""" Provides functionality for compiling PyTorch's autograd (automatic differentiation) system. This module implements compiled autograd, which traces and optimizes backward pass computations at runtime. The key components are: - AutogradCompilerInstance: Traces and compiles autograd graphs using FX - Context managers (_enable/_disable): Control when compiled autograd is active - Utility functions: Support graph manipulation, tensor operations, and hooks Compiled autograd can significantly improve backward pass performance by removing Python overhead and enabling additional optimizations. It works by capturing backward computations into an FX graph that can be compiled and optimized, while maintaining the same semantics as eager mode autograd. """ import contextlib import functools import itertools import operator import time from collections import Counter, defaultdict from collections.abc import Callable, Generator, Sequence from typing import Any, TYPE_CHECKING import torch import torch.utils._pytree as pytree from torch._dispatch.python import enable_python_dispatcher from torch._dynamo.external_utils import ( call_accumulate_grad, call_backward, call_hook, FakeCompiledAutogradEngine, unwrap_maybe_dynamic_int, ) from torch._dynamo.source import GetItemSource, LocalSource from torch._dynamo.utils import ( counters, get_chromium_event_logger, lazy_format_graph_code, set_locals_to_steal, ) from torch._functorch._aot_autograd.runtime_wrappers import ( AutogradLazyBackwardCompileInfo, CachedAutogradLazyBackwardCompileInfo, ) from torch._guards import compile_context, CompileContext, CompileId, Source from torch._logging import getArtifactLogger, trace_structured from torch._prims_common import clone_preserve_strides from torch._subclasses import FakeTensorMode from torch._subclasses.fake_tensor import FakeTensor from torch.fx import GraphModule from torch.fx.experimental._backward_state import BackwardState from torch.fx.experimental.proxy_tensor import ( disable_autocast_cache, disable_proxy_modes_tracing, fetch_object_proxy, ProxyTorchDispatchMode, PythonKeyTracer, track_tensor_tree, ) from torch.fx.experimental.symbolic_shapes import DimDynamic, ShapeEnv from torch.fx.traceback import preserve_node_meta, set_stack_trace from torch.types import FloatLikeType, IntLikeType from torch.utils._ordered_set import OrderedSet from torch.utils._traceback import CapturedTraceback if TYPE_CHECKING: from torch.fx.proxy import Proxy TURN_OFF_MSG = """You can turn off compiled autograd by either: 1. Moving the unsupported autograd call outside of the torch.compile'd region. 2. Wrapping the unsupported autograd call in the torch._dynamo.compiled_autograd._disable() context manager. 3. Setting torch._dynamo.config.compiled_autograd=False for the torch.compile call containing the unsupported autograd call. 4. Setting torch._dynamo.config.compiled_autograd=False at the start of the program.""" compiled_autograd_log = getArtifactLogger(__name__, "compiled_autograd") verbose_log = getArtifactLogger(__name__, "compiled_autograd_verbose") def snapshot_verbose_logging_enabled() -> bool: return torch._logging._internal.log_state.is_artifact_enabled( "compiled_autograd_verbose" ) def snapshot_cudagraph_enabled() -> bool: return torch._inductor.config.triton.cudagraphs def maybe_clone(x: torch.Tensor | None) -> torch.Tensor | None: if x is not None: return clone_preserve_strides(x) return x def extract_bw_module(CompiledFunction: Any) -> Callable[..., Any]: if isinstance( CompiledFunction._lazy_backward_info, AutogradLazyBackwardCompileInfo ): return CompiledFunction._lazy_backward_info.bw_module elif isinstance( CompiledFunction._lazy_backward_info, CachedAutogradLazyBackwardCompileInfo ): with torch._subclasses.fake_tensor.unset_fake_temporarily(): return CompiledFunction._lazy_backward_info.bw_module_fn() else: raise AssertionError( "Unexpected Lazy Backward Compilation Info Type. Please file an issue." ) # Note: [Anomaly Mode Semantics in Compiled Autograd] # In the eager autograd engine, anomaly mode is able to detect NaNs # after each node. This is useful, because the executed code with # and without anomaly mode are the same. So assuming determinism, # a NaN in regular mode should also happen in anomaly mode. # # With torch.compile, following eager semantics would require inserting # runtime asserts to check for NaNs, which could prevent some fusions. # This results in different code being run with and without anomaly mode. # So different semantics are needed, this implementation below will check # for NaNs at the end of the autograd call, instead of after each node class NaNChecker: def __init__(self, accumulate_grad: bool) -> None: self.accumulate_grad = accumulate_grad self.params_indices: list[int] = [] self.params_to_check: dict[str, torch.Tensor] = {} self.output_names: list[str] = [] def prep_with_graph(self, graph: torch.fx.Graph) -> None: inputs_node = next(iter(graph.nodes)) acc_grad_nodes = graph.find_nodes( op="call_function", target=call_accumulate_grad ) output_nodes = graph.find_nodes(op="output")[0].args[0] if self.accumulate_grad != bool(acc_grad_nodes): raise AssertionError( f"accumulate_grad mismatch: expected {self.accumulate_grad}, " f"got {bool(acc_grad_nodes)} acc_grad_nodes" ) if self.accumulate_grad != (not output_nodes): raise AssertionError( f"accumulate_grad mismatch with output_nodes: expected {self.accumulate_grad}, " f"got {not output_nodes}" ) for node in acc_grad_nodes: param_node = node.args[0] # AccumulateGrad always saves a reference to the param # so Compiled Autograd will always lift the param and # this should always be true if param_node.target is not operator.getitem: raise AssertionError( f"Expected param_node target to be operator.getitem, got {param_node.target}" ) if param_node.args[0] is not inputs_node: # type: ignore[possibly-undefined] raise AssertionError("Expected param_node.args[0] to be inputs_node") if not isinstance(param_node.args[1], int): raise AssertionError( f"Expected param_node.args[1] to be int, got {type(param_node.args[1])}" ) self.params_indices.append(param_node.args[1]) self.output_names = [node.name for node in output_nodes] def prep_with_inputs(self, inputs: tuple[torch.Tensor, ...]) -> None: if not self.accumulate_grad: # Using .grad, nothing to prep return # Using .backward, we must check existing grads on params if any for idx in self.params_indices: grad = inputs[idx].grad if grad is not None: if torch.isnan(grad).any(): raise AssertionError( f"Compiled autograd running under anomaly mode with inputs[{idx}] already " f"having NaN gradient. This is not supported. {TURN_OFF_MSG}" ) self.params_to_check[f"inputs[{idx}]"] = inputs[idx] def check(self, out: tuple[torch.Tensor, ...]) -> None: if self.accumulate_grad: # Using .backward, graph outputs are empty if out: raise AssertionError( "Expected empty output for accumulate_grad backward" ) nan_params: list[str] = [] for inputs_str, param in self.params_to_check.items(): if param.grad is None: # not true for autograd.grad raise AssertionError( "Expected param.grad to be non-None for backward mode" ) if torch.isnan(param.grad).any(): nan_params.append(inputs_str) if nan_params: raise RuntimeError( f"Compiled Autograd returned NaN gradients for parameters: {','.join(nan_params)}." ) else: # Using .grad, graph outputs are grads nan_grads: list[str] = [] for i, grad in enumerate(out): if torch.isnan(grad).any(): nan_grads.append(self.output_names[i]) if nan_grads: raise RuntimeError( f"Compiled Autograd returned NaN gradients for output nodes: {','.join(nan_grads)}." ) # We lazily bind "functional backward" variants for PyTorch built-in autograd # nodes to this class. Example: torch._dynamo.compiled_autograd.ops.MulBackward0 # Each "functional backward" is bound the first time the node's apply_with_saved # function is called. It's possible to avoid lazy binding and instead bind # all of this upfront (perhaps at import time) via codegen changes. class OpNamespace: def __init__(self) -> None: self.custom_function_name_counter: Counter[str] = Counter() def add( self, name: str, fn: Callable[..., Any], is_custom_function: bool, is_traceable: bool, ) -> str: if is_custom_function: name = "CppNode" + name count = self.custom_function_name_counter[name] self.custom_function_name_counter[name] += 1 name = f"{name}{count}" if hasattr(self, name): raise AssertionError(f"Op with name '{name}' already exists") result = Op(name, fn, is_custom_function) if is_traceable: # pyrefly: ignore [deprecated] setattr(self, name, torch._dynamo.allow_in_graph(result)) else: # C++ autograd function was not marked as traceable # Dynamo can't dry run it at compile time, so must fallback to eager @torch._dynamo.disable # type: ignore[misc] def run_non_traceable_cpp_in_eager(*args: Any, **kwargs: Any) -> Any: return result(*args, **kwargs) setattr(self, name, run_non_traceable_cpp_in_eager) return name def get(self, name: str) -> Any: return getattr(self, name) class Op: def __init__( self, name: str, fn: Callable[..., Any], is_custom_function: bool ) -> None: self.fn = fn self.is_custom_function = is_custom_function self.__name__ = name self.__module__ = "torch._dynamo.compiled_autograd.ops" def __call__(self, *args: Any, **kwargs: Any) -> Any: return self.fn(*args, **kwargs) def __repr__(self) -> str: return self.__module__ + "." + self.__name__ ops = OpNamespace() _graph_placeholders = ["inputs", "sizes", "scalars", "hooks", "packed_data"] _impure_targets = OrderedSet( [ call_hook, call_backward, FakeCompiledAutogradEngine._exec_final_callbacks_stub, call_accumulate_grad, ] ) COMPILE_COUNTER = itertools.count() def make_compile_context(compiled_autograd_id: int) -> Any: return compile_context( CompileContext( CompileId( compiled_autograd_id=compiled_autograd_id, frame_id=None, frame_compile_id=None, ) ) ) class AutogradCompilerInstance: def __init__(self, compiler_fn: Callable[..., Any]) -> None: self.compiler_fn = compiler_fn self.stack = contextlib.ExitStack() self.close = self.stack.close self.shape_env = ShapeEnv() self.fake_tensor_mode = FakeTensorMode( allow_fallback_kernels=True, allow_non_fake_inputs=True, shape_env=self.shape_env, ) self.fx_tracer = PythonKeyTracer() self.proxy_mode = ProxyTorchDispatchMode(self.fx_tracer, "symbolic") self.hooks_proxy: Proxy | None = None def wrap_fake(self, x: torch.Tensor, source: Source | None) -> FakeTensor: if not isinstance(x, torch.Tensor): raise AssertionError(f"Expected torch.Tensor, got {type(x)}") return self.fake_tensor_mode.from_tensor(x, source=source) @staticmethod def source(name: str, idx: Any) -> GetItemSource: return GetItemSource(LocalSource(name), idx) def begin_capture( self, inputs: list[torch.Tensor], sizes: list[int], scalars: list[IntLikeType | FloatLikeType], origins: list[list[tuple[int, str]]], accumulate_grad: bool, check_nans: bool, ) -> tuple[str, list[torch.Tensor], list[IntLikeType], list[FloatLikeType]]: counters["compiled_autograd"]["captures"] += 1 self.id = next(COMPILE_COUNTER) self.aot_id_counter: dict[int, int] = defaultdict(int) self.compile_context = make_compile_context(self.id) self.compile_context.__enter__() self.nan_checker = NaNChecker(accumulate_grad) if check_nans else None self.start_time_ns = time.time_ns() get_chromium_event_logger().log_event_start( "compiled_autograd", self.start_time_ns, {"graph_id": self.id}, log_pt2_compile_event=True, ) self.fx_tracer.root = torch.nn.Module() self.fx_tracer.graph = torch.fx.Graph(tracer_cls=PythonKeyTracer) self.fx_tracer.tensor_attrs = {} self.symnode_proxy_lookup = {} ( args_proxy, self.sizes_proxy, self.scalars_proxy, self.hooks_proxy, self.packed_data_proxy, ) = ( self.fx_tracer.create_proxy("placeholder", name, (), {}) for name in _graph_placeholders ) self.stack.enter_context(preserve_node_meta()) inputs_origins, sizes_origins, scalars_origins = origins # Turn on PythonDispatcher during initial trace to make it identifiable # that tracing is happening, which is needed to prevent hashing symints self.stack.enter_context(enable_python_dispatcher()) # tensor inputs to fake tensors x = inputs[0] # mypy will complain about unbound x try: for idx, x in enumerate(inputs): inputs[idx] = self.wrap_fake(x, self.source("inputs", idx)) except Exception as e: raise NotImplementedError( f"Found tensor of type {type(x)}, which is not supported by FakeTensorMode. {TURN_OFF_MSG}" ) from e self.bind_objects_to_proxies(inputs, args_proxy, inputs_origins) # size inputs to symints sym_sizes = [ self.shape_env.create_unspecified_symint_and_symbol( val, self.source("sizes", idx), DimDynamic.DYNAMIC, ) for idx, val in enumerate(sizes) ] # We want to mark every size as dynamic, but since there's no way to # mark a primitive `int` as dynamic, we need to wrap it in a tensor. # In the graph, we unwrap it with `unwrap_maybe_dynamic_int` back into a primitive. proxies = [self.sizes_proxy[i] for i in range(len(sym_sizes))] # type: ignore[index] for i, symint in enumerate(sym_sizes): proxies[i] = self.fx_tracer.create_proxy( "call_function", unwrap_maybe_dynamic_int, (proxies[i],), {}, ) if not isinstance(symint, int): self.symnode_proxy_lookup[symint.node] = proxies[i] proxies = self.bind_objects_to_proxies(sym_sizes, proxies, sizes_origins) for idx, val in enumerate(scalars): source = self.source("scalars", idx) if isinstance(val, int): scalars[idx] = self.shape_env.create_unspecified_symint_and_symbol( val, source, DimDynamic.DYNAMIC, ) elif isinstance(val, float): scalars[idx] = self.shape_env.create_symfloatnode( self.shape_env.create_unspecified_symbol( val, source=source, dynamic_dim=DimDynamic.DYNAMIC, ), hint=val, source=source, ) else: raise AssertionError("Unexpected scalar type: ", type(val)) self.bind_objects_to_proxies(scalars, self.scalars_proxy, scalars_origins) for i, symval in enumerate(scalars): self.symnode_proxy_lookup[symval.node] = self.scalars_proxy[i] # type: ignore[union-attr] # TODO(jansel): are all these modes needed? self.stack.enter_context(self.fake_tensor_mode) self.stack.enter_context(self.proxy_mode) self.stack.enter_context(disable_autocast_cache()) # Needed to make sure we don't accidentally specialize any symbols if self.fake_tensor_mode.shape_env is None: raise AssertionError("fake_tensor_mode.shape_env must not be None") env = self.fake_tensor_mode.shape_env self.stack.enter_context( torch.fx.experimental.symbolic_shapes._suppress_guards(env) ) # pyrefly: ignore [bad-return] return ( str(CompileContext.current_compile_id()), inputs, sym_sizes, scalars, # type: ignore[return-value] ) def log_compile_reasons( self, compile_reasons: list[str], ) -> None: if not compile_reasons: raise AssertionError("compile_reasons must not be empty") trace_structured( "artifact", metadata_fn=lambda: { "name": "compiled_autograd_compile_reasons", "encoding": "json", }, payload_fn=lambda: compile_reasons, ) def proxy_call_aot_backward( self, pinputs: Sequence[Any], psaved_tensors: Sequence[torch.Tensor], saved_tensors: Sequence[torch.Tensor], pctx: Any, ctx: Any, maybe_backward_state_idx: int | None, opaque_object_indices: list[int], ) -> Sequence[Any]: # The AOTBackward call consists of three things: the prologue, the # backward graph, and the epilogue. # Our strategy is: # - allow_in_graph the prologue (in the CA graph and Dynamo graph), # - copy-paste the backward graph into the CA graph so that CA passes and Dynamo can see it # - trace directly through the epilogue. Anything that gets baked in is # constant metadata (for example, metadata about the number of outputs, or removing # RNG arguments or effect tokens). # If Dynamo graph capture were better, then we could add a node for the prologue # into the CA graph and have Dynamo trace into it. psymints = [self.to_proxy(e) for e in ctx._get_compiled_autograd_symints()] popaque_objects = [ self.hooks_proxy[idx] # type: ignore[index] for idx in opaque_object_indices ] # NOTE: we should only close over constants CompiledFunction = ctx._forward_cls bw_module = extract_bw_module(CompiledFunction) metadata = CompiledFunction.metadata maybe_subclass_metadata = CompiledFunction.maybe_subclass_metadata aot_id = CompiledFunction._aot_id bw_prologue_fn = CompiledFunction._bw_prologue_fn bw_epilogue_fn = CompiledFunction._bw_epilogue_fn del CompiledFunction if torch.is_grad_enabled(): for output_alias_info in metadata.output_info: if output_alias_info.requires_grad: raise RuntimeError( "torch.compile does not currently support higher order gradients." ) @torch._dynamo.allow_in_graph # type: ignore[misc] def call_aot_bwd_prologue( ctx_saved_tensors: Sequence[torch.Tensor], ctx_symints: Sequence[IntLikeType], ctx_opaque_objs: Sequence[Any], flat_args: Sequence[Any], ) -> Any: return bw_prologue_fn( ctx_saved_tensors, ctx_symints, ctx_opaque_objs, flat_args, ) pgrads = self.fx_tracer.create_proxy( kind="call_function", # pyrefly: ignore [bad-argument-type] target=call_aot_bwd_prologue, args=( psaved_tensors, psymints, popaque_objects, pinputs, ), kwargs={}, ) pbackward_state = None if maybe_backward_state_idx is not None: pbackward_state = self.hooks_proxy[maybe_backward_state_idx] # type: ignore[index] # Copy-paste the AOT backward graph into the compiled autograd graph def copy_paste_aot_backward_graph() -> list[torch.Tensor]: def num_inputs(graph: torch.fx.Graph) -> int: num_args = 0 for node in graph.nodes: if node.op == "placeholder": num_args += 1 continue else: break return num_args # set up the proxy inputs to bw_module # the calling convention is: [*symints, *args (primals and tangents), backward_state] num_args = num_inputs(bw_module.graph) # type: ignore[attr-defined] pall_args = [ pgrads[i] for i in range(num_args - int(pbackward_state is not None)) ] # replace the symints with our symints symints = ctx._get_compiled_autograd_symints() if len(symints) != len(ctx.symints): raise AssertionError( f"symints length mismatch: {len(symints)} vs {len(ctx.symints)}" ) psymints = [self.to_proxy(e) for e in symints] pall_args[: len(symints)] = psymints # Add backward_state if pbackward_state is not None: pall_args.append(pbackward_state) # run over all nodes of the aot_backward graph. # copy and paste them all into the compiled autograd graph. args_idx = 0 # pyrefly: ignore [implicit-any] value_remap = {} poutputs: list[torch.fx.Proxy] | None = None # names of nodes must appear only once in the fx.Graph # dedup AOT backwards that appear multiple times deduped_aot_id = str(aot_id) if self.aot_id_counter[aot_id]: deduped_aot_id += f"_{self.aot_id_counter[aot_id]}" self.aot_id_counter[aot_id] += 1 def make_unique(node_name: str) -> str: # make it both informative and unique return f"aot{deduped_aot_id}_{node_name}" for node in bw_module.graph.nodes: # type: ignore[attr-defined] if node.op == "placeholder": ph = pall_args[args_idx].node ph.name = make_unique(node.name) value_remap[node] = ph args_idx += 1 elif node.op == "output": if len(node.args) != 1: raise AssertionError( f"Expected output node to have 1 arg, got {len(node.args)}" ) poutputs = [ torch.fx.Proxy(value_remap[n], self.fx_tracer) if isinstance(n, torch.fx.Node) else n for n in node.args[0] ] elif node.op == "get_attr": name = node.target qualname = self.fx_tracer.get_fresh_qualname(name) setattr(self.fx_tracer.root, qualname, getattr(bw_module, name)) result = self.fx_tracer.create_node("get_attr", qualname, (), {}) result.name = make_unique(node.name) value_remap[node] = result elif node.op == "call_function": if node.target is torch.ops.aten.view.default: # this aot bwd graph is being lazily compiled # we must manually apply the view_to_reshape post grad pass # since it was already applied to the aot fwd, and baked into the gradients node.target = torch.ops.aten.reshape.default result = self.fx_tracer.graph.node_copy( node, lambda n: value_remap[n] ) result.name = make_unique(node.name) value_remap[node] = result elif node.op == "call_module": name = node.target qualname = self.fx_tracer.get_fresh_qualname(name) setattr(self.fx_tracer.root, qualname, getattr(bw_module, name)) result = self.fx_tracer.graph.node_copy( node, lambda n: value_remap[n] ) result.target = qualname value_remap[node] = result else: raise AssertionError("shouldn't get here") if poutputs is None: raise AssertionError("No output node found in backward graph") # In general we don't know what the shapes of the outputs are, so allocate # some dummy sizes for them. def dummy() -> torch.Tensor: with disable_proxy_modes_tracing(): return torch.zeros(0, 0, 0, 0, 123) outputs = [ dummy() if isinstance(o, torch.fx.Proxy) else o for o in poutputs ] self.bind_objects_to_proxies(outputs, poutputs) return outputs outputs = copy_paste_aot_backward_graph() if maybe_subclass_metadata is not None: def proxy_subclass_constructor( subclass_meta: Any, is_runtime: bool, unwrapped_args: Sequence[Any], ) -> torch.Tensor: @torch._dynamo.allow_in_graph # type: ignore[misc] def make_subclass(*unwrapped_args: Any) -> Any: return subclass_meta.creation_fn( unwrapped_args, is_runtime=is_runtime ) punwrapped_args = pytree.tree_map(self.to_proxy, unwrapped_args) poutput = self.fx_tracer.create_proxy( kind="call_function", # pyrefly: ignore [bad-argument-type] target=make_subclass, args=tuple(punwrapped_args), kwargs={}, ) output = self.allocate_dummy() self.bind_objects_to_proxies([output], [poutput]) return output results = bw_epilogue_fn( outputs, make_subclass_override=proxy_subclass_constructor ) else: results = bw_epilogue_fn(outputs) presults = pytree.tree_map(self.to_proxy, results) return presults def proxy_call_backward( self, inputs: Sequence[Any], output_metadatas: Sequence[Any], saved_tensors: Sequence[torch.Tensor], backward_idx: int, ctx: torch.autograd.function.BackwardCFunction, maybe_backward_state_idx: int | None, opaque_object_indices: list[int], ) -> tuple[torch.Tensor | None, ...]: if self.hooks_proxy is None: raise AssertionError("hooks_proxy must be set before proxy_call_backward") pctx = self.hooks_proxy[backward_idx] # type: ignore[index] pinputs = self.to_proxy(inputs) psaved_tensors = self.to_proxy(saved_tensors) if hasattr(ctx._forward_cls, "_aot_id"): # type: ignore[attr-defined] # AOT backward proxies = self.proxy_call_aot_backward( pinputs, psaved_tensors, saved_tensors, pctx, ctx, maybe_backward_state_idx, opaque_object_indices, ) else: if getattr(ctx._forward_cls, "boxed_grads_call", False): # type: ignore[attr-defined] raise RuntimeError( f"boxed_grads_call=True on {ctx._forward_cls.__name__} " # type: ignore[attr-defined] "is not supported with compiled autograd. " ) proxies = self.fx_tracer.create_proxy( kind="call_function", target=call_backward, args=( pctx, psaved_tensors, *pinputs, ), kwargs={}, ) if proxies is None: raise AssertionError("proxies must not be None after backward call") with disable_proxy_modes_tracing(): # create fake Tensors grad_ins: list[torch.Tensor | None] = [] for idx, output_metadata in enumerate(output_metadatas): if output_metadata is None or proxies[idx] is None: grad_ins.append(None) continue layout, device, dtype, size = output_metadata grad_ins.append( torch.empty(size=size, dtype=dtype, layout=layout, device=device) ) self.bind_objects_to_proxies(grad_ins, proxies) return tuple(grad_ins) def call_copy_slices_prologue( self, inputs: Sequence[Any], base_sizes: Sequence[Any], base_strides: Sequence[Any], base_storage_offset: Any, view_sizes: Sequence[Any], view_strides: Sequence[Any], view_storage_offset: Any, ) -> Sequence[torch.Tensor]: args = ( inputs, self.to_proxy(base_sizes), self.to_proxy(base_strides), self.to_proxy(base_storage_offset), self.to_proxy(view_sizes), self.to_proxy(view_strides), self.to_proxy(view_storage_offset), ) return self.proxy_call(copy_slices_prologue, args, [None] * 3) def call_copy_slices_epilogue( self, needs_input_grad: Sequence[bool], result: torch.Tensor, res: Sequence[Any], grad_slice: torch.Tensor, ) -> Sequence[torch.Tensor]: return self.proxy_call( copy_slices_epilogue, (needs_input_grad, result, res, grad_slice), [None] * len(needs_input_grad), ) def allocate_dummy(self) -> torch.Tensor: with disable_proxy_modes_tracing(): # Weird quantity so it's easy to grep return torch.zeros([0, 123456789]) def allocate_dummy_like(self, x: torch.Tensor) -> torch.Tensor: with disable_proxy_modes_tracing(): return torch.empty_like(x) def bind_function( self, fn_name: str, fn: Callable[..., Any], is_custom_function: bool, is_traceable: bool, ) -> str: """Binds ops.fn_name = fn""" return ops.add(fn_name, fn, is_custom_function, is_traceable) def apply_functional( self, fn_name: str, grads: Sequence[Any], args: Any, output_metadata: Sequence[Any], ) -> Sequence[torch.Tensor]: """Proxies a call to ops.fn_name(grads, *args) into the graph""" op = ops.get(fn_name) return self.proxy_call(op, (grads, *args), output_metadata) def proxy_call( self, fn: Callable[..., Any], args: Any, output_metadata: Sequence[Any] ) -> Sequence[torch.Tensor]: """Proxies a call to fn(*args) into the graph""" proxy_args = pytree.tree_map(lambda e: self.to_proxy(e), args) proxy_out = self.fx_tracer.create_proxy( "call_function", fn, args=proxy_args, kwargs={} ) result = [self.allocate_dummy() for _ in output_metadata] self.bind_objects_to_proxies(result, [proxy_out[i] for i in range(len(result))]) return result def validate_outputs( self, _: Any, outputs: Sequence[Any], args: Any, output_metadata: Sequence[Any] ) -> Sequence[torch.Tensor]: """Proxies a call to ops.validate_outputs(outputs, *args) into the graph""" op = ops.get("validate_outputs") proxy_args = pytree.tree_map(self.to_proxy, (outputs, *args)) new_proxy_outputs = self.fx_tracer.create_proxy( "call_function", op, args=proxy_args, kwargs={} ) if len(output_metadata) != len(outputs): raise AssertionError( f"output_metadata length ({len(output_metadata)}) != outputs length ({len(outputs)})" ) self.bind_objects_to_proxies(outputs, new_proxy_outputs) return outputs def accumulate(self, old_var: Any, new_var: Any) -> torch.Tensor: old_var_proxy = self.to_proxy(old_var) new_var_proxy = self.to_proxy(new_var) proxy_out = self.fx_tracer.create_proxy( "call_function", torch.add, args=(old_var_proxy, new_var_proxy), kwargs={} ) result = self.allocate_dummy() self.bind_objects_to_proxies([result], [proxy_out]) return result def accumulate_grad( self, variable: torch.Tensor, variable_grad: torch.Tensor | None, grad: torch.Tensor | bool, has_post_hooks: bool | None = None, ) -> torch.Tensor | None: old_call = has_post_hooks is None if has_post_hooks is None: # Backward compatibility for extensions built before current grad # became an explicit capture argument. has_post_hooks = bool(grad) grad = variable_grad # type: ignore[assignment] variable_grad = variable.grad proxy_out = self.fx_tracer.create_proxy( "call_function", call_accumulate_grad, args=( self.to_proxy(variable), self.to_proxy(variable_grad), self.to_proxy(grad), has_post_hooks, ), kwargs={}, ) result = self.allocate_dummy_like(variable) self.bind_objects_to_proxies([result], [proxy_out]) variable.grad = result if old_call: return None return result def proxy_call_hook( self, hook: Callable[..., Any], *args: Any, **kwargs: Any ) -> torch.fx.Proxy: return self.fx_tracer.create_proxy( "call_function", call_hook, ( hook, *[self.to_proxy(x) for x in args], ), kwargs, ) def unpack_hook(self, hook_id: int, data_id: int) -> torch.Tensor: if self.hooks_proxy is None: raise AssertionError("hooks_proxy must be set before unpack_hook") hook = self.hooks_proxy[hook_id] # type: ignore[index] data = self.packed_data_proxy[data_id] # type: ignore[index] proxy = self.proxy_call_hook( hook, data, hook_type="unpack_hook", ) out = self.allocate_dummy() self.bind_objects_to_proxies([out], [proxy]) return out def tensor_pre_hook( self, inputs: list[torch.Tensor], hook_id: int, i: int ) -> list[torch.Tensor]: if self.hooks_proxy is None: raise AssertionError("hooks_proxy must be set before tensor_pre_hook") hook = self.hooks_proxy[hook_id] # type: ignore[index] proxy = self.proxy_call_hook( hook, inputs[i], hook_type="tensor_pre_hook", ) with disable_proxy_modes_tracing(): inputs[i] = maybe_clone(inputs[i]) # type: ignore[assignment] self.bind_objects_to_proxies([inputs[i]], [proxy]) return inputs def cpp_tensor_pre_hook( self, inputs: list[torch.Tensor], hook_id: int, i: int ) -> list[torch.Tensor]: proxy = self.fx_tracer.create_proxy( "call_function", torch._C._dynamo.compiled_autograd.call_cpp_tensor_pre_hooks, (hook_id, self.to_proxy(inputs[i])), {}, ) with disable_proxy_modes_tracing(): inputs[i] = maybe_clone(inputs[i]) # type: ignore[assignment] self.bind_objects_to_proxies([inputs[i]], [proxy]) return inputs def pre_hook(self, inputs: Sequence[Any], hook_id: int) -> list[torch.Tensor]: if self.hooks_proxy is None: raise AssertionError("hooks_proxy must be set before pre_hook") hook = self.hooks_proxy[hook_id] # type: ignore[index] proxies = self.proxy_call_hook( hook, inputs, hook_type="pre_hook", ) with disable_proxy_modes_tracing(): inputs = [maybe_clone(x) for x in inputs] self.bind_objects_to_proxies(inputs, proxies) return inputs def post_hook( self, outputs: list[torch.Tensor], inputs: Sequence[torch.Tensor], hook_id: int ) -> list[torch.Tensor]: if self.hooks_proxy is None: raise AssertionError("hooks_proxy must be set before post_hook") hook = self.hooks_proxy[hook_id] # type: ignore[index] proxies = self.proxy_call_hook( hook, outputs, inputs, hook_type="post_hook", ) with disable_proxy_modes_tracing(): outputs = [maybe_clone(x) for x in outputs] # type: ignore[misc] self.bind_objects_to_proxies(outputs, proxies) return outputs def post_acc_grad_hook( self, input: torch.Tensor, hook_id: int ) -> list[torch.Tensor]: if not isinstance(input, torch.Tensor): raise AssertionError(f"Expected torch.Tensor, got {type(input)}") if self.hooks_proxy is None: raise AssertionError("hooks_proxy must be set before post_acc_grad_hook") hook = self.hooks_proxy[hook_id] # type: ignore[index] proxy = self.proxy_call_hook( hook, input, hook_type="post_acc_grad_hook", ) with disable_proxy_modes_tracing(): res = [maybe_clone(input)] self.bind_objects_to_proxies(res, [proxy]) return res # type: ignore[return-value] # Note: [Compiled autograd and cudagraphs] # Eager autograd backward implements scalars as 0-dim tensors, see DivBackward0::other_. # When compiled autograd traces those nodes, it lifts the scalar tensors, resulting in a graph # with some cpu 0-dim tensor inputs. To prevent the entire graph from skipping cudagraph, we move the # scalar tensors to cuda. This works because ATen/prims ops will accept cuda 0-dim tensors too. def move_graph_nodes_to_cuda(self, graph: torch.fx.Graph) -> list[int]: to_move: dict[int, torch.fx.Node] = {} has_cuda_inputs = False nodes = list(graph.nodes) if nodes[0].target != "inputs": raise AssertionError( f"Expected first node target to be 'inputs', got '{nodes[0].target}'" ) inputs = nodes[0] inputs_users = list(inputs.users.keys()) # input access nodes should immediately follow placeholder nodes first_getitem_idx = len(_graph_placeholders) if nodes[first_getitem_idx] != inputs_users[0]: raise AssertionError("First getitem node does not match first inputs user") last_getitem_idx = first_getitem_idx + len(inputs_users) - 1 if nodes[last_getitem_idx] != inputs_users[-1]: raise AssertionError("Last getitem node does not match last inputs user") # getitem nodes on inputs for i, node in enumerate(inputs_users): if not has_cuda_inputs and node.meta["val"].device.type == "cuda": has_cuda_inputs = True continue is_cpu = node.meta["val"].device.type == "cpu" is_scalar = len(node.meta["val"].size()) == 0 if is_cpu and is_scalar: node_users = list(node.users.keys()) # We can only move the cpu scalar if it is not exposed to user code. if all( ( isinstance(user.target, torch._ops.OpOverload) and user.target.namespace in ("prims", "aten") ) or ( isinstance(user.target, Op) and not user.target.is_custom_function ) for user in node_users ): # all users are prims/aten, can move safely to_move[i] = node # only move cpu scalars to cuda if there were cuda activations in this graph, # this is to handle the case where cudagraphs is enabled on a cpu-only graph if has_cuda_inputs: for node in to_move.values(): verbose_log.debug("Moving node %s from cpu to cuda", node) node.meta["val"] = node.meta["val"].cuda() # return runtime indices we need to move to cuda return list(to_move.keys()) return [] def is_sym_node(self, node: Any) -> bool: return ( isinstance(node, torch.fx.Node) and node.op == "call_function" and node.target in [torch.ops.aten.sym_size.int, torch.ops.aten.sym_numel.default] ) def dce(self) -> None: # Most of these removed nodes would have been removed during Dynamo and AOTDispatch # Remove some of these nodes earlier to improve compilation speed # Dynamo guards will error instead of creating aliasing guards unless we unpack them in the graph unpack_nodes: OrderedSet[torch.fx.Node] = OrderedSet() i: int | None = None for i, node in enumerate(self.fx_tracer.graph.find_nodes(op="placeholder")): unpack_nodes.update(node.users.keys()) if i != len(_graph_placeholders) - 1: raise AssertionError( f"Expected {len(_graph_placeholders) - 1} placeholders, got {i}" ) def is_impure(node: torch.fx.Node) -> bool: if node in unpack_nodes or ( node.op == "call_function" and node.target in _impure_targets ): return True return node.is_impure() before = len(self.fx_tracer.graph.nodes) self.fx_tracer.graph.eliminate_dead_code(is_impure) after = len(self.fx_tracer.graph.nodes) verbose_log.debug("DCE removed %d nodes", before - after) def remove_unused_sizes(self) -> set[int]: used_sizes = [] unused_sizes = [] # seek placeholder, should be at nodes[1] it = iter(self.fx_tracer.graph.nodes) next(it) sizes_node = next(it) if sizes_node.name != "sizes": raise AssertionError( f"Expected placeholder 'sizes', got '{sizes_node.name}'" ) for getitem_node in sizes_node.users: if getitem_node.target is not operator.getitem: raise AssertionError( f"Expected operator.getitem target, got {getitem_node.target}" ) if getitem_node.users: used_sizes.append(getitem_node) else: # remove from the graph unused_sizes.append(getitem_node) used_sizes_idx: set[int] = set() for used in used_sizes: if not isinstance(used.args, tuple): raise AssertionError( f"Expected used.args to be tuple, got {type(used.args)}" ) if used.args[0] != sizes_node: raise AssertionError("Expected used.args[0] to be sizes_node") if not isinstance(used.args[1], int): raise AssertionError( f"Expected used.args[1] to be int, got {type(used.args[1])}" ) next_size_idx = len(used_sizes_idx) # used later reindex the runtime sizes arg used_sizes_idx.add(used.args[1]) # reindex the graph used.args = (used.args[0], next_size_idx) for unused in unused_sizes: self.fx_tracer.graph.erase_node(unused) return used_sizes_idx def create_graph_module(self, id: str) -> GraphModule: return GraphModule(self.fx_tracer.root, self.fx_tracer.graph, id) def end_capture(self, outputs: Any) -> tuple[Callable[..., Any], Any]: self.fx_tracer.create_proxy( "call_function", FakeCompiledAutogradEngine._exec_final_callbacks_stub, (), {}, ) self.stack.close() self.fx_tracer.create_node( "output", "output", (self.fx_tracer.create_arg(self.to_proxy(outputs)),), {}, ) runtime_inputs_to_move: list[int] = [] if snapshot_cudagraph_enabled(): runtime_inputs_to_move = self.move_graph_nodes_to_cuda(self.fx_tracer.graph) # We traced using dummy tensors. Delete all the metadata of the dummy tensors. # It's probably better to refactor this class to use a different tracer # than the make_fx tracer, but that is a larger change. for node in self.fx_tracer.graph.nodes: for field in ["tensor_meta", "example_value", "val"]: if field in node.meta: del node.meta[field] trace_structured( "artifact", metadata_fn=lambda: { "name": "compiled_autograd_graph_pre_reordering", "encoding": "string", }, payload_fn=lambda: GraphModule( self.fx_tracer.root, self.fx_tracer.graph, f"CompiledAutograd{self.id}PreReordering", ).print_readable(print_output=False), ) self.delay_unpack_hook_nodes() self.reorder_tensor_pre_hook_nodes() self.reorder_pre_hook_nodes_to_schedule_asap() self.reorder_accumulate_grad_nodes() self.reorder_pre_hook_nodes_to_mimic_eager() self.reorder_post_acc_grad_hook_nodes() self.reorder_post_hook_nodes() # TODO(yf225): work around: remove dead codes like `sym_size` and `sym_numel` which are not used downstream. e.g. # ``` # sym_numel_default = torch.ops.aten.sym_numel.default(sum_109); sum_109 = None # eq_115 = 16 == sym_numel_default; sym_numel_default = eq_115 = None # sym_size_int_39 = torch.ops.aten.sym_size.int(getitem_112, 1); getitem_112 = None # eq_116 = 16 == sym_size_int_39; eq_116 = None # eq_117 = 16 == sym_size_int_39; sym_size_int_39 = eq_117 = None # ``` # Proper fix is Richard's Python compiled autograd effort which will avoid calling make_fx and # should prevent these ops from going into the CA graph. self.dce() if self.nan_checker: self.nan_checker.prep_with_graph(self.fx_tracer.graph) # keep only sizes that are actually used in the graph used_sizes_idx = self.remove_unused_sizes() graph = self.create_graph_module(f"CompiledAutograd{self.id}") set_locals_to_steal(graph, ["inputs"]) lazy_graph_code = lazy_format_graph_code( "Compiled autograd graph", graph, include_device=True, include_stride=True, colored=True, ) compiled_autograd_log.info("%s", lazy_graph_code) verbose_log.debug("%s", lazy_graph_code) trace_structured( "compiled_autograd_graph", payload_fn=lambda: graph.print_readable(print_output=False), ) def runtime_wrapper( compiled_fn: Callable[..., Any], inputs: Any, sizes: Any, scalars: Any, hooks: Any, packed_inputs: Any, ) -> tuple[Any, Any]: global in_compiled_autograd_region try: in_compiled_autograd_region = True if self.nan_checker: self.nan_checker.prep_with_inputs(inputs) filtered_sizes = [] for idx, integer in enumerate(sizes): if idx in used_sizes_idx: # can't create negative size if integer > 0: filtered_sizes.append(torch.empty(0, integer)) torch._dynamo.maybe_mark_dynamic(filtered_sizes[-1], 1) else: filtered_sizes.append(integer) for i in runtime_inputs_to_move: inputs[i] = inputs[i].pin_memory().cuda(non_blocking=True) with _disable(), make_compile_context(self.id): out = compiled_fn( inputs, filtered_sizes, scalars, hooks, packed_inputs ) if self.nan_checker: self.nan_checker.check(out) return out finally: in_compiled_autograd_region = False get_chromium_event_logger().log_event_end( "compiled_autograd", time.time_ns(), {"graph_id": self.id}, self.start_time_ns, log_pt2_compile_event=True, ) self.compile_context.__exit__(None, None, None) return runtime_wrapper, self.compiler_fn(graph) @staticmethod def get_all_nodes(args: Sequence[Any]) -> list[torch.fx.Node]: # filter out non-Node args, like None nodes = [n for n in args if type(n) is torch.fx.Node] return nodes @staticmethod def is_placeholder(node: torch.fx.Node) -> bool: if node.op == "placeholder" or ( node.op == "call_function" and node.target is operator.getitem and node.args[0].op == "placeholder" # type: ignore[union-attr, arg-type] ): return True return False def reorder_accumulate_grad_nodes(self) -> None: """ Usage of AOTAutograd causes all the accumulate_grad_ nodes to get pushed to the end of the graph. This differs from eager mode, which schedules them as soon as possible. This pass attempts to reorder the graph to mimic eager behavior. """ for node in self.fx_tracer.graph.find_nodes( op="call_function", target=call_accumulate_grad ): param_node, variable_grad_node, grad_node = ( node.args[0], node.args[1], node.args[2], ) getitem_node = None if ( isinstance(grad_node, torch.fx.Node) and grad_node.target is operator.getitem ): getitem_node = grad_node grad_node = getitem_node.args[0] arg = max( self.get_all_nodes([param_node, variable_grad_node, grad_node]) ) # last arg if arg is not node.prev and not self.is_placeholder(arg): arg.append(node) if getitem_node is not None: arg.append(getitem_node) def delay_unpack_hook_nodes(self) -> None: """ We can delay unpack hooks until they are needed, even later than in the eager autograd engine. """ for node in self.fx_tracer.graph.find_nodes( op="call_function", target=call_hook ): if node.kwargs.get("hook_type", None) != "unpack_hook": continue first_user = min(node.users) first_user.prepend(node) def reorder_tensor_pre_hook_nodes(self) -> None: """ Usage of AOTAutograd causes all the tensor_pre_hook nodes to get pushed to the end of the graph. This differs from eager mode, which schedules them as soon as possible. This pass attempts to reorder the graph to mimic eager behavior. """ for node in self.fx_tracer.graph.find_nodes( op="call_function", target=call_hook ): if node.kwargs.get("hook_type", None) != "tensor_pre_hook": continue getitem_node = node.args[0] input_node = node.args[1] # tensor_pre_hook handle only one grad tensor if input_node is not node.prev and not self.is_placeholder(input_node): input_node.append(getitem_node) getitem_node.append(node) def reorder_pre_hook_nodes_to_schedule_asap(self) -> None: """ In this function, we schedule the pre hooks as soon as possible. This does not match eager behavior (schedule pre hook right before its registered node), but it can make acc grad be scheduled properly when the pre hooks are registered to them. After reordering acc grad node, we will reorder the pre hooks again to mimic eager behavior. """ for node in self.fx_tracer.graph.find_nodes( op="call_function", target=call_hook ): if node.kwargs.get("hook_type", None) != "pre_hook": continue getitem_node = node.args[0] # pre_hook handle a tuple of grad tensors input_nodes = self.get_all_nodes(node.args[1]) to_remove = [] to_append = [] hook_block = [node] # contain the hook and hook args getitem for n in input_nodes: if n.op == "call_function" and n.target is operator.getitem: to_append.append(n.args[0]) to_remove.append(n) hook_block.append(n) for a, b in zip(to_remove, to_append): input_nodes.remove(a) input_nodes.append(b) # type: ignore[arg-type] arg = max(input_nodes) # last input if arg is not node.prev and not self.is_placeholder(arg): arg.append(getitem_node) for n in hook_block: getitem_node.append(n) def reorder_pre_hook_nodes_to_mimic_eager(self) -> None: """ Usage of AOTAutograd causes all the pre_hook nodes to get pushed to the end of the graph. This differs from eager mode, which schedules them right before their registered node execution. This pass attempts to reorder the graph to mimic eager behavior. """ pre_hooks = [] for node in self.fx_tracer.graph.find_nodes( op="call_function", target=call_hook ): if node.kwargs.get("hook_type", None) != "pre_hook": continue pre_hooks.append(node) for node in reversed(pre_hooks): hook_getitem_node = node.args[0] users = list(node.users.keys()) if len(users) == 0: continue # users are all getitem ops and they are used by same registered node if not all( user.op == "call_function" and user.target is operator.getitem for user in users ): raise AssertionError( "Expected all users of pre_hook to be operator.getitem call_function nodes" ) registered_node = next(iter(users[0].users.keys())) if registered_node is not node.next: registered_node.prepend(hook_getitem_node) registered_node.prepend(node) for getitem in users: registered_node.prepend(getitem) def reorder_post_acc_grad_hook_nodes(self) -> None: """ Usage of AOTAutograd causes all the post_acc_grad_hook nodes to get pushed to the end of the graph. This differs from eager mode, which schedules them as soon as possible. This pass attempts to reorder the graph to mimic eager behavior. """ post_acc_grad_hooks = [] for node in self.fx_tracer.graph.find_nodes( op="call_function", target=call_hook ): if node.kwargs.get("hook_type", None) != "post_acc_grad_hook": continue post_acc_grad_hooks.append(node) # nodes in post_acc_grad_hooks are in topo order. For hooks registered # to same node, we should keep their relative order for node in reversed(post_acc_grad_hooks): getitem_node = node.args[0] param_node = node.args[1] # post_acc_grad_hook handle one param # find the corresponding acc_grad node acc_grad_node = None for n in list(param_node.users.keys()): if n.op == "call_function" and n.target is call_accumulate_grad: acc_grad_node = n break if acc_grad_node is None: raise AssertionError( "post_acc_grad_hook must have corresponding acc grad node" ) # append post_acc_grad_hook after acc_grad node acc_grad_node.append(getitem_node) getitem_node.append(node) def reorder_post_hook_nodes(self) -> None: """ Usage of AOTAutograd causes all the post_hook nodes to get pushed to the end of the graph. This differs from eager mode, which schedules them as soon as possible. This pass attempts to reorder the graph to mimic eager behavior. """ post_hooks = [] for node in self.fx_tracer.graph.find_nodes( op="call_function", target=call_hook ): if node.kwargs.get("hook_type", None) != "post_hook": continue post_hooks.append(node) for node in reversed(post_hooks): getitem_node = node.args[0] output_nodes = node.args[1] input_nodes = node.args[2] if len(output_nodes) > 0: continue # pyrefly: ignore [implicit-any] input_nodes_and_users = [] input_nodes_and_users.extend(list(input_nodes)) for input_node in input_nodes: input_nodes_and_users.extend( user for user in list(input_node.users.keys()) if not ( user.op == "call_function" and user.target is call_hook and node.kwargs.get("hook_type", None) == "post_hook" ) ) arg = max(input_nodes_and_users) # last input users if arg.op == "call_function" and arg.target is call_accumulate_grad: param_node = arg.args[0] post_acc_grad_hook_node = None for n in list(param_node.users.keys()): if ( n.op == "call_function" and n.target is call_hook and n.kwargs.get("hook_type", None) == "post_acc_grad_hook" ): post_acc_grad_hook_node = n if post_acc_grad_hook_node is not None: post_acc_grad_hook_node.append(getitem_node) getitem_node.append(node) continue if arg is not node.prev and not self.is_placeholder(arg): arg.append(getitem_node) getitem_node.append(node) def to_proxy(self, t: Any) -> Any: if t is None: return None if isinstance(t, list): return [self.to_proxy(x) for x in t] if isinstance(t, tuple): return tuple(self.to_proxy(x) for x in t) if isinstance(t, (torch.SymInt, torch.SymFloat)): return self.symnode_proxy_lookup[t.node] if not isinstance(t, torch.Tensor): # constant types like device, dtype, str return t proxy_tensor = fetch_object_proxy(self.fx_tracer, t) if not isinstance( proxy_tensor, torch.fx.experimental.proxy_tensor._ProxyTensor ): raise AssertionError(f"Expected _ProxyTensor, got {type(proxy_tensor)}") return proxy_tensor.proxy def bind_objects_to_proxies( self, objects: Sequence[Any], proxies: Any, origins: list[tuple[int, str]] | None = None, ) -> Sequence[Any]: if isinstance(proxies, torch.fx.Proxy): if origins: if len(origins) != len(objects): raise AssertionError( f"origins length ({len(origins)}) != objects length ({len(objects)})" ) bound_proxies = [] for i in range(len(objects)): nodecall_index, node_name = origins[i] self.set_node_origin(node_name, nodecall_index, None) bound_proxies.append(proxies[i]) # type: ignore[index] proxies = bound_proxies else: proxies = [proxies[i] for i in range(len(objects))] # type: ignore[index] if len(objects) != len(proxies): raise AssertionError( f"objects length ({len(objects)}) != proxies length ({len(proxies)})" ) track_tensor_tree(objects, proxies, constant=None, tracer=self.fx_tracer) return proxies def bind_backward_state(self, index: int) -> BackwardState: if self.hooks_proxy is None: raise AssertionError("hooks_proxy must be set before bind_backward_state") proxy = self.hooks_proxy[index] # type: ignore[index] bw_state = BackwardState() track_tensor_tree(bw_state, proxy, constant=None, tracer=self.fx_tracer) return bw_state def set_node_origin( self, node_name: str, nodecall_index: int, pyobj: torch.autograd.Function | None, ) -> None: maybe_aot_id = "" if pyobj is not None: forward_cls = pyobj._forward_cls # type: ignore[attr-defined] if hasattr(forward_cls, "_aot_id"): # backward was created by AOT Dispatcher if forward_cls._lazy_backward_info is None: raise RuntimeError( """This compiled backward function was saved by AOTAutogradCache, which does not support compiled autograd. Please turn off AOTAutogradCache using `TORCHINDUCTOR_AUTOGRAD_CACHE=0`.""" ) maybe_aot_id = forward_cls._aot_id new_code = f"{node_name}{maybe_aot_id} (NodeCall {nodecall_index})" raw_stack_trace = CapturedTraceback.extract().format()[-1] new_stack_trace = raw_stack_trace.replace( "raw_stack_trace = CapturedTraceback.extract().format()[-1]", new_code ) set_stack_trace(new_stack_trace) # state of the autograd engine dispatch, kept in sync by enable/disable context managers compiled_autograd_enabled = False # global flag to check if compiled autograd is enabled but Dynamo stance is "force_eager" compiled_autograd_enabled_force_eager = False # global flag to check if we are processing graphs produced from a compiled autograd graph in_compiled_autograd_region = False active_disable_ctx = False depth = 0 @contextlib.contextmanager def _enable( compiler_fn: Callable[..., Any], dynamic: bool = True, ignore_active_disable_ctx: bool = True, ) -> Generator[None, None, None]: # The entrypoint to enable CA. # It is recommended to enable via `torch._dynamo.config.compiled_autograd = True` rather # than using this context manager directly. If you are torch.compiling the corresponding # forward pass, make sure they are wrapped under this context as well. # # Example: # def train(model, inputs, target): # compiled_model = torch.compile(model) # pred = compiled_model(data) # loss = compute_loss(pred, target) # loss.backward() # # with _enable(compiler_fn): # train(model, inputs, target) # # Inputs: # - compiler_fn: The wrapper that will consume the compiled autograd graph, e.g. `torch.compile` # - dynamic: Whether compiled autograd will treat tensors in the autograd graph (params, activations) as dynamic. # This doesn't affect the dynamic configuration of the compilation wrapper. if not ignore_active_disable_ctx and active_disable_ctx: yield else: if dynamic: if type(dynamic) is not bool: raise AssertionError( f"Expected dynamic to be bool, got {type(dynamic)}" ) from torch._dynamo import eval_frame if eval_frame._stance.stance == "force_eager": # If user explicitly sets Dynamo stance to "force_eager", we want Compiled Autograd # to fall back to eager as well. global compiled_autograd_enabled_force_eager compiled_autograd_enabled_force_eager = True try: yield finally: compiled_autograd_enabled_force_eager = False else: # we need to import this, because user might not have imported it if they directly use this context manager # we need to lazily import it, because of circular dependencies if torch.cuda.is_available(): from torch._inductor import cudagraph_trees # noqa: F401 ( prior_compiler, prior_dynamic, ) = torch._C._dynamo.compiled_autograd.set_autograd_compiler( functools.partial(AutogradCompilerInstance, compiler_fn), dynamic ) if snapshot_verbose_logging_enabled(): torch._C._dynamo.compiled_autograd.set_verbose_logger(verbose_log) # type:ignore[arg-type] global compiled_autograd_enabled compiled_autograd_enabled = True global depth prior_depth = depth depth += 1 try: with torch.autograd.set_multithreading_enabled(False): yield finally: if not prior_compiler: compiled_autograd_enabled = False torch._C._dynamo.compiled_autograd.set_autograd_compiler( prior_compiler, prior_dynamic ) depth -= 1 if depth != prior_depth: raise AssertionError( "Nested Compiled Autograd Contexts must return before their parent context" ) @contextlib.contextmanager def _disable() -> Generator[None, None, None]: ( prior_compiler, prior_dynamic, ) = torch._C._dynamo.compiled_autograd.set_autograd_compiler(None, False) global compiled_autograd_enabled compiled_autograd_enabled = False global active_disable_ctx if not active_disable_ctx: active_disable_ctx = True try: yield finally: if prior_compiler: compiled_autograd_enabled = True active_disable_ctx = False torch._C._dynamo.compiled_autograd.set_autograd_compiler( prior_compiler, prior_dynamic ) # return to starting state of a new process def reset() -> None: global compiled_autograd_enabled compiled_autograd_enabled = False if in_compiled_autograd_region: raise AssertionError("Cannot reset while in_compiled_autograd_region is True") torch._C._dynamo.compiled_autograd.set_autograd_compiler(None, False) torch._C._dynamo.compiled_autograd.set_verbose_logger(None) torch._C._dynamo.compiled_autograd.clear_cache() global COMPILE_COUNTER COMPILE_COUNTER = itertools.count() # Reimplementation of part of CopySlices::apply in Python. # The shared code is really similar so we're not going to try to deduplicate. def copy_slices_prologue( inputs: Sequence[torch.Tensor], base_sizes: Sequence[IntLikeType], base_strides: Sequence[IntLikeType], base_storage_offset: IntLikeType, view_sizes: Sequence[IntLikeType], view_strides: Sequence[IntLikeType], view_storage_offset: IntLikeType, ) -> list[torch.Tensor]: grad = inputs[0] result = grad.new_empty_strided(base_sizes, base_strides) if grad is None: raise AssertionError("grad must not be None in copy_slices_prologue") result.copy_(grad) offset = view_storage_offset - base_storage_offset grad_slice = result.as_strided(view_sizes, view_strides, offset) return [result, grad_slice, grad_slice.clone(memory_format=torch.contiguous_format)] # Reimplementation of part of CopySlices::apply in Python. # The shared code is really similar so we're not going to try to deduplicate. def copy_slices_epilogue( needs_input_grad: Sequence[bool], result: torch.Tensor, res: Sequence[torch.Tensor | None], grad_slice: torch.Tensor, ) -> list[torch.Tensor | None]: grad_inputs: list[torch.Tensor | None] = [None] * len(needs_input_grad) for i in range(len(needs_input_grad)): if needs_input_grad[i]: if res[i] is None: continue if i == 0: to_copy = res[i] if to_copy is None: raise AssertionError( "res[0] must not be None for needs_input_grad[0]" ) grad_slice.copy_(to_copy) grad_inputs[i] = result else: grad_inputs[i] = res[i] return grad_inputs