mirror of
https://github.com/pytorch/pytorch.git
synced 2025-10-20 21:14:14 +08:00
This PR implements tracing of with contexts with TorchFunction modes which have the default enter/exit behavior (ie pushing/popping the mode) Typically the bytecode for a context manager looks like this during a graph break: 1. graph call 2. enter context 3. unsupported code 4. exit context 5. resume call resume fn structure: 1. enter context 2. jump ... 3. exit context The issue with torch function modes is that side effects will replay any mutations to the torch function stack performed during tracing. So, we do not need to enter and exit around the unsupported code in the original function (doing so would result in a duplicate torch function mode entry during execution of the unsupported code), and we don't need to enter again in the resume function (the mode that was pushed from the side effects bytecode would still be on the stack). So for torch function modes the structure of our output code is this: 1. graph call 2. mutate tf mode stack to replay mutations 4. unsupported code 5. on exception restore stack 6. resume function Then our resume fn looks like this: 1. no-op enter torch function mode 2. jump 3. exit tf mode To implement the no-op enter of the torch function mode I added torch function mode in polyfill which no-op enters, but normally exits. This is needed because we still want to trace the with context in the resume function, and exit properly (the exit instructions will still be in the function, so we need to generate instructions to set up the context). Separately from the bytecode, dynamo also tracks contexts on the block stack, which is how the SETUP_* instructions are implemented. Naturally at a graph break, we exit these block stacks to properly reset the contexts entirely, so that we can re-enter around the unsupported code soundly. However once again, in the torch function mode case, in the event of a graph we do not want to perform any exit side effects because we want to preserve the state of the mode stack as is so that we will properly update the stack with bytecode mentioned in the first section. If we exited here, dynamo would pop the mode off of the symbolic stack, and not update the true python torch function mode stack with the suffix bytecode. All in all, for torch function modes we enter exactly once, update the global torch function mode stack with side effects bytecode, re-read this stack when compiling the resume function, and exit exactly once in the resume function. This matches the semantics of eager exactly. Pull Request resolved: https://github.com/pytorch/pytorch/pull/135422 Approved by: https://github.com/williamwen42 ghstack dependencies: #134732, #133137, #135443, #135444
411 lines
11 KiB
Python
411 lines
11 KiB
Python
# mypy: allow-untyped-defs
|
|
import contextlib
|
|
import dis
|
|
import functools
|
|
import logging
|
|
import os.path
|
|
import random
|
|
import re
|
|
import sys
|
|
import types
|
|
import unittest
|
|
from typing import List, Optional, Sequence, Union
|
|
from unittest.mock import patch
|
|
|
|
import torch
|
|
from torch import fx
|
|
from torch._dynamo.output_graph import OutputGraph
|
|
|
|
from . import config, eval_frame, optimize_assert, reset
|
|
from .bytecode_transformation import (
|
|
create_instruction,
|
|
debug_checks,
|
|
is_generator,
|
|
transform_code_object,
|
|
)
|
|
from .guards import CheckFunctionManager, CompileId, GuardedCode
|
|
from .utils import same
|
|
|
|
|
|
np: Optional[types.ModuleType] = None
|
|
try:
|
|
import numpy as np
|
|
except ModuleNotFoundError:
|
|
np = None
|
|
|
|
|
|
unsupported = eval_frame.unsupported
|
|
three = 3
|
|
|
|
log = logging.getLogger(__name__)
|
|
|
|
|
|
def clone_me(x):
|
|
if x is None:
|
|
return None
|
|
return x.detach().clone().requires_grad_(x.requires_grad)
|
|
|
|
|
|
def remove_optimized_module_prefix(name) -> str:
|
|
return re.sub(r"^_orig_mod[.]", "", name)
|
|
|
|
|
|
def collect_results(model, prediction, loss, example_inputs):
|
|
results = []
|
|
results.append(prediction)
|
|
results.append(loss)
|
|
# if isinstance(loss, torch.Tensor) and loss.item() > 1:
|
|
# log.warning(
|
|
# f"High loss value alert - {loss:.2f}. Can result in unstable gradients."
|
|
# )
|
|
|
|
grads = {}
|
|
params = {}
|
|
for name, param in model.named_parameters():
|
|
if isinstance(model, eval_frame.OptimizedModule):
|
|
name = remove_optimized_module_prefix(name)
|
|
param_copy = param
|
|
grad = param.grad
|
|
# Treat None and zero grad as same
|
|
if param.grad is None:
|
|
grad = torch.zeros_like(param)
|
|
grads[name + ".grad"] = grad
|
|
params[name] = param_copy
|
|
results.append(grads)
|
|
results.append(params)
|
|
buffers = {}
|
|
for name, buffer in model.named_buffers():
|
|
if isinstance(model, eval_frame.OptimizedModule):
|
|
name = remove_optimized_module_prefix(name)
|
|
buffers[name] = buffer
|
|
results.append(buffers)
|
|
for example in example_inputs:
|
|
if isinstance(example, (tuple, list)):
|
|
for inp in example:
|
|
if isinstance(inp, torch.Tensor):
|
|
results.append(inp.grad)
|
|
else:
|
|
if isinstance(example, torch.Tensor):
|
|
results.append(example.grad)
|
|
return results
|
|
|
|
|
|
def requires_bwd_pass(out):
|
|
if isinstance(out, torch.Tensor):
|
|
return out.requires_grad
|
|
elif isinstance(out, (list, tuple)):
|
|
return any(requires_bwd_pass(x) for x in out)
|
|
elif out is None:
|
|
return False
|
|
elif isinstance(out, int):
|
|
return False
|
|
raise NotImplementedError("Don't know how to reduce", type(out))
|
|
|
|
|
|
def reduce_to_scalar_loss(out):
|
|
"""Reduce the output of a model to get scalar loss"""
|
|
if isinstance(out, torch.Tensor):
|
|
# Mean does not work on integer tensors
|
|
return out.sum() / out.numel()
|
|
elif isinstance(out, (list, tuple)):
|
|
return sum(reduce_to_scalar_loss(x) for x in out) / len(out)
|
|
elif type(out).__name__ in (
|
|
"MaskedLMOutput",
|
|
"Seq2SeqLMOutput",
|
|
"CausalLMOutputWithCrossAttentions",
|
|
):
|
|
return reduce_to_scalar_loss(out.logits)
|
|
elif type(out).__name__ == "SquashedNormal":
|
|
return out.mean.sum()
|
|
elif isinstance(out, dict):
|
|
return sum(reduce_to_scalar_loss(value) for value in out.values()) / len(
|
|
out.keys()
|
|
)
|
|
raise NotImplementedError("Don't know how to reduce", type(out))
|
|
|
|
|
|
def debug_dir() -> str:
|
|
path = os.path.join(os.path.dirname(__file__), "../debug")
|
|
if not os.path.exists(path):
|
|
os.mkdir(path)
|
|
return path
|
|
|
|
|
|
def debug_dump(name, code: types.CodeType, extra="") -> None:
|
|
with open(os.path.join(debug_dir(), name), "w") as fd:
|
|
fd.write(
|
|
f"{dis.Bytecode(code).info()}\n\n{dis.Bytecode(code).dis()}\n\n{extra}\n"
|
|
)
|
|
|
|
|
|
def debug_insert_nops(
|
|
frame, cache_size, hooks, _, *, skip: int = 0
|
|
) -> Optional[GuardedCode]:
|
|
"""used to debug jump updates"""
|
|
|
|
def insert_nops(instructions, code_options):
|
|
instructions.insert(0, create_instruction("NOP"))
|
|
instructions.insert(0, create_instruction("NOP"))
|
|
|
|
if is_generator(frame.f_code):
|
|
return None
|
|
|
|
debug_checks(frame.f_code)
|
|
code = transform_code_object(frame.f_code, insert_nops)
|
|
graph = OutputGraph(
|
|
code_options={},
|
|
compiler_fn=None,
|
|
root_tx=None,
|
|
export=False,
|
|
export_constraints=None,
|
|
frame_state={"_id": 0},
|
|
# TODO: shouldn't this be f_locals/f_globals from frame?
|
|
local_scope=locals(),
|
|
global_scope=globals(),
|
|
f_code=frame.f_code,
|
|
torch_function_mode_stack=[],
|
|
)
|
|
|
|
return GuardedCode(code, CheckFunctionManager(graph).check_fn, CompileId(0, 0))
|
|
|
|
|
|
class CompileCounter:
|
|
def __init__(self):
|
|
self.frame_count = 0
|
|
self.op_count = 0
|
|
|
|
def __call__(self, gm: torch.fx.GraphModule, example_inputs: List[torch.Tensor]):
|
|
self.frame_count += 1
|
|
for node in gm.graph.nodes:
|
|
if "call" in node.op:
|
|
self.op_count += 1
|
|
return gm.forward
|
|
|
|
def clear(self):
|
|
self.frame_count = 0
|
|
self.op_count = 0
|
|
|
|
|
|
class CompileCounterWithBackend:
|
|
def __init__(self, backend):
|
|
self.frame_count = 0
|
|
self.op_count = 0
|
|
self.backend = backend
|
|
self.graphs = []
|
|
|
|
def __call__(self, gm: torch.fx.GraphModule, example_inputs: List[torch.Tensor]):
|
|
from .backends.registry import lookup_backend
|
|
|
|
self.frame_count += 1
|
|
for node in gm.graph.nodes:
|
|
if "call" in node.op:
|
|
self.op_count += 1
|
|
self.graphs.append(gm)
|
|
return lookup_backend(self.backend)(gm, example_inputs)
|
|
|
|
|
|
# Equivalent to backend="eager", but also records graphs that
|
|
# we can assert on
|
|
class EagerAndRecordGraphs:
|
|
def __init__(self):
|
|
self.graphs = []
|
|
|
|
def __call__(self, gm: torch.fx.GraphModule, example_inputs: List[torch.Tensor]):
|
|
self.graphs.append(gm)
|
|
return gm.forward
|
|
|
|
|
|
def strip_comment(code) -> str:
|
|
code = str(code)
|
|
return re.sub(r"(?m)^ *#.*\n?", "", code)
|
|
|
|
|
|
def remove_trailing_space(code) -> str:
|
|
return "\n".join([line.rstrip() for line in code.split("\n")])
|
|
|
|
|
|
def normalize_gm(gm_str) -> str:
|
|
# strip comments as comments have path to files which may differ from
|
|
# system to system.
|
|
return remove_trailing_space(strip_comment(gm_str))
|
|
|
|
|
|
def empty_line_normalizer(code: str) -> str:
|
|
"""
|
|
Normalize code: remove empty lines.
|
|
"""
|
|
normal_code = re.sub(r"[\r\n]+", "\n", code)
|
|
return normal_code
|
|
|
|
|
|
def standard_test(
|
|
self,
|
|
fn,
|
|
nargs,
|
|
expected_ops=None,
|
|
expected_ops_dynamic=None,
|
|
expected_frame_count=1,
|
|
):
|
|
if not config.assume_static_by_default and expected_ops_dynamic is not None:
|
|
expected_ops = expected_ops_dynamic
|
|
|
|
actual = CompileCounter()
|
|
|
|
args1 = [torch.randn(10, 10) for _ in range(nargs)]
|
|
args2 = [torch.randn(10, 10) for _ in range(nargs)]
|
|
correct1 = fn(*args1)
|
|
correct2 = fn(*args2)
|
|
reset()
|
|
opt_fn = optimize_assert(actual)(fn)
|
|
val1a = opt_fn(*args1)
|
|
val2a = opt_fn(*args2)
|
|
val1b = opt_fn(*args1)
|
|
val2b = opt_fn(*args2)
|
|
reset()
|
|
self.assertTrue(same(val1a, correct1))
|
|
self.assertTrue(same(val1b, correct1))
|
|
self.assertTrue(same(val2a, correct2))
|
|
self.assertTrue(same(val2b, correct2))
|
|
self.assertEqual(actual.frame_count, expected_frame_count)
|
|
if expected_ops is not None:
|
|
self.assertEqual(actual.op_count, expected_ops)
|
|
|
|
|
|
def dummy_fx_compile(gm: fx.GraphModule, example_inputs):
|
|
return gm.forward
|
|
|
|
|
|
def format_speedup(speedup, pvalue, is_correct=True, pvalue_threshold=0.1):
|
|
if not is_correct:
|
|
return "ERROR"
|
|
if pvalue > pvalue_threshold:
|
|
return f"{speedup:.3f}x SAME"
|
|
return f"{speedup:.3f}x p={pvalue:.2f}"
|
|
|
|
|
|
def rand_strided(
|
|
size: Sequence[int],
|
|
stride: Sequence[int],
|
|
dtype: torch.dtype = torch.float32,
|
|
device: Union[str, torch.device] = "cpu",
|
|
extra_size: int = 0,
|
|
):
|
|
needed_size = (
|
|
sum((shape - 1) * stride for shape, stride in zip(size, stride))
|
|
+ 1
|
|
+ extra_size
|
|
)
|
|
if dtype.is_floating_point:
|
|
if dtype.itemsize == 1:
|
|
"""
|
|
normal distribution kernel is not implemented for fp8..
|
|
Workaround that by creating a fp16 tensor and then cast.
|
|
"""
|
|
buffer = torch.randn(needed_size, dtype=torch.float16, device=device).to(
|
|
dtype=dtype
|
|
)
|
|
else:
|
|
buffer = torch.randn(needed_size, dtype=dtype, device=device)
|
|
else:
|
|
buffer = torch.zeros(size=[needed_size], dtype=dtype, device=device)
|
|
return torch.as_strided(buffer, size, stride)
|
|
|
|
|
|
def _make_fn_with_patches(fn, *patches):
|
|
@functools.wraps(fn)
|
|
def _fn(*args, **kwargs):
|
|
with contextlib.ExitStack() as stack:
|
|
for module, attr, val in patches:
|
|
stack.enter_context(patch.object(module, attr, val))
|
|
|
|
return fn(*args, **kwargs)
|
|
|
|
return _fn
|
|
|
|
|
|
def make_test_cls_with_patches(
|
|
cls, cls_prefix, fn_suffix, *patches, xfail_prop=None, decorator=lambda x: x
|
|
):
|
|
DummyTestClass = type(f"{cls_prefix}{cls.__name__}", cls.__bases__, {})
|
|
DummyTestClass.__qualname__ = DummyTestClass.__name__
|
|
|
|
for name in dir(cls):
|
|
if name.startswith("test_"):
|
|
fn = getattr(cls, name)
|
|
if not callable(fn):
|
|
setattr(DummyTestClass, name, getattr(cls, name))
|
|
continue
|
|
new_name = f"{name}{fn_suffix}"
|
|
new_fn = _make_fn_with_patches(fn, *patches)
|
|
new_fn.__name__ = new_name
|
|
if xfail_prop is not None and hasattr(fn, xfail_prop):
|
|
new_fn = unittest.expectedFailure(new_fn)
|
|
setattr(DummyTestClass, new_name, decorator(new_fn))
|
|
# NB: Doesn't handle slots correctly, but whatever
|
|
elif not hasattr(DummyTestClass, name):
|
|
setattr(DummyTestClass, name, getattr(cls, name))
|
|
|
|
return DummyTestClass
|
|
|
|
|
|
# test Python 3.11+ specific features
|
|
def skipIfNotPy311(fn):
|
|
if sys.version_info >= (3, 11):
|
|
return fn
|
|
return unittest.skip(fn)
|
|
|
|
|
|
def skipIfNotPy312(fn):
|
|
if sys.version_info >= (3, 12):
|
|
return fn
|
|
return unittest.skip(fn)
|
|
|
|
|
|
def xfailIfPy312(fn):
|
|
if sys.version_info >= (3, 12):
|
|
return unittest.expectedFailure(fn)
|
|
return fn
|
|
|
|
|
|
def skipIfPy312(fn):
|
|
if sys.version_info >= (3, 12):
|
|
return unittest.skip(fn)
|
|
return fn
|
|
|
|
|
|
def requiresPy310(fn):
|
|
if sys.version_info >= (3, 10):
|
|
return fn
|
|
else:
|
|
unittest.skip(fn)
|
|
|
|
|
|
# Controls tests generated in test/inductor/test_torchinductor_dynamic_shapes.py
|
|
# and test/dynamo/test_dynamic_shapes.py
|
|
def expectedFailureDynamic(fn):
|
|
fn._expected_failure_dynamic = True
|
|
return fn
|
|
|
|
|
|
# Controls tests generated in test/inductor/test_torchinductor_codegen_dynamic_shapes.py
|
|
def expectedFailureCodegenDynamic(fn):
|
|
fn._expected_failure_codegen_dynamic = True
|
|
return fn
|
|
|
|
|
|
# Controls test generated in test/inductor/test_cpp_wrapper.py
|
|
def expectedFailureDynamicWrapper(fn):
|
|
fn._expected_failure_dynamic_wrapper = True
|
|
return fn
|
|
|
|
|
|
def reset_rng_state(use_xla=False):
|
|
torch.manual_seed(1337)
|
|
random.seed(1337)
|
|
if np:
|
|
np.random.seed(1337)
|
|
if use_xla:
|
|
import torch_xla.core.xla_model as xm
|
|
|
|
xm.set_rng_state(1337, str(xm.xla_device()))
|