exec_wildcard_import(self):
# Test that globals are not carried over from frame to frame
def fn1():
exec("from torch import *")
def fn2():
x = torch.zeros(4)
for i in range(5):
x = x + i
return x
def fn3():
fn1()
return fn2()
ref = fn3()
opt_fn3 = torch._dynamo.optimize("eager")(fn3)
res = opt_fn3()
self.assertTrue(same(ref, res))
def test_with_on_graph_break_inst(self):
def reversible(x):
print("Hello world") # Cause graph break so inline fails
return torch.sin(torch.cos(x))
def fn(x):
with torch.enable_grad():
a = torch.sin(x)
b = reversible(a)
c = torch.sigmoid(b)
c.sum().backward()
return x.grad
x = torch.randn(3, requires_grad=True)
x.grad = None
with torch.no_grad():
ref = fn(x)
x.grad = None
opt_fn = torch._dynamo.optimize("eager")(fn)
with torch.no_grad():
res = opt_fn(x)
self.assertTrue(same(ref, res))
def test_with_on_graph_break_nested(self):
def reversible(x):
torch._dynamo.graph_break() # Cause graph break so inline fails
return torch.sin(torch.cos(x))
def fn(x):
# nested context manager failed previously
with torch.no_grad():
with torch.enable_grad():
a = torch.sin(x)
b = reversible(a)
c = torch.sigmoid(b)
c.sum().backward()
return x.grad
x = torch.randn(3, requires_grad=True)
x.grad = None
with torch.no_grad():
ref = fn(x)
x.grad = None
opt_fn = torch._dynamo.optimize("eager")(fn)
with torch.no_grad():
res = opt_fn(x)
self.assertTrue(same(ref, res))
# https://github.com/pytorch/torchdynamo/issues/1446
def test_grad_mode_carrying_correct_state_after_graph_break(self):
def fn(x):
with torch.no_grad():
y = x * 3
print("Break")
z = x + 2
return y, z
x = torch.randn(3, requires_grad=True)
opt_fn = torch._dynamo.optimize("eager")(fn)
y, z = opt_fn(x)
self.assertFalse(y.requires_grad)
self.assertFalse(z.requires_grad)
def test_abc_setattr(self):
# tests that we correctly bail out of __setattr__ calls
# TODO: does not ensure ABC classes are correctly inferred as ClassVariables
# (doesn't test the fix for 'super()')
class BaseModule(torch.nn.Module, ABC):
def blah(self, x):
return x + 1
class Derived(BaseModule):
def __setattr__(self, name, value) -> None:
super().__setattr__(name, value)
def forward(self, x):
# expect a graph break on __setattr__
self.foo = 0
return self.blah(x)
def blah(self, x):
return super().blah(x)
x = torch.randn(3, requires_grad=True)
mod = Derived()
opt_mod = torch._dynamo.optimize("eager")(mod)
opt_mod(x)
# Not sure what this test is testing. It was earlier graph breaking on
# __dict__, so the counter >= 2. With __dict__ support, there is no
# graph break.
self.assertGreaterEqual(torch._dynamo.utils.counters["frames"]["ok"], 1)
self.assertGreaterEqual(torch._dynamo.utils.counters["frames"]["total"], 1)
@torch._dynamo.config.patch("suppress_errors", True)
def test_guard_fail_tensor_bool(self):
@torch._dynamo.disable(recursive=False)
def fn():
condition_shape = (5, 5)
dtypes = (torch.bool,)
shapes = (
(),
(5,),
(1, 5),
)
tensors = [
torch.empty(shape, dtype=dtype).fill_(17)
for shape, dtype in itertools.product(shapes, dtypes)
]
x_vals = (5.0, *tensors)
y_vals = (6.0, *tensors)
@torch._dynamo.disable
def get_expected(condition, x, y):
x_np = x.cpu().numpy() if isinstance(x, torch.Tensor) else x
y_np = y.cpu().numpy() if isinstance(y, torch.Tensor) else y
return torch.from_numpy(
np.where(condition.cpu().numpy(), x_np, y_np)
).to(common_dtype)
for x, y in zip(x_vals, y_vals):
condition = torch.empty(*condition_shape, dtype=torch.bool).bernoulli_()
common_dtype = torch.result_type(x, y)
def check_equal(condition, x, y):
# NumPy aggressively promotes to double, hence cast to output to correct dtype
expected = get_expected(condition, x, y)
result = torch.where(condition, x, y)
assert torch.allclose(expected, result)
check_equal(condition, x, y)
check_equal(condition, y, x)
fn()
opt_fn = torch._dynamo.optimize("eager")(fn)
opt_fn()
def test_guard_fail_nested_tuple(self):
def fn(args):
return torch.ones(()), args[0] * 2
# This adds a tensor check on args[1][0] and args[1][1]
args1 = (torch.ones(1), (torch.ones(1), torch.ones(1)))
args2 = (torch.ones(1), torch.ones(1))
opt_fn = torch._dynamo.optimize("eager")(fn)
ref = opt_fn(args1)
res = opt_fn(args2)
self.assertTrue(same(ref, res))
def test_nullcontext1(self):
@torch.compile(fullgraph=True, backend="eager")
def fn(x, ctx):
x = x.sin()
with ctx:
x = x.cos()
x = x.sin()
return x
y = torch.randn(10)
self.assertTrue(same(fn(y, contextlib.nullcontext()), y.sin().cos().sin()))
def test_nullcontext2(self):
@torch.compile(fullgraph=True, backend="eager")
def fn(x, ctx):
x = x.sin()
with ctx():
x = x.cos()
x = x.sin()
return x
y = torch.randn(10)
self.assertTrue(same(fn(y, contextlib.nullcontext), y.sin().cos().sin()))
def test_no_grad_inline(self):
@torch.no_grad()
def a(x):
return x.sin()
@torch.compile(backend="eager", fullgraph=True)
def b(x):
return a(x).cos()
y = torch.randn(10)
self.assertTrue(same(b(y), y.sin().cos()))
@skipIfWindows(
msg="torch._dynamo.exc.TorchRuntimeError: Failed running call_function <class 'torch.LongTensor'>(*(FakeTensor(..., size=(10,), dtype=torch.int32),), **{}):" # noqa: B950
)
def test_longtensor_list(self):
for partition in [0, 5, 10]:
@torch._dynamo.disable
def rand_gen():
rand_vals = [random.randint(5, 10) for _ in range(10)]
# List of tensors mixed with np.arrays
return list(np.array(rand_vals[:partition])) + [
torch.tensor(val) for val in rand_vals[partition:]
]
def fn(x):
random_list = rand_gen()
z = torch.LongTensor(random_list)
return x * z
x = torch.ones(10) * 2
random.seed(0)
ref0 = fn(x)
ref1 = fn(x)
random.seed(0)
opt_fn = torch._dynamo.optimize("eager")(fn)
res0 = opt_fn(x)
res1 = opt_fn(x)
self.assertTrue(same(ref0, res0))
self.assertTrue(same(ref1, res1))
def test_primtorch(self):
@torch._dynamo.optimize("eager")
def fn(x):
torch._refs.abs(x)
fn(torch.randn(3))
@unittest.expectedFailure
# inline_call [('inline in skipfiles: bind ...python3.10/inspect.py', 1)]
def test_primtorch_no_graph_break(self):
@torch._dynamo.optimize("eager", nopython=True)
def fn(x):
torch._refs.abs(x)
fn(torch.randn(3))
def test_torch_tensor_ops_no_graph_break(self):
@torch._dynamo.optimize("eager", nopython=True)
def fn(x):
torch.Tensor.abs_(x)
fn(torch.randn(3))
@unittest.skipIf(
not isinstance(torch.ops.aten.abs, torch._ops.OpOverloadPacket),
"old pt doesn't work",
)
def test_torch_ops_aten(self):
# Picked an op that doesn't show up in the default list
@torch._dynamo.optimize("eager", nopython=True)
def fn(x):
return torch.ops.aten.absolute(x)
fn(torch.randn(3))
def test_hf_gelu_inline(self):
class GELUActivation(nn.Module):
def __init__(self) -> None:
super().__init__()
self.act = nn.functional.gelu
def forward(self, input):
return self.act(input)
@torch._dynamo.optimize("eager", nopython=True)
def fn(x):
return GELUActivation()(x)
y = torch.randn(10)
self.assertTrue(same(fn(y), nn.functional.gelu(y)))
@torch._dynamo.optimize("eager", nopython=True)
def fn_returns(x):
return GELUActivation(), x + 1
act, _ = fn_returns(y)
self.assertIsInstance(act, GELUActivation)
self.assertIs(act.act, nn.functional.gelu)
self.assertTrue(hasattr(act, "_buffers")) # check that __init__ got called
def test_dropout_inline(self):
@torch._dynamo.optimize("eager")
def fn(x):
return torch.nn.Dropout(0.1)(x)
y = torch.randn(10)
torch.manual_seed(1337)
ref = nn.functional.dropout(y, 0.1)
torch.manual_seed(1337)
res = fn(y)
self.assertTrue(same(ref, res))
def test_setitem_boolean_mask_diff(self):
def fn(x, b, y):
x = x.clone()
x[b] = y
return x
opt_fn = torch._dynamo.optimize("aot_eager")(fn)
x = torch.randn(4, requires_grad=True)
b = torch.tensor([True, False, True, False])
y = torch.randn(2, requires_grad=True)
opt_fn(x, b, y)
def test_setitem_tuple_boolean_mask_diff(self):
def fn(x, b, y):
x = x.clone()
x[:, b] = y
return x
opt_fn = torch._dynamo.optimize("aot_eager")(fn)
x = torch.randn(8, 4, requires_grad=True)
b = torch.tensor([True, False, True, False])
y = torch.randn(2, requires_grad=True)
opt_fn(x, b, y)
def test_torch_tensor_ops(self):
def fn(x):
return torch.Tensor.abs_(x)
x = torch.randn(3)
opt_fn = torch._dynamo.optimize("eager", nopython=True)(fn)
y = fn(x)
y_ = opt_fn(x)
self.assertTrue(same(y, y_))
def test_guard_ordering_shape_fail(self):
# If a function which takes a tensor has an inner function which
# is compiled and generates a guard on its shape,
# they are evaluated in the wrong order. So if on a subsequent call
# an int is passed instead of a tensor, guard evaluation will crash
# with a "no attribute: shape" error
m = MockModule()
opt_m = torch._dynamo.optimize("eager")(m)
opt_m.fn(torch.ones((5, 5)))
opt_m.fn(-3)
def test_tensor_isinstance_tuple(self):
@torch._dynamo.optimize("eager")
def fn():
t = torch.ones(5, 5)
if not isinstance(t, (int, torch.Tensor)):
msg = str.format(
"{0} is not an instance of {1}",
type(t),
(int, torch.Tensor),
)
raise ValueError(msg)
return True
fn()
def test_isinstance_dtype(self):
@torch._dynamo.optimize("eager", nopython=True)
def fn(x):
isinstance(torch.bfloat16, torch.dtype)
return x
fn(torch.randn(3))
def test_isinstance_storage(self):
@torch._dynamo.optimize("eager")
def fn(x):
f = bytearray([0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x10, 0x40])
bools = torch.BoolStorage.from_buffer(f, "big")
assert isinstance(bools, torch.BoolStorage)
return x
fn(torch.randn(3))
def test_issue111522(self):
@torch.compile(backend="eager", fullgraph=True)
def f(x, y):
return x + y.a
class A:
a = 2
self.assertEqual(f(torch.zeros(2), A()), torch.full([2], 2.0))
del A.a
# graph break on missing attr
with self.assertRaises(torch._dynamo.exc.Unsupported):
f(torch.zeros(2), A())
def test_dict_list_values(self):
def inner_fn(args):
return [x[1].shape for x in args]
@torch._dynamo.optimize("eager")
def fn(tensors):
return inner_fn(zip(itertools.count(), tensors["args"]))
fn({"args": [torch.ones(5, 5), torch.ones(5, 6), torch.ones(5, 7)]})
fn({"args": [torch.ones(5, 5)]})
def test_dict_iter(self):
class MyMod(torch.nn.Module):
def forward(self, x):
z = {"my": 1, "const": 2, "dict": 3, "variable": 4}
tot = 0
for key in z:
tot += z[key]
return tot
x = torch.tensor([0])
model = MyMod()
opt_model = torch._dynamo.optimize("eager", nopython=True)(model)
y = opt_model(x)
self.assertEqual(y, 10)
def test_sort_out(self):
dtype = torch.float32
device = "cpu"
def fn():
tensor = torch.randn((3, 5), dtype=dtype, device=device)[:, 0]
values1 = torch.tensor(0, dtype=dtype, device=device)
indices1 = torch.tensor(0, dtype=torch.long, device=device)
torch.sort(tensor, out=(values1, indices1))
self.assertEqual(values1.stride(), (1,))
self.assertEqual(indices1.stride(), (1,))
fn()
opt_fn = torch._dynamo.optimize("eager")(fn)
opt_fn()
def test_sort_out2(self):
class MyModule(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.sorted = torch.nn.Buffer(torch.ones(4, 4))
self.indices = torch.nn.Buffer(torch.ones(4, 4, dtype=torch.long))
def forward(self, x):
torch.sort(x, out=(self.sorted, self.indices))
return (x + 1, self.sorted, self.indices)
x = torch.randn(4, 4)
m = MyModule()
ref = m(x)
opt_m = torch._dynamo.optimize("eager")(m)
res = opt_m(x)
self.assertTrue(same(ref, res))
def test_sigmoid_out(self):
dtype = torch.float32
device = "cpu"
def fn():
inp = torch.randn((3, 5), dtype=dtype, device=device)
out1 = torch.tensor(0, dtype=dtype, device=device)
torch.sigmoid(inp, out=out1)
self.assertEqual(out1.numel(), 15)
fn()
opt_fn = torch._dynamo.optimize("eager")(fn)
opt_fn()
def test_sigmoid_out2(self):
class MyModule(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.base = torch.nn.Buffer(torch.ones(4, 4))
def forward(self, x):
torch.sigmoid(x, out=self.base)
return x + self.base
x = torch.randn(4, 4)
m = MyModule()
ref = m(x)
opt_m = torch._dynamo.optimize("eager")(m)
res = opt_m(x)
self.assertTrue(same(ref, res))
def test_out_root_cell_shape_change(self):
@torch.compile(backend="eager")
def fn():
out = torch.empty(0)
def run():
x = torch.zeros(3, 5)
torch.sigmoid(x, out=out)
return out.size()
return run()
res = fn()
self.assertEqual((3, 5), res)
def test_out_nested_cell_shape_change(self):
@torch.compile(backend="eager")
def fn():
def run():
x = torch.zeros(3, 5)
out = torch.empty(0)
def capture():
return out # Force `out` to be a nested cell
torch.sigmoid(x, out=out)
return out.size()
return run()
res = fn()
self.assertEqual((3, 5), res)
def test_out_root_cell_tuple_shape_change(self):
@torch.compile(backend="eager")
def fn():
out1 = torch.empty(0)
out2 = torch.empty(0, dtype=torch.long)
def run():
x = torch.zeros(3, 5)
torch.sort(x, out=(out1, out2))
return out1.size(), out2.size()
return run()
res = fn()
self.assertEqual(((3, 5), (3, 5)), res)
def test_out_nested_cell_tuple_shape_change(self):
@torch.compile(backend="eager")
def fn():
def run():
x = torch.zeros(3, 5)
out1 = torch.empty(0)
out2 = torch.empty(0, dtype=torch.long)
def capture():
# Force `out1` and `out2` to be nested cells
return out1, out2
torch.sort(x, out=(out1, out2))
return out1.size(), out2.size()
return run()
res = fn()
self.assertEqual(((3, 5), (3, 5)), res)
def test_slice_into_list_mutable(self):
class Mod(torch.nn.Module):
def forward(self, listy):
x = listy[3:5]
for i in range(10):
z = torch.abs(torch.randn(10)) + 1
x[0] = z
return x
m = Mod()
listy = [torch.randn(10)] * 10
cnt = torch._dynamo.testing.CompileCounter()
opt_m = torch._dynamo.optimize(cnt, nopython=True)(m)
opt_m.forward(listy)
self.assertEqual(cnt.frame_count, 1)
@torch._dynamo.config.patch(capture_scalar_outputs=True)
def test_issue111918(self):
cnt = CompileCounter()
@torch.compile(backend=cnt, dynamic=True)
def fn(x):
x = x + 1
y = x.item()
if y > 2:
return x * 2
else:
return x * 3
x = torch.tensor([3.0])
fn(x)
self.assertEqual(cnt.frame_count, 2)
self.assertEqual(cnt.op_count, 4)
torch._dynamo.reset()
fn = torch.compile(fn, fullgraph=True, backend="eager")
with self.assertRaises(torch._dynamo.exc.UserError):
fn(x)
def test_vdd_duplicate_error(self):
def fn(a, dt):
keys = list(dt._jt_dict.keys())
p = torch.cos(dt._jt_dict[keys[0]]._value)
q = torch.sin(a)
r = torch.sigmoid(dt._jt_dict[keys[0]]._value)
return p + q + r
class Value:
def __init__(self) -> None:
self._value = torch.randn(4)
class Sample:
def __init__(self) -> None:
self._jt_dict = {}
self._jt_dict["POSITION_ID"] = Value()
a = torch.randn(4)
sample = Sample()
ref = fn(a, sample)
optimized_fn = torch._dynamo.optimize("eager", nopython=True)(fn)
res = optimized_fn(a, sample)
self.assertTrue(same(ref, res))
def test_specialized_stride(self):
def f():
e = torch.empty(4)
x = e[::2]
return x.stride()
self.assertEqual(f(), torch._dynamo.optimize("eager")(f)())
def test_out_none(self):
# https://github.com/pytorch/pytorch/issues/92814
def fn(input):
return torch.nn.functional.normalize(input, dim=0, out=None)
x = torch.rand([1])
self.assertEqual(fn(x), torch._dynamo.optimize("eager")(fn)(x))
def test_multi_import(self):
if not has_detectron2():
raise unittest.SkipTest("requires detectron2")
@torch._dynamo.optimize("eager", nopython=True)
def to_bitmasks(boxes):
from detectron2.layers.mask_ops import (
_paste_masks_tensor_shape,
paste_masks_in_image,
)
if (
paste_masks_in_image is not None
and _paste_masks_tensor_shape is not None
):
return boxes + 1
self.assertTrue((to_bitmasks(torch.zeros(10)) == torch.ones(10)).all())
def test_multi_dot_import(self):
def fn1(x):
return torch.sin(x)
def fn(x):
import torch.fx
_ = torch.fx.symbolic_trace(fn1)
return x * 2
x = torch.randn(10)
fn(x)
cnt = torch._dynamo.testing.CompileCounter()
opt_fn = torch._dynamo.optimize(cnt)(fn)
opt_fn(x)
self.assertEqual(cnt.frame_count, 1)
def test_relative_import(self):
try:
from . import utils as _ # noqa: F401
def fn(x):
from .utils import tensor_for_import_testing
return x * 2 * tensor_for_import_testing
except ImportError:
def fn(x):
from utils import tensor_for_import_testing
return x * 2 * tensor_for_import_testing
x = torch.randn(10)
fn(x)
cnt = torch._dynamo.testing.CompileCounter()
opt_fn = torch._dynamo.optimize(cnt, nopython=True)(fn)
opt_fn(x)
self.assertEqual(cnt.frame_count, 1)
def test_relative_import_no_modulename(self):
try:
from . import utils as _ # noqa: F401
def fn(x):
from . import utils
return x * 2 * utils.tensor_for_import_testing
except ImportError:
def fn(x):
import utils
return x * 2 * utils.tensor_for_import_testing
x = torch.randn(10)
fn(x)
cnt = torch._dynamo.testing.CompileCounter()
opt_fn = torch._dynamo.optimize(cnt, nopython=True)(fn)
opt_fn(x)
self.assertEqual(cnt.frame_count, 1)
def test_bigbird_unsqueeze_inplace(self):
def fn(reshape_2):
view_2 = reshape_2.clone()
view_2.unsqueeze_(2)
cat_11 = torch.cat([view_2], dim=2)
view_13 = cat_11.view((2, 12, 64, -1))
return (view_13,)
x = torch.randn(2, 12, 64, 64, requires_grad=True)
ref = fn(x)
opt_fn = torch._dynamo.optimize("aot_eager")(fn)
res = opt_fn(x)
self.assertTrue(same(ref, res))
def test_issue1466_size_aot_autograd(self):
def fn(x):
# do a tensor op and a size compute
y = x * 2
x_size = x.size()
# trigger a graph break
print("arf")
# use the tensor op and size compute
z = y.view(x_size) + 1
return z
x = torch.randn(2, 3, requires_grad=True)
ref = fn(x)
opt_fn = torch._dynamo.optimize("aot_eager")(fn)
res = opt_fn(x)
self.assertTrue(same(ref, res))
def test_ellipsis(self):
class Repro(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.lnorm = torch.nn.LayerNorm(
(256,), eps=1e-06, elementwise_affine=True
)
self.linear = torch.nn.Linear(
in_features=256, out_features=256, bias=True
)
def forward(self, cat_10):
lnorm = self.lnorm(cat_10)
getitem_64 = lnorm[
(slice(None, None, None), slice(0, 1, None), Ellipsis)
]
linear = self.linear(getitem_64)
return (linear,)
args = [torch.randn(2, 197, 256)]
mod = Repro()
opt_mod = torch._dynamo.optimize("eager", nopython=True)(mod)
self.assertTrue(same(mod(*args), opt_mod(*args)))
def test_reinplacing(self):
class MockModule(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.self_layoutlm_embeddings_x_position_embeddings = (
torch.nn.Embedding(1024, 768)
)
self.self_layoutlm_embeddings_y_position_embeddings = (
torch.nn.Embedding(1024, 768)
)
def forward(self, getitem_1, getitem_2, add):
self_layoutlm_embeddings_x_position_embeddings = (
self.self_layoutlm_embeddings_x_position_embeddings(getitem_1)
)
self_layoutlm_embeddings_y_position_embeddings = (
self.self_layoutlm_embeddings_y_position_embeddings(getitem_2)
)
add_1 = add + self_layoutlm_embeddings_x_position_embeddings
add_2 = add_1 + self_layoutlm_embeddings_y_position_embeddings
return (add_2,)
mod = MockModule()
opt_mod = torch._dynamo.optimize("aot_eager_decomp_partition")(mod)
args = [
((2, 512), (2048, 4), torch.int64, "cpu", False),
((2, 512), (2048, 4), torch.int64, "cpu", False),
((2, 512, 768), (393216, 768, 1), torch.float32, "cpu", True),
]
args = [
rand_strided(sh, st, dt, dev).requires_grad_(rg)
for (sh, st, dt, dev, rg) in args
]
self.assertTrue(same_two_models(mod, opt_mod, args))
def test_optimized_deepcopy(self):
# See https://github.com/pytorch/pytorch/pull/88629
class Foo(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.fc = torch.nn.Linear(in_features=2, out_features=3, bias=True)
def forward(self, x):
return self.fc(x)
mod = Foo()
opt_mod = torch._dynamo.optimize("eager")(mod)
args = [torch.randn(1, 2)]
self.assertTrue(same_two_models(mod, opt_mod, args))
def test_class_member(self):
class Foo(torch.nn.Module):
a = 4
b = torch.ones(3, 4)
def __init__(self) -> None:
super().__init__()
self.c = 4
def forward(self, x):
return x.cos() + self.a + self.b + self.c
mod = Foo()
opt_mod = torch._dynamo.optimize("eager", nopython=True)(mod)
args = (torch.randn(3, 4),)
self.assertTrue(same(mod(*args), opt_mod(*args)))
def test_named_buffers(self):
class Foo(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.x = torch.nn.Buffer(torch.ones(3))
self.y = torch.nn.Buffer(torch.ones(3))
def forward(self, inp):
res = 0
for name, buffer in self.named_buffers():
res += buffer.sum()
return inp.cos() + res
mod = Foo()
opt_mod = torch._dynamo.optimize("eager", nopython=True)(mod)
args = (torch.randn(3, 4),)
self.assertTrue(same(mod(*args), opt_mod(*args)))
def test_requires_grad_guards_with_grad_mode1(self):
def f(x):
if x.requires_grad:
return x + 1
else:
return x + 2
x = torch.ones(2, requires_grad=True)
f_compiled = torch.compile(f)
with torch.no_grad():
# compile an inference graph
f_compiled(x)
# Test: we should fail guards and recompile (even though it's still an inference graph)
out_ref = f(x.detach())
out = f_compiled(x.detach())
self.assertEqual(out_ref, out)
self.assertEqual(out_ref.requires_grad, out.requires_grad)
def test_requires_grad_guards_with_grad_mode2(self):
x = torch.ones(2, requires_grad=True)
x_ref = x.clone().detach().requires_grad_(True)
m = torch.nn.Linear(2, 2)
m_compiled = torch.compile(m)
with torch.no_grad():
# compile an inference graph
m_compiled(x)
# Test: we should fail guards and recompile a training graph
out_ref = m(x_ref)
out = m_compiled(x)
self.assertEqual(out_ref, out)
self.assertEqual(out_ref.requires_grad, out.requires_grad)
def test_is_symbolic_tracing(self):
# Ensure no graph break here
def fn(x):
if is_fx_tracing_test():
return x * 2
return x * 4
a = torch.randn(4)
ref = fn(a)
opt_fn = torch._dynamo.optimize("eager", nopython=True)(fn)
res = opt_fn(a)
self.assertTrue(same(ref, res))
def test_tokenization(self):
from collections import UserDict
class BatchEncoding(UserDict):
"""
Copied from tokenization
"""
def __init__(
self,
data,
):
super().__init__(data)
def __getattr__(self, item: str):
try:
return self.data[item]
except KeyError as e:
raise AttributeError from e
def tokenization(x):
encoding = BatchEncoding({"key": x})
return encoding["key"]
opt_fn = torch._dynamo.optimize("eager")(tokenization)
x = torch.rand((1, 4))
ref = tokenization(x)
res = opt_fn(x)
self.assertTrue(same(ref, res))
def test_modules(self):
class Foo(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.fc = torch.nn.Linear(4, 3)
def forward(self, inp):
res = torch.zeros(3, 3)
for mod in self.modules():
res += self.fc(inp)
return res
mod = Foo()
args = (torch.ones(3, 4),)
cnt = torch._dynamo.testing.CompileCounter()
opt_mod = torch._dynamo.optimize(cnt, nopython=True)(mod)
self.assertTrue(same(mod(*args), opt_mod(*args)))
self.assertEqual(cnt.op_count, 5)
self.assertEqual(cnt.frame_count, 1)
def test_omegaconf_listconfig_iter(self):
obj = ListConfig()
x = torch.zeros(2)
def fn():
y = x
for i in obj:
y += i
return y
expected = fn()
actual = torch.compile(fn, fullgraph=True, backend="eager")()
self.assertEqual(actual, expected)
def test_user_defined_iter(self):
class MyIter:
def __init__(self) -> None:
self.i = 0
def __iter__(self):
return self
def __next__(self):
if self.i < 3:
self.i += 1
return self.i
raise StopIteration
@torch.compile(backend="eager", fullgraph=True)
def fn(x):
for i in MyIter():
x += i
return x
self.assertEqual(fn(torch.zeros(1)), torch.full([1], 6.0))
def test_stop_iteration_reconstruct(self):
@torch.compile(backend="eager", fullgraph=True)
def fn(x):
return x.sin(), StopIteration(1, 2, 3)
_, res = fn(torch.ones(1))
self.assertEqual(str(res), str(StopIteration(1, 2, 3)))
def test_tensor_data_kwarg(self):
# https://github.com/pytorch/pytorch/issues/96278
def f():
return torch.tensor(data=[[1.0, -1.0]])
cnt = torch._dynamo.testing.CompileCounter()
opt_fn = torch._dynamo.optimize(cnt, nopython=True)(f)
self.assertTrue(same(f(), opt_fn()))
self.assertEqual(cnt.frame_count, 1)
@requires_cuda
def test_norm_dtype(self):
def foo(_stack0):
getitem = _stack0[(slice(None, None, None), -1)]
_stack0 = None
normalize = torch.nn.functional.normalize(getitem, p=2, dim=1)
getitem = None
return (normalize,)
args = [((2, 50, 256), (1, 256, 1), torch.float16, "cuda", False)]
args = [
rand_strided(sh, st, dt, dev).requires_grad_(rg)
for (sh, st, dt, dev, rg) in args
]
opt_foo = torch._dynamo.optimize("aot_eager_decomp_partition")(foo)
with torch.cuda.amp.autocast(enabled=True):
ref = foo(*args)[0]
res = foo(*args)[0]
self.assertEqual(ref.dtype, res.dtype)
self.assertTrue(same(res, ref))
def test_for_loop_graph_break(self):
def inner(x):
return torch.sin(x)
def fn(x):
for _ in range(100):
inner(x)
torch._dynamo.graph_break()
return x
cnt = torch._dynamo.testing.CompileCounter()
opt_fn = torch._dynamo.optimize(cnt)(fn)
x = torch.randn(4)
opt_fn(x)
self.assertEqual(cnt.frame_count, 1)
self.assertEqual(cnt.op_count, 1)
def test_for_loop_graph_break_before(self):
# Checks that the backedge is calculated correctly
def inner(x):
return torch.sin(x)
def fn(x):
torch._dynamo.graph_break()
for _ in range(100):
inner(x)
return x
cnt = torch._dynamo.testing.CompileCounter()
opt_fn = torch._dynamo.optimize(cnt)(fn)
x = torch.randn(4)
opt_fn(x)
self.assertEqual(cnt.frame_count, 1)
self.assertEqual(cnt.op_count, 100)
def test_avoid_dupe_specialization(self):
def f(x, y):
return (x + y) * 1
opt_f = torch._dynamo.optimize("aot_eager")(f)
for b in [True, False]:
x = torch.randn(4, requires_grad=b)
y = torch.randn(4, requires_grad=b)
self.assertEqual(f(x, x), opt_f(x, x))
self.assertEqual(f(x, y), opt_f(x, y))
def test_validate_model_kwargs(self):
cnt = CompileCounter()
def f1(a, b):
return torch.sin(a) + torch.cos(b)
@torch.compile(backend=cnt, fullgraph=True)
def f2(**kwargs):
_validate_model_kwargs(f1, kwargs)
return f1(**kwargs)
x = torch.randn(10)
y = torch.randn(10)
self.assertEqual(f2(a=x, b=y), f1(x, y))
self.assertEqual(cnt.frame_count, 1)
self.assertEqual(cnt.op_count, 3)
def test_swin_base_tensor_attr(self):
class Foo(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
# NB: not a parameter or buffer
self.t = torch.randn(3)
def forward(self, x):
return x + torch.cat((self.t, self.t))
mod = Foo()
opt_mod = torch._dynamo.optimize("eager")(mod)
args = [torch.randn(6)]
self.assertTrue(same_two_models(mod, opt_mod, args))
opt_mod(*args)
def test_pointless_graph_removal(self):
cnt = torch._dynamo.testing.CompileCounter()
@torch.compile(backend=cnt)
def fn(x):
with torch.no_grad():
torch._dynamo.graph_break()
return x + 1
fn(torch.randn(4))
self.assertEqual(cnt.frame_count, 1)
self.assertEqual(cnt.op_count, 3)
def test_output_aliases_intermediate(self):
def f(x):
intermediate = x.mul(2)
return intermediate.view(-1), intermediate
opt_f = torch._dynamo.optimize("aot_eager")(f)
for b in [True, False]:
x = torch.randn(4, requires_grad=b)
out = f(x)
out_test = opt_f(x)
self.assertEqual(out[0], out_test[0])
self.assertEqual(out[1], out_test[1])
self.assertEqual(out[0].requires_grad, out_test[0].requires_grad)
self.assertEqual(out[1].requires_grad, out_test[1].requires_grad)
# test that the aliasing relationship of outputs is preserved
out[0].mul_(2)
out_test[0].mul_(2)
self.assertEqual(out[0], out_test[0])
self.assertEqual(out[1], out_test[1])
def test_while_loop_graph_break(self):
# Repro of tacotron2 cache_size_recompilation
def inner(x):
return torch.sin(x)
def fn(x):
i = 20
while i > 10:
x = inner(x)
i -= 1
torch._dynamo.graph_break()
return x
cnt = torch._dynamo.testing.CompileCounter()
opt_fn = torch._dynamo.optimize(cnt)(fn)
x = torch.randn(4)
opt_fn(x)
self.assertEqual(cnt.frame_count, 1)
self.assertEqual(cnt.op_count, 1)
def test_nested_while_loop_graph_break(self):
def inner_loop(x):
i = 3
while i > 0:
i -= 1
x += 1
torch._dynamo.graph_break()
return x
def inner(x):
inner_loop(x)
return torch.sin(x)
def fn(x):
i = 20
while i > 10:
x = inner(x)
i -= 1
torch._dynamo.graph_break()
return x
cnt = torch._dynamo.testing.CompileCounter()
opt_fn = torch._dynamo.optimize(cnt)(fn)
x = torch.randn(4)
opt_fn(x)
self.assertEqual(cnt.frame_count, 1)
self.assertEqual(cnt.op_count, 1)
def test_while_loop_graph_break_inside_call_function(self):
# Repro of huggingface graph break inside loop in `get_parameter_dtype`.
# Skip only the inner frame that has loop that contains graph break.
def inner(x):
for i in range(3):
x += 1
torch._dynamo.graph_break()
return x
def fn(x):
x += 2
inner(x)
x += 3
return x
cnt = torch._dynamo.testing.CompileCounter()
opt_fn = torch._dynamo.optimize(cnt)(fn)
x = torch.randn(4)
opt_fn(x)
self.assertEqual(cnt.frame_count, 2)
self.assertEqual(cnt.op_count, 2)
def test_exception_in_dynamo_handling(self):
hit_handler = False
# See https://github.com/pytorch/pytorch/pull/96488
@contextlib.contextmanager
def ctx():
try:
yield
except RuntimeError:
nonlocal hit_handler
hit_handler = True
@torch._dynamo.optimize("eager")
def f():
with ctx():
h()
def h():
raise RuntimeError("boof")
# Should not error
f()
self.assertTrue(hit_handler)
def test_generator_dealloc(self):
# See https://github.com/pytorch/pytorch/pull/96488
#
# NB: yes, [(...)] is intentional, this is a list containing a
# generator
generator_box = [(x for x in [1, 2, 3])]
counter = torch._dynamo.testing.CompileCounter()
def g(x):
return x + 2
# TODO: This test is pretty delicate. To test if it's actually doing
# anything, rebuild eval_frame.c with '#define TORCHDYNAMO_DEBUG 1'
# and then look at the logs for:
#
# TRACE[_custom_eval_frame:650] begin <genexpr> test_repros.py 2276 -1 0 0
# TRACE[_custom_eval_frame:664] throw <genexpr>
#
# This means we're actually hitting the relevant codepath
# NB: Make sure we don't actually Dynamo this frame; if we do Dynamo
# this frame, Dynamo actually DOES understand list.clear and will
# arrange for the generator deallocation to happen when the eval frame
# handler is disabled, which will prevent the bug from happening (we
# specifically want to trigger the generator deallocation WHILE the
# dynamo eval frame handler is active), as that will cause the
# generator to become exhausted and trigger the throw_flag == TRUE
# case.
@torch._dynamo.disable(recursive=False)
def f(x):
generator_box.clear()
return g(x)
self.assertNoUnraisable(
lambda: torch._dynamo.optimize(counter)(f)(torch.randn(3))
)
# Make sure the x + 2 is captured (a previous incorrect implementation
# of this fix would have disabled the eval frame callback, which means
# g wouldn't get traced
self.assertEqual(counter.op_count, 1)
def test_error_return_without_exception_set(self):
# https://github.com/pytorch/pytorch/issues/93781
@torch.compile
def f():
_generator_type = type(_ for _ in ())
self.assertNoUnraisable(f)
def common_merge_criteria_processor_list(self, list_cls, fullgraph):
cnt = CompileCounter()
@torch.compile(backend=cnt, fullgraph=fullgraph)
def f(x, left, right):
combined = _merge_criteria_processor_list(left, right)
return combined(x)
l1 = list_cls([torch.nn.ReLU(), torch.nn.Sigmoid()])
l2 = list_cls([])
input = torch.randn(16)
result = f(input, l1, l2)
self.assertEqual(result, l1(input))
self.assertEqual(cnt.frame_count, 1)
self.assertEqual(cnt.op_count, 2)
cnt.clear()
l3 = list_cls([torch.nn.SiLU()])
expected = l3(l1(input))
result = f(input, l1, l3)
self.assertEqual(len(l1), 3)
self.assertEqual(result, expected)
self.assertEqual(cnt.frame_count, 1)
self.assertEqual(cnt.op_count, 3)
def test_merge_criteria_processor_list1(self):
self.common_merge_criteria_processor_list(CustomList1, False)
def test_merge_criteria_processor_list2(self):
self.common_merge_criteria_processor_list(CustomList2, True)
def test_restricted_list_subclass1(self):
cnt = CompileCounter()
@torch.compile(backend=cnt, fullgraph=True)
def fn(a, b):
l = CustomList2()
l.extend([True])
l.append(a)
l.extend([b])
l.pop(0)
l.append(l.length_times_10())
return sum(l)
x = torch.randn(10)
y = torch.randn(10)
self.assertEqual(fn(x, y), x + y + 20)
self.assertEqual(cnt.op_count, 3)
def test_restricted_list_subclass2(self):
cnt = CompileCounter()
@torch.compile(backend=cnt, fullgraph=True)
def fn(a, b):
l1 = CustomList2([a + 1])
l2 = CustomList2([b + 2])
l1.extend(l2)
return l1
x = torch.randn(10)
y = torch.randn(10)
z = fn(x, y)
self.assertEqual(type(z), CustomList2)
self.assertEqual(len(z), 2)
self.assertEqual(z.length_times_10(), 20)
self.assertEqual(list(z), [x + 1, y + 2])
def test_restricted_list_subclass3(self):
cnt = CompileCounter()
@torch.compile(backend=cnt, fullgraph=True)
def fn(a: CustomList2, b: CustomList2):
a.extend(b)
a.append_twice(b[2] + 1)
a.append(b[3] + 2)
return b
x = torch.randn(10)
y = torch.randn(10)
l = CustomList2([x, y])
self.assertIs(fn(l, l), l)
self.assertEqual(len(l), 7)
self.assertIs(l[0], x)
self.assertIs(l[1], y)
self.assertIs(l[2], x)
self.assertIs(l[3], y)
self.assertEqual(l[4], x + 1)
self.assertIs(l[5], l[4])
self.assertEqual(l[6], y + 2)
def test_rewrite_assert_with_msg(self):
def f(x):
b = x.sin()
assert x[0] == 3, "First dim need to be 3"
return x.cos() + b
args = (torch.Tensor([3, 4, 5]),)
cnt = torch._dynamo.testing.CompileCounter()
opt_f = torch._dynamo.optimize(cnt, nopython=True)(f)
self.assertTrue(same(f(*args), opt_f(*args)))
self.assertEqual(cnt.op_count, 6)
self.assertEqual(cnt.frame_count, 1)
exported, _ = torch._dynamo.export(f)(torch.Tensor([3, 4, 5]))
self.assertTrue(same(exported(*args), f(*args)))
def test_list_aliasing(self):
cnt = CompileCounter()
@torch.compile(backend=cnt, fullgraph=True)
def fn(a):
a.append(torch.sin(a[0]))
return a
x = torch.randn(10)
l = [x]
self.assertIs(fn(l), l)
self.assertEqual(len(l), 2)
self.assertIs(l[0], x)
self.assertEqual(l[1], torch.sin(x))
self.assertEqual(cnt.frame_count, 1)
self.assertEqual(cnt.op_count, 1)
def test_not_rewrite_assert_for_other_errors(self):
def f(x):
b = x.sin()
if not x.sum() <= 3:
raise ValueError("input sum needs to be 3")
return x.cos() + b
args = (torch.Tensor([3, 4, 5]),)
opt_fn = torch._dynamo.optimize("eager")(f)
with self.assertRaisesRegex(ValueError, "input sum needs to be 3"):
opt_fn(*args)
def test_rewrite_assert_dont_change_bytecode(self):
def fn(x):
with torch.no_grad():
assert x.max() < 5, f"invalid max {x.max()}"
x = torch.sin(x)
return x
x = torch.ones(4)
opt_fn = torch._dynamo.optimize("eager")(fn)
self.assertTrue(same(fn(x), opt_fn(x)))
def test_rewrite_assert_without_msg(self):
def f(x):
b = x.sin()
assert x[0] == 3
return x.cos() + b
args = (torch.Tensor([3, 4, 5]),)
exported, _ = torch._dynamo.export(f)(torch.Tensor([3, 4, 5]))
self.assertTrue(same(exported(*args), f(*args)))
with self.assertRaisesRegex(RuntimeError, "assertion error"):
exported(torch.Tensor([5, 6, 7]))
def test_rewrite_assert_with_non_string_msg(self):
def f(x):
b = x.sin()
assert x[0] == 2, x.size()
return x.cos() + b
torch._dynamo.utils.counters.clear()
args = torch.Tensor([3, 4, 5])
opt_f = torch._dynamo.optimize("eager")(f)
with self.assertRaisesRegex(AssertionError, "torch.Size"):
opt_f(args)
self.assertEqual(
torch._dynamo.utils.counters["graph_break"][
"assert with non-string message"
],
1,
)
def test_rewrite_assert_noop(self):
def f(x):
b = x.sin()
assert True
assert x.dtype == torch.float32
return x.cos() + b
args = (torch.Tensor([3, 4, 5]),)
exported, _ = torch._dynamo.export(f)(torch.Tensor([3, 4, 5]))
self.assertTrue(same(exported(*args), f(*args)))
cnt = torch._dynamo.testing.CompileCounter()
opt_f = torch._dynamo.optimize(cnt, nopython=True)(f)
self.assertTrue(same(f(*args), opt_f(*args)))
# torch._assert shouldn't be in the graph
self.assertEqual(cnt.op_count, 3)
self.assertEqual(cnt.frame_count, 1)
exported, _ = torch._dynamo.export(f)(torch.Tensor([4, 4, 5]))
self.assertTrue(same(exported(*args), f(*args)))
def test_size_typematch(self):
def f(x, y):
if isinstance(x, torch.Size):
return y + 1
else:
return y + 2
y = torch.zeros(1)
x1 = torch.Size((3,))
x2 = (3,)
cnt = torch._dynamo.testing.CompileCounter()
opt_f = torch._dynamo.optimize(cnt, nopython=True)(f)
self.assertTrue(same(f(x1, y), opt_f(x1, y)))
self.assertTrue(same(f(x2, y), opt_f(x2, y)))
self.assertEqual(cnt.frame_count, 2)
def test_dict_subclass_contains(self):
# pattern from huggingface
class ClassInstantier(collections.OrderedDict):
pass
@torch.compile(fullgraph=True, backend="eager")
def f(x, d):
if "key1" in d:
x = x + 2
if "key2" in d:
x = x + 4
x = x + 8
return x
result = f(torch.ones(8), ClassInstantier({"key1": torch.ones(8)}))
self.assertTrue(same(result, torch.full([8], 11.0)))
result = f(torch.ones(8), ClassInstantier({"key2": torch.ones(8)}))
self.assertTrue(same(result, torch.full([8], 13.0)))
def test_hf_classinstantier(self):
# hf activations.py
class ClassInstantier(collections.OrderedDict):
def __getitem__(self, key):
content = super().__getitem__(key)
cls, kwargs = content if isinstance(content, tuple) else (content, {})
return cls(**kwargs)
ACT2CLS = ClassInstantier(
{
"relu": (nn.ReLU, {"inplace": False}),
"tanh": nn.Tanh,
}
)
@torch.compile(fullgraph=True, backend="eager")
def f(x, act):
return ACT2CLS[act](x)
y = torch.randn(10)
self.assertTrue(same(f(y, "tanh"), torch.tanh(y)))
self.assertTrue(same(f(y, "relu"), torch.relu(y)))
def test_ephemeral_module(self):
# hf activations.py
class ReLUSquaredActivation(nn.Module):
def forward(self, input):
relu_applied = torch.nn.functional.relu(input)
squared = torch.square(relu_applied)
return squared
@torch.compile(fullgraph=True, backend="eager")
def f(x):
x = x + 0.2
x = ReLUSquaredActivation()(x)
x = x + 1
return x
y = torch.randn(10)
self.assertTrue(same(f(y), ReLUSquaredActivation()(y + 0.2) + 1))
def test_inplace_unsqueeze_input(self):
def backend(gm, example_inputs):
self.assertEqual(example_inputs[-1].size(), torch.Size([1, 3, 4]))
return gm
@torch.compile(backend=backend)
def fn(x):
x.unsqueeze_(0)
return x + 1
inputs = [torch.randn(3, 4)]
self.assertEqual(fn(*inputs).size(), torch.Size([1, 3, 4]))
self.assertEqual(inputs[0].size(), torch.Size([1, 3, 4]))
def test_batchnorm_e2e(self):
class Repro(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.bn = torch.nn.BatchNorm2d(
64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True
)
self.conv1 = torch.nn.Conv2d(
64,
64,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1),
bias=False,
)
def forward(self, x):
x1 = self.bn(x)
x2 = self.conv1(x1)
out = torch.nn.functional.relu(x2)
return (out,)
torch.manual_seed(1337)
m_ref = Repro()
m_test = deepcopy(m_ref)
@torch._dynamo.optimize("aot_eager_decomp_partition")
def compiled_fn(x):
return m_test(x)
x_ref = torch.randn(2, 64, 32, 32, requires_grad=True)
x_test = x_ref.clone()
# Loop multiple times: each iteration the running_mean/var on batchnorm will update,
# which changes the output of the next iteration
for _ in range(3):
ref = m_ref(x_ref)
res = compiled_fn(x_test)
self.assertTrue(same(ref, res))
for r in ref:
if r.requires_grad:
r.sum().backward()
for r in res:
if r.requires_grad:
r.sum().backward()
for param_ref, param_test in zip(m_ref.parameters(), m_test.parameters()):
self.assertTrue(same(param_ref, param_test))
# Assert running_mean/var
for buffer_ref, buffer_test in zip(m_ref.buffers(), m_test.buffers()):
self.assertTrue(same(buffer_ref, buffer_test))
@torch._dynamo.config.patch("assume_static_by_default", False)
def test_dynamic_shapes_right_side(self):
def f(x):
return torch.ones(5 * x.shape[0])
inp = torch.randn(6, 5)
gm, _ = torch._dynamo.export(f, aten_graph=True)(torch.randn(4, 5))
self.assertEqual(gm(inp).shape, f(inp).shape)
@torch._dynamo.config.patch("specialize_int", False)
def test_maybe_multiply_symint(self):
# https://github.com/pytorch/pytorch/issues/97346
from torch._functorch.aot_autograd import aot_module_simplified
def my_aot_compiler(gm, example_inputs):
def my_compiler(gm, example_inputs):
return gm.forward
# Invoke AOTAutograd
return aot_module_simplified(gm, example_inputs, fw_compiler=my_compiler)
def my_example(t1, t2, d):
out = torch.add(t1, t2, alpha=d)
return out
compiled_fn = torch.compile(backend=my_aot_compiler, dynamic=True)(my_example)
t1 = torch.arange(3, dtype=torch.float32).requires_grad_(True)
t2 = torch.arange(3, dtype=torch.float32).requires_grad_(True)
ra = compiled_fn(t1, t2, 5)
self.assertEqual(ra, torch.tensor([0.0, 6.0, 12.0]))
ra = compiled_fn(t1, t2, 6)
self.assertEqual(ra, torch.tensor([0.0, 7.0, 14.0]))
def test_build_map_unpack_with_call(self):
def forward_with_cond_scale(x, t, cond_scale, self_cond, other1, other2):
return x.sin() + t + cond_scale + self_cond + other1 + other2
@torch.compile(backend="eager", fullgraph=True)
def fn(x):
d1 = dict(other1=5)
d2 = dict(other2=4)
text_cond = {**d1, **d2}
return forward_with_cond_scale(x, 1, cond_scale=2, self_cond=3, **text_cond)
self.assertTrue(same(fn(torch.ones(4)), torch.ones(4).sin() + 15))
@torch._dynamo.config.patch(verbose=True)
def test_graph_break_unsupported_fake(self):
counter = torch._dynamo.testing.CompileCounter()
@torch._dynamo.optimize(counter)
def f(x):
return torch.ops.test_sample.foo(x + 1) + 1
f(torch.randn(3))
self.assertEqual(counter.op_count, 2)
self.assertEqual(counter.frame_count, 2)
def test_delattr(self):
class MyObj:
def __init__(self, a, b):
self.a = a
self.b = b
@torch.compile(backend="eager", fullgraph=True)
def fn(x, obj):
del obj.a
obj.c = x + 1
del obj.c
tmp = MyObj(x + 2, x + 3)
del tmp.b
if hasattr(obj, "a"):
return x + 1
return tmp
x = torch.zeros([])
obj1 = MyObj(x, x)
obj2 = fn(x, obj1)
self.assertFalse(hasattr(obj1, "a"))
self.assertFalse(hasattr(obj1, "c"))
self.assertFalse(hasattr(obj2, "b"))
self.assertEqual(obj1.b.item(), 0)
self.assertEqual(obj2.a.item(), 2)
def test_delattr_raises(self):
class MyObj:
def __init__(self, a, b):
self.a = a
self.b = b
@torch.compile(backend="eager")
def fn(x, obj):
del obj.a
x = x + 1
obj.a # will raise
return x
x = torch.zeros([])
obj1 = MyObj(x, x)
self.assertRaises(AttributeError, lambda: fn(x, obj1))
def test_delsubscr(self):
@torch.compile(backend="eager")
def fn(x):
del x["a"]
y = x["b"] + 1
return y
x = {"a": torch.tensor([1]), "b": torch.tensor([1])}
result = fn(x)
self.assertFalse(hasattr(x, "a"))
self.assertEqual(result.item(), 2)
def test_delsubscr_raises(self):
@torch.compile(backend="eager")
def fn(x):
del x["a"]
y = x["a"] + 1 # should raise KeyError
return y
x = {"a": torch.tensor([1]), "b": torch.tensor([1])}
self.assertRaises(KeyError, lambda: fn(x))
def test_attached_attribute_in_dir(self):
class MyModule(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.linear = torch.nn.Linear(16, 16)
self.relu = torch.nn.ReLU()
def forward(self, x):
return self.relu(self.linear(x))
mod = torch.compile(MyModule(), backend="eager")
mod.is_compiled = True
self.assertTrue("is_compiled" in dir(mod))
@torch._dynamo.config.patch("automatic_dynamic_shapes", False)
def test_dynamic_shapes_implicit_guard(self):
def f(x):
y = x * x.size(x.shape[0])
torch.sum(y, [y.shape[0]])
return y
cnt = torch._dynamo.testing.CompileCounter()
opt_fn = torch._dynamo.optimize(cnt, nopython=True)(f)
opt_fn(torch.randn(3, 1, 1, 1, 1))
self.assertEqual(cnt.frame_count, 1)
def test_dalle2_maybe(self):
def normalize(x):
return x.cos()
@torch.compile(backend="eager", fullgraph=True)
def fn(x, normalize_img):
lowres_cond_img = x.sin()
lowres_cond_img = maybe(normalize_img)(lowres_cond_img)
return lowres_cond_img
self.assertEqual(fn(torch.ones([]), normalize), torch.ones([]).sin().cos())
def test_functools_wraps(self):
def cool_name(x):
return x.sin()
@torch.compile(backend="eager", fullgraph=True)
def fn(x):
y = x.cos()
@functools.wraps(cool_name)
def uncool_name():
return cool_name(y)
return uncool_name
result = fn(torch.ones([]))
self.assertEqual(result.__name__, "cool_name")
self.assertEqual(result(), torch.ones([]).cos().sin())
def test_dynamic_shapes_float_guard(self):
def f(x):
return torch.nn.functional.dropout(x, x.shape[0] / 6)
cnt = torch._dynamo.testing.CompileCounter()
opt_fn = torch._dynamo.optimize(cnt, nopython=True)(f)
opt_fn(torch.randn(3))
self.assertEqual(cnt.frame_count, 1)
@torch._dynamo.config.patch(capture_scalar_outputs=True)
def test_tensor_item(self):
def f(x, y):
val = y.item()
return x.sum() + val
gm, _ = torch._dynamo.export(
f,
aten_graph=True,
)(
torch.zeros(6, 4),
torch.tensor(1),
)
self.assertEqual(
f(torch.zeros(6, 4), torch.tensor(1)),
gm(torch.zeros(6, 4), torch.tensor(1)),
)
self.assertEqual(
f(torch.zeros(6, 4), torch.tensor(2)),
gm(torch.zeros(6, 4), torch.tensor(2)),
)
def test_dataclass_init_with_default_factory_with_inputs(self):
@dataclasses.dataclass
class DClass:
sharding_contexts: Any = dataclasses.field(default_factory=list)
a: int = 1
def fn(x, inp_list):
d = DClass(inp_list)
d.sharding_contexts.append(x.sin() + d.a)
return d
x = torch.randn(4)
inp_list1 = [1, 2, 3]
inp_list2 = [2, 3, 4]
inp_list3 = [1, 2]
ref1 = fn(x, inp_list1)
ref2 = fn(x, inp_list2)
ref3 = fn(x, inp_list3)
cnt = torch._dynamo.testing.CompileCounter()
opt_fn = torch.compile(fn, fullgraph=True)
opt_ret1 = opt_fn(x, inp_list1)
opt_ret2 = opt_fn(x, inp_list2)
opt_ret3 = opt_fn(x, inp_list3)
self.assertEqual(ref1.sharding_contexts, opt_ret1.sharding_contexts)
self.assertEqual(ref2.sharding_contexts, opt_ret2.sharding_contexts)
self.assertEqual(ref3.sharding_contexts, opt_ret3.sharding_contexts)
def test_list_index(self):
for i, list_type in enumerate(
(
list,
tuple,
torch.Size,
collections.deque,
namedtuple("FourElems", "one two three four", defaults=[0, 0, 0, 0]),
)
):
torch._dynamo.reset()
for index in ([], [2], [0, 3]):
def f(t):
if i == 4: # namedtuple
xs = list_type(1, 2, 3, 4)
else:
xs = list_type([1, 2, 3, 4])
res = xs.index(3, *index)
return t + res
res = torch._dynamo.optimize(backend="eager", nopython=True)(f)(
torch.zeros(1)
)
self.assertEqual(res, torch.tensor([2.0]))
def test_list_index_not_found(self):
def f(t):
xs = ["bar", "foo", "baz", "buzz"]
res = xs.index("non-existent")
return t + res
# Raising ValueError from item not found is unsupported
with self.assertRaises(
torch._dynamo.exc.Unsupported,
):
torch._dynamo.optimize(backend="eager", nopython=True)(f)(torch.zeros(1))
def test_list_index_tensor_unsupported(self):
for index in ([], [2], [0, 3]):
def f(t):
xs = [torch.tensor([i]) for i in range(4)]
res = xs.index(torch.tensor([2]), *index)
return t + res
with self.assertRaisesRegex(
torch._dynamo.exc.UserError, "Dynamic control flow is not supported"
):
torch._dynamo.optimize(backend="eager", nopython=True)(f)(
torch.zeros(1)
)
def test_hf_xsoftmax_inference(self):
def fn(input, mask):
return XSoftmax.apply(input + 1, mask, 1) + 2
fn_opt = torch.compile(fn, backend="eager", fullgraph=True)
inputs = [
torch.randn(4, 10),
torch.randn(4, 10) < 0,
]
expected = fn(*inputs)
actual = fn_opt(*inputs)
self.assertTrue(same(actual, expected))
@mock.patch("torch._dynamo.config.guard_nn_modules", True)
def test_hf_xsoftmax_training(self):
from torch._dynamo.utils import counters
counters.clear()
def fn(input, mask):
return XSoftmax.apply(input, mask, 1)
cnt = torch._dynamo.testing.CompileCounter()
fn_opt = torch.compile(fn, backend=cnt, fullgraph=False)
torch.manual_seed(1234)
inputs1 = [
torch.randn(4, 10, requires_grad=True),
torch.randn(4, 10) < 0,
]
torch.manual_seed(1234)
inputs2 = [
torch.randn(4, 10, requires_grad=True),
torch.randn(4, 10) < 0,
]
expected = fn(*inputs1)
actual = fn_opt(*inputs2)
self.assertTrue(same(actual, expected))
self.assertEqual(dict(counters["frames"]), {"total": 1, "ok": 1})
self.assertEqual(cnt.op_count, 2)
self.assertEqual(cnt.frame_count, 1)
cnt.clear()
counters.clear()
expected.sum().backward()
actual.sum().backward()
self.assertTrue(same(inputs1[0].grad, inputs2[0].grad))
# currently we don't capture the backwards frame
self.assertEqual(cnt.frame_count, 0)
self.assertEqual(cnt.op_count, 0)
self.assertEqual(dict(counters["frames"]), {})
self.assertEqual(dict(counters["graph_break"]), {})
def test_autograd_function_graph_break(self):
class MySin(torch.autograd.Function):
@staticmethod
def forward(ctx, x):
torch._dynamo.graph_break()
ctx.save_for_backward(x)
return x.sin()
@staticmethod
def backward(ctx, gx):
(x,) = ctx.saved_tensors
return gx * x.cos()
x = torch.randn([], requires_grad=True)
@torch.compile(backend="eager")
def fn(x):
return MySin.apply(x)
y = fn(x)
self.assertEqual(y, x.sin())
(gx,) = torch.autograd.grad(y, x)
self.assertEqual(gx, x.cos())
def test_jit_trace_errors(self):
@torch.compile(backend="eager", dynamic=True)
def f(x):
return x + 1
with self.assertRaises(RuntimeError):
torch.jit.trace(f, torch.randn(3))
@torch._dynamo.config.patch("assume_static_by_default", False)
def test_tensor_split(self):
def f(x):
return torch.split(x, x.shape[0] // 2, dim=0)[0]
gm, _ = torch._dynamo.export(
f,
aten_graph=True,
)(
torch.zeros(6, 4),
)
self.assertEqual(f(torch.ones(8, 4)), gm(torch.ones(8, 4)))
def test_optim_state_references_cleared(self):
model = torch.nn.Linear(2048, 2048, bias=False)
x = torch.ones(2048)
state_ref = 0
optimizer = torch.optim.Adadelta(model.parameters(), lr=0.01)
def opt_step():
optimizer.step()
compiled_opt_step = torch._dynamo.optimize("eager")(opt_step)
def compiled_model_step(x):
optimizer.zero_grad()
y = model(x)
torch.sum(y).backward()
compiled_opt_step()
compiled_model_step(x)
# Picked "square_avg" arbitrarily to check that
# optimizer state tensors are deallocated
state_ref = weakref.ref(
optimizer.state[optimizer.param_groups[0]["params"][0]]["square_avg"]
)
optimizer = None
self.assertIsNone(state_ref())
def test_grad_references_cleared(self):
model = torch.nn.Linear(2048, 2048, bias=False)
x = torch.ones(2048)
optimizer = torch.optim.Adadelta(model.parameters(), lr=0.01)
def opt_step():
optimizer.step()
compiled_opt_step = torch._dynamo.optimize("eager")(opt_step)
def compiled_model_step(x):
optimizer.zero_grad(True)
y = model(x)
torch.sum(y).backward()
compiled_opt_step()
compiled_model_step(x)
param_grad_ref = weakref.ref(next(iter(model.parameters())).grad)
optimizer.zero_grad(True)
self.assertIsNone(param_grad_ref())
def test_batch_encoding_clone_inputs(self):
class BatchEncoding(dict):
"""
Copied from test_tokenization
"""
def __init__(
self,
data,
):
super().__init__(data)
def __getattr__(self, item: str):
try:
return self.data[item]
except KeyError as e:
raise AttributeError from e
encoding = BatchEncoding({"key": torch.rand((1, 4))})
cloned_encoding = torch._dynamo.utils.clone_inputs(encoding)
self.assertTrue(type(cloned_encoding) is not dict)
def test_iadd_graph_break(self):
def fn(x):
a = ()
x = torch.sin(x)
a += (x,)
return a
x = torch.randn(4)
ref = fn(x)
opt_fn = torch._dynamo.optimize("eager", nopython=True)(fn)
res = opt_fn(x)
self.assertTrue(same(ref, res))
def test_odict_get_item_index_name(self):
d = {float: torch.float32, np.float16: torch.float16}
@torch.compile(backend="eager")
def f(x, y1, y2):
return torch.zeros(5, dtype=d[y1]), torch.zeros(5, dtype=d[y2])
f(torch.zeros(4), float, np.float16)
def test_dedup_global(self):
@torch.compile()
def f():
return _GLOBAL_CPU_TENSOR + _GLOBAL_CPU_TENSOR
self.assertEqual(f(), _GLOBAL_CPU_TENSOR + _GLOBAL_CPU_TENSOR)
def test_randint_out_dynamic(self):
def randint_fn(high, size, out):
return torch.randint(high, size, out=out)
opt_model = torch.compile(randint_fn)
out1 = torch.empty(10, dtype=torch.int32)
opt_model(17, (10,), out1)
out2 = torch.empty(12, dtype=torch.int32)
opt_model(17, (12,), out2)
@requires_cuda
def test_guard_default_device(self):
try:
torch.set_default_device("cuda")
counter = torch._dynamo.testing.CompileCounter()
@torch._dynamo.optimize(counter)
def f():
x = torch.randn(3)
return x * 2
self.assertEqual(f().device.type, "cuda")
self.assertEqual(counter.frame_count, 1)
torch.set_default_device("cpu")
self.assertEqual(f().device.type, "cpu")
self.assertEqual(counter.frame_count, 2)
finally:
torch.set_default_device(None)
def test_list_self_reference(self):
# Issue - https://github.com/pytorch/pytorch/issues/100150
root = []
root[:] = [root, root, None, None]
@torch._dynamo.optimize("eager")
def test_bug():
return root
test_bug()
def test_hf_bigbird_unsqueeze(self):
def torch_bmm_nd(inp_1, inp_2, ndim=None):
torch._dynamo.graph_break()
return torch.bmm(inp1, inp2)
def fn(inp1, inp2, inp3, inp4, c):
a = torch_bmm_nd(inp1, inp2, 4)
a.unsqueeze_(2)
a = a * 2
b = torch_bmm_nd(inp3, inp4, 4)
b.unsqueeze_(2)
l = a + b
out = torch.cat([a, b, c], dim=2)
return out, l
inp1 = torch.rand(1, 64, 448)
inp2 = torch.rand(1, 448, 64)
inp3 = torch.rand(1, 64, 448)
inp4 = torch.rand(1, 448, 64)
c = torch.rand(1, 64, 1, 64)
cnt = torch._dynamo.testing.CompileCounter()
opt_fn = torch._dynamo.optimize(cnt)(fn)
opt_fn(inp1, inp2, inp3, inp4, c)
self.assertEqual(cnt.frame_count, 3)
def test_torch_variable_type(self):
# from torchvision
def check_type(obj, types_or_checks):
for type_or_check in types_or_checks:
if (
isinstance(obj, type_or_check)
if isinstance(type_or_check, type)
else type_or_check(obj)
):
return True
return False
opt_check_type = torch._dynamo.optimize("eager")(check_type)
ref = check_type(torch.randn(4), [torch.Tensor])
res = opt_check_type(torch.randn(4), [torch.Tensor])
self.assertEqual(ref, res)
# Test for https://github.com/pytorch/pytorch/issues/103132
@torch._dynamo.config.patch("assume_static_by_default", False)
def test_inference_mode_dynamic_shapes(self):
class Repro(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
def forward(self, param):
z = torch.matmul(param, param)
return z
model = Repro()
# Need a 3d tensor to actually cause the error:
# we go down a path of the C++ matmul decomp that calls sizes().
inp = torch.randn(4, 4, 4, requires_grad=True)
model = torch.compile(model, backend="aot_eager", dynamic=True)
with torch.inference_mode():
model(inp)
def test_kwargs_out_list_variable(self):
class Repro(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
def forward(self, param):
z = torch.frexp(**param)
return z
model = Repro()
params = {"input": torch.tensor([[0.0, 1, 2, 4]])}
params["out"] = [
torch.empty(0, dtype=torch.float32), # mantissa
torch.empty(0, dtype=torch.int32), # exponent
]
model = torch.compile(model, backend="eager")
mantissa, exponent = model(params)
ref_mantissa = torch.tensor([[0.0000, 0.5000, 0.5000, 0.5000]])
ref_exponent = torch.tensor([[0, 1, 2, 3]], dtype=torch.int32)
self.assertEqual(ref_mantissa, mantissa)
self.assertEqual(ref_exponent, exponent)
@torch._dynamo.config.patch(capture_scalar_outputs=True)
def test_split_with_sizes_aot_autograd(self):
def fn(result, split_sizes):
rs = torch.ops.aten.split_with_sizes(result, split_sizes.tolist())
return rs
example_inputs = (
torch.randn(32, requires_grad=True),
torch.tensor((7, 16, 9)),
)
actual = torch.compile(fn, fullgraph=True, backend="aot_eager")(*example_inputs)
expected = fn(*example_inputs)
self.assertEqual(actual, expected)
def test_unspecialized_nn_module_with_torch_variable_attribute(self):
"""
In this case self.fn = something that should be a TorchVariable.
When it's not a TorchVariable, dynamo tries to trace through and fails.
This makes sure that the self.fn is handled as a TorchVariable.
"""
class UserModule(torch.nn.Module):
torchdynamo_force_dynamic = True # forced to be a UnspecializedNNModule
def __init__(self, fn):
super().__init__()
self.fn = fn
def forward(self, **inp):
return self.fn(**inp)
inputs = {
"input": torch.randn([2, 9]).uniform_(0, 1),
"target": torch.randn([2, 9]).uniform_(0, 1),
"reduction": "mean",
}
mod = UserModule(torch.nn.functional.binary_cross_entropy)
ref = mod(**inputs)
res = torch._dynamo.optimize("eager", nopython=True)(mod)(**inputs)
self.assertEqual(ref, res)
def test_call_finally_python_3_8(self):
# Issue - https://github.com/pytorch/pytorch/issues/97811
def make_fn(g):
def fn():
while True:
try:
print(g)
break
except Exception as _:
break
return torch.compile(fn, backend="eager")
make_fn(None)()
def test_call_finally_python_3_8_2(self):
def f(x):
while x:
try:
pass
except Exception as _:
continue
torch.compile(f, backend="eager")(0)
def test_call_finally_opcode_python_3_8(self):
def fn():
try:
return torch.zeros(4)
finally:
return torch.ones(4) # noqa: SIM107, B012
result = torch.compile(fn, backend="aot_eager")()
self.assertEqual(result, torch.ones(4))
def test_string_format(self):
s = "temp{i}"
@torch.compile(backend="eager", fullgraph=True)
def fn(x):
if s.format(i=4) == "temp4":
return torch.sin(x)
return torch.cos(x)
x = torch.randn(4)
self.assertEqual(fn(x), torch.sin(x))
# Repro of torch._dynamo.exc.InternalTorchDynamoError: 'NoneType' object has no attribute 'guards'
# due to bad empty list handling
def test_empty_list_contains_with_jump(self):
def fn(x, l):
if x in l:
return x.cos()
return x.sin()
counter = CompileCounter()
compiled_fn = torch._dynamo.optimize(counter)(fn)(torch.randn([2, 2]), [])
self.assertEqual(counter.frame_count, 1)
def test_graph_break_on_jit_isinstance(self):
@torch.compile(backend="eager")
def fn(x):
if torch.jit.isinstance(x, List[str]):
return x * 2
return x
opt_fn = torch.compile(fn, backend="eager")
x = torch.rand(4)
self.assertTrue(same(fn(x), opt_fn(x)))
def test_add_sub_alpha_out(self):
inp = torch.randn(2, 3, 4)
other = 1
alpha = 2
for op in [torch.add, torch.sub]:
out = torch.zeros(2, 3, 4)
compile_out = torch.zeros(2, 3, 4)
op(inp, other, alpha=alpha, out=out)
compiled_fn = torch.compile(op, dynamic=True)
compiled_fn(inp, other, alpha=alpha, out=compile_out)
self.assertTrue(same(out, compile_out))
def test_negative_shape_guard(self):
def fn(x):
if x.size() != (5, 1, 2, 3):
return x.cos()
return x.sin()
counter = torch._dynamo.testing.CompileCounter()
opt_fn = torch.compile(fn, backend=counter, dynamic=True)
x = torch.ones(5, 1, 3, 4)
x2 = torch.ones(5, 1, 2, 3)
self.assertEqual(fn(x), opt_fn(x))
self.assertEqual(fn(x2), opt_fn(x2))
self.assertEqual(counter.frame_count, 2)
@torch._dynamo.config.patch(capture_scalar_outputs=True)
def test_deferred_runtime_asserts(self):
@torch.compile(fullgraph=True)
def f(x):
y = x.item()
torch._check_is_size(y)
if y >= 0:
return x * 2
else:
return x * 3
f(torch.tensor([3]))
self.assertRaises(RuntimeError, lambda: f(torch.tensor([-2])))
def test_addr_alpha_beta_out(self):
inp = torch.randn(2, 3)
vec1 = torch.randn(2)
vec2 = torch.randn(3)
alpha = 2
beta = 5
out = torch.zeros(2, 3)
compile_out = torch.zeros(2, 3)
torch.addr(inp, vec1, vec2, alpha=alpha, beta=beta, out=out)
compiled_fn = torch.compile(torch.addr, dynamic=True)
compiled_fn(inp, vec1, vec2, alpha=alpha, beta=beta, out=compile_out)
self.assertTrue(same(out, compile_out))
def test_setattr_requires_grad_graph_breaks(self):
def fn(x):
z = x + 4
x.requires_grad = True
y = x * z
return y
for backend in ["count", "eager", "aot_eager"]:
if backend == "count":
backend = CompileCounter()
opt_fn = torch.compile(fn, backend=backend)
eager = torch.zeros(5)
compiled = eager.clone()
out_eager = fn(eager)
out_opt = opt_fn(compiled)
self.assertEqual(out_eager, out_opt)
out_eager.sum().backward()
out_opt.sum().backward()
self.assertEqual(eager, compiled)
if isinstance(backend, CompileCounter):
self.assertEqual(backend.frame_count, 2) # graph breaks
def test_dynamic_shapes_double_not_equal(self):
# https://github.com/pytorch/pytorch/issues/113393
def fn(x):
if x.size() != (5, 1, 2, 3):
return x.cos()
return x.sin()
opt_fn = torch.compile(fn, backend="eager")
x = torch.ones(5, 1, 2, 3)
x2 = torch.ones(5, 1, 3, 4)
self.assertEqual(fn(x), opt_fn(x))
self.assertEqual(fn(x2), opt_fn(x2))
def test_inductor_no_recursionerror_on_for_loops(self):
def forward(x):
for _ in range(10000):
x = 1.0 * x
return x
self.assertTrue(
same(torch.compile(forward)(torch.tensor([1.0])), torch.tensor([1.0]))
)
def test_user_defined_object_callable(self):
# https://github.com/pytorch/pytorch/issues/114019
class MyCallable:
def __call__(self, x):
return x + 1
def fn(x):
# Create in graph - will not have source
return MyCallable()(x)
fn_opt = torch.compile(fn, backend="eager", fullgraph=True)
self.assertEqual(fn_opt(torch.zeros(1)), fn(torch.zeros(1)))
@torch._dynamo.config.patch(log_compilation_metrics=True)
def test_many_views_with_mutation(self):
# When symbolic storage offsets were added in #113734, tensors_definitely_do_not_overlap
# began adding shape guards - a quadratic amount relative to the number of inputs.
# Test this configuration, and test that a reasonable number of guards are added.
# Note, when dynamic shapes are turned on, this test fails and we still get quadratic guards.
def fn(x):
x[0].relu_()
return torch.cat(x).sum()
AMT = 32
src = torch.rand(16 * (AMT + 1))
x = [src.as_strided((4, 4), (4, 1), 3 + 16 * i) for i in range(AMT)]
torch._dynamo.reset()
torch._dynamo.utils.clear_compilation_metrics()
res = torch.compile(fn, backend="aot_eager")(x)
all_metrics = torch._dynamo.utils.get_compilation_metrics()
total_guards = sum(metric.guard_count for metric in all_metrics)
self.assertLess(total_guards, AMT * 8)
total_shape_env_guards = sum(
metric.shape_env_guard_count for metric in all_metrics
)
self.assertLess(total_shape_env_guards, AMT * 8)
# https://github.com/pytorch/pytorch/issues/118799
def test_subclass_graph_output_repro(self):
@torch._dynamo.allow_in_graph
def to_subclass(x):
return TwoTensor(x.clone(), x.clone())
def f(x):
tmp_subclass = to_subclass(x)
return tmp_subclass.view(-1)
x = torch.ones(2)
out_ref = f(x)
out_test = torch.compile(f, backend="aot_eager")(x)
self.assertEqual(out_ref, out_test)
def test_numpy_tobytes_no_error(self):
def fn(x):
x += 1
z = x.tobytes()
x += 1
return z
cnt = torch._dynamo.testing.CompileCounter()
opt_fn = torch._dynamo.optimize(cnt)(fn)
opt_arg, arg = np.array([1, 2]), np.array([1, 2])
self.assertEqual(opt_fn(opt_arg), fn(arg))
self.assertEqual(cnt.frame_count, 2)
def test_numpy_not_ndarray_recompiles(self):
import torch
def fn(x=None):
if x is None:
x = np.ones(3)
elif isinstance(x, int):
x = np.ones(6)
elif isinstance(x, str):
x = np.ones(9)
return x**2
cnt = torch._dynamo.testing.CompileCounter()
opt_fn = torch._dynamo.optimize(cnt)(fn)
x = np.zeros((2, 2))
self.assertEqual(opt_fn(x), fn(x))
self.assertEqual(cnt.frame_count, 1)
self.assertEqual(opt_fn(), fn())
self.assertEqual(cnt.frame_count, 2)
self.assertEqual(opt_fn(10), fn(10))
self.assertEqual(cnt.frame_count, 3)
self.assertEqual(opt_fn("10"), fn("10"))
self.assertEqual(cnt.frame_count, 4)
@parametrize(
"backend",
["eager", "aot_eager", "inductor"],
)
@parametrize(
"func_name",
["func1", "func2", "func3"],
)
def test_tensor_set_data(self, backend, func_name):
# https://github.com/pytorch/pytorch/issues/113030
def func1(x, y):
x.data = y
x.add_(1)
return x
def func2(x, y):
x.data = y
y.data = torch.zeros([0])
return x
def func3(x, y):
z = x
x.data = y
y.data = torch.zeros([0])
return torch.tensor(x is z)
funcs = {"func1": func1, "func2": func2, "func3": func3}
func = funcs[func_name]
if backend != "eager" and func is func1:
# add_ not working w/ aot_autograd?
return
torch._dynamo.reset()
cnt = torch._dynamo.testing.CompileCounterWithBackend(backend)
compiled_fn = torch.compile(func, backend=cnt, fullgraph=True)
requires_grad = func is not func1
for i in range(0, 5):
# Inputs
eager_a = torch.ones([6], requires_grad=requires_grad)
compiled_a = torch.ones([6], requires_grad=requires_grad)
eager_b = torch.ones([6], requires_grad=requires_grad)
compiled_b = torch.ones([6], requires_grad=requires_grad)
# Eager
out_eager = func(eager_a, eager_b)
# Compiled
out_compiled = compiled_fn(compiled_a, compiled_b)
self.assertEqual(eager_a, compiled_a)
self.assertEqual(eager_b, compiled_b)
self.assertTrue(torch.equal(out_eager, out_compiled))
# func1 hits a leaf Variable that requires grad is being used in an in-place operation
if requires_grad:
bwd_inp_eager = torch.randn([6])
bwd_inp_compiled = torch.clone(bwd_inp_eager)
eager_a.backward(bwd_inp_eager)
compiled_a.backward(bwd_inp_compiled)
self.assertEqual(eager_a.grad, compiled_a.grad)
# Prove guarding works - we run the compiled_fn 5 times
# frame_count should stay at 1.
self.assertEqual(cnt.frame_count, 1)
@unittest.skipIf(
TEST_WITH_ROCM or not PLATFORM_SUPPORTS_FLASH_ATTENTION,
"flash attention not supported",
)
def test_flash_attn_backward_mixed_strides(self):
# in this repro, "grad_out" and "value" are transposed tensors,
# but "key" and "value" are contiguous
def gen_inputs(device):
return (
torch.randn(
2, 513, 16, 64, dtype=torch.float16, device=device
).transpose(1, 2),
torch.randn(2, 16, 513, 64, dtype=torch.float16, device=device),
torch.randn(2, 16, 513, 64, dtype=torch.float16, device=device),
torch.randn(
2, 513, 16, 64, dtype=torch.float16, device=device
).transpose(1, 2),
torch.randn(2, 16, 513, 64, dtype=torch.float16, device=device),
torch.randn(2, 16, 513, device=device),
None,
None,
513,
513,
0.0,
False,
torch.tensor(1, dtype=torch.int64),
torch.tensor(1, dtype=torch.int64),
)
inps_cuda = gen_inputs("cuda")
inps_meta = gen_inputs("meta")
(
out1_ref,
out2_ref,
out3_ref,
) = torch.ops.aten._scaled_dot_product_flash_attention_backward(
*inps_cuda, scale=0.125
)
from torch._meta_registrations import meta__scaled_dot_product_flash_backward
out1_test, out2_test, out3_test = meta__scaled_dot_product_flash_backward(
*inps_meta, scale=0.125
)
self.assertEqual(out1_ref.shape, out1_test.shape)
self.assertEqual(out1_ref.stride(), out1_test.stride())
self.assertEqual(out2_ref.shape, out2_test.shape)
self.assertEqual(out2_ref.stride(), out2_test.stride())
self.assertEqual(out3_ref.shape, out3_test.shape)
self.assertEqual(out3_ref.stride(), out3_test.stride())
def test_user_ctor_ctx_manager(self):
class UserCtxManager:
def __enter__(self):
return 1
def __exit__(self, exc_type, exc_val, exc_tb):
pass
def fn(x, y):
ucm = UserCtxManager()
return x * x
cnt = torch._dynamo.testing.CompileCounter()
opt_fn = torch._dynamo.optimize(cnt, nopython=True)(fn)
x = torch.rand([2, 2])
opt_fn(x, x)
self.assertExpectedInline(cnt.frame_count, """1""")
@torch._dynamo.config.patch(capture_scalar_outputs=True)
def test_unbacked_arange_in_bounds(self):
# see https://github.com/pytorch/pytorch/issues/113002
class PaddingNet(nn.Module):
def __init__(self) -> None:
super().__init__()
def forward(self, lengths):
max_seq_len = lengths.max().item()
row_vector = torch.arange(0, max_seq_len, 1)
matrix = torch.unsqueeze(lengths, dim=-1)
mask = row_vector < matrix
mask = mask.type(torch.float32)
mask_3d_btd = mask[:, :, None]
return mask_3d_btd
model = PaddingNet()
lengths = torch.tensor([5, 4, 4, 4], dtype=torch.int32)
cnt = torch._dynamo.testing.CompileCounter()
opt_fn = torch._dynamo.optimize(cnt, nopython=True)(model)
opt_fn(lengths)
self.assertEqual(cnt.frame_count, 1)
def test_overlapping_inputs_with_dynamic_shapes_error(self):
@torch.compile(backend="aot_eager")
def fn(a, b, c, d, e, f):
a.mul_(2)
b.mul_(2)
c.mul_(2)
d.mul_(2)
e.mul_(2)
f.mul_(2)
base = torch.ones(2, 20)
a = base[:, 0:2]
b = base[:, 2:4]
c = base[:, 4:6]
d = base[:, 6:8]
e = base[:, 8:10]
f = base[:, 10:12]
f2 = base[:, 10:14]
out = fn(a, b, c, d, e, f)
with self.assertRaisesRegex(
AssertionError, "is being compiled with dynamic shapes"
):
out2 = fn(a, b, c, d, e, f2)
def test_user_ctor_ctx_manager_custom_init(self):
class UserCtxManager:
def __init__(self, x):
x[0] = 10
def __enter__(self):
return 1
def __exit__(self, exc_type, exc_val, exc_tb):
pass
def fn(x, y):
ucm = UserCtxManager(y)
return x * y[0]
cnt = torch._dynamo.testing.CompileCounter()
opt_fn = torch._dynamo.optimize(cnt, nopython=True)(fn)
x = torch.rand([2, 2])
self.assertEqual(opt_fn(x, [5]), fn(x, [5]))
self.assertExpectedInline(cnt.frame_count, """1""")
def test_user_ctor_ctx_manager_custom_init_graph_break(self):
counter = [0]
class UserCtxManager:
def __init__(self, k):
k[0] += 1
def __enter__(self):
return 1
def __exit__(self, exc_type, exc_val, exc_tb):
pass
def fn(x, counter):
x = x * x
ucm = UserCtxManager(counter)
return x * x
cnt = torch._dynamo.testing.CompileCounter()
opt_fn = torch._dynamo.optimize(cnt)(fn)
x = torch.rand([2, 2])
self.assertEqual(opt_fn(x, counter), fn(x, counter))
self.assertEqual(counter[0], 2)
for i in range(0, 10):
opt_fn(x, counter)
self.assertEqual(counter[0], 12)
if torch._dynamo.config.assume_static_by_default:
self.assertExpectedInline(cnt.frame_count, """2""")
else:
self.assertExpectedInline(cnt.frame_count, """1""")
@unittest.expectedFailure
def test_many_overlapping_inputs_does_not_explode_guards(self):
from torch._dynamo.backends.common import aot_autograd
# Before, this was (9702, 0)
num_shape_guards = None
num_aot_guards = None
num_compiles = 0
def guard_count_backend(gm, *args):
nonlocal num_shape_guards
nonlocal num_aot_guards
nonlocal num_compiles
num_shape_guards = len(
torch._guards.TracingContext.try_get().fake_mode.shape_env.guards
)
num_aot_guards = len(
torch._guards.TracingContext.try_get().guards_context.aotautograd_guards
)
num_compiles += 1
return gm
aot_guard_counter = aot_autograd(fw_compiler=guard_count_backend)
@torch.compile(backend=aot_guard_counter, dynamic=True)
def f(*args):
for a in args:
a.add_(1)
x = torch.ones(1000, requires_grad=True)
args = x.split(10)
with torch.no_grad():
f(*args)
# In this example, there were 4950 guards (roughly (# tensors) ^ 2 // 2),
# because every pair of aliased inputs needs a guard.
self.assertTrue(num_aot_guards < 5000)
# But there are no dynamic shape guards.
self.assertEqual(num_shape_guards, 0)
# don't recompile
with torch.no_grad():
f(*args)
self.assertEqual(num_compiles, 1)
def test_issue134451(self):
class BoundingBox2DIndex(IntEnum):
_X = 0
_Y = 1
_HEADING = 2
_LENGTH = 3
_WIDTH = 4
@classmethod
def size(cls):
return 5
@classmethod
@property
def X(cls):
return cls._X
@classmethod
@property
def Y(cls):
return cls._Y
@classmethod
@property
def HEADING(cls):
return cls._HEADING
@classmethod
@property
def LENGTH(cls):
return cls._LENGTH
@classmethod
@property
def WIDTH(cls):
return cls._WIDTH
@classmethod
@property
def POINT(cls):
# assumes X, Y have subsequent indices
return slice(cls._X, cls._Y + 1)
@classmethod
@property
def STATE_SE2(cls):
# assumes X, Y, HEADING have subsequent indices
return slice(cls._X, cls._HEADING + 1)
class SimpleModel(nn.Module):
def __init__(self):
super().__init__()
self._mlp_states = nn.Sequential(
nn.Linear(10, 20),
nn.ReLU(),
nn.Linear(20, BoundingBox2DIndex.size()),
)
def forward(self, x):
agent_states = self._mlp_states(x)
agent_states[..., BoundingBox2DIndex.POINT] = (
agent_states[..., BoundingBox2DIndex.POINT].tanh() * 32
)
agent_states[..., BoundingBox2DIndex.HEADING] = (
agent_states[..., BoundingBox2DIndex.HEADING].tanh() * torch.pi
)
return agent_states
model = SimpleModel().eval()
input_tensor = torch.randn(1, 10, dtype=torch.float32)
opt = torch.compile(model.eval(), backend="eager", fullgraph=True)
actual = opt(input_tensor)
try:
expected = model(input_tensor)
except Exception as e:
raise unittest.SkipTest("eager failed, requires Python>=3.12") from e
self.assertEqual(actual, expected)
def test_invalid_seq_unpack(self):
def myfn(arg):
(a, b) = arg
def fn():
return myfn((1, 2, 3))
try:
torch.compile(fn)()
except ValueError:
pass
else:
self.fail("expected exception")
def test_megablocks_moe(self):
try:
from megablocks.layers import moe
from megablocks.layers.arguments import Arguments
except ImportError as e:
raise unittest.SkipTest("requires megablocks") from e
bs, sl, hs, num_experts, top_k = (16, 1024, 512, 1, 1)
args = Arguments(
hidden_size=hs,
ffn_hidden_size=hs * 2,
moe_num_experts=num_experts,
moe_capacity_factor=1,
moe_top_k=top_k,
)
moe_mlp = moe.MoE(args)
moe_mlp.cuda(torch.cuda.current_device()).half()
x = torch.randn(sl, bs, hs).cuda().half()
out1, _ = moe_mlp(x)
out2, _ = torch.compile(moe_mlp, backend="eager")(x)
self.assertEqual(out1, out2)
def test_udf_classes_reconstruction(self):
def fn(x):
o = T(5)
return o.x + x
opt_fn = torch.compile(fn, backend="eager")
T = IncByOne
x = torch.randn(4)
self.assertEqual(fn(x), opt_fn(x))
# This should recompile
T = IncByTwo
self.assertEqual(fn(x), opt_fn(x))
def test_contains_range_constprop(self):
def fn(x):
# dynamo should const prop to False
if 3 in range(0, 10):
return x + 1
else:
return x + 2
opt_fn = torch.compile(fn, backend="eager")
x = torch.zeros(4)
self.assertEqual(fn(x), opt_fn(x))
# https://github.com/pytorch/pytorch/issues/104505
def test_as_strided_on_base_with_mutation_works(self):
def foo(a):
f = a.as_strided((2,), (1,), 0)
f.add_(1.0)
return a
a = torch.randn(2, 4)
a_ref = a.clone()
out_ref = foo(a_ref)
f_compiled = torch.compile(foo, backend="aot_eager")
out = f_compiled(a)
self.assertEqual(out_ref, out)
self.assertEqual(a_ref, a)
# https://github.com/pytorch/pytorch/issues/104505
def test_as_strided_on_existing_view_banned(self):
def foo(a):
e = a.diagonal()
f = e.as_strided((2,), (1,), 0)
f.add_(1.0)
return a
a = torch.randn(2, 4)
a_ref = a.clone()
out_ref = foo(a_ref)
f_compiled = torch.compile(foo, backend="aot_eager")
with self.assertRaisesRegex(
RuntimeError,
"encountered a mutation on a view chain of length 2, where view 1 was an as_strided",
):
out = f_compiled(a)
def test_dont_aggressively_write_assert(self):
record_graph = torch._dynamo.testing.EagerAndRecordGraphs()
@torch.compile(dynamic=True, backend=record_graph)
def f(x):
assert x.shape[0] > 3
assert x[0].sum() > 0
assert 1 % (x.shape[0] // 2) != 0
assert 32 * (x.shape[0] // 2) ** 2 - 16 * (x.shape[0] // 2) != 0
return x.cos()
f(torch.ones(6, 4))
graph = record_graph.graphs[0]
# It is bit annoying that we generate useless statements for
# shape guards, but DCE should be able to remove them since t
# there is no backed assert on them. The reason this is ok is
# because dynamo will only skip the assert statement, but not
# the instructions before it.
self.assertExpectedInline(
str(graph.code).strip(),
"""\
def forward(self, s0 : torch.SymInt, s1 : torch.SymInt, L_x_ : torch.Tensor):
l_x_ = L_x_
getitem_2 = l_x_[0]
sum_1 = getitem_2.sum(); getitem_2 = None
gt_1 = sum_1 > 0; sum_1 = None
_assert_async = torch._assert_async(gt_1, 'assertion error'); gt_1 = _assert_async = None
cos = l_x_.cos(); l_x_ = None
return (cos,)""",
)
for node in graph.graph.nodes:
if "example_value" in node.meta and isinstance(
node.meta["example_value"], torch._subclasses.fake_tensor.FakeTensor
):
shape_env = node.meta["example_value"].fake_mode.shape_env
lower_ranges = [val.lower for val in shape_env.var_to_range.values()]
self.assertTrue(lower_ranges == [4, 2])
@torch.compile(dynamic=True, backend=record_graph)
def f_fail(x):
assert x.shape[0] < 3
# We graph-break here, so the failure should be eager
with self.assertRaisesRegex(AssertionError, ""):
f_fail(torch.ones(6, 4))
def test_detectron2_instances_cat(self):
class Instances:
def __init__(self, image_size: Tuple[int, int], **kwargs: Any):
self._image_size = image_size
self._fields: Dict[str, Any] = {}
for k, v in kwargs.items():
self.set(k, v)
@property
def image_size(self) -> Tuple[int, int]:
return self._image_size
def __setattr__(self, name: str, val: Any) -> None:
if name.startswith("_"):
super().__setattr__(name, val)
else:
self.set(name, val)
def __getattr__(self, name: str) -> Any:
if name == "_fields" or name not in self._fields:
raise AttributeError(
f"Cannot find field '{name}' in the given Instances!"
)
return self._fields[name]
def __len__(self) -> int:
for v in self._fields.values():
# use __len__ because len() has to be int and is not friendly to tracing
return v.__len__()
raise NotImplementedError("Empty Instances does not support __len__!")
def set(self, name: str, value: Any) -> None:
with warnings.catch_warnings(record=True):
data_len = len(value)
if len(self._fields):
assert (
len(self) == data_len
), f"Adding a field of length {data_len} to a Instances of length {len(self)}"
self._fields[name] = value
def get(self, name: str) -> Any:
return self._fields[name]
@staticmethod
def cat(instance_lists: List["Instances"]) -> "Instances":
assert all(isinstance(i, Instances) for i in instance_lists)
assert len(instance_lists) > 0
if len(instance_lists) == 1:
return instance_lists[0]
image_size = instance_lists[0].image_size
if not isinstance(
image_size, torch.Tensor
): # could be a tensor in tracing
for i in instance_lists[1:]:
assert i.image_size == image_size
ret = Instances(image_size)
for k in instance_lists[0]._fields.keys():
values = [i.get(k) for i in instance_lists]
v0 = values[0]
if isinstance(v0, torch.Tensor):
values = torch.cat(values, dim=0)
elif isinstance(v0, list):
values = list(itertools.chain(*values))
elif hasattr(type(v0), "cat"):
values = type(v0).cat(values)
else:
raise ValueError(
f"Unsupported type {type(v0)} for concatenation"
)
ret.set(k, values)
return ret
instances = [
Instances((16, 16), a=torch.randn(16, 16), b=torch.randn(16, 16))
for _ in range(3)
]
@torch.compile(backend="eager", fullgraph=True)
def fn(instances):
return instances[0].cat(instances)
actual = fn(instances)
expected = instances[0].cat(instances)
self.assertEqual(type(actual), type(expected))
self.assertEqual(actual.__dict__, expected.__dict__)
def test_weakref_construction(self):
def fn(x, y):
x_weak = weakref.ref(x)
return x_weak() * y
x = torch.randn(4)
y = torch.randn(4)
ref = fn(x, y)
opt_fn = torch.compile(fn, backend="eager", fullgraph=True)
res = opt_fn(x, y)
self.assertEqual(ref, res)
def test_weakref(self):
def fn(x_weak, weight, y):
if x_weak is not None and x_weak() is not weight:
return torch.sin(y)
return torch.cos(y)
weight = torch.randn(4)
y = torch.randn(4)
x_weak = weakref.ref(weight)
ref = fn(x_weak, weight, y)
opt_fn = torch.compile(fn, backend="eager", fullgraph=True)
res = opt_fn(x_weak, weight, y)
self.assertEqual(ref, res)
def test_weakref_reconstruct(self):
def fn(x_weak, weight, y):
y = torch.sin(y)
referent = x_weak()
torch._dynamo.graph_break()
if referent is not weight:
return torch.sin(y)
return torch.cos(y)
weight = torch.randn(4)
y = torch.randn(4)
x_weak = weakref.ref(weight)
ref = fn(x_weak, weight, y)
cnt = torch._dynamo.testing.CompileCounter()
opt_fn = torch.compile(fn, backend=cnt)
res = opt_fn(x_weak, weight, y)
self.assertEqual(ref, res)
self.assertEqual(cnt.frame_count, 2)
def test_weakref_del(self):
def fn(x_weak, y):
x = x_weak()
if x is not None:
return torch.sin(y)
return torch.cos(y)
weight = torch.randn(4)
x_weak = weakref.ref(weight)
y = torch.randn(4)
opt_fn = torch.compile(fn, backend="eager", fullgraph=True)
ref = fn(x_weak, y)
res = opt_fn(x_weak, y)
self.assertEqual(ref, res)
del weight
gc.collect()
ref = fn(x_weak, y)
res = opt_fn(x_weak, y)
self.assertEqual(ref, res)
# @torch._functorch.config.patch(
# recompute_views=True,
# )
# def test_storage_resize_forward_full_graph(self):
# class TestModule(torch.nn.Module):
# def __init__(self) -> None:
# super().__init__()
# self.param = torch.nn.Parameter(torch.randn(4, 4))
# def forward(self, x):
# self.param.untyped_storage().resize_(
# self.param.numel() * self.param.itemsize
# )
# with torch.no_grad():
# torch._foreach_copy_([self.param], [x])
# out = torch.matmul(self.param, self.param)
# self.param.untyped_storage().resize_(0)
# return out
# def post_accumulate_grad_hook(param):
# param.untyped_storage().resize_(0)
# # Beginning of backward, resize and put data into the param
# def pre_backward_hook(module, grad) -> None:
# module.param.untyped_storage().resize_(
# self.param.numel() * self.param.itemsize
# )
# with torch.no_grad():
# # simulates loading data into param from allgather
# module.param.fill_(2)
# def post_forward_hook(module, args, output):
# output.register_hook(functools.partial(pre_backward_hook, module))
# x = torch.randn(4, 4)
# mod_ref = TestModule()
# mod_test = deepcopy(mod_ref)
# # Start the param off with zero storage size to mimic fsdp
# mod_ref.param.untyped_storage().resize_(0)
# mod_test.param.untyped_storage().resize_(0)
# # Resize storage at beginning of backward
# # Free storage at end of backward
# mod_ref.register_forward_hook(post_forward_hook, prepend=False)
# mod_ref.param.register_post_accumulate_grad_hook(post_accumulate_grad_hook)
# mod_test.register_forward_hook(post_forward_hook, prepend=False)
# mod_test.param.register_post_accumulate_grad_hook(post_accumulate_grad_hook)
# mod_test = torch.compile(mod_test, backend=aot_graph_capture_backend)
# out_ref = mod_ref(x)
# out_test = mod_test(x)
# self.assertExpectedInline(
# str(fw_graph[0].code.strip()),
# """\
# def forward(self, primals_1, primals_2):
# _foreach_copy = torch.ops.aten._foreach_copy.default([primals_1], [primals_2]); primals_1 = primals_2 = None
# getitem = _foreach_copy[0]; _foreach_copy = None
# mm = torch.ops.aten.mm.default(getitem, getitem)
# return [mm, getitem]""",
# )
# self.assertEqual(out_ref, out_test)
def test_super_in_staticmethod(self):
class A:
@staticmethod
def foo():
return super().__init__()
def fn(obj):
return obj.foo()
obj = A()
try:
fn(obj)
except Exception as e:
orig_str = str(e)
self.assertIn("no arguments", orig_str)
try:
torch.compile(backend="eager")(fn)(obj)
except Exception as e:
compiled_str = str(e)
self.assertEqual(orig_str, compiled_str)
def test_super_staticmethod(self):
class Parent:
@staticmethod
def greet():
return 5
class Child(Parent):
@staticmethod
def greet(x):
return x * super(Child, Child).greet()
child = Child()
def fn(x):
return child.greet(x)
opt_fn = torch.compile(fn, backend="eager", fullgraph=True)
x = torch.ones(4)
ref = fn(x)
res = opt_fn(x)
self.assertEqual(ref, res)
def test_super_diamond(self):
class A:
def __init__(self):
super().__init__()
self.a = 5
class Nothing:
pass
class B(Nothing, A):
def __init__(self):
super().__init__()
self.b = 10
def run(self, x):
return self.a * self.b * x
def fn(x):
b = B()
return b.run(x)
opt_fn = torch.compile(fn, backend="eager", fullgraph=True)
x = torch.randn(4)
ref = fn(x)
res = opt_fn(x)
self.assertEqual(ref, res)
def test_vc_bumped_in_inference_graph(self):
@torch.compile
def f(x):
return x.mul_(2)
x = torch.randn(4)
vc_before = x._version
f(x)
vc_after = x._version
self.assertTrue(vc_after > vc_before)
def test_nn_module_callable(self):
class M(nn.Module):
def forward(self, x):
return x.sin()
def f(m):
return callable(m)
res = torch.compile(f, fullgraph=True)(M())
self.assertTrue(res)
def test_stk_sdd_is_transposed(self):
trigger_graph_break = False
def _is_transposed(x):
return (
not x.is_contiguous()
and x.stride()[0] == 1
and x.stride()[1] == x.size()[0]
)
class SDD(torch.autograd.Function):
@staticmethod
def forward(ctx, lhs, rhs):
ctx.save_for_backward(lhs, rhs)
out = torch.full_like(lhs, 1.0, dtype=lhs.dtype, device=lhs.device)
return out
@staticmethod
def backward(ctx, dy):
saved_tensors = ctx.saved_tensors
lhs, rhs = saved_tensors[:2]
trans_a = _is_transposed(lhs)
trans_b = _is_transposed(rhs)
dlhs = None
if ctx.needs_input_grad[0]:
dlhs = torch.full_like(lhs, 1.0 if trans_a else 2.0)
drhs = None
if ctx.needs_input_grad[1]:
drhs = torch.full_like(rhs, 1.0 if trans_b else 2.0)
if trigger_graph_break:
if _is_transposed(dy):
return dlhs + 1, drhs + 1, None, None
return dlhs, drhs, None, None
x1 = torch.randn((8, 8), requires_grad=True)
y1 = torch.randn((8, 8)).transpose(0, 1).requires_grad_(True)
x2 = torch.randn((8, 8), requires_grad=True)
y2 = torch.randn((8, 8)).transpose(0, 1).requires_grad_(True)
SDD.apply(x1, y1).sum().backward()
@torch.compile(backend="eager", fullgraph=True)
def fn():
return SDD.apply(x2, y2)
fn().sum().backward()
self.assertEqual(x1.grad, x2.grad)
self.assertEqual(y1.grad, y2.grad)
trigger_graph_break = True
with self.assertRaises(torch._dynamo.exc.Unsupported):
fn().sum().backward()
def test_partially_initialized_module_property(self):
class Matrix(torch.nn.Module):
def __init__(self, data):
super().__init__()
self._data = data
self.foo = 10 * self.blocking
@property
def data(self):
return self._data
@property
def blocking(self):
return self.data.shape[1]
@torch.compile(backend="eager", fullgraph=True)
def fn():
return Matrix(torch.randn(10, 20))
v = fn()
self.assertEqual(v.foo, 200)
self.assertEqual(v.data.shape, (10, 20))
self.assertEqual(type(v), Matrix)
def test_classmethod_with_slots(self):
class Mock:
__slots__ = ("_a",)
def __init__(self):
self._a = 2
@classmethod
def _m(cls):
return 3
def run(self, x):
return torch.sin(x) * self._a * self._m()
def fn(x):
mock = Mock()
return mock.run(x)
opt_fn = torch.compile(fn, backend="eager", fullgraph=True)
x = torch.randn(4)
self.assertEqual(fn(x), opt_fn(x))
def test_nn_parametrize(self):
class Module(nn.Module):
def __init__(self) -> None:
super().__init__()
self.param = torch.nn.Parameter(torch.randn(10, 10))
def forward(self, x):
return self.param @ x
class Parametrization(torch.nn.Module):
def forward(self, x):
return torch.sin(x)
m = Module()
torch.nn.utils.parametrize.register_parametrization(
m, "param", Parametrization()
)
sin_found = False
def backend(gm, _):
nonlocal sin_found
for node in gm.graph.nodes:
if node.target is torch.sin:
sin_found = True
return gm
opt_m = torch.compile(m, backend=backend, fullgraph=True)
inp = torch.randn(10, 10)
self.assertEqual(m(inp), opt_m(inp))
self.assertTrue(sin_found)
torch.nn.utils.parametrize.remove_parametrizations(m, "param")
sin_found = False
self.assertEqual(m(inp), opt_m(inp))
self.assertFalse(sin_found)
def test_nn_module_property_closure(self):
x = torch.randn(10, 10)
class Mod(torch.nn.Module):
@property
def y(self):
return torch.ones(10, 10) + x
def forward(self, x):
return x @ self.y
mod = Mod()
def fn(x):
return mod(x)
opt_fn = torch.compile(fn, backend="eager", fullgraph=True)
inp = torch.randn(10, 10)
self.assertEqual(fn(inp), opt_fn(inp))
def test_global_fn_mutation(self):
def foo(x, y):
return global_fn(x) + y
x = torch.ones(1)
y = torch.ones(1)
opt = torch.compile(foo, fullgraph=True, backend="eager")
self.assertEqual(opt(x, y), foo(x, y))
# Change global_fn
global global_fn
def new_fn(x):
return torch.cos(x)
global_fn = new_fn
self.assertEqual(opt(x, y), foo(x, y))
# ref https://github.com/pytorch/pytorch/issues/123974
def test_list_reverse(self):
def ladder(x):
trail = x.size(-1)
assert trail > 2
weights = []
for s in [trail, trail - 1, trail - 2]:
weights.append(torch.ones(s, s - 1))
for w in weights:
x = x @ w
weights.reverse()
for w in weights:
x = x @ w.t()
return x
data = torch.randn(3, 4)
opt_ladder = torch.compile(ladder, fullgraph=True, backend="eager")
self.assertEqual(opt_ladder(data), ladder(data))
def test_trace_functional_tensor_with(self):
from torch._subclasses.fake_tensor import FakeTensorMode
from torch._subclasses.functional_tensor import (
FunctionalTensor,
FunctionalTensorMode,
)
def f(a, tmp):
a_view = a.view(-1)
with torch.no_grad():
a.set_(tmp)
a_view.mul_(2)
return a + tmp
fake_mode = FakeTensorMode()
with FunctionalTensorMode():
inp = torch.ones(3, 3, requires_grad=True)
inp = fake_mode.from_tensor(inp, static_shapes=True)
inp = FunctionalTensor.to_functional(inp)
tmp = torch.ones(3, 3, requires_grad=True)
tmp = fake_mode.from_tensor(tmp, static_shapes=True)
tmp = FunctionalTensor.to_functional(tmp)
opt_f = torch.compile(f, backend="eager")
with self.assertRaisesRegex(
RuntimeError, "cannot mutate tensors with frozen storage"
):
opt_f(inp, tmp)
def test_const_dict_keyerror(self):
d = {}
def fn(x):
try:
y = d[0]
except KeyError:
y = 1
return x + y
opt_fn = torch.compile(fn, backend="eager")
inp = torch.randn(3, 3)
self.assertEqual(fn(inp), opt_fn(inp))
def test_dict_tag_guard(self):
class Foo:
def __init__(self) -> None:
self.scalar = 10
def fn(d, x):
return d["a"] * d["b"] * d["c"].scalar * x
foo = Foo()
d = {"a": 2, "b": 3, "c": foo}
opt_fn = torch.compile(fn, backend="eager")
inp = torch.randn(3, 3)
self.assertEqual(fn(d, inp), opt_fn(d, inp))
d["a"] = 4
self.assertEqual(fn(d, inp), opt_fn(d, inp))
# Check that recompilation happens
foo.scalar = 12
self.assertEqual(fn(d, inp), opt_fn(d, inp))
def test_nonconst_issubclass(self):
def fn(x):
if issubclass(x.__class__, np.ndarray):
return 1
return 0
opt_fn = torch.compile(fn, backend="eager")
opt_fn(np.ones([3, 3]))
def test_issue126128(self):
def fn():
x = torch.randn(1, 10)
y = torch.randn(10, 1)
return torch.mm(x, y).sum()
def fn2():
x = torch.randn(10, 100)
y = torch.randn(100, 10)
return torch.mm(x, y).sum()
with fresh_inductor_cache():
torch.compile(fn)()
torch.compile(fn2)()
def test_jit_script_defaults(self):
@torch.jit.script
def fast_cos(x, c: float = 2.0):
return torch.cos(x) * c
class Mod(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.fast_cos = fast_cos
def forward(self, x):
return self.fast_cos(x)
mod = Mod()
opt_mod = torch.compile(mod, backend="eager", fullgraph=True)
x = torch.randn(4)
self.assertEqual(mod(x), opt_mod(x))
def test_enum(self):
class ExplicitEnum(str, Enum):
@classmethod
def _missing_(cls, value):
raise ValueError(
f"{value} is not a valid {cls.__name__}, please select one of {list(cls._value2member_map_.keys())}"
)
class PaddingStrategy(ExplicitEnum):
LONGEST = "longest"
MAX_LENGTH = "max_length"
DO_NOT_PAD = "do_not_pad"
def fn(x):
a = PaddingStrategy("longest")
if a == PaddingStrategy.LONGEST:
return torch.sin(x)
return torch.cos(x)
x = torch.randn(3, 3)
opt_fn = torch.compile(fn, backend="eager", fullgraph=True)
self.assertEqual(fn(x), opt_fn(x))
def test_hasattr_builtin(self):
class MyClass:
foo: int = 1
def func(x, m):
if getattr(type(m), "foo", 0):
return x + MyClass.foo
return x
opt_func = torch.compile(func, backend="eager", fullgraph=True)
m = MyClass()
x = torch.zeros(())
self.assertEqual(func(x, m), opt_func(x, m))
self.assertEqual(func(x, 0), opt_func(x, 0))
def test_grad(self):
def fn(x, y):
x._grad = y
return x.grad.data
x = torch.randn(4, requires_grad=True)
y = torch.randn(4)
opt_fn = torch.compile(fn, backend="eager")
self.assertEqual(fn(x, y), opt_fn(x, y))
def test_nn_module_stack_bc(self):
from torch._dynamo.mutation_guard import GenerationTracker
def compiler(gm, *args):
module_stacks = [
node.meta.get("nn_module_stack", None) for node in gm.graph.nodes
]
module_stacks, _ = pytree.tree_flatten(module_stacks)
module_stacks = [x for x in module_stacks if isinstance(x, str)]
for stack in module_stacks:
self.assertTrue("_module" not in stack)
return gm.forward
class SubMod(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.linear = torch.nn.Linear(2, 2)
def forward(self, x):
return self.linear(x)
class Mod(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.submod1 = SubMod()
self.submod2 = SubMod()
def forward(self, x):
return self.submod1(x) + self.submod2(x)
mod = Mod()
opt_mod = torch.compile(mod, backend=compiler)
opt_mod(torch.randn(2, 2))
with torch._dynamo.config.patch(inline_inbuilt_nn_modules=True):
mod = Mod()
opt_mod = torch.compile(mod, backend=compiler)
opt_mod(torch.randn(2, 2))
# an example similar to Pippy usecase
mod = Mod()
GenerationTracker.tag(mod.submod1)
GenerationTracker.mark_class_dynamic(type(mod.submod1))
mod = Mod()
opt_mod = torch.compile(mod, backend=compiler)
opt_mod(torch.randn(2, 2))
def test_is_make_fx_tracing(self):
@torch.compile(backend="eager", fullgraph=True)
def fn(x):
torch.nn.modules.activation._is_make_fx_tracing()
return torch.sin(x)
fn(torch.rand(4))
def test_negative_floor_div_solve(self):
class CompiledClass(nn.Module):
def __init__(self) -> None:
super().__init__()
self.nums = torch.tensor([1, 2, 3, 4, 5, 6, 7, 8, 9, 10])
self.t = 5
def forward(self):
self.num = self.nums[self.t // 12]
self.t += 1
return self.num
m = CompiledClass()
m = torch.compile(m, backend="eager")
# the first call works
m()
# the second call causes a failure
m()
# https://github.com/pytorch/pytorch/issues/121621
def test_tensor_random(self):
def random_op(tensor, params):
res = tensor.random_(**params)
return res
random_op = torch.compile(random_op)
params = {"from": -10, "to": 10}
tensor = torch.randn([2, 3])
res = random_op(tensor, params)
# https://github.com/pytorch/pytorch/issues/131019
def test_tensor_uniform(self):
def uniform_op(tensor, params):
res = tensor.uniform_(**params)
return res
uniform_op = torch.compile(uniform_op)
params = {"from": -10, "to": 10}
tensor = torch.randn([2, 3])
res = uniform_op(tensor, params)
def test_data_attr_mutation_after_saved_for_bw(self):
def f(x):
out = x.sin()
x.data.mul_(2)
return out
x = torch.randn(4, requires_grad=True)
x_test = x.clone().detach().requires_grad_(True)
out = f(x)
out_test = torch.compile(f, backend="aot_eager")(x_test)
self.assertEqual(out, out_test)
out.sum().backward()
out_test.sum().backward()
self.assertEqual(x.grad, x_test.grad)
# https://github.com/pytorch/pytorch/issues/128072
def test_map_with_multiple_args(self):
def f(a, b):
return a[0] * b[0] + a[1] * b[1]
def gen_inps(len_x, len_y):
x = [torch.randn(5) for _ in range(len_x)]
y = [torch.randn(5) for _ in range(len_y)]
return x, y
def g(x, y):
return map(f, x, y)
opt_g = torch.compile(g, fullgraph=True, backend="eager")
inps = gen_inps(3, 3)
self.assertEqual(type(g(*inps)), type(opt_g(*inps)))
self.assertEqual(tuple(g(*inps)), tuple(opt_g(*inps)))
inps = gen_inps(3, 5)
self.assertEqual(type(g(*inps)), type(opt_g(*inps)))
self.assertEqual(tuple(g(*inps)), tuple(opt_g(*inps)))
def test_staticmethod_allow_in_graph(self):
class MyClass:
i = 3
@staticmethod
def foo_inner(x):
return torch.mul(x, MyClass.i)
# if dynamo inlines with fullgraph, will error
# verify that dynamo doesn't inline
@staticmethod
@torch._dynamo.allow_in_graph
def foo1(x):
torch._dynamo.graph_break()
return MyClass.foo_inner(x)
@torch.compile(backend="eager", fullgraph=True)
def f_bad(x):
return MyClass.foo1(x)
f_bad(torch.ones(2, 2))
def test_guard_with_tuple_mutation(self):
class Foo:
def __init__(self) -> None:
self.x = 10
foo = Foo()
d = {
"a": 2,
"b": (foo,),
}
def fn(x, d):
return x * d["a"] * d["b"][0].x
opt_fn = torch.compile(fn, backend="eager")
inp = torch.randn(3, 3)
self.assertEqual(fn(inp, d), opt_fn(inp, d))
d["b"][0].x = 12
self.assertEqual(fn(inp, d), opt_fn(inp, d))
def test_compile_complex_conj(self):
def f(x):
return torch.mul(x, 2j)
x_ref = torch.randn(4, 2, requires_grad=True)
x_test = x_ref.clone().detach().requires_grad_(True)
out_ref = f(torch.view_as_complex(x_ref))
out_test = torch.compile(f, backend="aot_eager")(torch.view_as_complex(x_test))
self.assertEqual(out_ref, out_test)
torch.view_as_real(out_ref).sum().backward()
torch.view_as_real(out_test).sum().backward()
self.assertEqual(x_ref.grad, x_test.grad)
# https://github.com/pytorch/pytorch/issues/132200
def test_partitioner_cse_respects_mutation_boundaries(self):
set_available = hasattr(torch.ops, "fsdp") and hasattr(torch.ops.fsdp, "set_")
if not set_available:
return
@torch.compile(backend="aot_eager_decomp_partition")
def f(x, l):
# z0 and z1 can be CSEd
z0 = x.sin()
z1 = x.sin()
y = x + 1
torch.ops.fsdp.copy_.default(x, y)
# z3 and z3 can be CSEd with each other,
# but *not* with z0/z1 (they cross a mutation boundary)
z2 = x.sin()
z3 = x.sin()
return z0, z1, z2, z3, l**2
x = torch.randn(3)
x_clone = x.clone()
l = torch.randn(3, requires_grad=True)
z0, z1, z2, z3, _ = f(x, l)
# the partitioner runs CSE. We expect that of the 4 sin() ops above:
# - the first 2 are CSE'd
# - the last 2 are CSE'd
# - the set_() op in the middle is a mutation barrier, preventing CSE
self.assertEqual(z0, (x_clone).sin())
self.assertEqual(z1, (x_clone).sin())
self.assertEqual(z2, (x_clone + 1).sin())
self.assertEqual(z3, (x_clone + 1).sin())
# https://github.com/pytorch/pytorch/issues/132197
def test_fsdp_set_input_mutation_applied_when_input_gets_no_gradients(self):
set_available = hasattr(torch.ops, "fsdp") and hasattr(torch.ops.fsdp, "set_")
if not set_available:
return
@torch.compile(backend="aot_eager_decomp_partition")
def f(x, l):
z = x.sin()
y = x + 1
# graph input has its storage mutated
torch.ops.fsdp.copy_.default(x, y)
z2 = x.sin()
return z2, l**2
x = torch.randn(3)
x_test = x.clone()
l = torch.randn(3, requires_grad=True)
result, _ = f(x, l)
result_test, _ = torch.compile(f, backend="aot_eager_decomp_partition")(
x_test, l
)
self.assertEqual(result, result_test)
self.assertEqual(x, x_test)
def test_changing_stride(self):
cnt = torch._dynamo.testing.CompileCounter()
@torch.compile(backend=cnt)
def fn(x, y):
return x * y
for i in range(1, 4):
x = torch.randn(4, i)
# create a view for i > 1
if i == 1:
x1 = x
else:
x1 = x[:, 0:1]
y = torch.randn(4, 1)
print(x1.shape, y.shape)
fn(x1, y)
self.assertTrue(cnt.frame_count <= 2)
@torch._dynamo.config.patch(guard_nn_modules=False)
@torch._dynamo.config.patch(inline_inbuilt_nn_modules=False)
def test_inlining_cornercase(self):
"""
nn.Modules can be mapped to either NNModuleVariable or UnspecializedNNModuleVariable. For NNModuleVariable, the
tensor attributes become part of the Dynamo graph. For unspecialized, they are lifted as inputs.
But there is a cornercase. Suppose you have NNModuleVariable with a submodule that is
UnspecializedNNModuleVariable. Today, Dynamo will still consider the submodule as specialized (courtesy of
guard.source().is_nn_module()). In retrospect, this is a mistake but there are dependencies of export and also
cudagraphs which make it harder to fix the corner case right away. The long term solution is
inline_inbuilt_nn_modules anyways, so we might have to live with this cornercase in the short term.
We are starting to annotate the source of each nn module more precisely - NNModuleVariable attribute is marked
as NNModuleSource, UnspecilaizedNNModuleVariable attribute is marked as UnspecializedNNModuleSource. But this
changes the behavior for the cornercase. And fails some tests which have unfortunately relied on this behavior.
To solve this, we tag the source only when inline_inbuilt_nn_module flag is turned on.
In this test, we purposely turn the flag off, testing that the tagging is disabled.
"""
class SubMod(torch.nn.Module):
def __init__(self):
super().__init__()
self.linear = torch.nn.Linear(1, 1)
self.a = torch.randn(1, 1)
self.counter = 0
self.multipliers = [2.2, 3.3]
def forward(self, x):
self.counter += 1
return (
self.linear(x) * self.a * self.multipliers[0] * self.multipliers[1]
)
class Mod(torch.nn.Module):
def __init__(self):
super().__init__()
self.submod = SubMod()
def forward(self, x):
return self.submod(x)
mod = Mod()
opt_mod = torch.compile(mod, backend="eager")
x = torch.randn(1, 1)
ref = mod(x)
res = opt_mod(x)
mod.submod.multipliers = [3.3, 4.4]
# Since guard_nn_modules is False, this will not recompile
with torch._dynamo.config.patch(error_on_recompile=True):
ref = mod(x)
res = opt_mod(x)
def test_optimized_module_training(self):
mod = torch.nn.Linear(3, 3)
mod.eval()
opt_mod = torch.compile(mod, backend="eager")
self.assertFalse(opt_mod.training)
opt_mod.train()
self.assertTrue(opt_mod.training)
self.assertTrue(mod.training)
mod.eval()
self.assertFalse(opt_mod.training)
@requires_cuda
def test_memleak_when_graph_input_has_tensor_attr(self):
@torch.compile(backend="eager")
def f(x):
x.add_(1)
mem_before = torch.cuda.memory_allocated()
x = torch.ones(2, device="cuda")
x.foo = torch.zeros(2, device="cuda")
f(x)
del x.foo
del x
mem_after = torch.cuda.memory_allocated()
self.assertEqual(mem_before, mem_after)
# check when non-tensor data structure attribute contains a tensor
@torch.compile(backend="eager")
def f(x):
x.add_(1)
mem_before = torch.cuda.memory_allocated()
x = torch.ones(2, device="cuda")
x.foo = [torch.zeros(2, device="cuda") for _ in range(5)]
f(x)
del x.foo
del x
mem_after = torch.cuda.memory_allocated()
self.assertEqual(mem_before, mem_after)
# check with tensor refcycle
@torch.compile(backend="eager")
def g(x, y):
return x + y
mem_before = torch.cuda.memory_allocated()
x = torch.ones(2, device="cuda")
y = torch.zeros(2, device="cuda")
x.foo = [y]
y.foo = [x]
g(x, y)
del x.foo
del y.foo
del x
del y
mem_after = torch.cuda.memory_allocated()
self.assertEqual(mem_before, mem_after)
def test_os_fspath(self):
@torch.compile(backend="eager", fullgraph=True)
def fn(x):
os.fspath(".")
return torch.sin(x)
fn(torch.randn(4))
@requires_cuda
# This test will fail as flip in combination with particular input lenghts
# produces weird results.
# This is under investigations in
# https://github.com/pytorch/pytorch/issues/131805
@unittest.skip("Skip this flip test for the moment. It is under investigation")
def test_flip_bad_accuracy(self):
import torch
import torch._dynamo.config
import torch._functorch.config
import torch._inductor.config
import torch._inductor.inductor_prims
import torch.fx.experimental._config
class Repro(torch.nn.Module):
def __init__(self):
super().__init__()
def forward(self, arg0_1):
rev = torch.ops.prims.rev.default(arg0_1, [0])
arg0_1 = None
slice_1 = torch.ops.aten.slice.Tensor(rev, 0, 0, -1, 2)
slice_2 = torch.ops.aten.slice.Tensor(rev, 0, 1, 9223372036854775807, 2)
add_1 = torch.ops.aten.add.Tensor(slice_1, slice_2)
slice_1 = slice_2 = None
slice_3 = torch.ops.aten.slice.Tensor(add_1, 0, 0, -1, 2)
slice_4 = torch.ops.aten.slice.Tensor(
add_1, 0, 1, 9223372036854775807, 2
)
add_2 = torch.ops.aten.add.Tensor(slice_3, slice_4)
slice_3 = slice_4 = None
slice_5 = torch.ops.aten.slice.Tensor(add_2, 0, 0, -1, 2)
slice_6 = torch.ops.aten.slice.Tensor(
add_2, 0, 1, 9223372036854775807, 2
)
add_3 = torch.ops.aten.add.Tensor(slice_5, slice_6)
slice_5 = slice_6 = None
slice_9 = torch.ops.aten.slice.Tensor(add_2, 0, 0, 1)
add_2 = None
unsqueeze = torch.ops.aten.unsqueeze.default(slice_9, 1)
slice_9 = None
unsqueeze_1 = torch.ops.aten.unsqueeze.default(add_3, 1)
add_3 = None
cat = torch.ops.aten.cat.default([unsqueeze, unsqueeze_1], 1)
unsqueeze = unsqueeze_1 = None
view = torch.ops.aten.view.default(cat, [2])
cat = None
slice_10 = torch.ops.aten.slice.Tensor(view, 0, 0, -1)
slice_11 = torch.ops.aten.slice.Tensor(
add_1, 0, 2, 9223372036854775807, 2
)
add_5 = torch.ops.aten.add.Tensor(slice_10, slice_11)
slice_10 = slice_11 = None
slice_12 = torch.ops.aten.slice.Tensor(add_1, 0, 0, 1)
add_1 = None
cat_1 = torch.ops.aten.cat.default([slice_12, add_5])
slice_12 = add_5 = None
unsqueeze_2 = torch.ops.aten.unsqueeze.default(cat_1, 1)
cat_1 = None
unsqueeze_3 = torch.ops.aten.unsqueeze.default(view, 1)
view = None
cat_2 = torch.ops.aten.cat.default([unsqueeze_2, unsqueeze_3], 1)
unsqueeze_2 = unsqueeze_3 = None
view_1 = torch.ops.aten.view.default(cat_2, [4])
cat_2 = None
slice_13 = torch.ops.aten.slice.Tensor(
rev, 0, 2, 9223372036854775807, 2
)
add_6 = torch.ops.aten.add.Tensor(view_1, slice_13)
slice_13 = None
slice_14 = torch.ops.aten.slice.Tensor(rev, 0, 0, 1)
rev = None
cat_3 = torch.ops.aten.cat.default([slice_14, add_6])
slice_14 = add_6 = None
constant_pad_nd = torch.ops.aten.constant_pad_nd.default(
view_1, [0, 1], 0.0
)
view_1 = None
unsqueeze_4 = torch.ops.aten.unsqueeze.default(cat_3, 1)
cat_3 = None
unsqueeze_5 = torch.ops.aten.unsqueeze.default(constant_pad_nd, 1)
constant_pad_nd = None
cat_4 = torch.ops.aten.cat.default([unsqueeze_4, unsqueeze_5], 1)
unsqueeze_4 = unsqueeze_5 = None
view_2 = torch.ops.aten.view.default(cat_4, [10])
cat_4 = None
slice_15 = torch.ops.aten.slice.Tensor(view_2, 0, 0, 9)
view_2 = None
rev_1 = torch.ops.prims.rev.default(slice_15, [0])
slice_15 = None
return (rev_1,)
mod = Repro()
x = torch.arange(9, device=torch.device("cuda"))
@torch.compile
def f(x):
return mod(x)
out = f(x)
self.assertEqual(torch.flip(torch.cumsum(torch.flip(x, [0]), 0), [0]), out[0])
# https://github.com/pytorch/pytorch/issues/88813
def test_return_value_duplication_tensor(self) -> None:
def fn(val: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
return val * 2, val * 2
x = torch.randn(2, requires_grad=True)
expect = fn(x)
self.assertNotEqual(
expect[0].untyped_storage().data_ptr(),
expect[1].untyped_storage().data_ptr(),
)
actual = torch.compile(fn, backend="aot_eager")(x)
self.assertNotEqual(
actual[0].untyped_storage().data_ptr(),
actual[1].untyped_storage().data_ptr(),
)
# https://github.com/pytorch/pytorch/issues/114344
def test_return_value_duplication_mixed_grad(self) -> None:
def fn(val: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
with torch.no_grad():
out0 = val + 1
out1 = val + 1
return out0, out1
x = torch.randn(2, requires_grad=True)
with torch.enable_grad():
expect = fn(x)
actual = torch.compile(fn, backend="aot_eager")(x)
self.assertEqual(expect[0].requires_grad, actual[0].requires_grad)
self.assertEqual(expect[1].requires_grad, actual[1].requires_grad)
# https://github.com/pytorch/pytorch/pull/134726#discussion_r1738774371
def test_return_value_duplication_scalar(self) -> None:
def fn(val: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
x, y = val * 2, val * 2
return x[0], y[0]
x = torch.randn(2, requires_grad=True)
expect = fn(x)
self.assertNotEqual(
expect[0].untyped_storage().data_ptr(),
expect[1].untyped_storage().data_ptr(),
)
actual = torch.compile(fn, backend="aot_eager")(x)
self.assertNotEqual(
actual[0].untyped_storage().data_ptr(),
actual[1].untyped_storage().data_ptr(),
)
def test_torch_compile_in_compile_frame(self):
def gn(x, c=None):
if c is None:
c = 2
return c * x
def outer_func(x):
return torch.compile(gn, backend="eager")(x)
compile_outer = torch.compile(outer_func, backend="eager", fullgraph=True)
x = torch.randn(4)
ref = outer_func(x)
res = compile_outer(x)
self.assertEqual(ref, res)
# https://github.com/pytorch/pytorch/issues/136640
def test_inductor_dynamic_shapes_broadcasting(self) -> None:
def fn(x, y):
x_view = x.view(-1, 4)
y_view = y.view(-1, 4)
return x_view * y_view
x = torch.randn(4)
y = torch.randn(8)
out_ref = fn(x, y)
out_test = torch.compile(fn, dynamic=True)(x, y)
self.assertEqual(out_ref, out_test)
# https://github.com/pytorch/pytorch/issues/119162
def test_inductor_rng_default_dtype(self) -> None:
@torch.compile
def fn():
tmp = torch.randn(4, 4, dtype=torch.bfloat16)
return tmp
try:
old = torch.get_default_dtype()
torch.set_default_dtype(torch.bfloat16)
out = fn()
finally:
torch.set_default_dtype(old)
# output dtype should be float32
self.assertEqual(out.dtype, torch.bfloat16)
@unittest.skipIf(not HAS_MSGSPEC, "missing msgspec package")
def test_c_defined_metaclass(self):
class User(msgspec.Struct):
"""A new type describing a User"""
name: str
value: int
def fn(x):
u = User("alice", 10)
return x * u.value
x = torch.randn(4)
opt_fn = torch.compile(fn, backend="eager")
self.assertEqual(fn(x), opt_fn(x))
@unittest.skipIf(not HAS_OMEGACONG, "missing omegaconf package")
def test_omegaconf_dictconfig(self):
def fn(cfg, x):
a = cfg["foo"].a * x
b = cfg.bar["b"] * a
cfg.__dict__["baz"] = 4
return b * cfg.baz
config = OmegaConf.create({"foo": {"a": 3}, "bar": {"b": 5}})
x = torch.randn(4)
opt_fn = torch.compile(fn, backend="eager", fullgraph=True)
ref = fn(config, x)
cloned_config = copy.deepcopy(config)
res = opt_fn(cloned_config, x)
self.assertEqual(fn(config, x), opt_fn(config, x))
self.assertEqual(cloned_config.baz, 4)
# https://github.com/pytorch/pytorch/issues/136257
def test_overwriting_params(self):
class M(torch.nn.Module):
def __init__(self):
super().__init__()
self.fc1 = torch.nn.Linear(2, 2)
self.fc2 = torch.nn.Linear(2, 2)
def forward(self, x):
x = self.fc1(x)
x = self.fc2(x)
return x
class ZeROOrderedDict(collections.OrderedDict):
def __init__(self, parent_module=None, *args, **kwargs):
"""A replacement for ``collections.OrderedDict`` to detect external ZeRO params.
Args:
parent_module (``collections.OrderedDict``): the collection to replace
"""
super().__init__(*args, **kwargs)
self._parent_module = parent_module
def __getitem__(self, key):
param = super().__getitem__(key)
# Params can be registered as None (e.g., bias)
if param is None:
return param
# do something here
return param
def inject_parameters(module, cls):
for module in module.modules(): # noqa: B020
if cls == ZeROOrderedDict:
new_param = cls(parent_module=module)
else:
new_param = cls()
for key, param in module._parameters.items():
new_param[key] = param
module._parameters = new_param
model = M()
inject_parameters(model, ZeROOrderedDict)
model = torch.compile(model, backend="eager", fullgraph=True)
x = torch.ones(2)
with torch.no_grad():
y = model(x)
def test_typed_dict(self):
class LlavaImagePixelInputs(TypedDict):
type: Literal["pixel_values"]
data: torch.Tensor
"""Shape: `(batch_size, num_channels, height, width)`"""
def fn(x, y):
obj = LlavaImagePixelInputs(type=int, data=y)
out = x * obj["data"]
obj["data"] = 3
return out * obj["data"]
x, y = torch.randn(4), torch.randn(4)
ref = fn(x, y)
opt_fn = torch.compile(fn, backend="eager", fullgraph=True)
res = opt_fn(x, y)
self.assertEqual(ref, res)
def test_typed_dict_total(self):
class LlavaImagePixelInputs(TypedDict):
type: Literal["pixel_values"]
data: torch.Tensor
"""Shape: `(batch_size, num_channels, height, width)`"""
def fn(x, y):
obj = LlavaImagePixelInputs(data=y, total=False)
return x * obj["data"]
x, y = torch.randn(4), torch.randn(4)
ref = fn(x, y)
opt_fn = torch.compile(fn, backend="eager", fullgraph=True)
res = opt_fn(x, y)
self.assertEqual(ref, res)
@skipIfPy312 # listcomp bytecode is optimized
def test_listcomp(self):
class Module(torch.nn.Module):
def __init__(self):
super().__init__()
self._num = 4
@torch._dynamo.disable(recursive=False)
def forward(self, x):
values = [i * torch.cos(x) for i in range(self._num)]
return sum(values)
mod = Module()
def fn(x):
return mod(x)
cnt = torch._dynamo.testing.CompileCounter()
opt_fn = torch.compile(fn, backend=cnt)
x = torch.randn(4)
ref = fn(x)
res = opt_fn(x)
self.assertEqual(ref, res)
self.assertEqual(cnt.frame_count, 1)
# Ensure that the listcomp is fully compiled
self.assertEqual(cnt.op_count, 8)
# https://github.com/pytorch/pytorch/issues/140266
def test_distributions_subclass(self):
import torch
from torch.distributions import Categorical
class SubCateg(Categorical):
...
@torch.compile(backend="eager", fullgraph=True)
def make_dist_and_execute(t, d):
categ = d(logits=t)
a = categ.log_prob(categ.sample()) + categ.probs + categ.logits
return a
for _ in range(2):
make_dist_and_execute(torch.randn(10), SubCateg)
def test_tensor_split_within_device_cm(self):
@torch.compile(fullgraph=True)
def split(x):
return x.split(4, 0)
x = torch.zeros(12)
res = split(x)
with torch.device("cpu"):
self.assertEqual(res, split(x))
def test_method_overriding(self):
class DilateConv(torch.nn.Module):
def __init__(
self,
dilate_func=None,
):
super().__init__()
self.dilate_func = dilate_func
def forward(self, x):
return self.dilate_func() * torch.sin(x)
class MainModule(torch.nn.Module):
def __init__(self):
super().__init__()
self.mod = DilateConv(self.dilate_func)
self.a = 4
def dilate_func(self):
return self.a
def forward(self, x):
return self.mod(x)
mod = MainModule()
opt_mod = torch.compile(mod, backend="eager", fullgraph=True)
x = torch.randn(4)
ref = mod(x)
res = opt_mod(x)
self.assertEqual(ref, res)
instantiate_parametrized_tests(ReproTests)
if __name__ == "__main__":
from torch._dynamo.test_case import run_tests
run_tests()
Reference in New Issue View Git Blame Copy Permalink