# mypy: ignore-errors import contextlib import functools import inspect import itertools import logging import types from typing import Dict, List, Optional, TYPE_CHECKING import torch._C import torch.fx import torch.nn import torch.onnx.operators from torch._dynamo.utils import get_fake_value from torch._dynamo.variables import ConstantVariable from torch._dynamo.variables.base import VariableTracker from torch._dynamo.variables.builtin import BuiltinVariable from torch._dynamo.variables.functions import UserFunctionVariable from torch._dynamo.variables.tensor import SymNodeVariable from torch._guards import Source from torch._ops import HigherOrderOperator from torch.fx.passes.shape_prop import _extract_tensor_metadata from torch.utils import _pytree as pytree from .. import variables from ..exc import ( IncorrectUsage, UncapturedHigherOrderOpError, unimplemented, Unsupported, ) from ..source import AttrSource from ..utils import proxy_args_kwargs from .dicts import ConstDictVariable from .lazy import LazyVariableTracker from .lists import ListVariable, TupleVariable if TYPE_CHECKING: from torch._dynamo.symbolic_convert import InstructionTranslator log = logging.getLogger(__name__) def raise_hard_error_if_graph_break(reason): def deco(fn): @functools.wraps(fn) def graph_break_as_hard_error(*args, **kwargs): try: return fn(*args, **kwargs) except Unsupported as e: msg = " Scroll up to find out what causes the graph break." raise UncapturedHigherOrderOpError(reason + msg) from e return graph_break_as_hard_error return deco @contextlib.contextmanager def dynamo_enable_grad(tx: "InstructionTranslator", enable=True): from . import GradModeVariable org_value = torch.is_grad_enabled() try: GradModeVariable.create(tx, enable, initialized=True) yield finally: GradModeVariable.create(tx, org_value, initialized=True) def only_consist_of(var, types, allow_none=False): if isinstance(var, types): return True if allow_none and var.is_python_constant() and var.as_python_constant() is None: return True if isinstance(var, (TupleVariable, ListVariable)): return all(only_consist_of(item, types, allow_none) for item in var.items) if isinstance(var, ConstDictVariable): return all( only_consist_of(item, types, allow_none) for item in var.items.values() ) return False # A more read-able syntax sugar for creating a UserFunctionVariable for f # and run call_function on it. Make it return a function to preserve the calling # convention of the original f. def _make_inlined(tx: "InstructionTranslator", f): assert callable(f), "Expect f to be a python callable." def inline_call(*args, **kwargs): return UserFunctionVariable(f).call_function(tx, args, kwargs) return inline_call def _call_function_and_unflatten_output( tx, fn, args, kwargs, flat_example_value, ret_treespec ): from .builder import wrap_fx_proxy # Store the invocation as a call flat_variable = wrap_fx_proxy( tx=tx, proxy=tx.output.create_proxy( "call_function", fn, args=args, kwargs=kwargs, ), example_value=flat_example_value, ) # Transform variable back into a list (previously made into a tuple by # speculate_subgraph function) so as to respect the pytree API typing. flat_list_variable = BuiltinVariable(list).call_function(tx, [flat_variable], {}) return ( _make_inlined(tx, pytree.tree_unflatten)(flat_list_variable, ret_treespec) if ret_treespec else flat_variable ) def _assert_tensors_nonaliasing(inputs, outputs): input_tensor_ids = { id(t) for t in pytree.tree_leaves(inputs) if isinstance(t, torch.Tensor) } output_tensor_ids = { id(t) for t in pytree.tree_leaves(outputs) if isinstance(t, torch.Tensor) } assert input_tensor_ids.isdisjoint( output_tensor_ids ), "inputs to function body cannot alias outputs" def _check_supported_callable_arg( tx: "InstructionTranslator", func_var: VariableTracker, arg_name ): is_callable = ( BuiltinVariable(callable).call_function(tx, [func_var], {}).as_python_constant() ) if not is_callable: unimplemented(f"{arg_name} is of unsupported callable type {str(func_var)}.") def validate_args_and_maybe_create_graph_inputs( sub_args, tracer, tx, set_subgraph_inputs, description, sub_args_names=None, ): from . import AutogradFunctionContextVariable from .builder import wrap_fx_proxy_cls assert tracer.parent is not None if set_subgraph_inputs == "flatten_manual": flat_args, tree_spec = _make_inlined(tx, pytree.tree_flatten)( ListVariable(sub_args) ).unpack_var_sequence(tx) flat_inputs = validate_args_and_maybe_create_graph_inputs( flat_args.unpack_var_sequence(tx), tracer, tx, set_subgraph_inputs="manual", description=description, ) return _make_inlined(tx, pytree.tree_unflatten)( ListVariable(flat_inputs), tree_spec ).unpack_var_sequence(tx) else: if sub_args_names is not None: # Can be greater if user passes some args as kwargs assert len(sub_args_names) >= len(sub_args) args = [] for idx, a in enumerate(sub_args): assert isinstance(a, VariableTracker) if set_subgraph_inputs == "automatic": args.append(a) continue elif set_subgraph_inputs == "semi_automatic": if isinstance(a, AutogradFunctionContextVariable): arg_name = ( a.as_proxy().node.name if sub_args_names is None else sub_args_names[idx] ) tracer.create_graph_input(arg_name) elif a.maybe_fx_node() is not None: node = a.maybe_fx_node() arg_name = ( a.as_proxy().node.name if sub_args_names is None else sub_args_names[idx] ) new_proxy = tracer.create_graph_input(arg_name) example_value = ( node.meta["example_value"] if "example_value" in node.meta else None ) a = wrap_fx_proxy_cls( target_cls=type(a), tx=tx, proxy=new_proxy, example_value=example_value, ) args.append(a) continue if a.is_python_constant(): # This arg is not used in the body of the higher order op. # Currently, this new input is added to make the calls # happy, which expect a fixed number of arguments. In # future, we can clean this up. arg_name = ( "const_unused" if sub_args_names is None else f"const_unused_{sub_args_names[idx]}" ) tracer.create_graph_input(arg_name) new_arg = a # Weird special case, we probably want to delete it or fold it # into the next case (of `a` being placeable into a graph) elif isinstance(a, AutogradFunctionContextVariable): arg_name = ( a.as_proxy().node.name if sub_args_names is None else sub_args_names[idx] ) tracer.create_graph_input(arg_name) new_arg = a # If `a` can be put into a graph elif a.maybe_fx_node() is not None: node = a.maybe_fx_node() arg_name = node.name if sub_args_names is None else sub_args_names[idx] new_proxy = tracer.create_graph_input(arg_name) example_value = ( node.meta["example_value"] if "example_value" in node.meta else None ) new_arg = wrap_fx_proxy_cls( target_cls=type(a), tx=tx, proxy=new_proxy, example_value=example_value, ) # If `a` cannot be put into a graph else: # HOPs work much better if they use speculate_subgraph(set_subgraph_inputs="automatic"). unimplemented( f"{description} with body that accepts non-Tensors as input. " f"Got: {a.python_type()}" ) args.append(new_arg) return args # This helper function is used to make sure two graphs share the same input signature. For example, # in torch.cond, two branches might lift different set of tensors as inputs. This function helps to # dedup the inputs and modify the graphs to take the same set of inputs. def _merge_graph_inputs( l_graph, l_lifted_freevars, l_name, r_graph, r_lifted_freevars, r_name ): def dedup_and_sort_lifted_freevars(l_lifted_freevars, r_lifted_freevars): # The nn module attributes are guaranteed to be registered into the top-level graph module during # higher order op speculation. Therefore, get_attr nodes in two branches with the same # target refer to the same attribute and we can safely deduplicate them with their target. # # Note: ideally, dynamo should just create a single proxy for the same attribute of a nn module. But # true_branch and false_branch belong to two separate tracing contexts, they may register the same # attribute to top level seperately. This creates two get_attr proxies for the same attribute # that have different meta data such as stack_trace (one stack trace for the true_branch, # and the other for false_branch). It seems better to discard the proxy explicitly in cond # than make dynamo create a single proxy for the same get_attr target. def shared_getattrs(l_lifted_proxies, r_lifted_proxies): true_targets = { proxy.node.target: proxy for proxy in l_lifted_proxies if proxy.node.op == "get_attr" } l_shared_getattrs = {} r_shared_getattrs = {} for false_proxy in r_lifted_proxies: if ( false_proxy.node.op == "get_attr" and false_proxy.node.target in true_targets ): true_proxy = true_targets[false_proxy.node.target] l_shared_getattrs[true_proxy] = true_proxy r_shared_getattrs[false_proxy] = true_proxy return l_shared_getattrs, r_shared_getattrs l_shared_getattrs, r_shared_getattrs = shared_getattrs( l_lifted_freevars.keys(), r_lifted_freevars.keys() ) l_shared_freevars = (l_lifted_freevars.keys() & r_lifted_freevars.keys()).union( l_shared_getattrs.keys() ) r_shared_freevars = (l_lifted_freevars.keys() & r_lifted_freevars.keys()).union( r_shared_getattrs.keys() ) unique_l_freevars = l_lifted_freevars.keys() - l_shared_freevars unique_r_freevars = r_lifted_freevars.keys() - r_shared_freevars def _sort_by_name(vars): return sorted(vars, key=lambda var: var.node.name) return ( list(_sort_by_name(list(l_shared_freevars))), list(_sort_by_name(list(r_shared_freevars))), list(_sort_by_name(list(unique_l_freevars))), list(_sort_by_name(list(unique_r_freevars))), ) (l_shared, r_shared, unique_l, unique_r) = dedup_and_sort_lifted_freevars( l_lifted_freevars, r_lifted_freevars ) # Let's say we capture cond(pred, true_fn, false_fn, (x,)) # With set_graph_input set to automatic, # true_fn has lifted variables x, a, b, c # false_fn has lifted variables x, a, b, d # Then fixup_branch_inps make sure both branches have the same signature, i.e.: # - true_fn(x, a, b, c_true_branch, d_false_branch) # - false_fn(x, a, b, c_true_branch, d_false_branch) # # More formally, the signature has three parts in the following order: # 1. used in both branches: x, a, b # 2. only used in true branches: c, suffixed with _true_branch # 3. only used in false branches: d, suffixed with _false_branch # Within each part, we re-order the nodes by name to have a derterministic ordering for testing. def fixup_branch_inps(graph, lifted_freevars, shared, unique_l, unique_r): def _insert_or_replace_phs(new_args, name_suffix): for arg in new_args: new_ph = graph.placeholder(arg.node.name + name_suffix) # Override with new_ph if there exists a old placeholder. if arg in lifted_freevars: old_ph = lifted_freevars[arg].node old_ph.replace_all_uses_with(new_ph) # replace_all_uses_with doesn't clean users. Clean it mannually so that we could erase it. old_ph.users = {} graph.erase_node(old_ph) first_not_ph_node = next( node for node in graph.nodes if node.op != "placeholder" ) with graph.inserting_before(first_not_ph_node): _insert_or_replace_phs(shared, "") _insert_or_replace_phs(unique_l, "_" + l_name) _insert_or_replace_phs(unique_r, "_" + r_name) fixup_branch_inps(l_graph, l_lifted_freevars, l_shared, unique_l, unique_r) fixup_branch_inps(r_graph, r_lifted_freevars, r_shared, unique_l, unique_r) return l_graph, r_graph, l_shared, r_shared, unique_l, unique_r # See NOTE [HigherOrderOperator tracing design] for details of the design def speculate_subgraph( tx, f, sub_args, sub_kwargs, description, *, # source_target is the .value of HigherOrderOpVariable and is the # target of the proxy that we created for the higherOrderOperator. source_target=None, always_restore=False, enable_grad=None, # NOTE [argument `set_subgraph_inputs`] # set_subgraph_inputs controls what how to construct subgraphs' placeholders from sub_args. # 1. if your HOP supports arbitrary inputs, use set_subgraph_inputs="automatic" (most recommended). # 2. if your HOP supports only Tensor and symnode inputs, use set_subgraph_inputs="flatten_manual" (recommended). # If sub_args contain Pytree structure (e.g. dict/list/tuple/set), the sub_args will be flattened first. # Then the flattened args are manually set as subgraph's placeholders. # 3. if your HOP must preserve inputs that are not tensor or symnode as placeholders e.g. AutogradFunctionContextVariable # use set_subgraph_inputs="manual" (not recommended). We do not recommend it in general because it has the # restriction that user need to manually control how to create placeholders and VariableTrackers for the args. set_subgraph_inputs="automatic", restore_side_effects=True, should_flatten_outputs=False, # Pass in an originating tracer - this is needed for preserving context # across fwd-bwd for autograd.Function tracer=None, ): if sub_kwargs is None: sub_kwargs = {} assert set_subgraph_inputs in { "automatic", "semi_automatic", "flatten_manual", "manual", }, "Please use one of the supported set_subgraph_inputs options." # See NOTE [Temporary argument `set_subgraph_inputs`] if sub_kwargs and set_subgraph_inputs != "automatic": unimplemented("Use `set_subgraph_inputs=automatic` when passing `sub_kwargs`.") try: # ensure guards on args get installed in parent subgraph f, sub_args, sub_kwargs = LazyVariableTracker.realize_all( (f, sub_args, sub_kwargs), ) with tx.output.subtracer(source_target, tracer) as subtracer: sub_args_names = maybe_positional_arg_names(f) # User mismatch in the number of args. Will eventually lead to an error. if sub_args_names is not None and len(sub_args_names) < len(sub_args): sub_args_names = None args = validate_args_and_maybe_create_graph_inputs( sub_args, subtracer, tx, set_subgraph_inputs, description, sub_args_names, ) validate_args_and_maybe_create_graph_inputs( sub_kwargs.values(), subtracer, tx, set_subgraph_inputs="automatic", description=description, ) autograd_ctx = ( dynamo_enable_grad(tx, enable_grad) if enable_grad is not None else contextlib.nullcontext() ) # For handling side effects, we can make an argument that we don't # have to do anything here. The side effects infra does a good job # of graph breaking if we mutate any nonlocal or global variable # while subtracing. As a result if tracing succeeds, side effects # data structure will only contain read-only data structures that # are put there for tracking purposes. # But on the other hand, there is an argument that if we ever write # a new side effect in Dynamo which does not go through the side # effect infra, we can end up in bad state. # Therefore we restore the side effects after tracing. The catch is # that we have to special handle tensor variables. If we have seen a # nonlocal variable tensor during subtracing, we want to keep a # track of that tensor, so that later subtracing or the root tracer # itself does not create a new proxy for the already observed tensor # variable. if restore_side_effects: prev_side_effects = tx.output.side_effects.clone() with autograd_ctx: output = f.call_function(tx, args, sub_kwargs) if restore_side_effects: new_side_effects = tx.output.side_effects.clone() prev_side_effects.track_tensor_variables_from_runahead_side_effects( new_side_effects ) tx.output.side_effects = prev_side_effects treespec = None if should_flatten_outputs: # Flatten the speculated subgraph output. output, treespec = _make_inlined(tx, pytree.tree_flatten)( output ).unpack_var_sequence(tx) # Actually, transform the list (returned by flatten) into a tuple # for dynamo consistency. output = BuiltinVariable(tuple).call_function(tx, [output], {}) # Register output to graph # Modeled off of compile_and_call_fx_graph # TODO: support pytree output # We check always_restore because we dont use the output or side effects of always_restore code, # like bwd. if always_restore: # Nothing left to do here return (output, treespec), tx.output.graph, subtracer.lifted_freevars else: from . import TensorVariable if not only_consist_of(output, TensorVariable, allow_none=True): unimplemented( "HigherOrderOperator body's output must consist of tensors only" ) # The output proxies might not belong to this SubgraphTracer # (if they are free variables that were never lifted) # so lift them here. output_proxies = output.as_proxy() output_proxies = pytree.tree_map( subtracer.maybe_lift_tracked_freevar_to_input, output_proxies ) tx.output.create_node( "output", "output", (subtracer.create_arg((output_proxies,))), {}, ) graph = tx.output.graph graph.lint() lifted_freevars = subtracer.lifted_freevars return ( (output, treespec), graph, lifted_freevars, ) except Unsupported as ex: f_name = f"{type(f).__name__}" if isinstance(f, UserFunctionVariable): f_name = f.get_name() msg = ( f"speculate_subgraph: while introspecting {description}, we were unable " f"to trace function `{f_name}` into a single graph. This means " f"that Dynamo was unable to prove safety for this API and will " f"fall back to eager-mode PyTorch, which could lead to a slowdown." ) log.info(msg) log.info(ex) raise ex def make_attr(tx: "InstructionTranslator", name): node = tx.output.create_proxy( "get_attr", name, (), {}, ) return node def add_subgraph(tx: "InstructionTranslator", name, gm): next_name = None i = 0 while not next_name: candidate = f"{name}_{i}" if candidate in tx.output.nn_modules: i += 1 else: next_name = candidate gm.__name__ = next_name gm.torchdynamo_force_dynamic = False # This graph module is not present in the user space, so it can't be # accessed by a source. Set source=None. tx.output.register_attr_or_module(gm, next_name, source=None) return next_name class TorchHigherOrderOperatorVariable(VariableTracker): def __init__( self, value: HigherOrderOperator, source: Optional[Source] = None, **kwargs ) -> None: super().__init__(**kwargs) self.value = value self.source = source @staticmethod def make(value, source=None, **kwargs): if value.__name__ == "cond": return CondHigherOrderVariable(value, source, **kwargs) elif value.__name__ == "while_loop": return WhileLoopHigherOrderVariable(value, source, **kwargs) elif value.__name__ in ("map", "map_impl"): return MapHigherOrderVariable(value, source, **kwargs) elif value.__name__ == "executorch_call_delegate": return ExecutorchCallDelegateHigherOrderVariable(value, source, **kwargs) elif value.__name__ == "out_dtype": return OutDtypeHigherOrderVariable(value, source, **kwargs) elif value.__name__ == "wrap": return WrapHigherOrderVariable(value, source, **kwargs) elif value.__name__ == "hints_wrapper": return HintsWrapperHigherOrderVariable(value, source, **kwargs) elif value.__name__ == "flex_attention": return FlexAttentionHigherOrderVariable(value, source, **kwargs) elif value.__name__ in ( "wrap_activation_checkpoint", "tag_activation_checkpoint", ): return CheckpointHigherOrderVariable(value, source, **kwargs) elif value.__name__ == "_export_tracepoint": return ExportTracepointHigherOrderVariable(value, source, **kwargs) elif value.__name__ == "trace_wrapped": return TraceWrappedHigherOrderOperatorVariable(value, source, **kwargs) elif value.__name__ == "strict_mode": return StrictModeHigherOrderVariable(value, source, **kwargs) elif value.__name__ == "run_with_rng_state": return RunWithRNGStateHigherOrderVariable(value, source, **kwargs) elif value.__name__ == "associative_scan": return AssociativeScanHigherOrderVariable(value, source, **kwargs) elif value.__name__ == "call_torchbind": return CallTorchbindHigherOrderVariable(value, source, **kwargs) elif value.__name__ == "wrap_with_set_grad_enabled": return WrapWithSetGradEnabledHigherOrderVariable(value, source, **kwargs) elif value.__name__ == "auto_functionalized": return AutoFunctionalizeHigherOrderVariable(value, source, **kwargs) else: unimplemented(f"HigherOrderOperator {value.__name__}") def call_function( self, tx: "InstructionTranslator", args: List[VariableTracker], kwargs: Dict[str, VariableTracker], ) -> VariableTracker: unimplemented(f"HigherOrderOperator {self.value.__name__}") class CondHigherOrderVariable(TorchHigherOrderOperatorVariable): @raise_hard_error_if_graph_break( reason="Cond doesn't work unless it is captured completely with torch.compile." ) def call_function( self, tx: "InstructionTranslator", args: "List[VariableTracker]", kwargs: "Dict[str, VariableTracker]", ) -> "VariableTracker": from . import ListVariable, TensorVariable args, kwargs = LazyVariableTracker.realize_all((args, kwargs)) for i, k in enumerate(["pred", "true_fn", "false_fn", "operands"]): if v := kwargs.pop(k, None): assert i == len( args ), "did not provide the right number of non-keyword args" args.append(v) if kwargs: unimplemented(f"torch.cond: Got unexpected kwargs: {list(kwargs.keys())}") # TODO(voz): Support fake tensor dispatch for recursive # ops - see torch/dispatch/_dispatcher.py if len(args) != 4: unimplemented( f"Expected 4 arguments but got {len(args)}.\n" f"Usage: cond(pred, true_fn, false_fn, operands)", ) # Specialize into one of the branches since pred is constant if type(args[0]) is ConstantVariable: log.warning( "Pred is a Python constant. When used with torch.cond, it executes only one of the branches." " If you want torch.cond to perserve two branches, please make the predicate a boolean tensor or a SymBool." ) if args[0].as_python_constant(): return args[1].call_function(tx, args[3].unpack_var_sequence(tx), {}) else: return args[2].call_function(tx, args[3].unpack_var_sequence(tx), {}) # predicate if type(args[0]) not in (ConstantVariable, TensorVariable, SymNodeVariable): unimplemented( f"Expected pred to be bool or a boolean tensor with single " f"item but got {str(type(args[0]))} " f"with original python type {str(args[0].python_type())}.", ) # operands if not isinstance(args[3], (ListVariable, TupleVariable)): unimplemented( f"Expected a tuple but got {args[3].python_type()}", ) operands = args[3].unpack_var_sequence(tx) if not only_consist_of(args[3], (TensorVariable,)): unimplemented( "Expect operands to be a tuple of pytrees that only consists of tensor leaves." ) # branches _check_supported_callable_arg(tx, args[1], "true_fn") _check_supported_callable_arg(tx, args[2], "false_fn") # Our strategy for tracing the true/false branches of cond # are to checkpoint our graphstate, run the true branch, # roll it back to the checkpoint, and run the false # branch, and then merge the graphstates. Well, perhaps # "merge" is too strong a word: we mostly assert that # the resulting graphstates have to be the same. # # We only permit guards to diverge (we union the guards from # both branches). In particular, this means that side # effects are NOT permitted inside true/false branches; this # would be difficult to implement, because of the path # explosion problem. def speculate_branch(branch): # NB: 0 is predicate ix = 1 if branch else 2 # TODO: Support kwargs ( (ret_val, ret_treespec), ret_graph, ret_lifted_freevars, ) = speculate_subgraph( tx, args[ix], operands, {}, "cond", source_target=self.value, should_flatten_outputs=True, ) if not only_consist_of(ret_val, (TensorVariable,)): unimplemented( "Expected branches to return a possibly nested list/tuple/dict of tensors but it consists of non tensors.", ) return ret_val, ret_treespec, ret_graph, ret_lifted_freevars (true_r, true_treespec, true_graph, true_lifted_freevars) = speculate_branch( True ) true_nn_modules = dict(tx.output.nn_modules) ( false_r, false_treespec, false_graph, false_lifted_freevars, ) = speculate_branch(False) false_nn_modules = dict(tx.output.nn_modules) same_treespec = _make_inlined(tx, pytree.TreeSpec.__eq__)( true_treespec, false_treespec ) if not same_treespec.as_python_constant(): unimplemented("Expected branches to return the same pytree structure.") def diff_meta(tensor_vars1, tensor_vars2): assert all( isinstance(var, TensorVariable) for var in tensor_vars1 + tensor_vars2 ) all_diffs = [] for i, (var1, var2) in enumerate(zip(tensor_vars1, tensor_vars2)): # We check the meta data associated with meta["example_value"] meta1 = _extract_tensor_metadata( var1.proxy.node.meta["example_value"], include_contiguity=False ) meta2 = _extract_tensor_metadata( var2.proxy.node.meta["example_value"], include_contiguity=False ) if meta1 != meta2: all_diffs.append((f"pair{i}:", meta1, meta2)) return all_diffs if diffs := diff_meta( true_r.unpack_var_sequence(tx), false_r.unpack_var_sequence(tx) ): unimplemented( f"Expected branches to return tensors with same metadata. [(tensor_pair, difference)...]:{diffs}" ) ( true_graph, false_graph, true_shared, false_shared, unique_true, unique_false, ) = _merge_graph_inputs( true_graph, true_lifted_freevars, "true_branch", false_graph, false_lifted_freevars, "false_branch", ) true_name = add_subgraph( tx, "cond_true", torch.fx.GraphModule(true_nn_modules, true_graph), ) false_name = add_subgraph( tx, "cond_false", torch.fx.GraphModule(false_nn_modules, false_graph), ) true_node = make_attr(tx, true_name) false_node = make_attr(tx, false_name) p_args = ( args[0].as_proxy(), true_node, false_node, # We pick true_shared but it shouldn't matter true_shared + unique_true + unique_false, ) flat_example_value = pytree.tree_map_only( torch.fx.Proxy, lambda a: a.node.meta["example_value"], true_r.as_proxy(), ) return _call_function_and_unflatten_output( tx, torch.ops.higher_order.cond, p_args, {}, flat_example_value, true_treespec, ) class CallTorchbindHigherOrderVariable(TorchHigherOrderOperatorVariable): def __init__(self, hop, source, script_obj_var, method_name) -> None: super().__init__(hop, source) self.script_obj_var = script_obj_var self.method_name = method_name def call_function( self, tx: "InstructionTranslator", args: List[VariableTracker], kwargs: Dict[str, VariableTracker], ) -> VariableTracker: from .builder import wrap_fx_proxy args, kwargs = LazyVariableTracker.realize_all((args, kwargs)) args_proxy = [arg.as_proxy() for arg in args] kwargs_proxy = {k: v.as_proxy() for k, v in kwargs.items()} return wrap_fx_proxy( tx=tx, proxy=tx.output.create_proxy( "call_function", self.value, args=tuple( [self.script_obj_var.as_proxy(), self.method_name] + args_proxy ), kwargs=kwargs_proxy, ), ) class WhileLoopHigherOrderVariable(TorchHigherOrderOperatorVariable): @raise_hard_error_if_graph_break( reason="while_loop doesn't work unless it is captured completely with torch.compile." ) def call_function( self, tx: "InstructionTranslator", args: List[VariableTracker], kwargs: Dict[str, VariableTracker], ) -> VariableTracker: from . import TensorVariable args, kwargs = LazyVariableTracker.realize_all((args, kwargs)) for i, k in enumerate(["cond_fn", "body_fn", "operands"]): if v := kwargs.pop(k, None): assert i == len( args ), "did not provide the right number of non-keyword args" args.append(v) if kwargs: unimplemented( f"torch.while_loop: Got unexpected kwargs: {list(kwargs.keys())}" ) if len(args) != 4: unimplemented( f"Expected 4 arguments but got {len(args)}.\n" f"Usage: while_loop(cond_fn, body_fn, operands)", ) _check_supported_callable_arg(tx, args[0], "cond_fn") _check_supported_callable_arg(tx, args[1], "body_fn") # operands if not isinstance(args[2], (ListVariable, TupleVariable)): unimplemented( f"Expected a tuple but got {args[2].python_type()}", ) operands = args[2].unpack_var_sequence(tx) if not only_consist_of(args[2], (TensorVariable,)): unimplemented( "Expect operands to be a tuple of pytrees that only consists of tensor leaves." ) # additional inputs check if not isinstance(args[3], (ListVariable, TupleVariable)): unimplemented( f"Expected a tuple but got {args[3].python_type()}", ) additional_inputs = args[3].unpack_var_sequence(tx) ( (cond_r, cond_treespec), cond_graph, cond_lifted_freevars, ) = speculate_subgraph( tx, args[0], operands + additional_inputs, {}, "while_loop", source_target=self.value, set_subgraph_inputs="manual", ) cond_nn_modules = dict(tx.output.nn_modules) if not isinstance(cond_r, TensorVariable): unimplemented( f"Expected cond_fn to return a tensor but got {cond_r.python_type()}", ) cond_r_meta = _extract_tensor_metadata( cond_r.proxy.node.meta["example_value"], include_contiguity=False ) if not cond_r_meta.dtype == torch.bool or not cond_r_meta.shape == torch.Size( [] ): unimplemented( f"Expected cond_fn to return a tensor with shape (,) but got {cond_r_meta.shape}" ) ( (body_r, body_treespec), body_graph, body_lifted_freevars, ) = speculate_subgraph( tx, args[1], operands + additional_inputs, {}, "while_loop", source_target=self.value, set_subgraph_inputs="manual", should_flatten_outputs=True, ) ( cond_graph, body_graph, cond_shared, body_shared, cond_unique, body_unique, ) = _merge_graph_inputs( cond_graph, cond_lifted_freevars, "cond_fn", body_graph, body_lifted_freevars, "body_fn", ) # Note: cond_shared and body_shared refer to the same proxy in parent graph # so using either of them is OK. Use cond_shared as it doesnt matter. additional_lifted_inputs = cond_shared + cond_unique + body_unique body_nn_modules = dict(tx.output.nn_modules) cond_name = add_subgraph( tx, "cond_fn", torch.fx.GraphModule(cond_nn_modules, cond_graph), ) body_name = add_subgraph( tx, "body_fn", torch.fx.GraphModule(body_nn_modules, body_graph), ) cond_node = make_attr(tx, cond_name) body_node = make_attr(tx, body_name) p_args = ( cond_node, body_node, tuple([operand.as_proxy() for operand in operands]), tuple( [inp.as_proxy() for inp in additional_inputs] + additional_lifted_inputs ), ) flat_example_value = pytree.tree_map_only( torch.fx.Proxy, lambda a: a.node.meta["example_value"], body_r.as_proxy(), ) return _call_function_and_unflatten_output( tx, torch.ops.higher_order.while_loop, p_args, {}, flat_example_value, body_treespec, ) class AssociativeScanHigherOrderVariable(TorchHigherOrderOperatorVariable): @raise_hard_error_if_graph_break( reason="associative_scan must be captured completely with torch.compile." ) def call_function( self, tx: "InstructionTranslator", args: List[VariableTracker], kwargs: Dict[str, VariableTracker], ) -> VariableTracker: from .builder import SourcelessBuilder, wrap_fx_proxy args, kwargs = LazyVariableTracker.realize_all((args, kwargs)) def arg_extractor(combine_fn, input, dim): return combine_fn, input, dim combine_fn, input, dim = arg_extractor(*args, **kwargs) if input.python_type() != list: unimplemented( f"Expected input to be a list of tensors but got {input.python_type()}", ) assert isinstance(input, torch._dynamo.variables.lists.BaseListVariable) # Trace the subgraph # TODO: Fix these pointless new_empty calls appearing in the dynamo output graph. sub_args = [ leaf.call_method( tx, "new_empty", args=( SourcelessBuilder.create( tx, leaf.size if leaf.size is not None else BuiltinVariable(getattr) .call_function(tx, [leaf, ConstantVariable.create("shape")], {}) .items, ), ), kwargs={ "dtype": SourcelessBuilder.create(tx, leaf.dtype), "requires_grad": SourcelessBuilder.create(tx, leaf.requires_grad), }, ) for leaf in itertools.chain(input.items, input.items) ] ( (combine_result, combine_treespec), combine_graph, combine_lifted_freevars, ) = speculate_subgraph( tx, combine_fn, sub_args, sub_kwargs={}, description="scan_combine", source_target=self.value, set_subgraph_inputs="flatten_manual", ) if combine_lifted_freevars: unimplemented( f"Combine fn had unexpected freevars: {combine_lifted_freevars}" ) if combine_result.python_type() != list: unimplemented( f"Expected combine_fn to return a list if tensor but got {combine_result.python_type()}", ) input_proxy = input.as_proxy() combine_result_proxy = combine_result.as_proxy() for result, inp_proxy in zip(combine_result_proxy, input_proxy): inp_meta = inp_proxy.node.meta["example_value"] combine_result_meta = result.node.meta["example_value"] if combine_result_meta.device != inp_meta.device: unimplemented( f"Expected combine_fn to return a tensor on device {inp_meta.device} but " + f"got {combine_result_meta.device}" ) if combine_result_meta.dtype != inp_meta.dtype: unimplemented( f"Expected combine_fn to return a tensor of {inp_meta.dtype} but " + f"got {combine_result_meta.dtype}" ) combine_gm = torch.fx.GraphModule(dict(tx.output.nn_modules), combine_graph) combine_fn_name = add_subgraph(tx, "scan_combine", combine_gm) p_args = ( make_attr(tx, combine_fn_name), input_proxy, dim.as_proxy(), ) with tx.fake_mode: out_meta = tuple( inp_proxy.node.meta["example_value"].clone() for inp_proxy in input_proxy ) return wrap_fx_proxy( tx=tx, proxy=tx.output.create_proxy( "call_function", torch.ops.higher_order.associative_scan, p_args, {} ), example_value=out_meta, ) def non_single_tensor_return_unsupported(api, ret): from . import TensorVariable if not isinstance(ret, TensorVariable): raise Unsupported( f"{api} over function that returns something " f"other than one Tensor" ) class MapHigherOrderVariable(TorchHigherOrderOperatorVariable): def call_function( self, tx: "InstructionTranslator", args: List[VariableTracker], kwargs: Dict[str, VariableTracker], ) -> VariableTracker: from . import TensorVariable from .builder import wrap_fx_proxy_cls if len(kwargs) > 0: unimplemented( "torch.ops.higher_order.map: kwargs are not supported in the map operator." ) _check_supported_callable_arg(tx, args[0].realize(), "map_fn") assert type(args[1].realize()) is TensorVariable sample_shape = get_fake_value(args[1].as_proxy().node, tx).size() if len(sample_shape) < 1 or sample_shape[0] == 0: unimplemented( "map() operator doesn't support scalar or zero-sized tensors during tracing." ) # To get the example output from map() we will need to provide at least one sample to # the loop body. In our case we will always use xs[0], and our map() won't support zero # sized tensor during tracing. first_dim = wrap_fx_proxy_cls( target_cls=TensorVariable, tx=tx, proxy=args[1].as_proxy()[0] ) # TODO: Support kwargs ( (body_r, body_spec), body_graph, body_lifted_freevars, ) = speculate_subgraph( tx, args[0], [ first_dim, *args[2:], ], {}, "torch.ops.higher_order.map", source_target=self.value, set_subgraph_inputs="flatten_manual", should_flatten_outputs=True, ) subgraph_example_value = [ proxy.node.meta["example_value"] for proxy in body_r.as_proxy() ] with tx.output.fake_mode: # We need to expand the example output from map() so that it has # the same first dimension as the mapped input. # We also do a clone with contiguous_format. This is to be consistent with # eager semantic of map, which stacks the outputs. The result is contiguous # as a result of the stack operation. map_example_out = [ t.expand(sample_shape[0], *t.size()).clone( memory_format=torch.contiguous_format ) for t in subgraph_example_value ] body_nn_modules = dict(tx.output.nn_modules) body_name = add_subgraph( tx, "map_body", torch.fx.GraphModule(body_nn_modules, body_graph), ) body_node = make_attr(tx, body_name) p_args = ( body_node, [args[1].as_proxy()], [arg.as_proxy() for arg in args[2:]] + list(body_lifted_freevars.keys()), ) return _call_function_and_unflatten_output( tx, torch.ops.higher_order.map_impl, p_args, {}, map_example_out, body_spec ) class ExecutorchCallDelegateHigherOrderVariable(TorchHigherOrderOperatorVariable): def call_function( self, tx: "InstructionTranslator", args: "List[VariableTracker]", kwargs: "Dict[str, VariableTracker]", ) -> "VariableTracker": from .builder import wrap_fx_proxy # This is operator for delegation within Executorch which calls a # specific function in the given lowered module with the given # operators. The actual operator is defined in the Executorch codebase. # This is a bad hierarchical violation since # executorch_call_delegate sits at a higher level than dynamo, but # there's no real solution to this issue yet. if len(kwargs) > 0: unimplemented( "executorch_call_delegate: kwargs arguments were not enabled." ) lowered_module = tx.output.get_submodule(args[0].module_key) lowered_node = make_attr(tx, args[0].module_key) p_args = tuple(arg.as_proxy() for arg in args[1:]) real_sub_args = pytree.tree_map_only( torch.fx.Proxy, lambda a: get_fake_value(a.node, tx), p_args ) with tx.fake_mode: example_value = lowered_module.original_module.module()(*real_sub_args) # NOTE [Guaranteeing the 1-1 correspondence of FakeTensors and real tensors]: # executorch modules promise not to alias inputs and outputs. # Thus, output FakeTensors will correctly not alias input FakeTensors. _assert_tensors_nonaliasing(real_sub_args, example_value) p_args = (lowered_node,) + p_args # Store the invocation as a call return wrap_fx_proxy( tx=tx, proxy=tx.output.create_proxy( "call_function", self.value, args=tuple(p_args), kwargs={}, ), example_value=example_value, ) class FunctorchHigherOrderVariable(UserFunctionVariable): def call_function( self, tx: "InstructionTranslator", args: "List[VariableTracker]", kwargs: "Dict[str, VariableTracker]", ) -> "VariableTracker": if not torch._dynamo.config.capture_func_transforms: name = self.get_name() fn = { "grad_impl": "grad", "vmap_impl": "vmap", "vjp": "vjp", "jvp": "jvp", "jacrev": "jacrev", "jacfwd": "jacfwd", "hessian": "hessian", "linearize": "linearize", "functional_call": "functional_call", }.get(name) assert name is not None unimplemented( f"torch.func.{fn} capture is disabled, " "it can be turned on by setting " "`torch._dynamo.config.capture_func_transforms=True`" ) return super().call_function(tx, args, kwargs) class FunctionalCallVariable(FunctorchHigherOrderVariable): def call_function( self, tx, args: List[VariableTracker], kwargs: Dict[str, VariableTracker] ) -> VariableTracker: if not torch._dynamo.config.inline_inbuilt_nn_modules: unimplemented( "torch.func.functional_call capture is disabled, " "it can be turned on by setting " "`torch._dynamo.config.inline_inbuilt_nn_modules=True`" ) return super().call_function(tx, args, kwargs) class WrapHigherOrderVariable(TorchHigherOrderOperatorVariable): def create_wrapped_node( self, tx: "InstructionTranslator", args, kwargs, description ): # See NOTE [HigherOrderOperator tracing design] for more details ( (body_r, treespec), body_graph, body_lifted_freevars, ) = speculate_subgraph( tx, args[0], # function [*args[1:]], kwargs, description, source_target=self.value, should_flatten_outputs=True, ) body_gmod = torch.fx.GraphModule(tx.output.nn_modules, body_graph) body_name = add_subgraph( tx, "wrap_body", body_gmod, ) body_node = make_attr(tx, body_name) # Since, we call `speculate_subgraph` with `set_subgraph_inputs="automatic`, # all the arguments are lifted. lifted_args = tuple(arg for arg in body_lifted_freevars.keys()) proxy_args = (body_node,) + lifted_args example_value = pytree.tree_map_only( torch.fx.Proxy, lambda a: a.node.meta["example_value"], body_r.as_proxy(), ) return proxy_args, {}, example_value, body_r, treespec, body_gmod def call_function( self, tx: "InstructionTranslator", args: "List[VariableTracker]", kwargs: "Dict[str, VariableTracker]", ) -> "VariableTracker": # This flattens the kwargs into lifted args p_args, p_kwargs, example_value, body_r, treespec, _ = self.create_wrapped_node( tx, args, kwargs, "wrap" ) if len(p_kwargs) > 0: unimplemented("kwargs should have been flattened into lifted args") flat_example_value = pytree.tree_map_only( torch.fx.Proxy, lambda a: a.node.meta["example_value"], body_r.as_proxy(), ) return _call_function_and_unflatten_output( tx, self.value, tuple(p_args), p_kwargs, flat_example_value, treespec ) class WrapWithSetGradEnabledHigherOrderVariable(TorchHigherOrderOperatorVariable): """ This hop is not exposed to users but is inserted into the graph after export as a post-processing step. """ def call_function( self, tx: "InstructionTranslator", args: "List[VariableTracker]", kwargs: "Dict[str, VariableTracker]", ) -> "VariableTracker": args, kwargs = LazyVariableTracker.realize_all((args, kwargs)) if kwargs: unimplemented( f"wrap_with_set_grad_enabled: Got unexpected kwargs: {list(kwargs.keys())}" ) grad_enabled, fn_var, *rest_args = args if not isinstance(grad_enabled, ConstantVariable): unimplemented("grad_enabled must be a constant") _check_supported_callable_arg(tx, fn_var, "enable_grad_fn") with torch.set_grad_enabled(grad_enabled.as_python_constant()): ( (body_r, treespec), body_graph, body_lifted_freevars, ) = speculate_subgraph( tx, fn_var, [*rest_args], {}, "torch.ops.higher_order.wrap_with_set_grad_enabled", source_target=self.value, set_subgraph_inputs="manual", should_flatten_outputs=True, ) if len(body_lifted_freevars) > 0: unimplemented( f"wrap_with_set_grad_enabled: Got unexpected freevars {body_lifted_freevars}" ) body_gmod = torch.fx.GraphModule(tx.output.nn_modules, body_graph) body_name = add_subgraph( tx, "wrap_body", body_gmod, ) body_node = make_attr(tx, body_name) proxy_args = tuple( [ grad_enabled.as_python_constant(), body_node, ] + [operand.as_proxy() for operand in rest_args] ) example_value = pytree.tree_map_only( torch.fx.Proxy, lambda a: a.node.meta["example_value"], body_r.as_proxy(), ) return _call_function_and_unflatten_output( tx, self.value, proxy_args, {}, example_value, treespec ) class HintsWrapperHigherOrderVariable(TorchHigherOrderOperatorVariable): @raise_hard_error_if_graph_break( reason="Hints_wrapper doesn't work unless it is captured completely with torch.compile." ) def call_function( self, tx, args: "List[VariableTracker]", kwargs: "Dict[str, VariableTracker]" ) -> "VariableTracker": _check_supported_callable_arg(tx, args[0], "body_fn") # inputs if len(args) != 3: unimplemented( f"Expected 3 arguments but got {len(args)}.\n" f"Usage: hints_wrapper(body_fn, args, kwargs, hints).\n" f"kwargs required to be provided explicitly." ) if not isinstance(args[1], (ListVariable, TupleVariable)): unimplemented( f"Expected a tuple but got {args[1].python_type()}", ) operands = args[1].unpack_var_sequence(tx) if not isinstance(args[2], ConstDictVariable): unimplemented( f"Expected a dict but got {args[2].python_type()}", ) if "hints" not in kwargs: raise IncorrectUsage("hints_wrapper - key hints not provided") ( (body_r, treespec), body_graph, body_lifted_freevars, ) = speculate_subgraph( tx, args[0], # function operands, args[2].as_python_constant(), "hints_wrapper", source_target=self.value, should_flatten_outputs=True, ) body_gmod = torch.fx.GraphModule(tx.output.nn_modules, body_graph) body_name = add_subgraph( tx, "hints_wrapper_body", body_gmod, ) body_node = make_attr(tx, body_name) # Since, we call `speculate_subgraph` with `set_subgraph_inputs="automatic`, # all the arguments are lifted. lifted_args = tuple(arg for arg in body_lifted_freevars.keys()) p_args = (body_node, lifted_args, {}) p_kwargs = {} # add hints into p_kwargs p_kwargs["hints"] = kwargs["hints"].as_python_constant() flat_example_value = pytree.tree_map_only( torch.fx.Proxy, lambda a: a.node.meta["example_value"], body_r.as_proxy(), ) return _call_function_and_unflatten_output( tx, self.value, p_args, p_kwargs, flat_example_value, treespec ) class OutDtypeHigherOrderVariable(TorchHigherOrderOperatorVariable): def call_function( self, tx: "InstructionTranslator", args: "List[VariableTracker]", kwargs: "Dict[str, VariableTracker]", ) -> "VariableTracker": from .builder import wrap_fx_proxy if len(kwargs) > 0: unimplemented("out_dtype does not handle kwargs") p_args = tuple(arg.as_proxy() for arg in args) op = p_args[0] output_dtype = p_args[1] fake_sub_args = pytree.tree_map_only( torch.fx.Proxy, lambda a: a.node.meta["example_value"], p_args[2:] ) # This is a simplified implementation of this operator just for tracing. # Actual implementation may also first promote the arguments example_value = op(*fake_sub_args).to(dtype=output_dtype) # Store the invocation as a call return wrap_fx_proxy( tx=tx, proxy=tx.output.create_proxy( "call_function", self.value, args=tuple(p_args), kwargs={}, ), example_value=example_value, ) class StrictModeHigherOrderVariable(TorchHigherOrderOperatorVariable): @raise_hard_error_if_graph_break( reason="strict_mode HOO doesn't work unless it is captured completely with torch.compile." ) def call_function( self, tx: "InstructionTranslator", args: "List[VariableTracker]", kwargs: "Dict[str, VariableTracker]", ) -> "VariableTracker": callable = args[0] unpacked_sequence = args[1].unpack_var_sequence(tx) # TODO (tmanlaibaatar) support pytree here for arg in unpacked_sequence: if isinstance(arg, (ListVariable, TupleVariable, ConstDictVariable)): unimplemented("strict_mode HOO only works for flat inputs for now") if kwargs: unimplemented( f"strict_mode HOO received unexpected kwargs: {list(kwargs.keys())}" ) ( (ret_val, ret_treespec), ret_graph, ret_lifted_freevars, ) = speculate_subgraph( tx, args[0], unpacked_sequence, {}, "strict_mode", source_target=self.value, should_flatten_outputs=True, ) strict_mode_nn_modules = dict(tx.output.nn_modules) strict_mode_name = add_subgraph( tx, "strict_mode_body", torch.fx.GraphModule(strict_mode_nn_modules, ret_graph), ) strict_mode_node = make_attr(tx, strict_mode_name) p_args = ( strict_mode_node, tuple(arg for arg in ret_lifted_freevars.keys()), ) flat_example_value = pytree.tree_map_only( torch.fx.Proxy, lambda a: a.node.meta["example_value"], ret_val.as_proxy(), ) return _call_function_and_unflatten_output( tx, torch.ops.higher_order.strict_mode, p_args, {}, flat_example_value, ret_treespec, ) class CheckpointHigherOrderVariable(WrapHigherOrderVariable): def call_function( self, tx: "InstructionTranslator", args: List[VariableTracker], kwargs: Dict[str, VariableTracker], ) -> VariableTracker: from torch._higher_order_ops.wrap import TagActivationCheckpoint from torch.utils.checkpoint import noop_context_fn from .builder import wrap_fx_proxy context_fn = None if "context_fn" in kwargs and kwargs["context_fn"] != noop_context_fn: ctx = kwargs.pop("context_fn") if isinstance(ctx, torch._dynamo.variables.UserFunctionVariable): context_fn = ctx.fn elif isinstance( ctx, torch._dynamo.variables.functions.FunctoolsPartialVariable ): context_fn = ctx.as_python_constant() else: raise NotImplementedError( f"checkpoint not implemented for {type(ctx)} context_fn" ) checkpoint_kwargs, gmod_kwargs = TagActivationCheckpoint.divide_kwargs(kwargs) # Here we use checkpoint_kwargs (and not gmod kwargs). gmod_kwargs are # already flattened above and managed inside the fx graph. ( p_args, _, example_value, body_r, treespec, checkpointed_gmod, ) = self.create_wrapped_node( tx, args, gmod_kwargs, "torch.utils.checkpoint.checkpoint" ) if context_fn is not None: checkpointed_gmod.meta["_checkpoint_context_fn"] = context_fn _, checkpoint_kwargs = proxy_args_kwargs([], checkpoint_kwargs) # Store the invocation as a call variable = wrap_fx_proxy( tx=tx, proxy=tx.output.create_proxy( "call_function", self.value, args=tuple(p_args), kwargs=checkpoint_kwargs, ), example_value=example_value, ) if treespec is None: return variable # Transform variable back into a list (previously made into a tuple by # speculate_subgraph function) so as to respect the pytree API typing. variable = BuiltinVariable(list).call_function(tx, [variable], {}) return _make_inlined(tx, pytree.tree_unflatten)(variable, treespec) class ExportTracepointHigherOrderVariable(TorchHigherOrderOperatorVariable): def call_function( self, tx: "InstructionTranslator", args: "List[VariableTracker]", kwargs: "Dict[str, VariableTracker]", ) -> "VariableTracker": from .builder import wrap_fx_proxy p_args = tuple(arg.as_proxy() for arg in args) p_kwargs = {key: arg.as_proxy() for key, arg in kwargs.items()} return wrap_fx_proxy( tx=tx, proxy=tx.output.create_proxy( "call_function", self.value, args=p_args, kwargs=p_kwargs, ), example_value=None, ) class RunWithRNGStateHigherOrderVariable(TorchHigherOrderOperatorVariable): def call_function( self, tx: "InstructionTranslator", args: "List[VariableTracker]", kwargs: "Dict[str, VariableTracker]", ) -> "VariableTracker": from .builder import wrap_fx_proxy p_args = tuple(arg.as_proxy() for arg in args) p_kwargs = {key: arg.as_proxy() for key, arg in kwargs.items()} return wrap_fx_proxy( tx=tx, proxy=tx.output.create_proxy( "call_function", self.value, args=p_args, kwargs=p_kwargs, ), example_value=None, ) class AutoFunctionalizeHigherOrderVariable(TorchHigherOrderOperatorVariable): def call_function( self, tx, args: "List[VariableTracker]", kwargs: "Dict[str, VariableTracker]" ) -> "VariableTracker": from .builder import wrap_fx_proxy p_args = tuple(arg.as_proxy() for arg in args) p_kwargs = {key: arg.as_proxy() for key, arg in kwargs.items()} return wrap_fx_proxy( tx=tx, proxy=tx.output.create_proxy( "call_function", self.value, args=p_args, kwargs=p_kwargs, ), example_value=None, ) class TraceWrappedHigherOrderOperatorVariable(TorchHigherOrderOperatorVariable): """ Handles torch._dynamo._trace_wrapped_higher_order_op.inner_trace by unwrapping the higher order op and inlining through it. This op is created by dynamo to survive through AotAutograd, then unwrapped here in the call to dynamo from compiled autograd. """ def call_function( self, tx: "InstructionTranslator", args: "List[VariableTracker]", kwargs: "Dict[str, VariableTracker]", ) -> "VariableTracker": kwargs = dict(kwargs) fn = kwargs.pop("fn") return fn.call_function(tx, args, kwargs) class FlexAttentionHigherOrderVariable(TorchHigherOrderOperatorVariable): @staticmethod def normalize_to_args(args, kwargs): # input signature is (query, key, value, score_mod, block_mask, *other_buffers), # block_mask is a tuple, and we don't want to flatten it. # only flatten kwargs into lists flat_kwargs = pytree.tree_flatten(kwargs)[0] # Combine the flattened lists all_args = args + flat_kwargs return all_args def create_wrapped_node( self, tx: "InstructionTranslator", query: "VariableTracker", fn: "VariableTracker", fn_name: str, ): from torch._higher_order_ops.flex_attention import TransformGetItemToIndex from .builder import SourcelessBuilder tx: InstructionTranslator = tx def create_scalar(): return query.call_method( tx, "new_empty", (SourcelessBuilder.create(tx, []),), { "dtype": SourcelessBuilder.create(tx, torch.int32), }, ) bhmn = [create_scalar() for _ in range(4)] if fn_name == "score_mod": scores_require_grad: bool = query.requires_grad score = query.call_method( tx, "new_empty", (SourcelessBuilder.create(tx, []),), {"requires_grad": SourcelessBuilder.create(tx, scores_require_grad)}, ) new_args = [score, *bhmn] else: assert fn_name == "mask_fn", "Illegal function name: " + fn_name new_args = [*bhmn] with TransformGetItemToIndex(): ( (body_output, body_treespec), body_graph, body_lifted_freevars, ) = speculate_subgraph( tx, fn, new_args, {}, # expect only args no kwargs for now description=fn_name, source_target=self.value, set_subgraph_inputs="flatten_manual", ) body_name = add_subgraph( tx, fn_name, torch.fx.GraphModule(tx.output.nn_modules, body_graph), ) body_node = make_attr(tx, body_name) # It is possible that the score-mod function captures some free variables that are not # passed in as arguments. In this case, we need to lift them, which is handled by speculate_subgraph. # We then need to create proxies for this + the inputs. lifted_args = tuple(arg for arg in body_lifted_freevars.keys()) proxy_args = (body_node, lifted_args) return proxy_args def call_function( self, tx: "InstructionTranslator", args: "List[VariableTracker]", kwargs: "Dict[str, VariableTracker]", ) -> "VariableTracker": from .builder import wrap_fx_proxy ( query, key, value, score_mod, block_mask, scale, kernel_options, ) = self.normalize_to_args(args, kwargs) score_mod_node, score_mod_lifted_args = self.create_wrapped_node( tx, query, score_mod, "score_mod" ) mask_fn = block_mask.items[-1] if isinstance(mask_fn, ConstantVariable): mask_fn = UserFunctionVariable(torch.nn.attention._flex_attention._no_mask) mask_fn_node, mask_fn_lifted_args = self.create_wrapped_node( tx, query, mask_fn, "mask_fn" ) proxied_args = [ query, key, value, TupleVariable(block_mask.items[:-1], source=block_mask.source), scale, kernel_options, ] # Store the invocation as a call # Norm_kwargs contains the score_function and we dont want to proxy this because # Proxying user defined functions is not supported. inp_args, _ = proxy_args_kwargs(proxied_args, {}) query_meta = query.as_proxy().node.meta["example_value"] logsumexp_shape = query_meta.size()[:-1] # [B, H, M] with torch._guards.TracingContext.try_get().fake_mode: out_meta = torch.empty_like( query_meta, memory_format=torch.contiguous_format ) lse_meta = query_meta.new_empty(logsumexp_shape, dtype=torch.float32) example_value = (out_meta, lse_meta) # Compose the ordered HOO args: # - inp_args: [query, key, value, block_mask, scale, kernel_options] # - subgraph node: [score_mod, mask_fn_node] # - lifted args from tracing subgraph: [score_mod_other_buffers, mask_fn_other_buffers] _, _, _, inp_arg_block_mask, inp_arg_scale, inp_arg_kernel_options = inp_args block_mask = tuple(inp_arg_block_mask + (mask_fn_node,)) return wrap_fx_proxy( tx=tx, proxy=tx.output.create_proxy( "call_function", self.value, args=inp_args[:3] + ( score_mod_node, block_mask, inp_arg_scale, inp_arg_kernel_options, score_mod_lifted_args, mask_fn_lifted_args, ), kwargs={}, ), example_value=example_value, ) class AutogradFunctionApplyVariable(VariableTracker): def __init__(self, fwd_graph, bwd_graph, parent_source, **kwargs) -> None: super().__init__(**kwargs) self.fwd_graph = fwd_graph self.bwd_graph = bwd_graph self.parent_source = parent_source def call_function( self, tx: "InstructionTranslator", args: "List[VariableTracker]", kwargs: "Dict[str, VariableTracker]", ) -> "VariableTracker": from . import ( AutogradFunctionContextVariable, UserDefinedClassVariable, UserFunctionVariable, UserMethodVariable, ) from .builder import wrap_fx_proxy """ Consider the following: class MySin(torch.autograd.Function): @staticmethod def forward(ctx, x): ctx.save_for_backward(x) return x.sin() @staticmethod def backward(ctx, grad): x, = ctx.saved_tensors return grad * x.cos() We want the resulting graphs to look like: def fwd(ctx, x): # (output, saved tensors / attrs) return (x.sin(), [x]) # bwd(ctx, grad0, grad1, ..., gradn, *saved_tensors_or_attrs) def bwd(ctx, grad, x): return grad * x.cos() To accomplish this, we're going to: 1. Construct a ctx object 2. (fwd_out, _), fwd_graph, fwd_freevars = speculate_subgraph on MySin.forward (manually_set_inputs=True) 3. (bwd_out, _), bwd_graph, bwd_freevars = speculate_subgraph on MySin.backward, while manually setting the ctx and grad inputs. 4. Manually rewriting the fwd graph's output to be (output, stuff_that_gets_used in bwd_graph) Getting from 3 to 4 is pretty elegant: stuff_that_gets_used in bwd graph is just the bwd_freevars returned from speculate_subgraph, assuming MySin.backward doesn't capture any arguments. All these steps work if MySin.backward doesn't capture any values. This is a limitation in general that we should check for. """ prev_side_effects = tx.output.side_effects.clone() fwd_tracer = torch._dynamo.output_graph.SubgraphTracer( tx.output, parent=tx.output.current_tracer, source_target="autograd.Function", ) fwd_src = AttrSource(self.parent_source, member="forward") ctx = AutogradFunctionContextVariable.create(tx, args, kwargs) if isinstance(self.fwd_graph, types.FunctionType): fwd_fn = UserFunctionVariable(self.fwd_graph) fwd_args = [ctx, *args] elif isinstance(self.fwd_graph, types.MethodType): fwd_fn = UserMethodVariable( self.fwd_graph.__func__, UserDefinedClassVariable(self.fwd_graph.__class__), ) fwd_args = [fwd_fn.obj, ctx, *args] else: unimplemented("non-function or method") # Speculate subgraph on the fwd (fwd_out, _), fwd_graph, fwd_freevars = speculate_subgraph( tx, fwd_fn, fwd_args, kwargs, "autograd.Function", set_subgraph_inputs="semi_automatic", restore_side_effects=False, tracer=fwd_tracer, ) if ctx.mutable_local in tx.output.side_effects.store_attr_mutations: if ( "_materialize_non_diff_grads" in tx.output.side_effects.store_attr_mutations[ctx.mutable_local] ): unimplemented("NYI") bwd_tracer = torch._dynamo.output_graph.SubgraphTracer( tx.output, parent=fwd_tracer, source_target="autograd.Function", ) # Speculate subgraph on the backward. We make the # bwd tracer a child of the fwd tracer, because backward may rely on # tensors/attrs created in the fwd tracer. if isinstance(fwd_out, variables.BaseListVariable): bwd_args = [ctx, *fwd_out.items] else: bwd_args = [ctx, fwd_out] bwd_src = AttrSource(self.parent_source, member="backward") if isinstance(self.bwd_graph, types.FunctionType): bwd_fn = UserFunctionVariable(self.bwd_graph, source=bwd_src) elif isinstance(self.bwd_graph, types.MethodType): bwd_fn = UserMethodVariable( self.bwd_graph.__func__, UserDefinedClassVariable(self.bwd_graph.__class__), source=bwd_src, ) bwd_args = [bwd_fn.obj, *bwd_args] else: unimplemented("non-function or method") def is_strict_for(v: VariableTracker): if isinstance(v, variables.TensorVariable): # we can be more lax for stuff from forward return v.proxy.tracer is not fwd_tracer return True with tx.output.subtracer(fwd_fn, fwd_tracer), tx.strict_translation_mode( is_strict_for ): (bwd_out, _), bwd_graph, bwd_freevars = speculate_subgraph( tx, bwd_fn, bwd_args, kwargs, "autograd.Function", enable_grad=False, set_subgraph_inputs="manual", restore_side_effects=False, tracer=bwd_tracer, ) # TODO: assert that bwd_graph didn't capture values that were # not created inside fwd_graph. # TODO(oulgen): Ideally, we would not do a linear search for output # node but as things currently are there could be nodes after the # output node # This is bug prone as if there's code after the output node, then # graph.output will append the output at the very end # This might be a behavior difference # If users call ctx.mark_non_differentiable, we should capture these output tensors who # are marked as non-differentiable and pass them to ApplyTemplate # at torch._functorch.autograd_function.AutogradFunctionApply for reconstruction. non_differentiable_idx = [] if ctx.non_differentiable is not None: non_differentiable_set = set(ctx.non_differentiable) assert isinstance(fwd_out, variables.BaseListVariable) for i, x in enumerate(fwd_out.items): if ( isinstance(x, variables.TensorVariable) and x.as_proxy() in non_differentiable_set ): non_differentiable_idx.append(i) # Rewrite the output of fwd_graph to (output, stuff_necessary_for_bwd) for node in fwd_graph.find_nodes(op="output"): fwd_graph.erase_node(node) break # Because we lift the bwd_freevars as inputs of the bwd_graph, # we have to manually add the bwd_freevars as output of fwd_graph. # However, the bwd_freevars got from speculate_subgraph use the Proxies in the bwd_graph, # we need to convert them to Proxies in the fwd_graph and then generate new fwd_graph output. fwd_proxy_of_bwd_freevars = [] for k in bwd_freevars.keys(): if k in fwd_freevars: fwd_proxy_of_bwd_freevars.append(fwd_freevars[k]) else: fwd_proxy_of_bwd_freevars.append(k) new_fwd_graph_outputs = (fwd_out.as_proxy(), fwd_proxy_of_bwd_freevars) new_fwd_graph_outputs = pytree.tree_map(lambda x: x.node, new_fwd_graph_outputs) fwd_graph.output(new_fwd_graph_outputs) fwd_graph.lint() # Store fwd_body fwd_nn_modules = tx.output.tracing_context.module_context.copy_graphstate() fwd_name = add_subgraph( tx, "fwd_body", torch.fx.GraphModule(fwd_nn_modules.nn_modules, fwd_graph), ) fwd_node = make_attr(tx, fwd_name) # The type of original args can be arbitrary, but we only support basic type in FX graph. # So the speculated subgraph input includes original tensor args and the lifted freevars. # We need to filter out the original tensor args and concat them with the lifted freevars # to generate the proxy args for the FX call_function node. filtered_args = [] # A boolean list to mark if the type of corresponding argument is tensor. # This is used to determine if a FX node's argument should be an argument of # ApplyTemplate.forward and if we should skip the output from ApplyTemplate.backward # at torch._functorch.autograd_function.AutogradFunctionApply. args_tensor_mask = [False] * len(args) for i, arg in enumerate(args): if isinstance(arg, (variables.TensorVariable, variables.SymNodeVariable)): filtered_args.append(arg) args_tensor_mask[i] = True # Rewrite the output of bwd_graph to remove the grad output for the non-Tensor args. new_bwd_graph_outputs = None for node in bwd_graph.find_nodes(op="output"): bwd_graph.erase_node(node) break # The same as the above fwd proxies, we need to use the bwd proxies in the bwd_graph # if some of the output is from fwd_freevars. bwd_out_proxy = bwd_out.as_proxy() bwd_proxy_of_fwd_freevars = [] if isinstance(bwd_out_proxy, (tuple, list)): for k in bwd_out_proxy: if k in bwd_freevars: bwd_proxy_of_fwd_freevars.append(bwd_freevars[k]) else: bwd_proxy_of_fwd_freevars.append(k) else: if bwd_out_proxy in bwd_freevars: bwd_proxy_of_fwd_freevars = bwd_freevars[bwd_out_proxy] else: bwd_proxy_of_fwd_freevars = bwd_out_proxy # Remove bwd output for non-Tensor args. output_proxy = bwd_proxy_of_fwd_freevars if isinstance(output_proxy, (tuple, list)): new_bwd_graph_outputs = () for x, mask in zip(output_proxy, args_tensor_mask): if mask: new_bwd_graph_outputs = new_bwd_graph_outputs + (x,) else: assert x is None, f"Grad of non-Tensor arg {x} is not None." else: new_bwd_graph_outputs = output_proxy # Update the bwd graph output. new_bwd_graph_outputs = pytree.tree_map( lambda x: None if x is None else x.node, new_bwd_graph_outputs ) bwd_graph.output(new_bwd_graph_outputs) bwd_graph.lint() # Store bwd_body bwd_nn_modules = tx.output.tracing_context.module_context.copy_graphstate() bwd_name = add_subgraph( tx, "bwd_body", torch.fx.GraphModule(bwd_nn_modules.nn_modules, bwd_graph), ) bwd_node = make_attr(tx, bwd_name) tx.output.side_effects = prev_side_effects p_args = ( fwd_node, bwd_node, *([arg.as_proxy() for arg in filtered_args] + list(fwd_freevars.keys())), ) example_value = pytree.tree_map_only( torch.fx.Proxy, lambda a: a.node.meta["example_value"], fwd_out.as_proxy(), ) # Store the invocation as a call from torch._functorch.autograd_function import autograd_function_apply return wrap_fx_proxy( tx=tx, proxy=tx.output.create_proxy( "call_function", autograd_function_apply, args=p_args, kwargs={ "args_tensor_mask": args_tensor_mask, "non_differentiable_idx": non_differentiable_idx, }, ), example_value=example_value, ) def maybe_positional_arg_names(func): result = [] if not hasattr(func, "get_function"): return None try: fn = func.get_function() except (Unsupported, NotImplementedError): return None try: sig = inspect.signature(func.get_function()) except ValueError: return None for name, param in sig.parameters.items(): if param.kind is inspect.Parameter.VAR_POSITIONAL: return None if ( param.kind is inspect.Parameter.POSITIONAL_ONLY or param.kind is inspect.Parameter.POSITIONAL_OR_KEYWORD ): if name == "self": # FX graphs can't have a placeholder named self result.append("self_") else: result.append(name) return result