# mypy: allow-untyped-defs """ Utility functions and classes used throughout the TorchDynamo system. This module contains a collection of helper utilities used by various parts of Dynamo for: - Performance metrics collection and reporting - Compilation timing and debugging - Graph manipulation and tensor operations - Runtime guards and checks - Common data structure operations - Testing and development tools This is an internal module that provides shared functionality used across the Dynamo codebase. """ from __future__ import annotations import atexit import collections import contextlib import copy import dataclasses import datetime import dis import enum import functools import gc import importlib import inspect import itertools import json import linecache import logging import math import operator import os import re import sys import textwrap import threading import time import traceback import types import typing import uuid import warnings import weakref from collections import Counter, OrderedDict from contextlib import contextmanager from dataclasses import is_dataclass from functools import lru_cache from types import MethodWrapperType from typing import ( Any, Callable, cast, ClassVar, Generic, Optional, overload, TypeVar, Union, ) from typing_extensions import Literal, TypeIs import torch import torch._functorch.config import torch.fx.experimental.symbolic_shapes import torch.utils._pytree as pytree from torch import fx from torch._C import ( _instruction_counter, _len_torch_function_stack, _pop_torch_function_stack, _push_on_torch_function_stack, ) from torch._dispatch.python import enable_python_dispatcher from torch._dynamo.metrics_context import MetricsContext, RuntimeMetricsContext from torch._guards import CompileId, Source, TracingContext from torch._subclasses.meta_utils import is_sparse_compressed from torch._utils_internal import ( justknobs_check, log_chromium_event_internal, log_compilation_event, record_chromium_event_internal, signpost_event, ) from torch.fx._utils import _format_graph_code, lazy_format_graph_code from torch.monitor import _WaitCounter from torch.nn.modules.lazy import LazyModuleMixin from torch.utils._triton import has_triton, has_triton_package from torch.utils.hooks import RemovableHandle if typing.TYPE_CHECKING: from collections.abc import Generator, Iterable, Iterator, KeysView, ValuesView try: import numpy as np except ModuleNotFoundError: np = None # type: ignore[assignment] try: import torch._logging import torch._numpy as tnp from torch._guards import detect_fake_mode # noqa: F401n from torch._logging import LazyString from . import config # NOTE: Make sure `NP_SUPPORTED_MODULES` and `NP_TO_TNP_MODULE` are in sync. if np: NP_SUPPORTED_MODULES: tuple[types.ModuleType, ...] = ( np, np.fft, np.linalg, np.random, ) NP_TO_TNP_MODULE = { np: tnp, np.fft: tnp.fft, np.linalg: tnp.linalg, np.random: tnp.random, } else: NP_SUPPORTED_MODULES = () NP_TO_TNP_MODULE = {} from torch._subclasses.fake_tensor import FakeTensor, is_fake, maybe_get_fake_mode except ImportError: pass T = TypeVar("T") unpatched_nn_module_getattr = torch.nn.Module.__getattr__ unpatched_nn_module_call = torch.nn.Module.__call__ unpatched_nn_module_call_impl = torch.nn.Module._call_impl counters: collections.defaultdict[str, Counter[str]] = collections.defaultdict( collections.Counter ) optimus_scuba_log: dict[str, Any] = {} troubleshooting_url = ( "https://pytorch.org/docs/main/torch.compiler_troubleshooting.html" ) nnmodule_doc_url = "https://pytorch.org/docs/main/torch.compiler_nn_module.html" nnmodule_doc_url_msg = f"See {nnmodule_doc_url} for more information and limitations." log = logging.getLogger(__name__) # profiling compilation time by function compilation_time_metrics: dict[str, list[float]] = {} # This supports calculate_time_spent(), which reports cumulative times # across the process for any "phase" populated by dynamo_timed. Reset if # reset_frame_count() is called. cumulative_time_spent_ns: dict[str, float] = collections.defaultdict(float) timer_counter = itertools.count() # Abstraction on top of counters. class ReInplaceTrigger(enum.Enum): AUTO_FUNC_V1 = 1 AUTO_FUNC_V2 = 2 TRITON_OPS = 3 class ReinplaceCounters: _values: collections.defaultdict[str, int] = collections.defaultdict(int) # Track sizes of known not re-inplaced tensors (exclude dynamic shapes). @classmethod def add_missed_bytes(cls, trigger: ReInplaceTrigger, bytes: int): if bytes != 0: cls._values[f"missed_bytes_{trigger.name}"] += bytes # Track number of not re-inplaced tensors. @classmethod def add_missed_opportunities(cls, trigger: ReInplaceTrigger, count: int): if count != 0: cls._values[f"missed_tensors_{trigger}"] += count @classmethod def clear(cls): cls._values.clear() @classmethod def get_total_missed(cls): sum = 0 for trigger in ReInplaceTrigger: sum += cls._values.get(f"missed_tensors_{trigger}", 0) return sum @classmethod def get_total_missed_bytes(cls): sum = 0 for trigger in ReInplaceTrigger: sum += cls._values.get(f"missed_bytes_{trigger.name}", 0) return sum @classmethod def log(cls): # if not empty log. if cls._values: signpost_event("inductor", "reinplace_counters", cls._values) def tabulate( rows: Union[list[tuple[str, object]], list[list[object]]], headers: Union[tuple[str, ...], list[str]], ) -> str: try: import tabulate return tabulate.tabulate(rows, headers=headers) except ImportError: return "\n".join( ", ".join(map(str, row)) for row in itertools.chain([headers], rows) ) curr_frame = 0 # Note: Called for you by dynamo - you almost never ever want to invoke this yourself. def increment_frame() -> None: global curr_frame curr_frame = curr_frame + 1 # Note: Called for you by dynamo - you almost never ever want to invoke this yourself. def reset_frame_count() -> None: global curr_frame cumulative_time_spent_ns.clear() compilation_time_metrics.clear() curr_frame = 0 op_count = 0 def increment_op_count(cnt: int) -> None: global op_count op_count += cnt # Get the total time in seconds for each "phase" # For example, {'entire_frame_compile':8.574629999999999, 'backend_compile':5.26806} def calculate_time_spent() -> dict[str, float]: total_by_key = {} for phase, timing in cumulative_time_spent_ns.items(): total_by_key[phase] = timing / 1e9 total_by_key["total_wall_time"] = total_by_key.get( "entire_frame_compile", 0 ) + total_by_key.get("entire_backward_compile", 0) return total_by_key # Print a report of time spent so far # Ex: # TIMING: # entire_frame_compile:8.574629999999999 # backend_compile:5.26806 def print_time_report() -> None: total_by_key = calculate_time_spent() out = "TIMING:" for key, value in total_by_key.items(): out = f"{out} {key}:{round(value, 5)}" print(out) # Use the following singleton to capture and log CompilationMetrics. Entering the context # manager allocates a new record to be logged when it exits. (You should not need to use # this directly unless you introduce a new code path where compilation metrics would be # gathered). While compiling, use the setters or timer in MetricsContext to update fields # in the current context. For example: # # To set a single field once (use overwrite=True to overwrite): # get_metrics_context().set("metric_name", value) # # To set multiple fields at once (use overwrite=True to overwrite): # get_metrics_context().update({"name1": val1, "name2": val2}) # # To increment an integer field: # get_metrics_context().increment("metric_name", value) # # To record execution time, MetricsContext works with dynamo_timed: # def foo(...): # # Updates the "metric_us" field. # with dynamo_timed("metric", dynamo_compile_column_us="metric_us") # ... # _METRICS_CONTEXT: MetricsContext _RUNTIME_METRICS_CONTEXT: RuntimeMetricsContext def get_metrics_context() -> MetricsContext: return _METRICS_CONTEXT def get_runtime_metrics_context() -> RuntimeMetricsContext: return _RUNTIME_METRICS_CONTEXT class CompileEventLogLevel(enum.Enum): """ Enum that loosely corresponds with a "log level" of a given event. CHROMIUM_EVENT: Logs only to tlparse. COMPILE_EVENT: Logs to tlparse + PT2 Compile Events COMPILATION_METRIC: Logs to tlparse, PT2 Compile Events, and dynamo_compile """ CHROMIUM = 1 PT2_COMPILE = 2 COMPILATION_METRIC = 3 class CompileEventLogger: """ Helper class for representing adding metadata(i.e. columns) to various compile events. Use CompileEventLogger to add event data to: - Chromium events - PT2 Compile Events - CompilationMetrics This should be used in conjunction with dynamo_timed() and metrics contexts, which create timed spans and events. CompileEventLogger uses three log levels (described in CompileEventLogLevel), where each log level logs to all sources below it in the hierarchy. Example usages: - I want to log to an existing chromium event within dynamo timed: with dynamo_timed("my_event"): CompileEventLogger.chromium("my_event", foo=bar) - I want to log my event to both chromium + pt2_compile_events: with dynamo_timed("my_event", log_pt2_compile_event=True): CompileEventLogger.pt2_compile("my_event", foo=bar) - I want to add information to dynamo events and dynamo_compile CompileEventLogger.compilation_metric(foo=bar) """ @staticmethod def log_instant_event( event_name: str, metadata: dict[str, Any], time_ns: Optional[int] = None, log_level: CompileEventLogLevel = CompileEventLogLevel.CHROMIUM, ): if time_ns is None: time_ns = time.time_ns() chromium_log = get_chromium_event_logger() if log_level == CompileEventLogLevel.CHROMIUM: log_pt2_compile_event = False elif log_level == CompileEventLogLevel.PT2_COMPILE: log_pt2_compile_event = True else: raise RuntimeError( "Cannot log instant event at COMPILATION_METRIC level. Please choose one of CHROMIUM_EVENT or COMPILE_EVENT" ) chromium_log.log_instant_event( event_name, time_ns, metadata, log_pt2_compile_event ) @staticmethod def add_data( event_name: str, log_level: CompileEventLogLevel, overwrite: bool = False, **metadata: object, ): """ Centralized API for adding data to various events Log an event to a toplevel "dynamo" event or metrics context depending on log level. """ chromium_log = get_chromium_event_logger() pt2_compile_substack = chromium_log.get_pt2_compile_substack() if log_level == CompileEventLogLevel.CHROMIUM: chromium_log.add_event_data(event_name, **metadata) elif log_level == CompileEventLogLevel.PT2_COMPILE: pt2_compile_substack = chromium_log.get_pt2_compile_substack() if event_name not in pt2_compile_substack: raise RuntimeError( "Error: specified log level PT2_COMPILE, but the event %s" " is not logged to pt2_compile_events. Make sure the event is active and you passed " "log_pt2_compile_event=True to dynamo_timed", event_name, ) chromium_log.add_event_data(event_name, **metadata) else: assert log_level == CompileEventLogLevel.COMPILATION_METRIC top_event = chromium_log.get_outermost_event() if event_name != top_event: raise RuntimeError( "Log level is COMPILATION_METRIC, but event_name isn't the toplevel event. " "CompilationMetrics must be logged to the toplevel event. Consider using `log_toplevel_event_data` directly." ) metrics_context = get_metrics_context() if not metrics_context.in_progress(): raise RuntimeError( "No metrics context is in progress. Please only call this function within a metrics context." ) # TODO: should we assert that the keys of metadata are in CompilationMetrics? metrics_context.update(metadata, overwrite) chromium_log.add_event_data(event_name, **metadata) @staticmethod def add_toplevel( log_level: CompileEventLogLevel, overwrite: bool = False, **metadata: object ): """ Syntactic sugar for logging to the toplevel event """ top_event = get_chromium_event_logger().get_outermost_event() if top_event is None: raise RuntimeError( "No toplevel event active. Please only call this function within a dynamo_timed context." ) CompileEventLogger.add_data(top_event, log_level, overwrite, **metadata) @staticmethod def increment( event_name: str, log_level: CompileEventLogLevel, key: str, value: int ): """ Increments an existing field, or adds it """ chromium_log = get_chromium_event_logger() if ( log_level == CompileEventLogLevel.CHROMIUM or log_level == CompileEventLogLevel.PT2_COMPILE ): chromium_log.increment(event_name, key, value) else: assert log_level == CompileEventLogLevel.COMPILATION_METRIC top_event = chromium_log.get_outermost_event() if event_name != top_event: raise RuntimeError( "Log level is COMPILATION_METRIC, but event_name isn't the toplevel event. " "CompilationMetrics must be logged to the toplevel event. Consider using `increment_toplevel` directly." ) metrics_context = get_metrics_context() if not metrics_context.in_progress(): raise RuntimeError( "No metrics context is in progress. Please only call this function within a metrics context/dynamo_timed." ) metrics_context.increment(key, value) chromium_log.increment(event_name, key, value) @staticmethod def increment_toplevel( key: str, value: int = 1, log_level: CompileEventLogLevel = CompileEventLogLevel.COMPILATION_METRIC, ): """ Increments a value on the toplevel metric. By default, logs to metric. """ chromium_log = get_chromium_event_logger() top_event = chromium_log.get_outermost_event() if top_event is None: raise RuntimeError( "No toplevel event active. Please only call this function within a metrics context/dynamo_timed." ) CompileEventLogger.increment(top_event, log_level, key, value) @staticmethod def add_to_set( event_name: str, log_level: CompileEventLogLevel, key: str, value: Any ): """ Add metadata to a set of values with key . Creates a set if it doesn't exist. """ chromium_log = get_chromium_event_logger() if ( log_level == CompileEventLogLevel.CHROMIUM or log_level == CompileEventLogLevel.PT2_COMPILE ): chromium_log.add_to_set(event_name, key, value) else: assert log_level == CompileEventLogLevel.COMPILATION_METRIC top_event = chromium_log.get_outermost_event() if event_name != top_event: raise RuntimeError( "Log level is COMPILATION_METRIC, but event_name isn't the toplevel event. " "CompilationMetrics must be logged to the toplevel event. Consider using `add_to_set_metric` directly." ) metrics_context = get_metrics_context() if not metrics_context.in_progress(): raise RuntimeError( "No metrics context is in progress. Please only call this function within a metrics context/dynamo_timed." ) metrics_context.add_to_set(key, value) chromium_log.add_to_set(event_name, key, value) @staticmethod def add_to_set_toplevel( key: str, value: Any, log_level: CompileEventLogLevel = CompileEventLogLevel.COMPILATION_METRIC, ): """ Same as add to set, just does it automatically to the toplevel event instead of having to explicitly name it. Defaults to COMPILATION_METRIC log level. """ chromium_log = get_chromium_event_logger() top_event = chromium_log.get_outermost_event() if top_event is None: raise RuntimeError( "No toplevel event active. Please only call this function within a metrics context/dynamo_timed." ) CompileEventLogger.add_to_set(top_event, log_level, key, value) # Helper functions that are syntactic sugar @staticmethod def chromium(event_name: str, **metadata: object): """ Add to in chromium. Each key/value of metadata will appear in the chromium trace. should be the name of a timed event span passed to `dynamo_timed`. """ CompileEventLogger.add_data( event_name, CompileEventLogLevel.CHROMIUM, overwrite=False, **metadata ) @staticmethod def pt2_compile(event_name: str, **metadata: object): """ Add to in chromium and PT2 Compile Events. Each key/value of metadata will appear in the chromium trace. Each kwarg name becomes a column in PT2 Compile Events, with the corresponding kwarg value. should be the name of a timed event span passed to `dynamo_timed`, with log_to_pt2_compile_events=True. """ CompileEventLogger.add_data( event_name, CompileEventLogLevel.PT2_COMPILE, overwrite=False, **metadata ) @staticmethod def compilation_metric(overwrite: bool = False, **metadata: object): """ Add to the CompilationMetrics context. Also logs to PT2 Compile Events and chromium. Each key/value of metadata will appear in the chromium trace. Each kwarg name becomes a column in PT2 Compile Events and Dynamo Compile, with the corresponding kwarg value. """ CompileEventLogger.add_toplevel( CompileEventLogLevel.COMPILATION_METRIC, overwrite, **metadata ) @staticmethod def instant( event_name: str, metadata: dict[str, Any], time_ns: Optional[int] = None ): """ Log an instant event to chromium logs with name at time . The `args` field in Perfetto will point to metadata. should be a value obtained from time.time_ns(). """ CompileEventLogger.log_instant_event( event_name, metadata, time_ns, CompileEventLogLevel.CHROMIUM ) @staticmethod def try_add_pt2_compile(event_name: str, **metadata: object): """ Adds to an existing pt2_compile event, but silently returns if the event doesn't exist. This function is syntactic sugar for chromium_event_logger().try_add_event_data. """ chromium_log = get_chromium_event_logger() chromium_log.try_add_event_data(event_name, **metadata) @contextmanager def dynamo_timed( key: str, # TODO(masneral): Deprecate this param. phase_name: Optional[str] = None, log_pt2_compile_event: bool = False, metadata: Optional[dict[str, object]] = None, dynamo_compile_column_us: Optional[str] = None, dynamo_compile_runtime_column_us: Optional[str] = None, compile_id: Optional[CompileId] = None, is_forward: Optional[bool] = None, log_waitcounter: bool = False, ) -> Generator[Any, None, None]: """ dynamo_timed is a context manager By wrapping a function in dynamo_timed, we can get a few things: 1) Optionally log timings to pt2_compile_events. 2) Optionally log timings to CompilationMetrics (dynamo_compile). 3) Optionally log chromium events. 4) Optionally increment a WaitCounter. 5) Store a record in compilation_time_metrics For example: def _foo(...): with dynamo_timed("_foo"): ... Would show up as an entry in our timing dict: OrderedDict([('_foo', [0.083690, 0.23949, 3.1425e-05])]) This is extremely useful for granular debugging. Although it is tempting to use dynamo_timed as a decorator, please do not. In its decorator form it makes cProfile traces less useful as dynamo_timed suddenly becomes a bottleneck for lots of function calls (as only one parent pointer is recorded). Params: - key: key into compile_time_metrics. If phase_name is not provided, this is also the event name used for pt2_compile_events logs and chromium events. - phase_name: Optional override for the event name. - log_pt2_compile_event: Whether to log a pt2 compile event internally. - metadata: Extra metadata to put in pt2_compile_events. - dynamo_compile_column_us: If provided, updates the specified CompilationMetrics field to be logged to dyname_compile column. We expect all columns to be _us; therefore, the field name must end with "_us". - dynamo_compile_runtime_column_us: Like 'dynamo_compile_column_us', but should be used for those columns captured outside of a compile context, e.g., runtime autotuning. - compile_id: In the typical case, this parameter should not be needed. Use to supply the compile_id for those cases where we want to log a compile_id where it's not naturally available, e.g., for runtime autotuning. - is_forward: Optionally set an is_forward field for those logging destinations that support it. - log_waitcounter: If set, we'll log a waitcounter of the form "pytorch.dynamo_timed.{key}" """ # We're standardizing on microseconds for dynamo_compile timings. if dynamo_compile_column_us is not None: assert dynamo_compile_column_us.endswith("_us") # Only one of these should be set. assert dynamo_compile_column_us is None or dynamo_compile_runtime_column_us is None if phase_name: event_name = phase_name fn_name = key else: event_name = key fn_name = None if key not in compilation_time_metrics: compilation_time_metrics[key] = [] event_metadata = {} if metadata: event_metadata.update(metadata) if fn_name: event_metadata.update({"fn_name": fn_name}) if is_forward is not None: event_metadata.update({"is_backward": not is_forward}) chromium_log: ChromiumEventLogger = get_chromium_event_logger() start_ns = time.time_ns() chromium_log.log_event_start( event_name, start_ns, event_metadata, log_pt2_compile_event, compile_id ) try: with torch.profiler.record_function(f"{key} (dynamo_timed)"): if log_waitcounter: with _WaitCounter(f"pytorch.dynamo_timed.{key}").guard(): yield else: yield finally: end_ns = time.time_ns() time_spent_ns = end_ns - start_ns compilation_time_metrics[key].append(time_spent_ns / 1e9) chromium_log.log_event_end( event_name, end_ns, {}, start_ns, log_pt2_compile_event, compile_id ) if dynamo_compile_column_us: metrics_context = get_metrics_context() if metrics_context.in_progress(): metrics_context.increment( dynamo_compile_column_us, time_spent_ns // 1000 ) # TODO: the events that we capture in calculate_time_spent() seem a little # arbitrary. Currently, it's only those fields that are present in # CompilationMetrics (but note that we accumulate by the associated event # name, not the field name in CompilationMetrics). Do we want to keep it # this way? cumulative_time_spent_ns[event_name] += time_spent_ns if dynamo_compile_runtime_column_us: get_runtime_metrics_context().increment( dynamo_compile_runtime_column_us, time_spent_ns // 1000, extra={ "compile_id": compile_id, "is_runtime": True, "is_forward": is_forward, }, ) cumulative_time_spent_ns[event_name] += time_spent_ns @overload def compile_times(repr: Literal["str"], aggregate: bool = False) -> str: ... @overload def compile_times( repr: Literal["csv"], aggregate: bool = False ) -> tuple[list[str], list[object]]: ... def compile_times(repr="str", aggregate: bool = False): """ Get metrics about torchdynamo frontend/backend compilation times. Accumulates information from functions tagged with `dynamo_timed`. repr='str' returns a printable string for user interaction, and 'csv' returns headers, rows which can be logged for output aggregate causes values from multiple compilations (e.g. split graphs) to be accumulated into one value. If false, expect more than one value per metric. """ def fmt_fn(values, item_fn=lambda x: x): if aggregate: return item_fn(sum(values)) return ", ".join(map(item_fn, values)) if repr == "str": rows = [ (k, fmt_fn(compilation_time_metrics[k], item_fn=lambda x: f"{x:.4f}")) for k in compilation_time_metrics ] out = "TorchDynamo compilation metrics:\n" out += tabulate(rows, headers=("Function", "Runtimes (s)")) return out elif repr == "csv": values = [ fmt_fn(v, item_fn=lambda x: f"{x:.6f}") for v in compilation_time_metrics.values() ] headers = list(compilation_time_metrics.keys()) return headers, values return None @atexit.register def dump_compile_times() -> None: log.info(compile_times(repr="str", aggregate=True)) tensortype_to_dtype = { torch.FloatTensor: (torch.float32, torch.float), torch.DoubleTensor: (torch.float64, torch.double), torch.HalfTensor: (torch.float16, torch.half), torch.BFloat16Tensor: (torch.bfloat16,), torch.ByteTensor: (torch.uint8,), torch.CharTensor: (torch.int8,), torch.LongTensor: (torch.int64, torch.long), torch.IntTensor: (torch.int32, torch.int), torch.ShortTensor: (torch.int16, torch.short), torch.BoolTensor: (torch.bool,), } class DuplicateWarningChecker: def __init__(self, maxsize: int = 4096) -> None: self.maxsize = maxsize self.reset() def reset(self): self.set = OrderedDict() def add(self, key: Union[str, tuple[object, object]]) -> bool: if key in self.set: self.set.move_to_end(key, last=True) if not config.verbose: return False else: self.set[key] = None while len(self.set) > self.maxsize: self.set.popitem(last=False) return True graph_break_dup_warning_checker = DuplicateWarningChecker() def setup_compile_debug(): compile_debug = os.environ.get("TORCH_COMPILE_DEBUG", "0") == "1" if compile_debug: return add_file_handler() return contextlib.ExitStack() def reset_graph_break_dup_checker() -> None: graph_break_dup_warning_checker.reset() def add_file_handler(): log_path = os.path.join(get_debug_dir(), "torchdynamo") os.makedirs(log_path, exist_ok=True) log_file_handler = logging.FileHandler(os.path.join(log_path, "debug.log")) logger = logging.getLogger("torch._dynamo") logger.addHandler(log_file_handler) exitstack = contextlib.ExitStack() exitstack.callback(lambda: logger.removeHandler(log_file_handler)) return exitstack def setup_log_file(): exitstack = contextlib.ExitStack() if config.log_file_name is not None: log_file_handler = logging.FileHandler(config.log_file_name) for logger in torch._logging._internal.get_loggers(): logger.addHandler(log_file_handler) exitstack.callback(lambda: logger.removeHandler(log_file_handler)) return exitstack return exitstack def gen_record_file_name(exc, code) -> str: return f"{get_debug_dir()}/error_recordings/\ {code.co_name}_{type(exc).__name__}_{code.co_firstlineno}.rec" def write_record_to_file(filename: str, exec_record) -> None: try: if os.path.exists(filename): log.warning( "Unable to write execution record %s; file already exists.", filename ) else: os.makedirs(os.path.dirname(filename), exist_ok=True) with open(filename, "wb") as f: exec_record.dump(f) except Exception: log.exception("Unable to write execution record %s", filename) def count_calls(g: fx.Graph) -> int: c = 0 for n in g.nodes: if "call" in n.op: c += 1 return c def identity(x: T) -> T: return x def hashable(x): try: hash(x) return True except TypeError: return False # cannot hash writable memoryview object except ValueError: return False def nothing(*args, **kwargs): pass class ExactWeakKeyDictionary: """Similar to weakref.WeakKeyDictionary, but use `is`/`id` rather than `==` to compare equality""" def __init__(self): self.values = {} self.refs = {} def __getitem__(self, key): return self.values[id(key)] def get(self, key, default=None): return self.values.get(id(key), default) def __contains__(self, key): return id(key) in self.values def __setitem__(self, key, value): idx = id(key) if idx not in self.refs: self.refs[idx] = weakref.ref(key, lambda ref: self._remove_id(idx)) self.values[idx] = value def _remove_id(self, idx): if idx in self.values: del self.values[idx] if idx in self.refs: del self.refs[idx] def clear(self): self.refs.clear() self.values.clear() @overload def istype(obj: object, allowed_types: type[T]) -> TypeIs[T]: ... @overload def istype( obj: object, allowed_types: tuple[type[list[T]], type[tuple[T, ...]]] ) -> TypeIs[T]: ... @overload def istype(obj: object, allowed_types: Iterable[type]) -> bool: ... def istype(obj, allowed_types): """isinstance() without subclasses""" if isinstance(allowed_types, (tuple, list, set)): return type(obj) in allowed_types return type(obj) is allowed_types if sys.version_info >= (3, 12): # Some typing classes moved to C in 3.12, # which no longer have the _Final mixin. _builtin_final_typing_classes = ( typing.ParamSpecArgs, typing.ParamSpecKwargs, typing.ParamSpec, typing.TypeVar, typing.TypeVarTuple, typing.TypeAliasType, ) def is_typing(value): # _Final catches most of typing classes: # - Any # - Callable # - Union # ... # # NB: we intentionally ignore classes that inherit from Generic, since they # can be used as both TypingVariable as well as UserDefinedClassVariable. if sys.version_info >= (3, 12) and isinstance(value, _builtin_final_typing_classes): return True return isinstance(value, typing._Final) or value is typing.Generic # type: ignore[attr-defined] def is_numpy_int_type(value): if not np: return False return istype( value, ( np.int8, np.int16, np.int32, np.int64, np.uint8, np.uint16, np.uint32, np.uint64, ), ) def is_numpy_float_type(value): if not np: return False return istype( value, ( np.float16, np.float32, np.float64, ), ) def is_lru_cache_wrapped_function(value): return isinstance(value, functools._lru_cache_wrapper) and is_function( inspect.getattr_static(value, "__wrapped__") ) def is_function_or_wrapper(value): return is_function(value) or isinstance( value, (torch._ops.OpOverloadPacket, torch._ops.OpOverload) ) def is_function(value): return isinstance( value, ( types.FunctionType, types.BuiltinFunctionType, types.MethodDescriptorType, types.WrapperDescriptorType, ), ) cmp_name_to_op_mapping = { "__eq__": operator.eq, "__ne__": operator.ne, "__lt__": operator.lt, "__le__": operator.le, "__gt__": operator.gt, "__ge__": operator.ge, } cmp_name_to_op_str_mapping = { "__eq__": "==", "__ne__": "!=", "__lt__": "<", "__le__": "<=", "__gt__": ">", "__ge__": ">=", } def is_wrapper_or_member_descriptor(value): return isinstance( value, ( # set up by PyGetSetDef types.GetSetDescriptorType, # set by PyMethodDef, e.g. list.append types.MethodDescriptorType, # slots - list.__add__ types.WrapperDescriptorType, # set up by PyMemberDef types.MemberDescriptorType, # wrapper over C functions types.MethodWrapperType, ), ) def unwrap_if_wrapper(fn): return unwrap_with_attr_name_if_wrapper(fn)[0] def unwrap_with_attr_name_if_wrapper(fn): # TODO(anijain2305) - Investigate if we can get rid of this function # unpack @torch._dynamo.optimize()(fn) wrapped function if is_function(fn) and inspect.getattr_static(fn, "_torchdynamo_inline", False): fn = inspect.getattr_static(fn, "_torchdynamo_inline", fn) attr_name = "_torchdynamo_inline" else: attr_name = None return fn, attr_name def is_numpy_ndarray(value): if not np: return False return istype(value, np.ndarray) def istensor(obj): """Check of obj is a tensor""" tensor_list: tuple[type, ...] = ( torch.Tensor, torch.nn.Parameter, *config.traceable_tensor_subclasses, ) tensor_list = tensor_list + (torch._subclasses.FakeTensor,) return istype(obj, tensor_list) def is_lazy_module(mod): return isinstance(mod, LazyModuleMixin) @functools.lru_cache(4096) def print_once(*args): print(*args) def make_cell(val=None): """Some black magic to create a cell object that usually only exists in a closure""" x = val def f(): return x assert f.__closure__ is not None and len(f.__closure__) == 1 return f.__closure__[0] def proxy_args_kwargs(args, kwargs): try: proxy_args = tuple(arg.as_proxy() for arg in args) proxy_kwargs = {key: arg.as_proxy() for key, arg in kwargs.items()} return proxy_args, proxy_kwargs except NotImplementedError as e: from .exc import unimplemented_v2 from .variables.base import typestr unimplemented_v2( gb_type="Failed to convert args/kwargs to proxy", context=f"call_function args: {typestr(*args)} {typestr(*list(kwargs.values()))}", explanation="Missing `as_proxy()` implementation for some arg/kwarg.", hints=[], from_exc=e, ) def to_int_ms(v: Optional[float]) -> Optional[int]: return None if v is None else int(v * 1000) # float64 timestamp has a quarter microsecond precision in 2024, so while # this is suboptimal we shouldn't meaningfully lose precision def to_int_us(v: Optional[float]) -> Optional[int]: return None if v is None else int(v * 1_000_000) # Version field added to every log. Increment to make it easier to distinguish new # vs. old entries when you make a substantive change to how the logs are populated. LOG_FORMAT_VERSION = 3 @dataclasses.dataclass class CompilationMetrics: compile_id: Optional[str] = None frame_key: Optional[str] = None co_name: Optional[str] = None co_filename: Optional[str] = None co_firstlineno: Optional[int] = None cache_size: Optional[int] = None accumulated_cache_size: Optional[int] = None guard_count: Optional[int] = None shape_env_guard_count: Optional[int] = None graph_op_count: Optional[int] = None graph_node_count: Optional[int] = None graph_input_count: Optional[int] = None start_time: Optional[float] = None entire_frame_compile_time_s: Optional[float] = None backend_compile_time_s: Optional[float] = None inductor_compile_time_s: Optional[float] = None code_gen_time_s: Optional[float] = None fail_type: Optional[str] = None fail_reason: Optional[str] = None fail_user_frame_filename: Optional[str] = None fail_user_frame_lineno: Optional[int] = None non_compliant_ops: Optional[set[str]] = None compliant_custom_ops: Optional[set[str]] = None restart_reasons: Optional[set[str]] = None dynamo_time_before_restart_s: Optional[float] = None # Sometimes, we will finish analyzing a frame but conclude we don't want # to install any guarded code. True means we actually decided to install # a compiled frame has_guarded_code: Optional[bool] = None remote_cache_time_saved_s: Optional[float] = None structured_logging_overhead_s: Optional[float] = None config_suppress_errors: Optional[bool] = None config_inline_inbuilt_nn_modules: Optional[bool] = None specialize_float: Optional[bool] = None dynamo_config: Optional[str] = None is_forward: Optional[bool] = None num_triton_bundles: Optional[int] = None remote_fx_graph_cache_get_time_ms: Optional[int] = None remote_fx_graph_cache_put_time_ms: Optional[int] = None start_time_us: Optional[int] = None duration_us: Optional[int] = None dynamo_cumulative_compile_time_us: Optional[int] = None aot_autograd_cumulative_compile_time_us: Optional[int] = None inductor_cumulative_compile_time_us: Optional[int] = None inductor_code_gen_cumulative_compile_time_us: Optional[int] = None triton_compile_time_us: Optional[int] = None runtime_cudagraphify_time_us: Optional[int] = None runtime_triton_autotune_time_us: Optional[int] = None dynamo_compile_time_before_restart_us: Optional[int] = None cuda_synchronize_time_us: Optional[int] = None # TODO: instrument distributed_ephemeral_timeout_us: Optional[int] = None structured_logging_overhead_us: Optional[int] = None remote_fx_graph_cache_get_time_us: Optional[int] = None remote_fx_graph_cache_put_time_us: Optional[int] = None backward_cumulative_compile_time_us: Optional[int] = None end_time_us: Optional[int] = None pre_grad_pass_time_us: Optional[int] = None post_grad_pass_time_us: Optional[int] = None joint_graph_pass_time_us: Optional[int] = None log_format_version: int = LOG_FORMAT_VERSION inductor_config: Optional[str] = None remote_cache_version: Optional[int] = None inductor_fx_remote_cache_hit_count: Optional[int] = None inductor_fx_remote_cache_miss_count: Optional[int] = None inductor_fx_remote_cache_backend_type: Optional[str] = None inductor_fx_remote_cache_hit_keys: Optional[str] = None inductor_fx_remote_cache_miss_keys: Optional[str] = None cuda_version: Optional[str] = None triton_version: Optional[str] = None feature_usage: Optional[dict[str, bool]] = None compile_time_autotune_time_us: Optional[int] = None is_runtime: Optional[bool] = False gc_time_us: Optional[int] = None tensorify_float_attempt: Optional[bool] = None tensorify_float_success: Optional[bool] = None tensorify_float_failure: Optional[set[str]] = None guard_latency_us: Optional[float] = None recompile_reason: Optional[str] = None num_graph_breaks: Optional[int] = None triton_kernel_compile_times_us: Optional[str] = None ir_count: Optional[int] = None cudagraph_skip_reason: Optional[str] = None @classmethod def create(cls, metrics: dict[str, Any]): """ Factory method to create a CompilationMetrics from a dict of fields. Includes the logic to add legacy fields and any pre-processing, e.g., we transform some fields to comma-separated strings for scuba logging. """ def us_to_s(metric: Optional[int]) -> Optional[float]: return metric / 1e6 if metric is not None else None def us_to_ms(metric: Optional[int]) -> Optional[int]: return metric // 1000 if metric is not None else None def collection_to_str(metric: Optional[Any]) -> Optional[str]: def safe_str(item: Any) -> str: try: return str(item) except Exception: return "" if metric is None: return None if not isinstance(metric, (set, list)): return "" return ",".join(safe_str(item) for item in sorted(metric)) def collection_to_json_str(metric: Optional[Any]) -> Optional[str]: if metric is None: return None try: return json.dumps(list(metric)) except Exception: return "" # TODO: The following are legacy fields, populated from the fields that replace # them. Remove these when we decide we can really deprecate them. legacy_metrics = { "start_time": us_to_s(metrics.get("start_time_us")), "entire_frame_compile_time_s": us_to_s( metrics.get("dynamo_cumulative_compile_time_us") ), "backend_compile_time_s": us_to_s( metrics.get("aot_autograd_cumulative_compile_time_us") ), "inductor_compile_time_s": us_to_s( metrics.get("inductor_cumulative_compile_time_us") ), "code_gen_time_s": us_to_s( metrics.get("inductor_code_gen_cumulative_compile_time_us") ), "remote_cache_time_saved_s": us_to_s( metrics.get("distributed_ephemeral_timeout_us") ), "remote_fx_graph_cache_get_time_ms": us_to_ms( metrics.get("remote_fx_graph_cache_get_time_us") ), "remote_fx_graph_cache_put_time_ms": us_to_ms( metrics.get("remote_fx_graph_cache_put_time_us") ), "structured_logging_overhead_s": us_to_s( metrics.get("structured_logging_overhead_us") ), } all_metrics = {**legacy_metrics, **metrics} # Processing before logging: all_metrics["inductor_fx_remote_cache_hit_keys"] = collection_to_str( all_metrics.get("inductor_fx_remote_cache_hit_keys") ) all_metrics["inductor_fx_remote_cache_miss_keys"] = collection_to_str( all_metrics.get("inductor_fx_remote_cache_miss_keys") ) all_metrics["triton_kernel_compile_times_us"] = collection_to_json_str( all_metrics.get("triton_kernel_compile_times_us") ) compile_id = all_metrics.get("compile_id") all_metrics["compile_id"] = str(compile_id) if compile_id else None return cls(**all_metrics) DEFAULT_COMPILATION_METRICS_LIMIT = 64 _compilation_metrics: collections.deque[CompilationMetrics] = collections.deque( maxlen=DEFAULT_COMPILATION_METRICS_LIMIT ) def add_compilation_metrics_to_chromium(c: CompilationMetrics) -> None: """ These are the common fields in CompilationMetrics that existed before metrics_context, and aren't set by MetricsContext.set(). We add the subset of them that make sense in `dynamo`/toplevel events in PT2 Compile Events directly. If you're tempted to add to this list, consider using CompileEventLogger.compilation_metric() instead, which will automatically also add it to tlparse and PT2 Compile Events. TODO: Get rid of this function and replace it with CompileEventLogger directly instead. """ event_logger = get_chromium_event_logger() event_name = event_logger.get_outermost_event() if not event_name: return event_logger.add_event_data( event_name=event_name, frame_key=c.frame_key, co_name=c.co_name, co_filename=c.co_filename, co_firstlineno=c.co_firstlineno, cache_size=c.cache_size, accumulated_cache_size=c.accumulated_cache_size, guard_count=c.guard_count, shape_env_guard_count=c.shape_env_guard_count, graph_op_count=c.graph_op_count, graph_node_count=c.graph_node_count, graph_input_count=c.graph_input_count, fail_type=c.fail_type, fail_reason=c.fail_reason, fail_user_frame_filename=c.fail_user_frame_filename, fail_user_frame_lineno=c.fail_user_frame_lineno, # Sets aren't JSON serializable non_compliant_ops=list(c.non_compliant_ops) if c.non_compliant_ops is not None else None, compliant_custom_ops=list(c.compliant_custom_ops) if c.compliant_custom_ops is not None else None, restart_reasons=list(c.restart_reasons) if c.restart_reasons is not None else None, dynamo_time_before_restart_s=c.dynamo_time_before_restart_s, has_guarded_code=c.has_guarded_code, dynamo_config=c.dynamo_config, ) def _get_dynamo_config_for_logging() -> Optional[str]: def clean_for_json(d: dict[str, Any]) -> dict[str, Any]: blocklist = { "TYPE_CHECKING", "log_file_name", "verbose", "repro_after", "repro_level", "repro_forward_only", "repro_tolerance", "repro_ignore_non_fp", "same_two_models_use_fp64", "base_dir", "debug_dir_root", "_save_config_ignore", "log_compilation_metrics", "inject_BUILD_SET_unimplemented_TESTING_ONLY", "_autograd_backward_strict_mode_banned_ops", "reorderable_logging_functions", "ignore_logger_methods", "traceable_tensor_subclasses", "_custom_ops_profile", } return { key: sorted(value) if isinstance(value, set) else value for key, value in d.items() if key not in blocklist } config_dict = clean_for_json(config.get_config_copy()) return json.dumps(config_dict, sort_keys=True) def _scrubbed_inductor_config_for_logging() -> Optional[str]: """ Method to parse and scrub uninteresting configs from inductor config """ # TypeSafeSerializer for json.dumps() # Skips complex types as values in config dict class TypeSafeSerializer(json.JSONEncoder): def default(self, o): try: return super().default(o) except Exception: return "Value is not JSON serializable" keys_to_scrub: set[Any] = set() inductor_conf_str = None inductor_config_copy = ( torch._inductor.config.get_config_copy() if torch._inductor.config else None ) if inductor_config_copy is not None: try: for key, val in inductor_config_copy.items(): if not isinstance(key, str): keys_to_scrub.add(key) # Convert set() to list for json.dumps() if isinstance(val, set): inductor_config_copy[key] = list(val) # Evict unwanted keys for key in keys_to_scrub: del inductor_config_copy[key] # Stringify Inductor config inductor_conf_str = json.dumps( inductor_config_copy, cls=TypeSafeSerializer, skipkeys=True, sort_keys=True, ) except Exception: # Don't crash because of runtime logging errors inductor_conf_str = "Inductor Config is not JSON serializable" return inductor_conf_str def record_compilation_metrics( start_time_ns: int, end_time_ns: int, metrics: dict[str, Any], exc_type: Optional[type[BaseException]], exc_value: Optional[BaseException], ): if torch._inductor.utils.should_use_remote_fx_graph_cache(): try: from torch._inductor.fb.remote_cache import REMOTE_CACHE_VERSION remote_cache_version = REMOTE_CACHE_VERSION inductor_fx_remote_cache_backend_type = "_ManifoldCache" except ModuleNotFoundError: remote_cache_version = None inductor_fx_remote_cache_backend_type = None else: inductor_fx_remote_cache_backend_type = None remote_cache_version = None # Populate the compile_id from the metrics context if it's set. Otherwise, # look for it in the current compile context. compile_id = metrics.get("compile_id") if not compile_id: compile_id = torch._guards.CompileContext.current_compile_id() common_metrics = { "compile_id": compile_id, "start_time_us": start_time_ns // 1000, "end_time_us": end_time_ns // 1000, "duration_us": (end_time_ns - start_time_ns) // 1000, "fail_type": exc_type.__qualname__ if exc_type else None, "fail_reason": str(exc_value) if exc_value else None, "structured_logging_overhead_us": to_int_us( torch._logging.get_structured_logging_overhead() ), "dynamo_config": _get_dynamo_config_for_logging(), "inductor_config": _scrubbed_inductor_config_for_logging(), "cuda_version": torch.version.cuda, "triton_version": triton.__version__ if has_triton() else "", "remote_cache_version": remote_cache_version, "inductor_fx_remote_cache_backend_type": inductor_fx_remote_cache_backend_type, } compilation_metrics = CompilationMetrics.create({**common_metrics, **metrics}) _compilation_metrics.append(compilation_metrics) name = "compilation_metrics" if compilation_metrics.is_forward is False: name = "bwd_" + name if compilation_metrics.is_runtime is True: name = name + "_runtime" torch._logging.trace_structured( name, lambda: { k: list(v) if isinstance(v, set) else v for k, v in dataclasses.asdict(compilation_metrics).items() }, # NB: Because compilation metrics *includes* the logging overhead time, # we can't both *measure* the logging overhead of compilation metrics # without making it inconsistent with compilation metrics itself, so # we ignore the (hopefully small) time spent logging compilation metrics record_logging_overhead=False, # These may be runtime logs, e.g., runtime autotunning, so we provide # the CompileId from the compilation metrics in case it's not available # in the current trace. compile_id=compile_id, ) # If there's a chromium event in flight, add the CompilationMetrics to it. add_compilation_metrics_to_chromium(compilation_metrics) # Finally log the compilation metrics. if config.log_compilation_metrics: log_compilation_event(compilation_metrics) # record_compilation_metrics is called by the singleton MetricsContext exit handler. _METRICS_CONTEXT = MetricsContext(on_exit=record_compilation_metrics) _RUNTIME_METRICS_CONTEXT = RuntimeMetricsContext(on_exit=record_compilation_metrics) def set_compilation_metrics_limit(new_size: int) -> None: global _compilation_metrics while len(_compilation_metrics) > new_size: _compilation_metrics.popleft() new_deque = collections.deque(_compilation_metrics, maxlen=new_size) _compilation_metrics = new_deque def clear_compilation_metrics() -> None: global _compilation_metrics _compilation_metrics.clear() def get_compilation_metrics() -> list[CompilationMetrics]: return list(_compilation_metrics) class ChromiumEventLogger: """Logs chromium events to structured logs. tlparse will concatenate these into a perfetto UI link. See https://docs.google.com/document/d/1CvAClvFfyA5R-PhYUmn5OOQtYMH4h6I0nSsKchNAySU/preview#heading=h.yr4qxyxotyw for a specification of the Chromium Event JSON format. """ def get_stack(self) -> list[str]: """ The main event stack, with every chromium event. Logged to tlparse. """ if hasattr(self.tls, "stack"): return self.tls.stack else: self.tls.stack = [] return self.tls.stack def get_outermost_event(self) -> Optional[str]: """ Get the outermost event name (i.e. the longest running event) or None if the stack is empty. """ stack = self.get_stack() return stack[0] if stack else None def get_pt2_compile_substack(self): """ A smaller subset of the main stack that gets used to log PT2 Compile Events internally. """ if hasattr(self.tls, "pt2_compile_substack"): return self.tls.pt2_compile_substack else: self.tls.pt2_compile_substack = [] return self.tls.pt2_compile_substack def get_event_data(self) -> dict[str, Any]: if not hasattr(self.tls, "event_data"): self.tls.event_data = {} return self.tls.event_data def __init__(self): self.tls = threading.local() # Generate a unique id for this logger, which we can use in scuba to filter down # to a single python run. self.id_ = str(uuid.uuid4()) # TODO: log to init/id tlparse after I add support for it log.info("ChromiumEventLogger initialized with id %s", self.id_) def try_add_event_data(self, event_name: str, **kwargs) -> None: """ Same as add_event_data, but will silently not log if the event isn't in the stack. """ if event_name not in self.get_stack(): return self.add_event_data(event_name, **kwargs) def add_event_data( self, event_name: str, **kwargs, ) -> None: """ Adds additional metadata info to an in-progress event This metadata is recorded in the END event """ if event_name not in self.get_stack(): raise RuntimeError( f"Event {repr(event_name)} not in {self.get_stack()}. " "Cannot add metadata to events that aren't in progress. " "Please make sure the event has started and hasn't ended." ) event_data = self.get_event_data() if event_name not in event_data: event_data[event_name] = {} event_data[event_name].update(kwargs) def increment(self, event_name: str, key: str, value: int): """ Increment an integer event data field by the given amount """ if event_name not in self.get_stack(): raise RuntimeError( f"Event {repr(event_name)} not in {self.get_stack()}. " "Cannot add metadata to events that aren't in progress. " "Please make sure the event has started and hasn't ended." ) event_data = self.get_event_data() if event_name not in event_data: event_data[event_name] = {} if key not in event_data[event_name]: event_data[event_name][key] = 0 event_data[event_name][key] += value def add_to_set( self, event_name: str, key: str, value: Any, ): """ Add a value to a set within a event_name's metadata if it exists """ if event_name not in self.get_stack(): raise RuntimeError( f"Event {repr(event_name)} not in {self.get_stack()}. " "Cannot add metadata to events that aren't in progress. " "Please make sure the event has started and hasn't ended." ) event_data = self.get_event_data() if event_name not in event_data: event_data[event_name] = {} if key not in event_data[event_name]: event_data[event_name][key] = set() event_data[event_name][key].add(value) def log_event_start( self, event_name: str, time_ns: int, metadata: dict[str, Any], log_pt2_compile_event: bool = False, compile_id: Optional[CompileId] = None, ) -> None: """ Logs the start of a single event. :param str event_name Name of event to appear in trace :param time_ns Timestamp in nanoseconds :param metadata: Any extra metadata associated with this event :param log_pt2_compile_event: If True, log to pt2_compile_events :param compile_id: Explicit compile_id (rather than using the current context) """ compile_id = compile_id or torch._guards.CompileContext.current_compile_id() metadata["compile_id"] = str(compile_id) self._log_timed_event( event_name, time_ns, "B", metadata, ) self.get_stack().append(event_name) # Add metadata from start event self.add_event_data(event_name, **metadata) if log_pt2_compile_event: self.get_pt2_compile_substack().append(event_name) def reset(self) -> None: # We this on every compile in case a compile crashes or restarts and we haven't # cleared the stack. stack = self.get_stack() substack = self.get_pt2_compile_substack() stack.clear() substack.clear() event_data = self.get_event_data() event_data.clear() def log_event_end( self, event_name: str, time_ns: int, metadata: dict[str, Any], start_time_ns: int, log_pt2_compile_event: bool, compile_id: Optional[CompileId] = None, ) -> None: """ Logs the end of a single event. This function should only be called after log_event_start with the same event_name. :param event_name: Name of event to appear in trace :param time_ns: Timestamp in nanoseconds :param metadata: Any extra metadata associated with this event :param start_time_ns: The start time timestamp in nanoseconds :param log_pt_compile_event: If True, log to pt2_compile_events :param compile_id: Explicit compile_id (rather than using the current context) """ compile_id = compile_id or torch._guards.CompileContext.current_compile_id() metadata["compile_id"] = str(compile_id) # Grab metadata collected during event span all_event_data = self.get_event_data() if event_name in all_event_data: event_metadata = all_event_data[event_name] del all_event_data[event_name] else: event_metadata = {} # Add the passed in metadata event_metadata.update(metadata) event = self._log_timed_event( event_name, time_ns, "E", event_metadata, ) def pop_stack(stack): while event_name != stack[-1]: # If the event isn't the most recent one to end, pop # off the stack until it is. # Since event_name in self.stack, this pop is always safe log.warning( "ChromiumEventLogger: Detected overlapping events, fixing stack" ) stack.pop() event_stack = self.get_stack() # These stack health checks currently never happen, # but they're written this way to future proof any weird event # overlaps in the future. if event_name not in event_stack: # Something went wrong, we never called start on this event, # or it was skipped due to overlapping events below log.warning("ChromiumEventLogger: Start event not in stack, ignoring") return pop_stack(event_stack) if log_pt2_compile_event: pt2_compile_substack = self.get_pt2_compile_substack() pop_stack(pt2_compile_substack) log_chromium_event_internal( event, pt2_compile_substack, self.id_, start_time_ns ) # Pop actual event off of stack pt2_compile_substack.pop() # Finally pop the actual event off the stack event_stack.pop() def _log_timed_event( self, event_name: str, time_ns: int, phase: str, metadata: Optional[dict[str, Any]] = None, ) -> dict[str, Any]: """ Logs a timed event in chromium format. See log_event_start, log_event_end, etc. """ event = { "name": event_name, "ts": time_ns / 1000, # Chromium events are in micro seconds "args": metadata, "ph": phase, # These categories are needed in all chromium traces "cat": "dynamo_timed", "tid": 0, "pid": 0, # pid should be specified on all logs, we don't personally care about the actual process id } torch._logging.trace_structured( "chromium_event", payload_fn=lambda: event, suppress_context=False, expect_trace_id=False, # Not every chromium event will have a trace_id ) record_chromium_event_internal(event) return event def log_instant_event( self, event_name: str, time_ns: int, metadata: Optional[dict[str, Any]] = None, # By default, an instant event isn't logged internally, only to structured logging. log_pt2_compile_event: bool = False, ) -> None: """ Log an instant event with no associated duration. :param str event_name: Name of event to appear in trace :param int time_ns Timestamp in nanoseconds :param Optional[Dict[str, Any]] metadata: Any extra metadata associated with this event :param str cname optional color for the arrow in the trace """ if metadata is None: metadata = {} compile_id = str(torch._guards.CompileContext.current_compile_id()) metadata["compile_id"] = compile_id event = { "name": event_name, "ts": time_ns / 1000, "args": metadata, "ph": "i", # These categories are needed in all chromium traces "cat": "dynamo_timed", "tid": 0, "pid": 0, "s": "p", # We use "process" level instant events so they all appear on the same row in the trace. } torch._logging.trace_structured( "chromium_event", payload_fn=lambda: event, suppress_context=False, expect_trace_id=True, ) if log_pt2_compile_event: # Log an instant event with the same start and end time log_chromium_event_internal( event, self.get_pt2_compile_substack(), self.id_, time_ns ) CHROMIUM_EVENT_LOG: Optional[ChromiumEventLogger] = None def get_chromium_event_logger() -> ChromiumEventLogger: global CHROMIUM_EVENT_LOG if CHROMIUM_EVENT_LOG is None: CHROMIUM_EVENT_LOG = ChromiumEventLogger() return CHROMIUM_EVENT_LOG @contextmanager def chromium_event_timed( event_name: str, reset_event_log_on_exit: bool = False, log_pt2_compile_event: bool = False, ) -> Generator[Any, None, None]: """ Context manager that creates a chromium start and end event. Chromium event logging is integrated with dynamo_timed, so you probably want to use that instead. Use this context manager only if you want to avoid dynamo_timed. """ chromium_event_log = get_chromium_event_logger() chromium_start_time = time.time_ns() chromium_event_log.log_event_start( event_name, chromium_start_time, {}, log_pt2_compile_event, ) try: yield finally: chromium_event_log.log_event_end( event_name, time.time_ns(), {}, chromium_start_time, log_pt2_compile_event, ) if reset_event_log_on_exit: chromium_event_log.reset() @dataclasses.dataclass class CleanupHook: """Remove a global variable when hook is called""" scope: dict[str, Any] name: str def __call__(self, *args): # Make sure we're not shutting down if CleanupManager is not None: CleanupManager.count -= 1 del self.scope[self.name] @staticmethod def create(scope, name, val): assert name not in scope CleanupManager.count += 1 scope[name] = val return CleanupHook(scope, name) class CleanupManager(ExactWeakKeyDictionary): count = 0 instance: ClassVar[CleanupManager] def _remove_id(self, idx): for hook in self.values[idx]: hook() super()._remove_id(idx) CleanupManager.instance = CleanupManager() def clone_tensor(x): """Clone the tensor and its gradient""" y = x.clone().requires_grad_(x.requires_grad) if x.is_leaf and x.grad is not None: y.grad = x.grad.clone() return y def clone_input(x, *, dtype=None): """copy while preserving strides""" # TODO: this is questionable if is_fake(x): # this func fails on fake tensors in __torch_dispatch__ return x def torch_clone(x): y = torch.clone(x) if x.is_leaf: y.requires_grad_(x.requires_grad) if x.is_leaf and x.grad is not None: y.grad = clone_input(x.grad, dtype=dtype) if hasattr(x, "_dynamo_dynamic_indices"): y._dynamo_dynamic_indices = x._dynamo_dynamic_indices.copy() # type: ignore[attr-defined] return y with torch.no_grad(): if x.device.type == "xla": # Access data_ptr() for a xla tensor will cause crash return torch_clone(x) # Handle sparse storage (no stride). if x.layout is torch.sparse_coo: return torch.sparse_coo_tensor( torch_clone(x._indices()), torch_clone(x._values()), x.shape, is_coalesced=x.is_coalesced(), ) elif is_sparse_compressed(x): if x.layout in {torch.sparse_csr, torch.sparse_bsr}: compressed_indices = x.crow_indices() plain_indices = x.col_indices() else: compressed_indices = x.ccol_indices() plain_indices = x.row_indices() return torch.sparse_compressed_tensor( torch_clone(compressed_indices), torch_clone(plain_indices), torch_clone(x.values()), x.shape, layout=x.layout, ) needed_size = sum( (shape - 1) * stride for shape, stride in zip(x.size(), x.stride()) ) if x.is_quantized: result = torch.empty_quantized((needed_size + 32,), x) else: result = torch.empty( needed_size + 32, dtype=dtype or x.dtype, device=x.device ) cache_line_offset = ( (x.data_ptr() - result.data_ptr()) % 32 ) // x.element_size() result.as_strided_(x.size(), x.stride(), cache_line_offset) try: result.copy_(x.clone()) if x.is_leaf: result.requires_grad_(x.requires_grad) if x.is_leaf and x.grad is not None: result.grad = clone_input(x.grad, dtype=dtype) except RuntimeError: # RuntimeError: unsupported operation: more than one element of the written-to # tensor refers to a single memory location. Please clone() the tensor before # performing the operation. return torch_clone(x) if hasattr(x, "_dynamo_dynamic_indices"): result._dynamo_dynamic_indices = x._dynamo_dynamic_indices.copy() # type: ignore[attr-defined] return result def clone_inputs(example_inputs): res: Union[dict[Any, Any], list[Any]] if type(example_inputs) is dict: res = dict(example_inputs) for key, value in res.items(): if isinstance(value, tuple): res[key] = clone_inputs(value) else: assert isinstance(value, torch.Tensor), type(value) res[key] = clone_input(value) return res res = list(example_inputs) for i in range(len(res)): if isinstance(res[i], torch.Tensor): res[i] = clone_input(res[i]) return res def skip_frame_if_in_functorch_mode(val: torch.Tensor): try: val.data_ptr() # will throw for functorch tensors except RuntimeError as e: from .exc import SkipFrame # This will be GradTrackingTensor/BatchedTensor/etc functorch_subclass_name = re.sub(r"\(.*", "", repr(val)) raise SkipFrame( f"torch.compile cannot be run in context: {functorch_subclass_name}" ) from e @contextmanager def preserve_rng_state(): disable_functorch = torch._C._DisableFuncTorch disable_current_modes = torch.utils._python_dispatch._disable_current_modes with disable_current_modes(), disable_functorch(): rng_state = torch.clone(torch.random.get_rng_state()) skip_frame_if_in_functorch_mode(rng_state) if torch.cuda.is_available(): cuda_rng_state = torch.clone(torch.cuda.get_rng_state()) try: yield finally: with torch.utils._python_dispatch._disable_current_modes(): torch.random.set_rng_state(rng_state) if torch.cuda.is_available(): torch.cuda.set_rng_state(cuda_rng_state) # type: ignore[possibly-undefined] def is_jit_model(model0): return isinstance( model0, ( torch.jit._trace.TopLevelTracedModule, torch.jit._script.RecursiveScriptModule, torch.jit.ScriptFunction, torch.jit.ScriptModule, ), ) def torchscript(model, example_inputs, verbose=False): if is_jit_model(model): # already done? return model try: return torch.jit.trace(model, example_inputs) except Exception: try: return torch.jit.script(model) except Exception: if verbose: log.exception("jit error") else: log.error("Both torch.jit.trace and torch.jit.script failed") return None def getfile(obj): try: return inspect.getfile(obj) except (TypeError, OSError): return None def is_namedtuple(obj): """Test if an object is a namedtuple or a torch.return_types.* quasi-namedtuple""" return is_namedtuple_cls(type(obj)) def is_namedtuple_cls(cls): """Test if an object is a namedtuple or a (torch.return_types|torch.autograd.forward_ad).* quasi-namedtuple""" try: if issubclass(cls, tuple): module = getattr(cls, "__module__", None) if module in ("torch.return_types", "torch.autograd.forward_ad"): return True if isinstance(getattr(cls, "_fields", None), tuple) and callable( getattr(cls, "_make", None) ): # The subclassing style namedtuple can have an extra base `typing.Generic` bases = tuple(t for t in cls.__bases__ if t is not Generic) if bases == (tuple,): # This is a namedtuple type directly created by `collections.namedtuple(...)` return True if bases and any( ( # Subclass of namedtuple is_namedtuple_cls(t) # For subclasses of namedtuple, the __new__ method should not be customized and cls.__new__ is t.__new__ ) for t in bases ): return True except TypeError: pass return False @functools.lru_cache(1) def namedtuple_fields(cls) -> tuple[str, ...]: """Get the fields of a namedtuple or a torch.return_types.* quasi-namedtuple""" if cls is slice: return ("start", "stop", "step") assert issubclass(cls, tuple) if hasattr(cls, "_fields"): # normal namedtuples return cls._fields @dataclasses.dataclass class Marker: index: int # frustrating ones e.g. torch.return_types.max assert cls.__module__ == "torch.return_types" obj = cls(map(Marker, range(cls.n_fields))) fields: dict[str, int] = {} for name in dir(obj): if name[0] != "_" and isinstance(getattr(obj, name), Marker): fields[name] = getattr(obj, name).index assert len(fields) == cls.n_fields return tuple(sorted(fields, key=fields.get)) # type: ignore[arg-type] def checkpoint_params(gm): with torch.no_grad(): rng_state = torch.clone(torch.random.get_rng_state()) if torch.cuda.is_available(): cuda_rng_state = torch.clone(torch.cuda.get_rng_state()) saved_state = [ (param, param._version, torch.clone(param)) for param in itertools.chain(gm.parameters(), gm.buffers()) ] def restore(): with torch.no_grad(): torch.random.set_rng_state(rng_state) if torch.cuda.is_available(): torch.cuda.set_rng_state(cuda_rng_state) for param, version, original_value in saved_state: if param._version != version: param.copy_(original_value) return restore def timed(model, example_inputs, times=1): if torch.cuda.is_available(): synchronize = torch.cuda.synchronize else: synchronize = nothing synchronize() gc.collect() torch.manual_seed(1337) t0 = time.perf_counter() for _ in range(times): result = model(*example_inputs) synchronize() t1 = time.perf_counter() return result, t1 - t0 # type: ignore[possibly-undefined] def check_is_cuda(gm, example_inputs): return all(x.is_cuda for x in itertools.chain(example_inputs, gm.parameters(True))) @lru_cache(32) def rot_n_helper(n): assert n > 1 vars = [f"v{i}" for i in range(n)] rotated = reversed(vars[-1:] + vars[:-1]) fn = eval(f"lambda {','.join(vars)}: ({','.join(rotated)})") fn.__name__ = f"rot_{n}_helper" return fn common_constant_types: set[type] = { int, float, complex, bool, str, bytes, type(None), Ellipsis.__class__, NotImplemented.__class__, types.CodeType, # Commonly used immutable types from torch. torch.device, torch.dtype, torch.memory_format, torch.layout, torch.finfo, torch.iinfo, torch.nn.attention.SDPBackend, torch.cuda._CudaDeviceProperties, } if has_triton_package(): import triton common_constant_types.add(triton.language.dtype) """ Difference between is_safe_constant and common_constant_types. * common_constant_types: Constants would be wrapped by VariableBuilder.wrap_literal as ConstantVariable. * is_safe_constant: Constants can be loaded by LOAD_CONST bytecode. """ def is_safe_constant(v): if istype(v, (tuple, frozenset)): return all(map(is_safe_constant, v)) return isinstance( v, ( enum.Enum, type, torch.Size, typing._GenericAlias, # type: ignore[attr-defined] types.GenericAlias, ), ) or istype( v, common_constant_types | {slice}, ) @functools.lru_cache(None) def common_constants(): return { # We zero-one specialize shapes, so specialize these constants # too 0, 1, } def is_torch_sym(value): return isinstance(value, (torch.SymBool, torch.SymInt)) and not isinstance( value.node, torch.nested._internal.nested_int.NestedIntNode ) def is_int_specialization_case(value, source): from .source import is_from_defaults return not TracingContext.get().force_unspec_int_unbacked_size_like and ( # Assume integers from global variables want to be specialized not source.guard_source().is_local() # Assume that integers that came from NN modules want to be # specialized (as we don't expect users to be changing the # NN modules on the fly), unless explicitly disabled or ( source.guard_source().is_specialized_nn_module() and not config.allow_unspec_int_on_nn_module ) or ( source.guard_source().is_unspecialized_builtin_nn_module() and not config.allow_unspec_int_on_nn_module ) or is_from_defaults(source) # TODO: Delete this condition when rollout is done. NB: this # condition never evaluates True in open source or ( not justknobs_check("pytorch/dynamo:enable_unspecialize_zero_one_plain_int") and value in common_constants() ) ) def specialize_symnode(arg): from .variables import ConstantVariable, LazyVariableTracker, SymNodeVariable # Guard and specialize if isinstance(arg, LazyVariableTracker) and not arg.is_realized(): # Find if the arg would be realized as SymNodeVariable later on. If yes, # realize it and specialize. Else return the arg. source = arg.original_source() value = arg.original_value() is_symnode_vt = is_torch_sym(value) or ( not config.specialize_int and type(value) is int and not is_int_specialization_case(value, source) ) if not is_symnode_vt: return arg if isinstance(arg, SymNodeVariable): return ConstantVariable.create(arg.evaluate_expr()) return arg def guard_if_dyn(arg): from .variables import ConstantVariable arg = specialize_symnode(arg) if isinstance(arg, ConstantVariable): return arg.as_python_constant() return arg def check_constant_args(args, kwargs): return all(x.is_python_constant() for x in itertools.chain(args, kwargs.values())) def check_unspec_python_args(args, kwargs): from .variables.constant import ConstantVariable from .variables.tensor import UnspecializedPythonVariable unspec_count = 0 for x in itertools.chain(args, kwargs.values()): if isinstance(x, UnspecializedPythonVariable): unspec_count += 1 elif not isinstance(x, ConstantVariable): return False return unspec_count > 0 def check_unspec_or_constant_args(args, kwargs): # A fused version of: # return check_constant_args(args, kwargs) or check_unspec_python_args(args, kwargs) from .variables.tensor import UnspecializedPythonVariable for x in itertools.chain(args, kwargs.values()): if not (x.is_python_constant() or isinstance(x, UnspecializedPythonVariable)): return False return True def check_numpy_ndarray_args(args, kwargs): from .variables.tensor import NumpyNdarrayVariable return any( isinstance(x, NumpyNdarrayVariable) for x in itertools.chain(args, kwargs.values()) ) dict_keys: type[KeysView[Any]] = type({}.keys()) dict_values: type[ValuesView[Any]] = type({}.values()) odict_values: type[ValuesView[Any]] = type(OrderedDict().values()) tuple_iterator: type[Iterator[Any]] = type(iter(())) range_iterator: type[Iterator[Any]] = type(iter(range(0))) tuple_iterator_len = tuple_iterator.__length_hint__ # type: ignore[attr-defined] object_new = object.__new__ dict_new = dict.__new__ dict_methods = { method for method in itertools.chain(dict.__dict__.values(), OrderedDict.__dict__.values()) if callable(method) } tuple_new = tuple.__new__ tuple_methods = {method for method in tuple.__dict__.values() if callable(method)} list_methods = {method for method in list.__dict__.values() if callable(method)} list_getitem = list.__getitem__ str_methods = {method for method in str.__dict__.values() if callable(method)} def builtin_dict_keys(d): # Avoids overridden keys method of the dictionary assert isinstance(d, dict) return dict.keys(d) def get_items_from_dict(obj): # Get items without calling the user defined __getitem__ or keys method. assert isinstance(obj, dict) if istype(obj, (dict, OrderedDict)): return obj.items() elif isinstance(obj, OrderedDict): return [(k, OrderedDict.__getitem__(obj, k)) for k in OrderedDict.keys(obj)] else: return [(k, dict.__getitem__(obj, k)) for k in dict.keys(obj)] def nn_module_new(cls): obj = object_new(cls) torch.nn.Module.__init__(obj) return obj def product(it): return functools.reduce(operator.mul, it, 1) def tuple_iterator_getitem(it, index): _, (obj,), start = it.__reduce__() return obj[start + index] iter_next = next def normalize_range_iter(range_iter) -> tuple[int, int, int]: _, (range_obj,), maybe_idx = range_iter.__reduce__() # In 3.12+, `maybe_idx` could be None, and `range_obj.start` would've been # already incremented by the current index. start = range_obj.start + (maybe_idx or 0) stop = range_obj.stop step = range_obj.step return (start, stop, step) def to_subclass(t, cls): return t.as_subclass(cls) dict_getitem = dict.__getitem__ def dict_keys_getitem(d, n): # Call dict(d) to prevent calling overridden __iter__/keys dict_class = dict if isinstance(d, OrderedDict): dict_class = OrderedDict return next(itertools.islice(dict_class.keys(d), n, n + 1)) def enum_repr(value, local): # enum class can override __str__ method. Use __class__ and name attribute # to extract the class name and key name. name = value.__class__.__name__ val = value.name scope = "L" if local else "G" local_name = f'{scope}["{name}"].{val}' return local_name def set_example_value(node, example_value): # NB: example_value is a bit of a misnomer, because this is always a fake # tensor of some sort. Furthermore, these example values serve as the # runtime state of Dynamo tracing, which means if metadata mutation # occurs, the example_value gets directly updated (so you can't rely on # this to accurately reflect what the state of the value was at the time # the program was traced). node.meta["example_value"] = example_value shape_env = TracingContext.get().fake_mode.shape_env if ( symbol_to_path := torch.fx.experimental.symbolic_shapes.compute_unbacked_bindings( shape_env, example_value ) ): node.meta["unbacked_bindings"] = symbol_to_path def _get_fake_tensor(vt): fake_tensor = vt.as_proxy().node.meta.get("example_value") if not is_fake(fake_tensor): from . import graph_break_hints from .exc import unimplemented_v2 unimplemented_v2( gb_type="Cannot check Tensor object identity without its fake value", context=str(fake_tensor), explanation="TensorVariable is missing a fake example_value.", hints=[*graph_break_hints.DYNAMO_BUG], ) return fake_tensor def iter_contains(items, search, tx, check_tensor_identity=False): from .variables import ( BuiltinVariable, ConstantVariable, TensorVariable, VariableTracker, ) if search.is_python_constant(): found_const = any( x.is_python_constant() and x.as_python_constant() == search.as_python_constant() for x in items ) return ConstantVariable.create(found_const) must_check_tensor_id = False if check_tensor_identity and isinstance(search, TensorVariable): must_check_tensor_id = True # Match of Tensor means match of FakeTensor search = _get_fake_tensor(search) found: Optional[VariableTracker] = None for x in items: if must_check_tensor_id: if isinstance(x, TensorVariable): if search is _get_fake_tensor(x): # Object equivalence return ConstantVariable.create(True) else: check = BuiltinVariable(operator.eq).call_function(tx, [x, search], {}) if found is None: found = check else: found = BuiltinVariable(operator.or_).call_function( tx, [check, found], {} ) if found is None: found = ConstantVariable.create(False) return found def key_is_id(k): """Returns whether it indexes dictionaries using its id""" return isinstance(k, (torch.Tensor, torch.nn.Module, MethodWrapperType)) def key_to_id(value): return [id(k) if key_is_id(k) else k for k in value.keys()] def const_repr(x, *, local) -> str: from .trace_rules import is_builtin_callable if isinstance(x, (list, tuple)): elems_repr = ",".join(const_repr(s, local=local) for s in x) if isinstance(x, list): return f"[{elems_repr}]" else: assert isinstance(x, tuple) if len(x) == 1: return f"({elems_repr},)" else: return f"({elems_repr})" elif isinstance(x, enum.Enum): # To workaround repr(Enum) returning invalid global reference before python 3.11 # by calling enum_repr and removing quotes to render enum in guard code. return enum_repr(x, local=local).replace("'", "") elif is_builtin_callable(x): return x.__name__ elif isinstance(x, type): def fullname(o): klass = o.__class__ module = klass.__module__ if module == "builtins": return klass.__qualname__ # avoid outputs like 'builtins.str' return module + "." + klass.__qualname__ return fullname(x) else: return f"{x!r}" def dict_keys_repr(const_keys, *, local) -> str: keys_str = ",".join(const_repr(s, local=local) for s in const_keys) return "[" + keys_str + "]" GLOBAL_KEY_PREFIX = "__dict_key" from torch._subclasses import UnsupportedFakeTensorException # noqa: F401 def get_safe_global_name(tx, root, obj): # The global_mangled_class_name should be different for different # invocations of torch.compile. Otherwise, we can run into a situation # where multiple torch.compile invocations re-use the same global name, # but the global's lifetime is tied to the first invocation (and # may be deleted when the first torch.compile invocation is deleted) # We mangle it based off of the output_graph's id. return f"{root}_{id(obj)}_c{tx.output.compile_id}" def is_in(item: Any, *containers) -> bool: for container in containers: if item in container: return True return False def get_unique_name_wrt(prefix: str, *containers, requires_suffix=False) -> str: """ Return a name that starts with `prefix` and is not in any of the `containers` (e.g., map, set). """ if not requires_suffix and not is_in(prefix, *containers): return prefix for i in itertools.count(): candidate = f"{prefix}_{i}" if not is_in(candidate, *containers): return candidate raise AssertionError("unreachable") def wrap_fake_exception(fn): try: return fn() except UnsupportedFakeTensorException as e: from .exc import unimplemented_v2 msg = f"Encountered exception ({e.reason}) during fake tensor propagation." log.warning(msg) unimplemented_v2( gb_type="Fake tensor propagation exception", context=str(e.reason), explanation=msg, hints=[], from_exc=e, ) def deepcopy_to_fake_tensor(obj, fake_mode): with torch._subclasses.fake_tensor.FakeCopyMode(fake_mode): return wrap_fake_exception(lambda: copy.deepcopy(obj)) def rmse(ref, res): """ Calculate root mean squared error """ return torch.sqrt(torch.mean(torch.square(ref - res))) def same( ref, res, fp64_ref=None, cos_similarity=False, tol=1e-4, equal_nan=False, exact_dtype=True, relax_numpy_equality=False, ignore_non_fp=False, log_error=log.error, use_larger_multiplier_for_smaller_tensor=False, force_max_multiplier: bool = False, ): """Check correctness to see if ref and res match""" if fp64_ref is None: fp64_ref = ref if isinstance( ref, (list, tuple, collections.deque, torch.nn.ParameterList, torch.Size) ): assert isinstance(res, (list, tuple, collections.deque)), ( f"type mismatch {type(ref)} {type(res)}" ) if len(ref) != len(res): log_error("Length mismatch") return False return len(ref) == len(res) and all( same( ai, bi, fp64_refi, cos_similarity, tol, equal_nan, exact_dtype, relax_numpy_equality, ignore_non_fp, log_error=log_error, use_larger_multiplier_for_smaller_tensor=use_larger_multiplier_for_smaller_tensor, force_max_multiplier=force_max_multiplier, ) for ai, bi, fp64_refi in zip(ref, res, fp64_ref) ) elif type(ref).__name__ == "QuestionAnsweringModelOutput": # This skips checking accuracy for start_logits/end_logits. # Tentatively, start_logits/end_logits appear to be very prone to # inaccuracies and is somewhat subsumed by checking the loss. return same( ref.loss, res.loss, fp64_ref.loss, cos_similarity, tol, equal_nan, exact_dtype, relax_numpy_equality, ignore_non_fp, log_error=log_error, use_larger_multiplier_for_smaller_tensor=use_larger_multiplier_for_smaller_tensor, force_max_multiplier=force_max_multiplier, ) elif isinstance(ref, dict): assert isinstance(res, dict) assert set(ref.keys()) == set(res.keys()), ( f"keys mismatch {set(ref.keys())} == {set(res.keys())}" ) for k in sorted(ref.keys()): if not ( same( ref[k], res[k], fp64_ref[k], cos_similarity=cos_similarity, tol=tol, equal_nan=equal_nan, exact_dtype=exact_dtype, relax_numpy_equality=relax_numpy_equality, ignore_non_fp=ignore_non_fp, log_error=log_error, use_larger_multiplier_for_smaller_tensor=use_larger_multiplier_for_smaller_tensor, force_max_multiplier=force_max_multiplier, ) ): log_error("Accuracy failed for key name %s", k) return False return True elif isinstance(ref, set): assert isinstance(res, set) assert set(ref) == set(res), f"elements mismatch {set(ref)} == {set(res)}" return True elif isinstance(ref, (torch.Tensor, float)): assert not isinstance(ref, torch._subclasses.FakeTensor) assert not isinstance(res, torch._subclasses.FakeTensor) def to_tensor(t): return t if isinstance(t, torch.Tensor) else torch.tensor(t) ref, res, fp64_ref = (to_tensor(val) for val in (ref, res, fp64_ref)) if ref.is_sparse: assert res.is_sparse ref = ref.to_dense() res = res.to_dense() assert isinstance(res, torch.Tensor), f"type mismatch {type(ref)} {type(res)}" if exact_dtype: if ref.dtype != res.dtype: log_error("dtype mismatch %s, %s", ref.dtype, res.dtype) return False if ref.dtype == torch.bool: if ignore_non_fp: return True # triton stores bool as int8, so add this for more accurate checking r = torch.allclose( ref.to(dtype=torch.uint8), res.to(dtype=torch.uint8), atol=tol, rtol=tol, equal_nan=equal_nan, ) if not r: log_error("Accuracy failed: uint8 tensor did not match") return r if cos_similarity: ref = ref.flatten().to(torch.float32) res = res.flatten().to(torch.float32) if torch.allclose(ref, res, atol=tol, rtol=tol, equal_nan=True): # early exit that handles zero/nan better # cosine_similarity(zeros(10), zeros(10), dim=0) is 0 return True score = torch.nn.functional.cosine_similarity(ref, res, dim=0, eps=1e-6) if score < 0.99: log.warning("Similarity score=%s", score.detach().cpu().item()) return score >= 0.99 else: if not exact_dtype: ref = ref.to(res.dtype) # First try usual allclose if torch.allclose(ref, res, atol=tol, rtol=tol, equal_nan=equal_nan): return True # Check error from fp64 version if fp64_ref.dtype == torch.float64: # Fix a corner case that res and fp64_ref does not contains NaN and match (with loose tolerance) # while the ref contains NaN. In this case, RMSE should not match any ways. # But res is 'BETTER' than ref so we count it pass. # # This happens for Super_SloMo when loop ordering after fusion is enabled: # https://gist.github.com/shunting314/11f235c70f7db0d52718d26f4a701cab loose_tol = 1e-2 * 4 if ( not fp64_ref.isnan().any() and not res.isnan().any() and ref.isnan().any() and torch.allclose( fp64_ref.to(dtype=res.dtype), res, atol=loose_tol, rtol=loose_tol, equal_nan=equal_nan, ) ): return True ref_error = rmse(fp64_ref, ref).item() # ref unable to produce this with stable numerics in this precision, ignore if math.isnan(ref_error): log.warning( "Found nan in reference. Consider running in higher precision." ) res_error = rmse(fp64_ref, res).item() def get_multiplier(): # In some particular cases, we expect high difference in results. # At the moment one of this cases is inductor freezing bfloat16 convolution const folding. # In case of it the res_error is at least one order of magnitude higher. if force_max_multiplier: return 10.0 # In the case of using AMP (Automatic Mixed Precision), certain models have # failed the benchmark's correctness check. However, the end-to-end model's # accuracy when comparing AMP with FP32 is within a difference of less than 0.1%. # Thus, it's possible that the correctness check failures for these models are # false alarms. We use multiplier of 3 instead of 2 to avoid these false alarms. multiplier = ( 3.0 if res.dtype in (torch.float16, torch.bfloat16) else 2.0 ) if use_larger_multiplier_for_smaller_tensor and ( fp64_ref.numel() <= 10 ): multiplier = 10.0 elif use_larger_multiplier_for_smaller_tensor and ( fp64_ref.numel() <= 500 ): multiplier = 5.0 elif ( fp64_ref.numel() < 1000 or (ref.ndim == 4 and ref.shape[-1] == ref.shape[-2] == 1) # large tol means a benchmark has been specified as REQUIRE_HIGHER_TOLERANCE or tol >= 2 * 1e-2 ): # In the presence of noise, noise might dominate our error # metric for smaller tensors. # Similary, for 1x1 kernels, there seems to be high noise with amp. multiplier = 3.0 return multiplier multiplier = get_multiplier() passes_test = res_error <= (multiplier * ref_error + tol / 10.0) if ( not passes_test and equal_nan and math.isnan(ref_error) and math.isnan(res_error) # Some unit test for the accuracy minifier relies on # returning false in this case. and not torch._inductor.config.cpp.inject_relu_bug_TESTING_ONLY ): passes_test = True if not passes_test: log_error( "RMSE (res-fp64): %.5f, (ref-fp64): %.5f and shape=%s. res.dtype: %s, multiplier: %f, tol: %f" ", use_larger_multiplier_for_smaller_tensor: %d", res_error, ref_error, res.size(), res.dtype, multiplier, tol, use_larger_multiplier_for_smaller_tensor, ) return passes_test if ignore_non_fp: return True log_error("Accuracy failed: allclose not within tol=%s", tol) return False elif isinstance(ref, (str, int, type(None), bool, torch.device)): if ignore_non_fp: return True r = ref == res if not r: log_error("Accuracy failed (%s): %s != %s", type(ref), ref, res) return r elif is_numpy_int_type(ref) or is_numpy_float_type(ref): if relax_numpy_equality and not ( is_numpy_int_type(res) or is_numpy_float_type(res) ): ref = ref.item() r = (type(ref) is type(res)) and (ref == res) if not r: log_error("Accuracy failed (numpy): %s != %s", ref, res) return r elif is_numpy_ndarray(ref): return (type(ref) is type(res)) and same( torch.as_tensor(ref), torch.as_tensor(res), fp64_ref, cos_similarity=cos_similarity, tol=tol, equal_nan=equal_nan, exact_dtype=exact_dtype, relax_numpy_equality=relax_numpy_equality, ignore_non_fp=ignore_non_fp, log_error=log_error, use_larger_multiplier_for_smaller_tensor=use_larger_multiplier_for_smaller_tensor, ) elif type(ref).__name__ in ( "MaskedLMOutput", "Seq2SeqLMOutput", "CausalLMOutputWithCrossAttentions", "LongformerMaskedLMOutput", "Instances", "SquashedNormal", "Boxes", "Normal", "TanhTransform", "Foo", "Variable", ): assert type(ref) is type(res) return all( same( getattr(ref, key), getattr(res, key), getattr(fp64_ref, key), cos_similarity=cos_similarity, tol=tol, equal_nan=equal_nan, exact_dtype=exact_dtype, relax_numpy_equality=relax_numpy_equality, ignore_non_fp=ignore_non_fp, log_error=log_error, use_larger_multiplier_for_smaller_tensor=use_larger_multiplier_for_smaller_tensor, ) for key in ref.__dict__.keys() ) else: raise RuntimeError(f"unsupported type: {type(ref).__name__}") def format_func_info(code): short_filename = code.co_filename.split("/")[-1] return f"'{code.co_name}' ({short_filename}:{code.co_firstlineno})" @contextlib.contextmanager def disable_cache_limit(): prior = config.recompile_limit config.recompile_limit = sys.maxsize prior_acc_limit = config.accumulated_recompile_limit config.accumulated_recompile_limit = sys.maxsize try: yield finally: config.recompile_limit = prior config.accumulated_recompile_limit = prior_acc_limit # map from transformed code back to original user code orig_code_map = ExactWeakKeyDictionary() # keep a record of code_obj -> list of guard failure reasons for logging guard_failures: collections.defaultdict[Any, list[Any]] = collections.defaultdict(list) # Keep a record of graph break reasons for logging graph_break_reasons: list[torch._dynamo.output_graph.GraphCompileReason] = [] # keep record of compiled code, if we are in "error if recompile" # to track code that dynamo has compiled previously seen_code_map = ExactWeakKeyDictionary() # return same dir unless user changes config between calls @functools.lru_cache(None) def _get_debug_dir(root_dir): dir_name = ( "run_" + datetime.datetime.now().strftime("%Y_%m_%d_%H_%M_%S_%f") # use pid to avoid conflicts among ranks + "-pid_" + str(os.getpid()) ) return os.path.join(root_dir, dir_name) def get_debug_dir(): debug_root = config.debug_dir_root return _get_debug_dir(debug_root) def extract_fake_example_value(node, required=True): if "example_value" in node.meta and is_fake(node.meta["example_value"]): return node.meta["example_value"] elif required: from torch._dynamo.exc import unimplemented_v2 from . import graph_break_hints unimplemented_v2( gb_type="Missing FakeTensor example value", context=str(node), explanation=f"`FakeTensor` example value was required for {node} but not available.", hints=[*graph_break_hints.DYNAMO_BUG], ) else: return None def ensure_graph_fake(e, tx): assert maybe_get_fake_mode(e) is tx.fake_mode return e def get_fake_values_from_nodes(tx, nodes, allow_non_graph_fake): def visit(n: torch.fx.Node): if n.op == "call_function" and "example_value" not in n.meta: # fake tensor validity is checked inside get_fake_value using # ensure_graph_fake return get_fake_value(n, tx, allow_non_graph_fake) elif n.op == "get_attr" and "example_value" not in n.meta: assert n.target in tx.output.nn_modules gm = tx.output.nn_modules[n.target] assert isinstance(gm, torch.fx.GraphModule) return gm out = n.meta["example_value"] if not allow_non_graph_fake and isinstance(out, torch.Tensor): return ensure_graph_fake(out, tx) return out return torch.fx.node.map_arg(nodes, visit) def get_fake_value(node, tx, allow_non_graph_fake=False): """ Run the computation represented by `node` using fake tensors and return the result. allow_non_graph_fake: whether to allow the return result to be: 1. non-fake or 2. fake that is not created by this instance of Dynamo. If `True`, you must be prepared to deal with such return values, ideally by further wrapping them as this graph's fakes. """ from torch.utils._sympy.value_ranges import ValueRangeError from .exc import ( TorchRuntimeError, unimplemented_v2, Unsupported, UserError, UserErrorType, ) op = node.op # FX Node should always return the same fake value if "example_value" in node.meta and is_fake(node.meta["example_value"]): return node.meta["example_value"] args, kwargs = get_fake_values_from_nodes( tx, (node.args, node.kwargs), allow_non_graph_fake ) nnmodule = None if op == "call_method" and len(args) > 0 and isinstance(args[0], torch.nn.Module): # If the first argument is nn.Module, should copy to fake mode. args = (deepcopy_to_fake_tensor(args[0], tx.fake_mode),) + tuple(args[1:]) if op == "call_module": nnmodule = tx.output.nn_modules[node.target] if is_lazy_module(nnmodule) and hasattr(nnmodule, "_initialize_hook"): # In the case of a lazy module, we want to run # the pre-hooks which initialize it. # Afterwards, lazy module deletes its pre-hooks # to avoid treating it as lazy on subsequent recompile. nnmodule._infer_parameters(nnmodule, args) # no matter it's lazy module or not, we should copy to fake mode. nnmodule = deepcopy_to_fake_tensor(nnmodule, tx.fake_mode) if node.name in ["interpolate", "is_integer", "wrapped_gradient"] or any( isinstance(a, complex) for a in args ): # We need to specialize symfloats for now. Eventually we should do a tensorify pass in dynamo. args = tuple( float(arg) if isinstance(arg, torch.SymFloat) and arg.node.hint is not None else arg for arg in args ) try: with tx.fake_mode, enable_python_dispatcher(): ret_val = wrap_fake_exception( lambda: run_node(tx.output, node, args, kwargs, nnmodule) ) except Unsupported: raise except RuntimeError as e: cause: BaseException = e if e.__cause__ is not None: cause = e.__cause__ if isinstance( cause, torch._subclasses.fake_tensor.DataDependentOutputException ): # capture_scalar_outputs only works for these ops right now # see torch/_subclasses/fake_impls.py if cause.func in ( torch.ops.aten.item.default, torch.ops.aten._local_scalar_dense.default, ): # does this actually get triggered? hints = [ "Enable tracing of data-dependent output operators with " "`torch._dynamo.config.capture_scalar_outputs = True`", ] else: hints = [ "Consider wrapping the operator into a PyTorch-understood custom operator " "(see https://pytorch.org/tutorials/advanced/custom_ops_landing_page.html)", ] unimplemented_v2( gb_type="Data dependent operator", context=str(cause.func), explanation=f"Operator `{cause.func}` has a non-Tensor output " "whose value is dependent on the data of Tensor inputs.", hints=hints, ) elif isinstance( cause, torch._subclasses.fake_tensor.DynamicOutputShapeException ): if not torch._dynamo.config.capture_dynamic_output_shape_ops: unimplemented_v2( gb_type="Dynamic shape operator", context=str(cause.func), explanation=f"Operator `{cause.func}`'s output shape depends on input Tensor data.", hints=[ "Enable tracing of dynamic shape operators with " "`torch._dynamo.config.capture_dynamic_output_shape_ops = True`", ], ) else: unimplemented_v2( gb_type="Dynamic shape operator (no meta kernel)", context=str(cause.func), explanation=f"Operator `{cause.func}` does not have a meta kernel that supports dynamic output shapes", hints=[ "Please report an issue to PyTorch", ], ) elif isinstance( cause, torch._subclasses.fake_tensor.UnsupportedOperatorException ): op = cause.func import_suggestion = "" if isinstance(op, torch._ops.OpOverload): maybe_pystub = torch._C._dispatch_pystub( op._schema.name, op._schema.overload_name ) if maybe_pystub is not None: module, ctx = maybe_pystub import_suggestion = ( f"It's possible that the support was implemented in " f"module `{module}` and you may need to `import {module}`" f"({ctx}), otherwise " ) unimplemented_v2( gb_type="Operator does not support running with fake tensors", context=f"unsupported operator: {cause.func}", explanation="", hints=[ f"{import_suggestion}see " "https://docs.google.com/document/d/1GgvOe7C8_NVOMLOCwDaYV1mXXyHMXY7ExoewHqooxrs/edit#heading=h.64r4npvq0w0" " for how to fix", ], ) elif isinstance( cause, torch.fx.experimental.symbolic_shapes.GuardOnDataDependentSymNode ): raise UserError( # noqa: B904 UserErrorType.CONSTRAINT_VIOLATION, str(cause), case_name="constrain_as_size_example", ) elif isinstance(cause, ValueRangeError): raise UserError(UserErrorType.CONSTRAINT_VIOLATION, e.args[0]) from e elif isinstance(cause, TypeError) and "argument" in str(cause): unimplemented_v2( gb_type="TypeError when making fake tensor call", context=f"TypeError {node.target}: {cause}", explanation="", hints=[], ) raise TorchRuntimeError(str(e)).with_traceback(e.__traceback__) from None if not allow_non_graph_fake: _ = pytree.tree_map_only( torch.Tensor, functools.partial(ensure_graph_fake, tx=tx), ret_val ) return ret_val _current_node = threading.local() def get_current_node(): return getattr(_current_node, "value", None) @contextmanager def set_current_node(node): old = get_current_node() _current_node.value = node try: yield finally: _current_node.value = old def run_node(tracer, node, args, kwargs, nnmodule): """ Runs a given node, with the given args and kwargs. Behavior is dictated by a node's op. run_node is useful for extracting real values out of nodes. See get_real_value for more info on common usage. Note: The tracer arg is only used for 'get_attr' ops Note: The nnmodule arg is only used for 'call_module' ops Nodes that are not call_function, call_method, call_module, or get_attr will raise an AssertionError. """ op = node.op with set_current_node(node): def make_error_message(e): return ( f"Dynamo failed to run FX node with fake tensors: {op} {node.target}(*{args}, **{kwargs}): got " + repr(e) ) from .exc import Unsupported try: if op == "call_function": return node.target(*args, **kwargs) elif op == "call_method": if not hasattr(args[0], node.target): from .exc import unimplemented_v2 unimplemented_v2( gb_type="Missing attribute when running call_method node", context="", explanation=make_error_message("attribute not defined"), hints=[], ) return getattr(args[0], node.target)(*args[1:], **kwargs) elif op == "call_module": assert nnmodule is not None return nnmodule(*args, **kwargs) elif op == "get_attr": return tracer.output_graph.get_submodule(node.target) elif op == "placeholder": assert "example_value" in node.meta return node.meta["example_value"] except (NotImplementedError, UnsupportedFakeTensorException) as e: # NB: mimic how wrap_fake_exception does it from .exc import unimplemented_v2 hints = [] if isinstance(e, NotImplementedError): hints = [ "If the op is a PyTorch op, please file an issue to PyTorch.", ] unimplemented_v2( gb_type="NotImplementedError/UnsupportedFakeTensorException when running FX node", context="", explanation=make_error_message(e), hints=hints, from_exc=e, ) except Unsupported: raise except Exception as e: raise RuntimeError(make_error_message(e)).with_traceback( e.__traceback__ ) from e raise AssertionError(op) def get_real_value(node, tracer): """ Run the actual computation represented by `node` and return the result. This will execute any dependent nodes in the graph as well. """ from .exc import TorchRuntimeError cache = tracer.real_value_cache if node in cache: return cache[node] op = node.op args, kwargs = torch.fx.node.map_arg( # type: ignore[misc] (node.args, node.kwargs), lambda n: get_real_value(n, tracer), ) if op == "placeholder" and "grapharg" in node.meta: return node.meta["grapharg"].example if op == "call_module": nn_module = tracer.output_graph.nn_modules[node.target] if not is_lazy_module(nn_module): nn_module = copy.deepcopy(nn_module) else: # In the case of a lazy module, we want to run # the pre-hooks which initialize it nn_module(*args, **kwargs) else: nn_module = None try: real_value = run_node(tracer, node, args, kwargs, nn_module) cache[node] = real_value except RuntimeError as e: raise TorchRuntimeError(str(e)).with_traceback(e.__traceback__) from None return real_value def assert_no_fake_params_or_buffers(gm): from torch._subclasses.fake_tensor import FakeTensorConfig, is_fake def stack_or_hint(t): if FakeTensorConfig.debug: import traceback return f"FAKE TENSOR CREATION TRACEBACK: \n {traceback.format_list(t._debug_trace)}" else: return "Enable TORCH_FAKE_TENSOR_DEBUG=1 to get creation stack traces on fake tensors." for name, buffer in gm.named_buffers(): assert not is_fake(buffer), ( f"Unexpected fake buffer {name} {stack_or_hint(buffer)}" ) for name, param in gm.named_parameters(): assert not is_fake(param), ( f"Unexpected fake param {name} {stack_or_hint(param)}" ) def fqn(obj: Any): """ Returns the fully qualified name of the object. """ return f"{obj.__module__}.{obj.__qualname__}" def ifdynstaticdefault(count1, count2): if torch._dynamo.config.assume_static_by_default: return count1 else: return count2 def import_submodule(mod: types.ModuleType): """ Ensure all the files in a given submodule are imported """ for filename in sorted(os.listdir(os.path.dirname(cast(str, mod.__file__)))): if filename.endswith(".py") and filename[0] != "_": importlib.import_module(f"{mod.__name__}.{filename[:-3]}") def object_has_getattribute(value: Any): return class_has_getattribute(type(value)) def class_has_getattribute(cls: type): try: if isinstance( inspect.getattr_static(cls, "__getattribute__"), types.FunctionType, ): return True except AttributeError: pass return False def get_custom_getattr(value: Any, ignore_nn_module_getattr: bool = False): try: getattr_fn = inspect.getattr_static(type(value), "__getattr__") except AttributeError: getattr_fn = None if ignore_nn_module_getattr and getattr_fn is torch.nn.Module.__getattr__: # ignore this case of getattr getattr_fn = None return getattr_fn class TensorStaticReason(enum.Enum): PARAMETER = 2 NOT_TENSOR = 4 NN_MODULE_PROPERTY = 5 def tensor_static_reason_to_message(reason: TensorStaticReason): if reason == TensorStaticReason.PARAMETER: return "mark_dynamic on parameter, parameters are always static today." if reason == TensorStaticReason.NOT_TENSOR: return "mark_dynamic on a non tensor, how did this happen?" if reason == TensorStaticReason.NN_MODULE_PROPERTY: return "tensor is static because it is nn module associated." raise AssertionError(f"Illegal reason {reason}") def tensor_always_has_static_shape( tensor: Union[torch.Tensor, Any], is_tensor: bool, tensor_source: Source, ) -> tuple[bool, Optional[TensorStaticReason]]: """ Given a tensor, source, and is_tensor flag, determine if a shape should be static. Args: tensor - the real tensor to evaluate, parameters force a static shape. is_tensor - internal dynamo check, essentially "is_tensor": target_cls is TensorVariable, tensors not in a TensorVariable for whatever reason are forced static. Returns a tuple, where the first element is the bool of whether or not this tensor should have a static shape. The second element is a TensorStaticReason, useful for passing to tensor_static_reason_to_message if needed. """ from .source import is_from_unspecialized_param_buffer_source if ( tensor_source.guard_source().is_specialized_nn_module() or tensor_source.guard_source().is_unspecialized_builtin_nn_module() ) and config.force_nn_module_property_static_shapes: return True, TensorStaticReason.NN_MODULE_PROPERTY if ( type(tensor) is torch.nn.Parameter or is_from_unspecialized_param_buffer_source(tensor_source) ) and config.force_parameter_static_shapes: return True, TensorStaticReason.PARAMETER if not is_tensor: return True, TensorStaticReason.NOT_TENSOR return False, None def lazy_format_graph_tabular(fn_name, gm): def inner(): try: from tabulate import tabulate # TODO: Check that this is installed except ImportError: return ( "Tabulate module missing, please install tabulate to log the graph in tabular format, logging code instead:\n" + str(lazy_format_graph_code(fn_name, gm)) ) node_specs = [ [n.op, n.name, n.target, n.args, n.kwargs] for n in gm.graph.nodes ] graph_str = tabulate( node_specs, headers=["opcode", "name", "target", "args", "kwargs"] ) return _format_graph_code(fn_name, gm.forward.__code__.co_filename, graph_str) return LazyString(inner) def format_bytecode(prefix, name, filename, line_no, code): return f"{prefix} {name} {filename} line {line_no} \n{dis.Bytecode(code).dis()}\n" forward_hook_names = ["_forward_pre_hooks", "_forward_hooks"] backward_hook_names = ["_backward_pre_hooks", "_backward_hooks"] state_dict_hook_names = [ "_state_dict_pre_hooks", "_state_dict_hooks", "_load_state_dict_pre_hooks", "_load_state_dict_post_hooks", ] all_hook_names = forward_hook_names + backward_hook_names + state_dict_hook_names def nn_module_has_global_hooks(): # This is limited to backward hooks for now because NNModuleVariable # supports fwd hooks underneath. return len(torch.nn.modules.module._global_backward_hooks) or len( torch.nn.modules.module._global_backward_pre_hooks ) def nn_module_get_all_hooks( mod, check_forward_hooks=False, check_backward_hooks=False, check_state_dict_hooks=False, ): """ Sometimes its useful to differentiate between types of hooks such as forward/backward/pre hooks executed during module.__call__, and state_dict hooks which are executed separately. """ hook_dicts_to_check = [] check_all_hooks = ( not check_forward_hooks and not check_backward_hooks and not check_state_dict_hooks ) if check_forward_hooks or check_all_hooks: hook_dicts_to_check.extend(forward_hook_names) if check_backward_hooks or check_all_hooks: hook_dicts_to_check.extend(backward_hook_names) if check_state_dict_hooks: hook_dicts_to_check.extend(state_dict_hook_names) all_hooks = [] for hook_dict_name in hook_dicts_to_check: hooks = getattr(mod, hook_dict_name, []) for hook_name in hooks: hook = hooks[hook_name] all_hooks.append(hook) return all_hooks def nnmodule_has_hooks( mod, check_forward_hooks=False, check_backward_hooks=False, check_state_dict_hooks=False, ): """ Helper function to check if a module has any hooks attached to it. """ hooks = nn_module_get_all_hooks( mod, check_forward_hooks=check_forward_hooks, check_backward_hooks=check_backward_hooks, check_state_dict_hooks=check_state_dict_hooks, ) return bool(hooks) def to_numpy_helper(value): """Convert tensor and tnp.ndarray to numpy.ndarray.""" if is_fake(value): return value if isinstance(value, tnp.ndarray): return to_numpy_helper(value.tensor) elif isinstance(value, torch.Tensor): return value.numpy(force=True) elif isinstance(value, (tuple, list)): return type(value)(to_numpy_helper(obj) for obj in value) else: return value def numpy_to_tensor(value): """Convert tnp.ndarray to tensor, leave other types intact. If a list/tuple, loop through it to convert.""" assert np is not None if isinstance(value, np.ndarray): return torch.as_tensor(value) if isinstance(value, tnp.ndarray): return value.tensor elif isinstance(value, (tuple, list)): return type(value)(numpy_to_tensor(obj) for obj in value) else: return value class numpy_to_tensor_wrapper: def __init__(self, f): self.f = f self.__name__ = "wrapped_" + self.f.__name__ def __repr__(self) -> str: return f">" def __call__(self, *args, **kwargs): out = self.f(*args, **kwargs) return numpy_to_tensor(out) def numpy_attr_wrapper(obj, name): if isinstance(obj, tnp.ndarray): out = getattr(obj, name) return numpy_to_tensor(out) elif isinstance(obj, torch.Tensor): out = getattr(tnp.ndarray(obj), name) return numpy_to_tensor(out) class numpy_method_wrapper: """Convert obj from torch.Tensor to tnp.ndarray and call method. Then convert result back to torch.Tensor.""" def __init__(self, method: str): self.method = method self.__name__ = "wrapped_" + self.method def __repr__(self) -> str: return f">" def __call__(self, *args, **kwargs): obj = args[0] if isinstance(obj, torch.Tensor): obj = tnp.ndarray(obj) method_callable = getattr(obj, self.method) out = method_callable(*args[1:], **kwargs) return numpy_to_tensor(out) class numpy_operator_wrapper: """Implements dunder methods for tnp.ndarray via functions from the operator library""" def __init__(self, op: Callable[..., Any]): self.op = op self.__name__ = f"wrapped_{op.__name__}" def __repr__(self) -> str: return f">" def __call__(self, *args, **kwargs): assert not kwargs args = ( tnp.ndarray(arg) if isinstance(arg, torch.Tensor) else arg for arg in args ) out = self.op(*args) return numpy_to_tensor(out) def defake(x): if not isinstance(x, FakeTensor): return x size: torch._prims_common.ShapeType stride: torch._prims_common.StrideType if x._has_symbolic_sizes_strides: size = [] for s in x.size(): if isinstance(s, torch.SymInt): size.append(s.node.shape_env.size_hint(s.node.expr)) else: size.append(s) stride = [] for s in x.stride(): if isinstance(s, torch.SymInt): stride.append(s.node.shape_env.size_hint(s.node.expr)) else: stride.append(s) else: size = x.size() stride = x.stride() y = torch.empty_strided( size, stride, dtype=x.dtype, device=x.device, requires_grad=x.requires_grad, ) y.zero_() return y def is_utils_checkpoint(obj): # Lazy import to avoid circular dependencies import torch.utils.checkpoint return obj is torch.utils.checkpoint.checkpoint def is_invoke_subgraph(obj): from torch._higher_order_ops.invoke_subgraph import invoke_subgraph_placeholder return obj is invoke_subgraph_placeholder def build_invoke_subgraph_variable(**options): from .variables.higher_order_ops import TorchHigherOrderOperatorVariable return TorchHigherOrderOperatorVariable.make( torch._higher_order_ops.invoke_subgraph, **options, ) def build_checkpoint_variable(**options): import torch._higher_order_ops.wrap as higher_order_ops from .variables.higher_order_ops import TorchHigherOrderOperatorVariable # TODO - This is a temporary situation where we have two versions of # checkpointing implementation. We will converge on one and remove the other. activation_checkpoint_op: torch._ops.HigherOrderOperator = ( higher_order_ops.tag_activation_checkpoint ) if torch._functorch.config.functionalize_rng_ops: activation_checkpoint_op = higher_order_ops.wrap_activation_checkpoint return TorchHigherOrderOperatorVariable.make( activation_checkpoint_op, **options, ) def is_compile_supported(device_type): from .eval_frame import is_dynamo_supported compile_supported = is_dynamo_supported() if device_type == "cpu": pass elif device_type in ["cuda", "xpu"] and compile_supported: compile_supported = has_triton() else: compile_supported = False return compile_supported # The following 3.11 source code functions are adapted from # https://github.com/python/cpython/blob/v3.11.4/Lib/traceback.py # in order to output source code corresponding to bytecode in 3.11+. # We need our own versions since we want to support multiline expressions. def _fix_offset(str: str, offset: int) -> int: """ Convert byte offset `offset` of `str` into character offset. Byte offset is used for 3.11+ instruction column data. Takes things like unicode characters into consideration. Unchanged from CPython implementation. """ as_utf8 = str.encode("utf-8") return len(as_utf8[:offset].decode("utf-8", errors="replace")) @dataclasses.dataclass class _Anchors: # inclusive left_end_lineno: int left_end_offset: int right_start_lineno: int # exclusive right_start_offset: int def _extract_anchors_from_expr(segment: str) -> Optional[_Anchors]: """ Given source code `segment` corresponding to a bytecode instruction, determine: - for binary ops, the location of the binary op - for indexing, the location of the brackets. `segment` is expected to be a valid Python expression """ assert sys.version_info >= (3, 11) import ast try: # Without brackets, `segment` is parsed as a statement. # We expect an expression, so wrap `segment` in # brackets to handle multi-line expressions. tree = ast.parse("(\n" + segment + "\n)") except SyntaxError: return None if len(tree.body) != 1: return None lines = segment.split("\n") # get character index given byte offset def normalize(lineno, offset): return _fix_offset(lines[lineno], offset) # Gets the next valid character index in `lines`, if # the current location is not valid. Handles empty lines. def next_valid_char(lineno, col): while lineno < len(lines) and col >= len(lines[lineno]): col = 0 lineno += 1 assert lineno < len(lines) and col < len(lines[lineno]) return lineno, col # Get the next valid character index in `lines`. def increment(lineno, col): col += 1 lineno, col = next_valid_char(lineno, col) assert lineno < len(lines) and col < len(lines[lineno]) return lineno, col # Get the next valid character at least on the next line def nextline(lineno, col): col = 0 lineno += 1 lineno, col = next_valid_char(lineno, col) assert lineno < len(lines) and col < len(lines[lineno]) return lineno, col statement = tree.body[0] if isinstance(statement, ast.Expr): expr = statement.value if isinstance(expr, ast.BinOp): # ast gives locations for BinOp subexpressions, e.g. # ( left_expr ) + ( right_expr ) # left^^^^^ right^^^^^ # -2 since end_lineno is 1-indexed and because we added an extra # bracket to `segment` when calling ast.parse cur_lineno = cast(int, expr.left.end_lineno) - 2 cur_col = normalize(cur_lineno, expr.left.end_col_offset) cur_lineno, cur_col = next_valid_char(cur_lineno, cur_col) # Heuristic to find the operator character. # The original CPython implementation did not look for ), \, or #, # leading to incorrect anchor location, e.g. # (x) + (y) # ~~^~~~~~~ while (ch := lines[cur_lineno][cur_col]).isspace() or ch in ")\\#": if ch in "\\#": cur_lineno, cur_col = nextline(cur_lineno, cur_col) else: cur_lineno, cur_col = increment(cur_lineno, cur_col) # binary op is 1 or 2 characters long, on the same line right_col = cur_col + 1 if ( right_col < len(lines[cur_lineno]) and not (ch := lines[cur_lineno][right_col]).isspace() and ch not in "\\#" ): right_col += 1 # right_col can be invalid since it is exclusive return _Anchors(cur_lineno, cur_col, cur_lineno, right_col) elif isinstance(expr, ast.Subscript): # ast gives locations for value and slice subexpressions, e.g. # ( value_expr ) [ slice_expr ] # value^^^^^ slice^^^^^ # subscript^^^^^^^^^^^^^^^^^^^^ # find left bracket (first '[' after value) left_lineno = cast(int, expr.value.end_lineno) - 2 left_col = normalize(left_lineno, expr.value.end_col_offset) left_lineno, left_col = next_valid_char(left_lineno, left_col) while lines[left_lineno][left_col] != "[": left_lineno, left_col = increment(left_lineno, left_col) # find right bracket (final character of expression) right_lineno = cast(int, expr.end_lineno) - 2 right_col = normalize(right_lineno, expr.end_col_offset) return _Anchors(left_lineno, left_col, right_lineno, right_col) elif isinstance(expr, ast.Call): # ( func_expr ) (args, kwargs) # func^^^^^ # call^^^^^^^^^^^^^^^^^^^^^^^^ # find left bracket (first '(' after func) left_lineno = cast(int, expr.func.end_lineno) - 2 left_col = normalize(left_lineno, expr.func.end_col_offset) left_lineno, left_col = next_valid_char(left_lineno, left_col) while lines[left_lineno][left_col] != "(": left_lineno, left_col = increment(left_lineno, left_col) # find right bracket (final character of expression) right_lineno = cast(int, expr.end_lineno) - 2 right_col = normalize(right_lineno, expr.end_col_offset) return _Anchors(left_lineno, left_col, right_lineno, right_col) return None def get_instruction_source_311(code: types.CodeType, inst: dis.Instruction) -> str: """ Python 3.11+ only. Returns lines of source code (from code object `code`) corresponding to `inst`'s location data, and underlines relevant code to `inst`. Example: CALL on `g`: f(g( ^^ h(x))) ^^^^^ We need our own implementation in < 3.13 since `format_frame_summary` in Python's `traceback` module doesn't handle multi-line expressions (and their anchor extraction code is not completely correct). """ if sys.version_info >= (3, 13): # multiline traceback implemented in 3.13+ frame_summary = traceback.FrameSummary( code.co_filename, inst.positions.lineno, code.co_name, end_lineno=inst.positions.end_lineno, colno=inst.positions.col_offset, end_colno=inst.positions.end_col_offset, ) result = traceback.format_list([frame_summary])[0] # remove first line containing filename info result = "\n".join(result.splitlines()[1:]) # indent lines with original indentation orig_lines = [ linecache.getline(code.co_filename, lineno).rstrip() for lineno in range(inst.positions.lineno, inst.positions.end_lineno + 1) ] orig_lines_dedent = textwrap.dedent("\n".join(orig_lines)).splitlines() indent_len = len(orig_lines[0]) - len(orig_lines_dedent[0]) indent = orig_lines[0][:indent_len] result = textwrap.indent(textwrap.dedent(result), indent) return result assert inst.positions is not None if inst.positions.lineno is None: return "" # The rstrip + "\n" pattern is used throughout this function to handle # linecache.getline errors. Error lines are treated as empty strings "", but we want # to treat them as blank lines "\n". first_line = linecache.getline(code.co_filename, inst.positions.lineno).rstrip() if inst.positions.end_lineno is None: return first_line if inst.positions.col_offset is None or inst.positions.end_col_offset is None: return first_line # character index of the start of the instruction start_offset = _fix_offset(first_line, inst.positions.col_offset) # character index of the end of the instruction # compute later since end may be a different line end_offset = None # expression corresponding to the instruction so we can get anchors segment = "" # underline markers to be printed - start with `~` marker and replace with `^` later markers = [] # Compute segment and initial markers if inst.positions.end_lineno == inst.positions.lineno: end_offset = _fix_offset(first_line, inst.positions.end_col_offset) segment = first_line[start_offset:end_offset] markers.append(" " * start_offset + "~" * (end_offset - start_offset)) else: segment = first_line[start_offset:] + "\n" markers.append(" " * start_offset + "~" * (len(first_line) - start_offset)) last_line = linecache.getline( code.co_filename, inst.positions.end_lineno ).rstrip() end_offset = _fix_offset(last_line, inst.positions.end_col_offset) for lineno in range(inst.positions.lineno + 1, inst.positions.end_lineno): line = linecache.getline(code.co_filename, lineno).rstrip() segment += line + "\n" # don't underline leading spaces num_spaces = len(line) - len(line.lstrip()) markers.append(" " * num_spaces + "~" * (len(line) - num_spaces)) segment += last_line[:end_offset] num_spaces = len(last_line) - len(last_line.lstrip()) markers.append(" " * num_spaces + "~" * (end_offset - num_spaces)) anchors: Optional[_Anchors] = None try: anchors = _extract_anchors_from_expr(segment) except AssertionError: pass # replace `~` markers with `^` where necessary if anchors is None: markers = [marker.replace("~", "^") for marker in markers] else: # make markers mutable mutable_markers: list[list[str]] = [list(marker) for marker in markers] # anchor positions do not take start_offset into account if anchors.left_end_lineno == 0: anchors.left_end_offset += start_offset if anchors.right_start_lineno == 0: anchors.right_start_offset += start_offset # Turn `~`` markers between anchors to `^` for lineno in range(len(markers)): for col in range(len(mutable_markers[lineno])): if lineno < anchors.left_end_lineno: continue if lineno == anchors.left_end_lineno and col < anchors.left_end_offset: continue if ( lineno == anchors.right_start_lineno and col >= anchors.right_start_offset ): continue if lineno > anchors.right_start_lineno: continue if mutable_markers[lineno][col] == "~": mutable_markers[lineno][col] = "^" # make markers into strings again markers = ["".join(marker) for marker in mutable_markers] result = "" for i in range(len(markers)): result += ( linecache.getline(code.co_filename, inst.positions.lineno + i).rstrip() + "\n" ) result += markers[i] + "\n" return result def get_static_address_type(t): if isinstance(t, torch.Tensor): return getattr(t, "_dynamo_static_input_type", None) return None def is_rng_state_getter_or_setter(value): getters = ( # The following two functions are not identical, so don't remove anyone! torch._C.Generator.get_state, torch.default_generator.get_state, torch.get_rng_state, torch.cuda.get_rng_state, ) setters = ( torch._C.Generator.set_state, torch.default_generator.set_state, torch.set_rng_state, torch.cuda.set_rng_state, ) return value in (*setters, *getters) def is_tensor_base_attr_getter(value): return ( isinstance(value, types.MethodWrapperType) and value.__name__ == "__get__" and value.__self__.__objclass__ is torch._C._TensorBase # type: ignore[attr-defined] ) def is_torch_function_object(value): return hasattr(value, "__torch_function__") def has_torch_function(vt: torch._dynamo.variables.base.VariableTracker) -> bool: from torch._dynamo.variables import UserDefinedObjectVariable from torch._dynamo.variables.torch_function import TensorWithTFOverrideVariable # Note on lazy vars: The value will either be realized or not throughout the course of execution # if the value has a torch function, it will eventually be realized so we can realize it here # if the value does not have a torch function, it may or may not be realized # if it is realized it will be used and guards will be installed properly # if it is not used, guards won't be installed, and it doesn't matter # if the value has a torch function or not, so we should *not* realize it. # NB: We technically know that if is_realized is False, LazyVariableTracker has the peek_value method # but mypy does not unfortunately if vt.is_realized() or ( hasattr(vt, "peek_value") and hasattr(vt.peek_value(), "__torch_function__") ): if isinstance(vt, TensorWithTFOverrideVariable): return True return isinstance(vt, UserDefinedObjectVariable) and hasattr( vt.value, "__torch_function__" ) return False # see note [Tensor Fakification and Symbol Caching] def to_fake_tensor(t, fake_mode): symbolic_context = None source = None if tracing_context := torch._guards.TracingContext.try_get(): if t in tracing_context.tensor_to_context: symbolic_context = tracing_context.tensor_to_context[t] source = symbolic_context.tensor_source return fake_mode.from_tensor( t, static_shapes=False, symbolic_context=symbolic_context, source=source ) # NB: this works for both classes and instances def is_frozen_dataclass(value): return ( not object_has_getattribute(value) and not class_has_getattribute(value) and is_dataclass(value) and hasattr(value, "__dataclass_params__") and hasattr(value.__dataclass_params__, "frozen") and value.__dataclass_params__.frozen ) def get_first_attr(obj, *attrs): """ Return the first available attribute or throw an exception if none is present. """ for attr in attrs: if hasattr(obj, attr): return getattr(obj, attr) raise AssertionError(f"{obj} does not has any of the attributes: {attrs}") @contextlib.contextmanager def maybe_enable_compiled_autograd(should_enable, fullgraph=True, dynamic=True): if not should_enable: yield else: def compiler_fn(gm): def inner_compiler(gm_, example_inputs_): torch._dynamo.utils.counters["compiled_autograd"]["compiles"] += 1 return torch._inductor.compile(gm_, example_inputs_) return torch.compile( gm, backend=inner_compiler, fullgraph=fullgraph, dynamic=dynamic ) with torch._dynamo.compiled_autograd._enable(compiler_fn) as ctx: yield ctx def invalid_removeable_handle(): # need a subclass so weakref works class Invalid(dict): # type: ignore[type-arg] pass return RemovableHandle(Invalid()) # Returns a "proxy" (new object with the same class and dict) for (non-GraphModule) nn.Module's. # Attribute changes to the original object/proxy will be reflected in the other. # This is useful for cases where we want a keep-alive reference to a module without increasing # its reference count. def nn_module_proxy(mod): if not isinstance(mod, torch.nn.Module): return mod if isinstance(mod, torch.fx.GraphModule): # Dynamo-generated GM's shouldn't contain user-created GM's return mod proxy = mod.__class__.__new__(mod.__class__) proxy.__dict__ = mod.__dict__ return proxy class GmWrapper(torch.nn.Module): def __init__(self, gm, unflatten_fn): super().__init__() self.gm = gm self.unflatten_fn = unflatten_fn def forward(self, *args): args: list[Any] = list(args) return self.gm(*self.unflatten_fn(args)) def flatten_graph_inputs(gm: torch.fx.GraphModule, inputs, compile_gm): """ Mutate inputs so that they are flat and wrap gm such that it accepts those inputs. This is needed for graphs that take bumpy inputs. """ inputs_idx_to_clear = [ i for i, node in enumerate(gm.graph.nodes) if node.op == "placeholder" and node.meta.get("steal_arg", False) ] if torch._dynamo.compiled_autograd.in_compiled_autograd_region: # fast path, avoid pytree overhead # compiled autograd inputs are always a list of tensors, maybe followed by symints assert inputs_idx_to_clear == [0] assert isinstance(inputs[0], list) boxed_inputs_count = len(inputs[0]) def flatten_fn(args): return args[0] + list(args[1:]) def unflatten_fn(flat_args): return (flat_args[:boxed_inputs_count], *flat_args[boxed_inputs_count:]) compiled_fn = compile_gm(GmWrapper(gm, unflatten_fn), flatten_fn(inputs)) else: # slow path, don't know inputs structure flat_inputs, spec = pytree.tree_flatten(inputs) unflatten_fn = functools.partial(pytree.tree_unflatten, treespec=spec) compiled_fn = compile_gm(GmWrapper(gm, unflatten_fn), flat_inputs) # note this doesn't check the spec, assuming it is the same flatten_fn = pytree.arg_tree_leaves def wrapper(*args): flat_args = flatten_fn(args) # flat_args is a new list, so we need to clear references from the old list for i in inputs_idx_to_clear: args[i].clear() # this call is boxed to avoid increasing refcount until we reach aot_module_simplified forward return compiled_fn(flat_args) return wrapper def get_locals_to_steal(maybe_gm): if not isinstance(maybe_gm, torch.fx.GraphModule) or not hasattr(maybe_gm, "meta"): return [] return maybe_gm.meta.get("locals_to_steal", []) def set_locals_to_steal(gm, locals_to_steal): gm.meta["locals_to_steal"] = locals_to_steal class Lit: def __init__(self, s): self.s = s def __repr__(self) -> str: return self.s warn_once_cache: set[str] = set() def warn_once(msg, stacklevel=1): # Dynamo causes all warnings.warn (in user code and in Dynamo code) to print all the time. # https://github.com/pytorch/pytorch/issues/128427. # warn_once is a workaround: if the msg has been warned on before, then we will not # warn again. # NB: it's totally ok to store a cache of all the strings: this is what warnings.warn does as well. if msg in warn_once_cache: return warn_once_cache.add(msg) warnings.warn(msg, stacklevel=stacklevel + 1) def strip_color_from_string(text): # This regular expression matches ANSI escape codes ansi_escape = re.compile(r"\x1B[@-_][0-?]*[ -/]*[@-~]") return ansi_escape.sub("", text) @contextlib.contextmanager def _disable_saved_tensors_hooks_during_tracing(): # See NOTE: [Deferring tensor pack/unpack hooks until runtime] try: prior = torch._C._autograd._saved_tensors_hooks_set_tracing(True) yield finally: torch._C._autograd._saved_tensors_hooks_set_tracing(prior) def is_parameter_freezing(): return torch._inductor.config.freezing and not torch.is_grad_enabled() def get_torch_function_mode_stack(): return [ get_torch_function_mode_stack_at(i) for i in range(_len_torch_function_stack()) ] def get_torch_function_mode_stack_at(ind): assert ind < _len_torch_function_stack() and ind >= 0 return torch._C._get_function_stack_at(ind) def set_torch_function_mode_stack(stack): for _ in range(_len_torch_function_stack()): _pop_torch_function_stack() for mode in stack: _push_on_torch_function_stack(mode) def clear_torch_function_mode_stack(): for _ in range(_len_torch_function_stack()): _pop_torch_function_stack() # call from C dynamo in order to inspect values in pdb def _breakpoint_for_c_dynamo(*args): breakpoint() def verify_guard_fn_signature(value): fn = value.__metadata_guard__ sig = inspect.signature(fn) if len(sig.parameters) != 2: from .exc import InternalTorchDynamoError raise InternalTorchDynamoError( "Tensor subclass method __metadata_guard__ must take exactly two subclass metadata arguments" ) if fn.__self__ != value.__class__: from .exc import InternalTorchDynamoError raise InternalTorchDynamoError( "Tensor subclass method __metadata_guard__ must be a classmethod" ) def does_not_override_dict_iter_methods(user_cls): return ( user_cls.items in (dict.items, OrderedDict.items) and user_cls.values in (dict.values, OrderedDict.values) and user_cls.keys in (dict.keys, OrderedDict.keys) and user_cls.__iter__ in (dict.__iter__, OrderedDict.__iter__) ) # Helper functions below are to prevent __torch_function__ # calls from happening in the middle of __torch_function__ # compiled bytecode # They will be skipped which is the desired result def call_size(x, i): @torch._dynamo.disable(recursive=True) def fn(x, i): return x.size(i) return fn(x, i) def call_stride(x, i): @torch._dynamo.disable(recursive=True) def fn(x, i): return x.stride(i) return fn(x, i) def call_storage_offset(x): @torch._dynamo.disable(recursive=True) def fn(x): return x.storage_offset() return fn(x) # Helper function to extract relevant parts of a tensor's __dict__ to store in node meta. # To avoid ref cycles, it's important that no tensors are present here, so leave those out. def _extract_tensor_dict(t): KEYS_TO_COPY = [ "_dynamo_static_input_type", "tag", ] tensor_dict = { key: copy.copy(t.__dict__[key]) for key in KEYS_TO_COPY if key in t.__dict__ } return tensor_dict # This is useful for reconstructing within the Dynamo graph the non-graph-input objects # whose lifetime is governed by the user. # e.g. torch.cuda.Event is a prime example. user_obj_id_to_weakref: dict[int, weakref.ReferenceType[object]] = {} def get_user_object_from_id(obj_id): obj = user_obj_id_to_weakref[obj_id]() assert obj is not None, "User object is no longer alive" return obj def store_user_object_weakref(obj): obj_id = id(obj) user_obj_id_to_weakref[obj_id] = weakref.ref(obj) class CompileTimeInstructionCounter: _counter: int = 0 _id: int = -1 _depth = 0 @classmethod def start(cls) -> None: cls._depth = cls._depth + 1 if cls._depth == 1: cls._id = _instruction_counter.start() @classmethod def end(cls) -> None: cls._depth = cls._depth - 1 if cls._depth == 0: cls._counter += _instruction_counter.end(cls._id) cls._id = -1 @classmethod def clear(cls) -> None: cls._counter = 0 @classmethod def value(cls) -> int: return cls._counter @classmethod @contextmanager def record(cls): try: if config.record_compile_time_instruction_count: cls.start() yield finally: if config.record_compile_time_instruction_count: cls.end() def set_feature_use(feature: str, usage: bool): """ Records whether we are using a feature Generally a feature is a JK. """ # Note that sometimes (tests etc...) we're not in a context which we can record into if get_metrics_context().in_progress(): get_metrics_context().set_key_value("feature_usage", feature, usage) _ddp_optimization_mode: tuple[str, ...] = ( "ddp_optimizer", "python_reducer", # experimental mode "no_optimization", ) def get_optimize_ddp_mode(): optimize_ddp = config.optimize_ddp if isinstance(optimize_ddp, bool): mode = "ddp_optimizer" if optimize_ddp else "no_optimization" elif isinstance(optimize_ddp, str): mode = optimize_ddp else: raise ValueError( f"Invalid dynamo config optimize_ddp type {type(optimize_ddp)=}" ) assert mode in _ddp_optimization_mode, ( f"Invalid dynamo config optimize_ddp value {mode=}" ) return mode