# mypy: allow-untyped-defs import copy import functools import heapq import itertools import logging import math import operator import os import os.path from collections import defaultdict from dataclasses import dataclass, replace from typing import Callable, Optional, TYPE_CHECKING, Union import torch import torch._inductor.inductor_prims import torch.distributed import torch.fx as fx import torch.utils._pytree as pytree from torch._functorch._activation_checkpointing.ac_logging_utils import ( create_structured_trace_for_min_cut_info, ) from torch.fx.experimental._backward_state import BackwardState from torch.fx.experimental.proxy_tensor import is_sym_node, py_sym_types from torch.fx.experimental.sym_node import magic_methods, method_to_operator from torch.fx.experimental.symbolic_shapes import ( find_symbol_binding_fx_nodes, free_symbols, hint_int, is_symbol_binding_fx_node, ) from torch.fx.passes import graph_drawer from torch.utils._ordered_set import OrderedSet from torch.utils.checkpoint import CheckpointPolicy from . import config from ._activation_checkpointing.graph_info_provider import GraphInfoProvider from ._activation_checkpointing.knapsack import ( dp_knapsack, greedy_knapsack, ilp_knapsack, ) from ._activation_checkpointing.knapsack_evaluator import KnapsackEvaluator from ._aot_autograd.logging_utils import get_aot_graph_name from ._aot_autograd.utils import get_cuda_generator_meta_val, is_with_effects from .compile_utils import fx_graph_cse, get_aten_target if TYPE_CHECKING: import sympy AOT_PARTITIONER_DEBUG: bool = config.debug_partitioner log: logging.Logger = logging.getLogger(__name__) aten = torch.ops.aten prims = torch.ops.prims @dataclass class OpTypes: """Class for keeping track of different operator categories""" fusible_ops: OrderedSet[Callable] compute_intensive_ops: OrderedSet[Callable] random_ops: OrderedSet[Callable] view_ops: OrderedSet[Callable] recomputable_ops: OrderedSet[Callable] def is_fusible(self, node: fx.Node): return get_aten_target(node) in self.fusible_ops def is_compute_intensive(self, node: fx.Node): return get_aten_target(node) in self.compute_intensive_ops def is_random(self, node: fx.Node): return get_aten_target(node) in self.random_ops def is_view(self, node: fx.Node): return get_aten_target(node) in self.view_ops def is_recomputable(self, node: fx.Node): return get_aten_target(node) in self.recomputable_ops @dataclass class NodeInfo: # Be careful about iterating over these explicitly, as their order may not # be deterministic inputs: list[fx.Node] _required_fw_nodes: OrderedSet[fx.Node] required_bw_nodes: OrderedSet[fx.Node] unclaimed_nodes: OrderedSet[fx.Node] fw_order: dict[fx.Node, int] @functools.cached_property def required_fw_nodes(self) -> list[fx.Node]: return sorted( (n for n in self._required_fw_nodes), key=lambda n: self.fw_order[n] ) def is_required_fw(self, n: fx.Node) -> bool: return n in self._required_fw_nodes def is_required_bw(self, n: fx.Node) -> bool: return n in self.required_bw_nodes def is_unclaimed(self, n: fx.Node) -> bool: return n in self.unclaimed_nodes def get_fw_order(self, n: fx.Node) -> int: assert n in self._required_fw_nodes, f"Node {n} not in fw nodes!" return self.fw_order[n] @dataclass class MinCutOptions: ban_if_used_far_apart: bool ban_if_long_fusible_chains: bool ban_if_materialized_backward: bool ban_if_not_in_allowlist: bool ban_if_reduction: bool def must_recompute(node: fx.Node) -> bool: return node.meta.get("recompute", None) in [ CheckpointPolicy.MUST_RECOMPUTE, CheckpointPolicy.PREFER_RECOMPUTE, ] def has_recomputable_ops(fx_g: fx.GraphModule) -> bool: for node in fx_g.graph.nodes: if must_recompute(node): return True return False def has_recomputable_rng_ops(fx_g: fx.GraphModule) -> bool: for node in fx_g.graph.nodes: if ( must_recompute(node) and hasattr(node.target, "tags") and torch.Tag.nondeterministic_seeded in node.target.tags ): return True return False def sym_node_size(node: fx.Node) -> int: if isinstance(node.meta["val"], (torch.SymInt, torch.SymBool)): return 1 assert isinstance(node.meta["val"], torch.SymFloat) return 4 class InvalidNodeBase: def __repr__(self): return "Invalid Node" InvalidNode = InvalidNodeBase() def _extract_graph_with_inputs_outputs( joint_graph: fx.Graph, inputs: list[fx.Node], outputs: list[fx.Node], subgraph: Optional[str] = None, ) -> fx.Graph: """ Given a graph, extracts out a subgraph that takes the specified nodes as inputs and returns the specified outputs. This includes specifying non-placeholder nodes as inputs. The general strategy is to initialize all inputs with proxies as we encounter them, and trace through the graph, only keeping values which take in valid proxies. Then, all dead code is eliminated. """ new_graph = fx.Graph() env = {} # Add new placeholder nodes in the order specified by the inputs for node in inputs: new_node = new_graph.placeholder(node.name) # Can't use node_copy here as we may be turning previous call_function into placeholders new_node.meta = node.meta env[node] = new_node for node in joint_graph.nodes: if _must_be_in_backward(node) and subgraph != "backward": env[node] = InvalidNode # type: ignore[assignment] continue if node in env: # Node must be one of our inputs. (Any member of env which wasn't an # input to start must have been created by this loop and won't be in # joint_graph.nodes). continue elif node.op == "placeholder": env[node] = InvalidNode # type: ignore[assignment] elif node.op == "call_function": all_args = pytree.arg_tree_leaves(*node.args, **node.kwargs) all_args = [ isinstance(env[x], InvalidNodeBase) for x in all_args if isinstance(x, fx.Node) ] if any(all_args): env[node] = InvalidNode # type: ignore[assignment] continue env[node] = new_graph.node_copy(node, lambda x: env[x]) elif node.op == "get_attr": env[node] = new_graph.node_copy(node, lambda x: env[x]) elif node.op == "output": pass output_values = [] for x in outputs: if isinstance(x, fx.Node): if x not in env: raise RuntimeError(f"Node {x} couldn't be found in env") assert not isinstance( env[x], InvalidNodeBase ), f"Node {x} was invalid, but is output" output_values.append(env[x]) else: output_values.append(x) new_graph.output(tuple(output_values)) new_graph.eliminate_dead_code() new_graph.lint() return new_graph def _is_primal(node: fx.Node) -> bool: return ( node.op == "placeholder" and "tangents" not in str(node.target) and not _is_bwd_seed_offset(node) and not _is_fwd_seed_offset(node) ) def _is_tangent(node: fx.Node) -> bool: return node.op == "placeholder" and "tangents" in str(node.target) def _is_bwd_seed_offset(node: fx.Node) -> bool: return node.op == "placeholder" and ( "bwd_seed" in str(node.target) or "bwd_base_offset" in str(node.target) ) def _is_fwd_seed_offset(node: fx.Node) -> bool: return node.op == "placeholder" and ( "fwd_seed" in str(node.target) or "fwd_base_offset" in str(node.target) ) def _is_backward_state(node: fx.Node) -> bool: return node.op == "placeholder" and isinstance(node.meta.get("val"), BackwardState) def _has_tag_is_backward(node: fx.Node) -> bool: return node.meta.get("partitioner_tag", None) == "is_backward" def _has_tag_must_be_in_backward(node: fx.Node) -> bool: return node.meta.get("partitioner_tag", None) == "must_be_in_backward" def _must_be_in_backward(node: fx.Node) -> bool: return _has_tag_must_be_in_backward(node) or ( _has_tag_is_backward(node) and is_with_effects(node) ) def _extract_fwd_bwd_outputs( joint_module: fx.GraphModule, *, num_fwd_outputs ) -> tuple[list[fx.Node], list[fx.Node]]: outputs = pytree.arg_tree_leaves( *(node.args for node in joint_module.graph.find_nodes(op="output")) ) fwd_outputs = outputs[:num_fwd_outputs] bwd_outputs = outputs[num_fwd_outputs:] return fwd_outputs, bwd_outputs def _remove_by_name(saved_values: list[fx.Node], name: str): for saved_value in saved_values: if saved_value.name == name: saved_values.remove(saved_value) break def _extract_fwd_bwd_modules( joint_module: fx.GraphModule, saved_values: list[fx.Node], saved_sym_nodes: list[fx.Node], *, num_fwd_outputs: int, ) -> tuple[fx.GraphModule, fx.GraphModule]: fwd_outputs, bwd_outputs = _extract_fwd_bwd_outputs( joint_module, num_fwd_outputs=num_fwd_outputs ) placeholders = joint_module.graph.find_nodes(op="placeholder") primal_inputs = [*filter(_is_primal, placeholders)] tangent_inputs = [*filter(_is_tangent, placeholders)] fwd_seed_offset_inputs = [*filter(_is_fwd_seed_offset, placeholders)] bwd_seed_offset_inputs = [*filter(_is_bwd_seed_offset, placeholders)] backward_state_inputs = [*filter(_is_backward_state, placeholders)] bwd_graph = _extract_graph_with_inputs_outputs( joint_module.graph, saved_sym_nodes + saved_values + tangent_inputs + bwd_seed_offset_inputs, bwd_outputs, "backward", ) for node in bwd_graph.find_nodes(op="placeholder"): # This is to filter out saved values that don't actually end up being used by the backwards pass if not node.users: _remove_by_name(saved_values, node.name) _remove_by_name(saved_sym_nodes, node.name) elif _is_backward_state(node): # BackwardState is saved directly _remove_by_name(saved_values, node.name) assert backward_state_inputs # Now that we have the finalized list of saved values, we need to ensure # we propagate all symbols which are referenced by backwards inputs. # These are not directly used in the graph but are required for downstream # sizevar assignment saved_symbols: OrderedSet[sympy.Symbol] = OrderedSet() saved_sym_nodes_binding = [] saved_sym_nodes_derived = [] # Some symbols may already be bound in the directly saved_sym_nodes, # keep track of them so we don't re-bind them for node in saved_sym_nodes: symbol = is_symbol_binding_fx_node(node) if symbol: saved_symbols.add(symbol) saved_sym_nodes_binding.append(node) else: saved_sym_nodes_derived.append(node) # Now go through all of the prospective backward inputs and track any # other symbols we need to bind symbol_bindings = find_symbol_binding_fx_nodes(joint_module.graph) for node in itertools.chain(saved_sym_nodes_derived, saved_values, tangent_inputs): if "val" not in node.meta: continue new_symbols = free_symbols(node.meta["val"]) - saved_symbols # NB: Deterministic order please! for s in sorted(new_symbols, key=lambda s: s.name): # NB: For well formed graphs, the symbol should always be present, # but we also have ways to produce ill-formed graphs, e.g., direct # make_fx usages, so don't choke in this case if s not in symbol_bindings: continue saved_sym_nodes_binding.append(symbol_bindings[s]) saved_symbols |= new_symbols # Update saved_sym_nodes that are now reordered to have all bindings at # front. This can also be used later on to figure out the position of saved # sym nodes in the output of fwd graph. saved_sym_nodes.clear() saved_sym_nodes.extend(saved_sym_nodes_binding + saved_sym_nodes_derived) # Now, we re-generate the fwd/bwd graphs. # NB: This might increase compilation time, but I doubt it matters fwd_graph = _extract_graph_with_inputs_outputs( joint_module.graph, primal_inputs + fwd_seed_offset_inputs, fwd_outputs + saved_values + saved_sym_nodes, "forward", ) bwd_graph = _extract_graph_with_inputs_outputs( joint_module.graph, saved_sym_nodes + saved_values + tangent_inputs + bwd_seed_offset_inputs + backward_state_inputs, bwd_outputs, "backward", ) fwd_module = fx._lazy_graph_module._make_graph_module(joint_module, fwd_graph) bwd_module = fx._lazy_graph_module._make_graph_module(joint_module, bwd_graph) return fwd_module, bwd_module def default_partition( joint_module: fx.GraphModule, _joint_inputs, *, num_fwd_outputs ) -> tuple[fx.GraphModule, fx.GraphModule]: """ Partitions the :attr:`joint_module` in a manner that closely resembles the behavior observed in the original ``.forward()`` and ``.backward()`` of the callable, i.e., the resulting forward graph contains those operators that are executed in the original ``.forward()`` callable passed to :func:`aot_function`. The default partitioner collects the operators that are between the forward inputs and the forward outputs. This helps in finding the tensors which have to be stashed for the backward pass. These stashed tensors become the output of the generated forward graph. The remaining operators are then placed in the backward graph. .. warning:: This API is experimental and likely to change. Args: joint_module(fx.GraphModule): The joint forward and backward graph. This is the result of AOT Autograd tracing. Returns: Returns the generated forward and backward Fx graph modules. """ if has_recomputable_ops(joint_module): return min_cut_rematerialization_partition( joint_module, _joint_inputs, num_fwd_outputs=num_fwd_outputs ) primal_inputs = list(filter(_is_primal, joint_module.graph.nodes)) fwd_seed_offset_inputs = list(filter(_is_fwd_seed_offset, joint_module.graph.nodes)) inputs = primal_inputs + fwd_seed_offset_inputs fwd_outputs, bwd_outputs = _extract_fwd_bwd_outputs( joint_module, num_fwd_outputs=num_fwd_outputs ) forward_only_graph = _extract_graph_with_inputs_outputs( joint_module.graph, inputs, fwd_outputs, "forward" ) forward_node_names = OrderedSet( node.name for node in forward_only_graph.nodes if node.op != "output" ) saved_values = [] saved_sym_nodes = [] for node in joint_module.graph.nodes: if node.name not in forward_node_names: continue if is_sym_node(node): # Symints must be kept separate from tensors so that PythonFunction only calls # save_for_backward on tensors and stashes symints in autograd .ctx saved_sym_nodes.append(node) elif "tensor_meta" not in node.meta and node.op == "call_function": # Since we can't save tuple of tensor values, we need to flatten out what we're saving users = node.users assert all(user.target == operator.getitem for user in users) saved_values.extend(users) else: backward_usages = [ n for n in node.users if n.name not in forward_node_names ] if "tensor_meta" in node.meta and all( is_sym_node(n) for n in backward_usages ): # If we have a tensor in the forward, where only its sizes/strides are needed in the backward, # and not the actual tensor data, # then it will be a lot cheaper to save only the sizes/strides, and not the actual tensor. # # Note that saving the tensor could also cause compilation problems: # If the user mutated an input in the forward and uses its sizes/strides in the backward, # then we would be obligated to clone the input before saving it to appease autograd. # (This is how we originally found this bug). saved_sym_nodes.extend(backward_usages) else: saved_values.append(node) saved_values = list(dict.fromkeys(saved_values).keys()) saved_sym_nodes = list(dict.fromkeys(saved_sym_nodes).keys()) return _extract_fwd_bwd_modules( joint_module, saved_values, saved_sym_nodes=saved_sym_nodes, num_fwd_outputs=num_fwd_outputs, ) INT_INF = int(1e6) def _tensor_nbytes(numel: int, dtype) -> int: return numel * dtype.itemsize def _size_of(node: fx.Node) -> int: def object_nbytes(x) -> int: if not isinstance(x, torch.Tensor): return 0 return _tensor_nbytes(hint_int(x.numel(), fallback=4096), x.dtype) if "val" in node.meta: val = node.meta["val"] if isinstance(val, py_sym_types): return 1 # NB: The fallback values here are meaningless, maybe we should respect # torch._inductor.config.unbacked_symint_fallback (but this is a # layering violation) elif isinstance(val, (list, tuple)): return sum(object_nbytes(n) for n in val) elif isinstance(val, dict): return sum(object_nbytes(n) for _, n in val.items()) elif isinstance(val, torch.Tensor): return object_nbytes(val) raise RuntimeError(f"Unknown metadata type {type(val)} on node {node}") if node.op == "get_attr" or node.target is torch.ops.aten._assert_scalar.default: return 0 raise RuntimeError( f"Node {node} didn't have `val` metadata; we should always have `val` metadata on the nodes." ) # Used for some investigative purposes def _count_ops(graph: fx.Graph): from collections import defaultdict cnt: dict[str, int] = defaultdict(int) for node in graph.nodes: if node.op == "call_function": cnt[node.target.__name__] += 1 log.info("%s", sorted(cnt.items(), key=lambda x: x[1], reverse=True)) @functools.lru_cache(None) def pointwise_ops(): ops = [] for attr_name in dir(torch.ops.aten): opoverloadpacket = getattr(torch.ops.aten, attr_name) if not isinstance(opoverloadpacket, torch._ops.OpOverloadPacket): continue for overload in opoverloadpacket.overloads(): op_overload = getattr(opoverloadpacket, overload) if torch.Tag.pointwise in op_overload.tags: # currently aot autograd uses packet not overload ops.append(opoverloadpacket) break return ops def sort_depths(args, depth_map: dict[fx.Node, int]) -> list[tuple[fx.Node, int]]: arg_depths = { arg: depth_map[arg] for arg in args if isinstance(arg, torch.fx.node.Node) } return sorted(arg_depths.items(), key=lambda x: x[1], reverse=True) def reordering_to_mimic_autograd_engine(gm: fx.GraphModule) -> fx.GraphModule: """ This pass finds the first bwd node in the graph (by looking at users of tangents) and then reorders the graph by walking from this node to all the way to the end of the graph. At each op in this traveral, we insert this op in a new graph and try to bring only the relevant subgraph from the other non-bwd edges relevant for this op. This closely mimics the behavior of autograd engine. Why is this pass required in the first place? This is an artifact of how partitioners work today. The starting point of partitioner is a joint graph, which is fwd and then bwd graph. In the case of checkpointing, we keep portions of fwd graph in their original place in the joint graph, while obtaining a bwd graph. As a result, the resulting bwd graph has copies of recomputed fwd subgraphs followed by the original bwd graph. If we run this naively, this leads to bad memory footprint, because the fwd subgraphs are live for way longer duration than necessary. This pass reorders the operations such that we prioritize the ops for the original bwd graph while only realizing those ops from the fwd graph that are necessary at any given point in the graph. """ new_graph = fx.Graph() env: dict[fx.Node, fx.Node] = {} # Add new placeholder nodes in the order specified by the inputs for node in gm.graph.find_nodes(op="placeholder"): env[node] = new_graph.node_copy(node, lambda x: env[x]) order = {} for idx, node in enumerate(gm.graph.nodes): order[node] = idx def insert_node_in_graph(node): cur_nodes = [node] insertable_nodes: OrderedSet[fx.Node] = OrderedSet() while len(cur_nodes) > 0: node = cur_nodes.pop() if node in insertable_nodes or node in env: continue insertable_nodes.add(node) # Bias traversal towards the nodes that have higher depth - prioritizes # critical path first. cur_nodes += node.all_input_nodes insertable_nodes = sorted(insertable_nodes, key=lambda n: order[n]) for node in insertable_nodes: env[node] = new_graph.node_copy(node, lambda x: env[x]) # Find first bwd node in the graph tangent_inputs = list(filter(_is_tangent, gm.graph.nodes)) first_node_in_bwd = None minimum_order = math.inf for tangent in tangent_inputs: for user in tangent.users: if order[user] < minimum_order: minimum_order = order[user] first_node_in_bwd = user # If gradInp does not depend upon gradOut, we may not find any nodes in the "backwards pass" if first_node_in_bwd is None: return gm # Build the graph op-by-op by starting from the node all the way to the end for node in list(gm.graph.nodes)[order[first_node_in_bwd] :]: insert_node_in_graph(node) # The output node is already built by the traversal. new_gm = torch.fx.GraphModule(gm, new_graph) return new_gm def apply_graphsafe_rng_functionalization( fw_module: torch.fx.GraphModule, bw_module: torch.fx.GraphModule, fw_node: torch.fx.Node, bw_node: torch.fx.Node, device: torch.device, rng_count: int, last_fwd_input: torch.fx.Node, last_bwd_input: torch.fx.Node, ): """ Note [CUDA Graph Safe RNG Functionalization] CUDA Graph capture doesn't work with get_rng_state and set_rng_state because these functions operate on CPU values, while CUDA Graph RNG capture uses on-device CUDA tensors. To solve this, we use graphsafe_set_state with a CUDA Generator registered to the CUDA Graph before capture begins. graphsafe_set_state updates the generator's pointer to reference a different GeneratorImpl, ensuring subsequent calls are correctly forwarded to the desired generator (and its cuda-tensor RNG state during graph capture). For each RNG operation's forward/backward pair: - We create two generators initialized with identical values - Each forward and backward call advances its respective generator equally - This keeps generators synchronized so forward and backward operations use matching RNG values When forward is called multiple times before backward (causing desynchronization): - We save the forward RNG state - We update the backward Generator's state before executing backward Before each CUDA Graph replay, replay_prologue updates captured RNG pointers with current states, ensuring backward Generator changes are reflected during replay. This function modifies both forward and backward computation graphs by: Creating RNG state placeholders for both passes Updating the forward node to use graph-safe RNG state Updating the backward node to use graph-safe RNG state For more details: https://github.com/pytorch/pytorch/issues/113541 """ device_idx = device.index assert device_idx is not None fw_graph = fw_module.graph bw_graph = bw_module.graph graphsafe_run_with_rng_state = torch._prims.rng_prims.graphsafe_run_with_rng_state # Handle forward pass # Note: [Generator arguments in AOTDispatcher] # Generator arguments in AOTDispatcher are added to support graphsafe rng # functionalization. See note above [CUDA Graph Safe RNG Functionalization] with fw_module.graph.inserting_after(last_fwd_input): fwd_rng_state = fw_module.graph.placeholder(f"fwd_rng_state_{rng_count}") fwd_rng_state.meta["val"] = get_cuda_generator_meta_val(device_idx) last_fwd_input = fwd_rng_state # Handle backward pass with bw_module.graph.inserting_after(last_bwd_input): bwd_rng_state = bw_module.graph.placeholder(f"bwd_rng_state_{rng_count}") # as above, clone so that meta val generator will not contain tensors bwd_rng_state.meta["val"] = get_cuda_generator_meta_val(device_idx) last_bwd_input = bwd_rng_state # Update forward node fw_kwargs = dict(fw_node.kwargs) fw_kwargs["rng_state"] = fwd_rng_state with fw_module.graph.inserting_after(fw_node): functional_fw_node = fw_graph.create_node( "call_function", graphsafe_run_with_rng_state, args=(fw_node.target, *fw_node.args), # type: ignore[arg-type] kwargs=fw_kwargs, ) fw_node.replace_all_uses_with(functional_fw_node) fw_graph.erase_node(fw_node) # Update backward node bwd_kwargs = dict(bw_node.kwargs) bwd_kwargs["rng_state"] = bwd_rng_state with bw_graph.inserting_before(bw_node): rng_output = bw_graph.create_node( "call_function", graphsafe_run_with_rng_state, args=(bw_node.target, *bw_node.args), # type: ignore[arg-type] kwargs=bwd_kwargs, ) bw_node.replace_all_uses_with(rng_output) bw_graph.erase_node(bw_node) return last_fwd_input, last_bwd_input def functionalize_rng_ops( joint_module: fx.GraphModule, fw_module: fx.GraphModule, bw_module: fx.GraphModule, num_sym_nodes: int, ) -> tuple[fx.GraphModule, fx.GraphModule]: # During user-driven activation checkpointing, we have to ensure that a rng # op in fwd yields the same output as the recomputed rng op in the bwd. To # do this, we use functionalize wrappers to wrap the random ops and share # rng state between the fwd and bwd graphs. # There are 3 main steps to do this # Step 1 - Construct a mapping of rng node between the fwd and its counterpart in bwd. # Step 2 - Modify the fwd pass such that # 1) Replace rand with run_and_save_rng_state wrapper # 2) Replace the users of the original op with the output[1] of this op. # 3) Collect all the rng_state - output[0] of each op, and make them # output nodes. Special care needs to be taken here because fwd outputs # has symints at the very end. # Step 3 - Modify the bwd pass such that # 1) Add the input nodes just before the tangents for the stashed rng states # 2) Replace rand with run_with_save_rng_state wrappers # 3) Use the stashed states as inputs to these ops # Unique id to generate name uid = itertools.count() def get_rng_ops(gmod): random_nodes = {} for node in gmod.graph.nodes: if ( node.op == "call_function" and hasattr(node.target, "tags") and torch.Tag.nondeterministic_seeded in node.target.tags ): random_nodes[node.name] = node return random_nodes def get_device(node) -> Optional[torch.device]: """ Check the example value of the node outputs to find the device type. """ if "val" not in node.meta: return None candidates = node.meta["val"] if not isinstance(candidates, tuple): candidates = (candidates,) for candidate in candidates: if isinstance(candidate, torch.Tensor): if candidate.device.type == "cuda": return candidate.device return torch.device("cpu") def get_sample_rng_state(device: Optional[torch.device]): if device is not None and device.type == "cuda": return torch.cuda.get_rng_state() return torch.get_rng_state() # Step 1 - Construct a mapping of rng node between the fwd and its counterpart in bwd. joint_graph_rng_ops = get_rng_ops(joint_module) fw_graph_rng_ops = get_rng_ops(fw_module) bw_graph_rng_ops = get_rng_ops(bw_module) recomputable_rng_ops_map = {} for node in joint_module.graph.nodes: if ( must_recompute(node) and hasattr(node.target, "tags") and torch.Tag.nondeterministic_seeded in node.target.tags ): base_node = joint_graph_rng_ops[node.name] fw_node = fw_graph_rng_ops[node.name] bw_node = bw_graph_rng_ops[node.name] recomputable_rng_ops_map[base_node] = {"fwd": fw_node, "bwd": bw_node} run_and_save_rng = torch._prims.rng_prims.run_and_save_rng_state run_with_rng_state = torch._prims.rng_prims.run_with_rng_state bw_tangent_start_node = None for node in bw_module.graph.find_nodes(op="placeholder"): if "tangent" in node.name: bw_tangent_start_node = node break if bw_tangent_start_node is None: raise RuntimeError( "Couldn't find tangent node in graph inputs. This is unexpected, please file a bug if you see this" ) fw_rng_state_outputs = [] last_fwd_input = next(reversed(fw_module.graph.find_nodes(op="placeholder"))) last_bwd_input = next(reversed(bw_module.graph.find_nodes(op="placeholder"))) devices = OrderedSet( get_device(node_pair["fwd"]) for node_pair in recomputable_rng_ops_map.values() ) devices.discard(torch.device("cpu")) # multiple cuda devices wont work with cudagraphs anyway, # fallback to non graphsafe rng checkpointing multi_cuda_devices = len(devices) > 1 # this changes numerics, so if fallback_random is set we will not use it ind_config = torch._inductor.config use_rng_graphsafe_rng_functionalization = ( config.graphsafe_rng_functionalization and not multi_cuda_devices and ( not ind_config.fallback_random or ind_config.test_configs.graphsafe_rng_func_ignores_fallback_random ) ) for rng_count, (base_node, node_pair) in enumerate( recomputable_rng_ops_map.items() ): # Step 2 - Modify the fwd pass such that fw_node = node_pair["fwd"] bw_node = node_pair["bwd"] device = get_device(fw_node) fw_graph = fw_module.graph bw_graph = bw_module.graph if ( use_rng_graphsafe_rng_functionalization and device is not None and device.type == "cuda" ): last_fwd_input, last_bwd_input = apply_graphsafe_rng_functionalization( fw_module, bw_module, fw_node, bw_node, device, rng_count, last_fwd_input, last_bwd_input, ) else: with fw_graph.inserting_before(fw_node): functional_fw_node = fw_graph.create_node( "call_function", run_and_save_rng, args=(fw_node.target, *fw_node.args), kwargs=fw_node.kwargs, ) state = fw_graph.create_node( "call_function", operator.getitem, args=(functional_fw_node, 0), kwargs={}, ) rng_output = fw_graph.create_node( "call_function", operator.getitem, args=( functional_fw_node, 1, ), kwargs={}, ) fw_node.replace_all_uses_with(rng_output) fw_graph.erase_node(fw_node) fw_rng_state_outputs.append(state) # Step 3 - Modify the bwd pass such that with bw_graph.inserting_before(bw_tangent_start_node): state_name = f"rng_state_output_{next(uid)}" bw_rng_state_node = bw_graph.placeholder(state_name) bw_rng_state_node.meta["val"] = get_sample_rng_state(device) with bw_graph.inserting_before(bw_node): rng_output = bw_graph.create_node( "call_function", run_with_rng_state, args=(bw_rng_state_node, bw_node.target, *bw_node.args), kwargs=bw_node.kwargs, ) bw_node.replace_all_uses_with(rng_output) bw_graph.erase_node(bw_node) # Add the rng states in the output of the fwd graph. AOT Autograd assumes # that symints are at the end of forward graph outputs. So, insert the new # rng states accordingly. if fw_rng_state_outputs: fw_output_node = next(iter(fw_module.graph.find_nodes(op="output"))) fw_outputs = fw_output_node.args[0] sym_node_start_idx = len(fw_outputs) - num_sym_nodes outputs = ( fw_outputs[:sym_node_start_idx] + tuple(fw_rng_state_outputs) + fw_outputs[sym_node_start_idx:] ) fw_module.graph.output(outputs) fw_module.graph.erase_node(fw_output_node) fw_module.recompile() bw_module.recompile() return fw_module, bw_module def cleanup_recompute_tags(joint_module: fx.GraphModule) -> fx.GraphModule: """ If there are two consecutive checkpointed blocks with no operator in between, we would still want to stash the tensor at the boundary of checkpointed blocks. The following pass makes the last output node non-recomputable to allow for that. """ for node in joint_module.graph.nodes: if must_recompute(node): for user in node.users: if ( must_recompute(user) and user.meta["ac_graph_id"] > node.meta["ac_graph_id"] ): node.meta["recompute"] = CheckpointPolicy.MUST_SAVE if node.meta.get("has_backward_hook", False) and not any( must_recompute(user) for user in node.users ): # If node is AC region output and has a backward hook on it, we intentionally choose to save it. # This is to work around circular dependencies in Traceable FSDP2+AC. # Example: # ``` # out = fully_shard(utils.checkpoint(module))(x) # norm_out = layer_norm(out) # ``` # Here there is a circular dependency: # 1. In backward, grad_input of layer_norm aka. `out_grad` is actually dependent on `out`. # 2. `out` depends on `out`'s backward hook created by FSDP2 (which does all-gather for `module` weights) # in order to be recomputed. # 3. `out`'s backward hook, as is the case for all eager backward hooks, depends on `out_grad` # -> circular dependency with (1)! # # Solution: check whether `out` has a backward hook, and if so, intentionally save `out` # in forward graph outputs. With this, we can break the above circular dependency. node.meta["recompute"] = CheckpointPolicy.MUST_SAVE return joint_module def solve_min_cut( joint_graph: fx.Graph, node_info: NodeInfo, min_cut_options: MinCutOptions, dont_ban: Optional[OrderedSet[fx.Node]] = None, ): if dont_ban is None: dont_ban = OrderedSet() op_types = get_default_op_list() if AOT_PARTITIONER_DEBUG: joint_module_ops = OrderedSet( str(node.target._overloadpacket) for node in joint_graph.nodes if node.op == "call_function" and hasattr(node.target, "_overloadpacket") ) ops_ignored = joint_module_ops - OrderedSet( str(i) for i in op_types.recomputable_ops ) log.info("Ops banned from re-materialization: %s", ops_ignored) def can_fuse_into_auto_functionalized(a, b): if b.target != torch.ops.higher_order.auto_functionalized: return False mutable_op = b.args[0] ( mutable_arg_names, _, ) = torch._higher_order_ops.auto_functionalize.get_mutable_args(mutable_op) for name in mutable_arg_names: arg = b.kwargs[name] if a is arg: return True if isinstance(arg, list): if a in arg: return True return False def can_fuse_into_triton_kernel_wrapper_functional(a, b): if b.target != torch.ops.higher_order.triton_kernel_wrapper_functional: return False mutable_arg_names = b.kwargs["tensors_to_clone"] for name in mutable_arg_names: arg = b.kwargs["kwargs"][name] if a is arg: return True return False def is_fusible(a, b): # We can perform "memory fusion" into a cat, but cat cannot be a # producer to a fusion if get_aten_target(b) == aten.cat: return True if can_fuse_into_auto_functionalized(a, b): return True if can_fuse_into_triton_kernel_wrapper_functional(a, b): return True if ( a.target is operator.getitem and a.args[0].target is torch.ops.higher_order.triton_kernel_wrapper_functional ): # if a is the output of a user triton kernel, # then (by default) we will not be able to fuse b into it return False return op_types.is_fusible(a) and op_types.is_fusible(b) try: import networkx as nx except ImportError as e: raise RuntimeError( "Need networkx installed to perform smart recomputation heuristics" ) from e def is_materialized_backwards(node): if op_types.is_view(node): return False cur_nodes = OrderedSet([node]) while len(cur_nodes) > 0: cur = cur_nodes.pop() for user in cur.users: if not node_info.is_required_fw(user) and not is_fusible(cur, user): return True if op_types.is_view(user): cur_nodes.add(user) return False def should_ban_recomputation(node): if node.op != "call_function": return False if node.target == operator.getitem: return False if node.meta.get("recompute", None) == CheckpointPolicy.MUST_SAVE: return True if config.recompute_views and op_types.is_view(node): return False if node.target in [aten.lift_fresh_copy.default, aten.lift_fresh.default]: return False if min_cut_options.ban_if_not_in_allowlist: if not op_types.is_recomputable(node): return True else: if op_types.is_random(node) or op_types.is_compute_intensive(node): return True # If a node *must* be materialized in the backwards pass, then we # should never recompute it. This is a pretty subtle point. In # general, the assumption we make is that recomputing a node in the # backwards pass is "free". However, if a node must be materialized # in the backwards pass, then recomputing it is never free. if min_cut_options.ban_if_materialized_backward and is_materialized_backwards( node ): log.debug("materialized backwards: %s %s", node, tuple(node.users)) return True # Arbitrary hack that sometimes seems to help things. The above # modification appears to have made this heuristic a lot less critical # for performance. # NB: As of PR #121692, this hack no longer seems necessary. if node.dist_from_bw < 1000 and node.dist_from_bw > config.max_dist_from_bw: return True # If the output of an op is 4x smaller (arbitrary choice), # then we don't allow recomputation. The idea here is that for # things like reductions, saving the output of the reduction is very # cheap/small, and it makes sure we don't do things like recompute # normalizations in the backwards. if min_cut_options.ban_if_reduction: input_tensors_size = sum( _size_of(i) for i in node.args if isinstance(i, fx.Node) ) output_size = _size_of(node) return output_size * 4 < input_tensors_size return False def is_materialized(node): if node.op == "placeholder": return True return not all(is_fusible(node, user) for user in node.users) def get_node_weight(node) -> float: mem_sz = _size_of(node) if config.recompute_views and op_types.is_view(node): # If `config.recompute_views=True`, we don't save views. This is generally # a good idea since views are free to recompute, and it makes it a bit simpler # to analyze. # NB: If they're not free to recompute (e.g. nested tensors)... I # think we should modify checks for view_ops to `is_view` and check # that. Basically, with nested tensors, `aten.view` is not a "view # op". return math.inf if isinstance(node.meta["val"], py_sym_types): # We never want to save symfloats if not isinstance(node.meta["val"], torch.SymInt): return INT_INF # Heuristic to bias towards nodes closer to the backwards pass # Complete guess about current value mem_sz = int(mem_sz * (1.1 ** max(min(node.dist_from_bw, 100), 1))) if is_materialized(node): return mem_sz else: return mem_sz * 2 nx_graph = nx.DiGraph() banned_nodes: OrderedSet[fx.Node] = OrderedSet() def ban_recomputation_if_allowed(node): if op_types.is_view(node): return False if node in dont_ban: return False # This bans recomputation of the node unless we've been forced not to by # user annotation if must_recompute(node): return False if "val" in node.meta and isinstance(node.meta["val"], torch.SymFloat): return False banned_nodes.add(node) # A node will only ever be recomputed if there is a path from an # ancestor of this node to the backwards path through this node that # doesn't go through any saved value. If this node is saved, then that # condition is not possible. nx_graph.add_edge("source", node.name + "_in", capacity=math.inf) return True for node in joint_graph.nodes: if node.op == "output": continue if node in node_info.required_bw_nodes: if node not in node_info.inputs: nx_graph.add_edge(node.name + "_in", "sink", capacity=math.inf) continue # If someone saves a input for backward as-is and backward # returns that tensor as-is as a grad input, then the node x would # be both a required_bw_node and an input. In this case we # (1) connect x_in to to the source, (2) x_out to the sink, and # (3) assign the proper weight to the x_in-x_out edge, so that # x would be part of cut nodes. A case where this happens is if # NestedTensor saves a offset tensor as part of the singleton int # in sizes. nx_graph.add_edge(node.name + "_out", "sink", capacity=math.inf) if must_recompute(node): # If user explicitly says they want to recompute a node, we honor it # by adding an inf-capacity edge from X_in to the sink. # This way, X_in node is guaranteed to be part of the subgraph that contains "sink" # after the cut, thus guaranteeing that X op will be recomputed. nx_graph.add_edge(node.name + "_in", "sink", capacity=math.inf) continue if _is_primal(node) or _is_fwd_seed_offset(node): ban_recomputation_if_allowed(node) # If a node can't be recomputed (too expensive or involves randomness), # we prevent it from being recomputed by adding an inf edge to the source # We only need to ban nodes in the fw pass, as those are the only ones that would be recomputed. if node_info.is_required_fw(node) and should_ban_recomputation(node): ban_recomputation_if_allowed(node) # Checks if a node is actually a tuple. Can be simplified to just an isinstance check if we always use faketensors. is_non_tensor_node = ( "val" not in node.meta and "tensor_meta" not in node.meta ) or ("val" in node.meta and not isinstance(node.meta["val"], torch.Tensor)) if is_sym_node(node): weight = float(sym_node_size(node)) elif is_non_tensor_node: weight = ( 0.0 if isinstance(node.meta.get("val"), BackwardState) else math.inf ) else: weight = get_node_weight(node) # Creates the weights on the "node" edge nx_graph.add_edge(node.name + "_in", node.name + "_out", capacity=weight) for user in node.users: nx_graph.add_edge(node.name + "_out", user.name + "_in", capacity=math.inf) # todo(chilli): This is the most questionable of the 3 heuristics for banning recompute. # Some example models to look at where this helps perf: poolformer_m36, # mixer_b16_224, cait_m36_384 # The "rough" idea here is that if you have some node that is used by both a # node nearby downstream as well as a node far downstream, if we recompute # both of the downstream nodes, we're unlikely to be able to fuse both # downstream nodes together. # Thus, we shouldn't aim to recompute far downstream nodes that depend on # this node. That intuition of "far downstream" is captured by whether # there's an unfusible op along the chain somewhere # It could probably be improved by properly analyzing what's going on in the # backwards pass instead of only relying on whether it's unfusible in the # forwards. def find_first_unfusible(start_nodes: list[fx.Node], max_range: int) -> int: """ Finds the first unfusible node in the chain of nodes starting from `start_nodes` and returns its position. """ sorted_nodes: list[tuple[int, fx.Node, bool]] = [] for n in start_nodes: heapq.heappush(sorted_nodes, (node_info.get_fw_order(n), n, True)) while len(sorted_nodes) > 0: _, node, node_is_fusible = heapq.heappop(sorted_nodes) if not node_is_fusible: return node_info.get_fw_order(node) for user in node.users: if node_info.is_required_fw(user): if node_info.get_fw_order(user) > max_range: continue val: tuple[int, fx.Node, bool] = ( node_info.get_fw_order(user), user, is_fusible(node, user), ) if val not in sorted_nodes: heapq.heappush(sorted_nodes, val) return max_range if min_cut_options.ban_if_used_far_apart: for used_node in node_info.required_fw_nodes: orders = [ node_info.get_fw_order(user) for user in used_node.users if node_info.is_required_fw(user) ] fw_users = [ user for user in used_node.users if node_info.is_required_fw(user) ] if len(orders) > 0: first_unfusible_use = find_first_unfusible(fw_users, max(orders)) for user in tuple(used_node.users): if ( node_info.is_required_fw(user) and node_info.get_fw_order(user) > first_unfusible_use and is_fusible(used_node, user) ): if user in banned_nodes: continue log.info( "used above/below fusible %s:(%s) -> %s -> %s:(%s)", used_node, node_info.get_fw_order(used_node), first_unfusible_use, user, node_info.get_fw_order(user), ) ban_recomputation_if_allowed(user) # This heuristic is fairly straightforward. The idea is that although it is # cheap to recompute bandwidth-bound ops, we don't want to end up in a situation # where we have a long chain of pointwise ops from the beginning to the end # of the model (like say, residual connections) # todo: I'm not totally sure why this heuristic matters. It's possible that this is # working around Inductor fusion decisions, or that it's a patch over # suboptimal partitioning decisions # Some models it improves perf on are cait_m36_384, mixer_b16_224, poolformer_m36 if min_cut_options.ban_if_long_fusible_chains: visited: OrderedSet[fx.Node] = OrderedSet() for start_node in joint_graph.nodes: if not node_info.is_required_fw(start_node): continue fusible: list[tuple[int, fx.Node]] = [ (node_info.get_fw_order(start_node), start_node) ] start_order = node_info.get_fw_order(start_node) while len(fusible) > 0: _, cur = heapq.heappop(fusible) if cur in visited: continue visited.add(cur) # 100 is arbitrary choice to try and prevent degenerate cases if ( node_info.get_fw_order(cur) > start_order + 100 and len(fusible) == 0 ): log.info( "too long %s %s %s %s", cur, start_node, node_info.get_fw_order(cur), node_info.get_fw_order(start_node), ) ban_recomputation_if_allowed(cur) break for user in cur.users: if ( node_info.is_required_fw(user) and is_fusible(cur, user) and user not in banned_nodes ): heapq.heappush(fusible, (node_info.get_fw_order(user), user)) try: cut_value, partition = nx.minimum_cut(nx_graph, "source", "sink") except Exception: log.info("Failed to compute min-cut on following graph:") log.info("\n".join(nx.readwrite.edgelist.generate_edgelist(nx_graph))) visualize_min_cut_graph(nx_graph) raise reachable, non_reachable = partition cutset: OrderedSet[tuple[str, str]] = OrderedSet() for u, nbrs in ((n, nx_graph[n]) for n in reachable): cutset.update((u, v) for v in nbrs if v in non_reachable) cut_nodes: OrderedSet[str] = OrderedSet() for node_in, node_out in cutset: assert node_in[:-3] == node_out[:-4] node_name = node_in[:-3] cut_nodes.add(node_name) name_to_node = get_name_to_node(joint_graph) # To make this stuff deterministic node_idx = {node: idx for idx, node in enumerate(joint_graph.nodes)} saved_values = sorted( (name_to_node[node] for node in cut_nodes), key=lambda x: node_idx[x] ) return saved_values, banned_nodes def visualize_min_cut_graph(nx_graph): import networkx as nx import pydot dot_format = nx.nx_pydot.to_pydot(nx_graph).to_string() dot_graph = pydot.graph_from_dot_data(dot_format)[0] for edge in dot_graph.get_edges(): weight = nx_graph[edge.get_source()][edge.get_destination()]["capacity"] # Set edge label to weight edge.set_label(str(weight)) # Color edges with weight 'inf' as red if weight == float("inf"): edge.set_color("red") log.info("Visualizing the failed graph to min_cut_failed.svg") dot_graph.write_svg("min_cut_failed.svg") def get_default_op_list() -> OpTypes: default_recomputable_ops: list[Callable] = [ aten.add, aten.sub, aten.div, aten.atan2, aten.mul, aten.max, aten.min, aten.pow, aten.remainder, aten.fmod, aten.__and__, aten.__or__, aten.__xor__, aten.__lshift__, aten.__rshift__, aten.eq, aten.ne, aten.ge, aten.gt, aten.le, aten.lt, aten.abs, aten.bitwise_not, aten.ceil, aten.floor, aten.frac, aten.neg, aten.relu, aten.round, aten.silu, aten.trunc, aten.log, aten.log10, aten.log1p, aten.log2, aten.lgamma, aten.exp, aten.expm1, aten.erf, aten.erfc, aten.cos, aten.acos, aten.cosh, aten.sin, aten.asin, aten.sinh, aten.tan, aten.atan, aten.tanh, aten.atanh, aten.sqrt, aten.rsqrt, aten.reciprocal, aten.sigmoid, aten.softplus, aten.threshold, aten.threshold_backward, aten.clamp, aten.where, aten.lerp, aten.addcmul, aten.gelu, aten.gelu_backward, aten.sum, aten.mean, aten._grad_sum_to_size, aten.sum_to_size, aten.amax, aten.to, aten.type_as, operator.getitem, aten.squeeze, aten.unsqueeze, aten.rsub, aten._to_copy, ] # noqa: E501,B950 recomputable_view_ops = [aten.squeeze, aten.unsqueeze, aten.alias] recomputable_view_ops += [ aten.view, aten.slice, aten.t, prims.broadcast_in_dim, aten.expand, aten.as_strided, aten.permute, aten.select, ] view_ops = recomputable_view_ops default_recomputable_ops += [ prims.div, prims.convert_element_type, aten.clone, aten._to_copy, aten.full_like, prims.var, prims.sum, aten.var, aten.std, prims.broadcast_in_dim, aten.select, aten._unsafe_view, aten.view, aten.expand, aten.slice, aten.reshape, aten.broadcast_tensors, aten.scalar_tensor, aten.ones, aten.new_zeros, aten.lift_fresh_copy, aten.arange, aten.triu, aten.var_mean, aten.isinf, aten.any, aten.full, aten.as_strided, aten.zeros, aten.empty, aten.empty_like, aten.argmax, aten.maximum, prims.iota, prims._low_memory_max_pool2d_offsets_to_indices, ] # noqa: E501,B950 # Natalia said that we should allow recomputing indexing :) default_recomputable_ops += [aten.index, aten.gather] default_recomputable_ops += view_ops default_recomputable_ops += pointwise_ops() default_recomputable_ops += [ aten.zeros_like, ] default_recomputable_ops += [method_to_operator(m) for m in magic_methods] recomputable_ops = OrderedSet(default_recomputable_ops) random_ops = OrderedSet([aten.native_dropout, aten.rand_like, aten.randn_like]) compute_intensive_ops = [ aten.mm, aten.convolution, aten.convolution_backward, aten.bmm, aten.addmm, aten._scaled_dot_product_flash_attention, aten._scaled_dot_product_efficient_attention, aten._flash_attention_forward, aten._efficient_attention_forward, aten.upsample_bilinear2d, aten._scaled_mm, ] # noqa: E501,B950 fusible_ops = recomputable_ops | random_ops return OpTypes( fusible_ops, OrderedSet(compute_intensive_ops), random_ops, OrderedSet(view_ops), recomputable_ops, ) def get_name_to_node(graph: fx.Graph): name_to_node = {} for node in graph.nodes: name_to_node[node.name] = node return name_to_node def _optimize_runtime_with_given_memory( joint_graph: fx.Graph, memory: list[float], runtimes: list[float], max_memory: float, node_info: NodeInfo, all_recomputable_banned_nodes: list[fx.Node], ) -> tuple[float, list[int], list[int]]: SOLVER = config.activation_memory_budget_solver if SOLVER == "greedy": return greedy_knapsack(memory, runtimes, max_memory) elif SOLVER == "ilp": return ilp_knapsack(memory, runtimes, max_memory) elif SOLVER == "dp": return dp_knapsack(memory, runtimes, max_memory) elif SOLVER == "dynamic_memory_budget_dp": log.warning( "dynamic_memory_budget_dp is an experimental solver. " "It does not guarantee performance improvements. " "Additionally, it is not guaranteed to be stable." ) graph_info_provider = GraphInfoProvider.inialize_from_graph( joint_graph=joint_graph, all_recomputable_banned_nodes=all_recomputable_banned_nodes, recorded_knapsack_input_memories=memory, recorded_knapsack_input_runtimes=runtimes, ) return dp_knapsack( memory, runtimes, KnapsackEvaluator( graph_info_provider=graph_info_provider, ).get_knee_point_memory_budget( knapsack_algo=dp_knapsack, max_mem_budget=max_memory, ), ) elif callable(SOLVER): saved_node_idx, recomp_node_idx = SOLVER( memory, joint_graph, max_memory, node_info, all_recomputable_banned_nodes ) return (0.0, saved_node_idx, recomp_node_idx) else: raise RuntimeError(f"Not aware of memory budget knapsack solver: {SOLVER}") from torch.utils._mode_utils import no_dispatch # replace symbols in size and strides with their hints without guarding. def _remove_symbols_without_guarding(x: torch.Tensor, fallback: int) -> torch.Tensor: shape = list(x.shape) def realize_symbol(d): return hint_int(d, fallback=fallback) shape = [realize_symbol(s) for s in shape] stride = [realize_symbol(s) for s in x.stride()] return x.new_empty_strided(shape, stride=stride) def estimate_runtime(node): RUNTIME_MODE = config.activation_memory_budget_runtime_estimator def materialize_arg(x): if isinstance(x, fx.Node) and isinstance(x.meta["val"], torch.Tensor): return _remove_symbols_without_guarding(x.meta["val"], fallback=4096) elif isinstance(x, fx.Node) and isinstance(x.meta["val"], torch.SymInt): return hint_int(x.meta["val"], fallback=4096) elif isinstance(x, fx.Node) and isinstance(x.meta["val"], torch.SymFloat): return 1.0 elif isinstance(x, fx.Node) and isinstance(x.meta["val"], torch.SymBool): return True else: return x if RUNTIME_MODE == "testing": return 1 elif RUNTIME_MODE == "profile": with no_dispatch(): from torch._inductor.runtime.benchmarking import benchmarker args, kwargs = pytree.tree_map(materialize_arg, (node.args, node.kwargs)) ms = benchmarker.benchmark_gpu(lambda: node.target(*args, **kwargs)) return ms elif RUNTIME_MODE == "flops": # todo(chilli): Normalize this to also return ms from torch.utils.flop_counter import FlopCounterMode args, kwargs = pytree.tree_map(materialize_arg, (node.args, node.kwargs)) with FlopCounterMode(display=False) as mode: node.target(*args, **kwargs) counted_flops = mode.get_total_flops() return max(counted_flops, 1) else: raise RuntimeError(f"Not aware of runtime estimator: {RUNTIME_MODE}") def choose_saved_values_set( joint_graph: fx.Graph, node_info: NodeInfo, memory_budget=1, ) -> list[fx.Node]: if memory_budget > 1 or memory_budget < 0: raise RuntimeError( f"The valid ranges for memory budget are 0 <= m <= 1. The provided value is {memory_budget}" ) min_cut_options = MinCutOptions( ban_if_used_far_apart=config.ban_recompute_used_far_apart, ban_if_long_fusible_chains=config.ban_recompute_long_fusible_chains, ban_if_materialized_backward=config.ban_recompute_materialized_backward, ban_if_not_in_allowlist=config.ban_recompute_not_in_allowlist, ban_if_reduction=config.ban_recompute_reductions, ) if config.aggressive_recomputation: min_cut_options = replace( min_cut_options, ban_if_used_far_apart=False, ban_if_long_fusible_chains=False, ban_if_materialized_backward=False, ban_if_not_in_allowlist=False, ) if memory_budget == 0: return node_info.inputs runtime_optimized_saved_values, _ = solve_min_cut( joint_graph, node_info, min_cut_options, ) # return runtime_optimized_saved_values if memory_budget == 1: return runtime_optimized_saved_values def estimate_activations_size(saved_values: list[fx.Node]) -> float: return sum(map(_size_of, saved_values)) / 1e9 min_act_size = estimate_activations_size(node_info.inputs) max_act_size = estimate_activations_size(runtime_optimized_saved_values) # The optimized choice is smaller than the inputs anyways if max_act_size <= min_act_size: return runtime_optimized_saved_values def get_normalized_size(sz): return (sz / 1e9) / (max_act_size - min_act_size) def get_mem_ratio(activations: list[fx.Node]): return (estimate_activations_size(activations) - min_act_size) / ( max_act_size - min_act_size ) more_aggressive_options = replace( min_cut_options, ban_if_used_far_apart=False, ban_if_long_fusible_chains=False, ban_if_materialized_backward=False, ) more_aggressive_saved_values, _ = solve_min_cut( joint_graph, node_info, more_aggressive_options ) if get_mem_ratio(more_aggressive_saved_values) < memory_budget: return more_aggressive_saved_values aggressive_options = replace( more_aggressive_options, ban_if_not_in_allowlist=False, ) aggressive_recomputation_saved_values, banned_nodes = solve_min_cut( joint_graph, node_info, aggressive_options ) if get_mem_ratio(aggressive_recomputation_saved_values) < memory_budget: return aggressive_recomputation_saved_values from torch._inductor.fx_utils import get_node_storage input_storages = OrderedSet(get_node_storage(node) for node in node_info.inputs) def get_recomputable_banned_nodes( banned_nodes: OrderedSet[fx.Node], ) -> list[fx.Node]: return [ i for i in banned_nodes if ( # Only allow recomputing nodes that are actually required for BW i.dist_from_bw < int(1e9) # type: ignore[attr-defined] and get_node_storage(i) not in input_storages ) ] recomputable_banned_nodes = get_recomputable_banned_nodes(banned_nodes) # sort first by name, to ensure determinism when multiple nodes have same size recomputable_banned_nodes = sorted(recomputable_banned_nodes, key=lambda x: x.name) # default: runtime_optimized_saved_values # more aggressive: more_aggressive_saved_values # full aggressive: aggressive_recomputation_saved_values all_recomputable_banned_nodes = sorted( recomputable_banned_nodes, key=_size_of, reverse=True ) if len(all_recomputable_banned_nodes) == 0: return node_info.inputs memories_banned_nodes = [ get_normalized_size(_size_of(i)) for i in all_recomputable_banned_nodes ] runtimes_banned_nodes = [ estimate_runtime(node) for node in all_recomputable_banned_nodes ] from torch.utils._mode_utils import no_dispatch def get_saved_values_knapsack(memory_budget, node_info, joint_graph): with no_dispatch(): ( expected_runtime, saved_node_idxs, recomputable_node_idxs, ) = _optimize_runtime_with_given_memory( joint_graph, memories_banned_nodes, runtimes_banned_nodes, max(memory_budget, 0), node_info, all_recomputable_banned_nodes, ) dont_ban: OrderedSet[fx.Node] = OrderedSet() for idx in recomputable_node_idxs: # if idx in all_recomputable_banned_nodes: try: dont_ban.add(all_recomputable_banned_nodes[idx]) except BaseException: pass assert dont_ban.issubset(all_recomputable_banned_nodes) saved_values, _ = solve_min_cut( joint_graph, node_info, aggressive_options, dont_ban, ) if AOT_PARTITIONER_DEBUG: create_structured_trace_for_min_cut_info( joint_graph=joint_graph, all_recomputable_banned_nodes=all_recomputable_banned_nodes, saved_node_idxs=saved_node_idxs, recomputable_node_idxs=recomputable_node_idxs, expected_runtime=expected_runtime, memories_banned_nodes=memories_banned_nodes, runtimes_banned_nodes=runtimes_banned_nodes, min_cut_saved_values=saved_values, ) return saved_values, expected_runtime if config.visualize_memory_budget_pareto: def estimate_for_budget(b): saved_values, expected_runtime = get_saved_values_knapsack( b, node_info=node_info, joint_graph=joint_graph ) return ( b, sum(runtimes_banned_nodes) - expected_runtime, get_mem_ratio(saved_values), ) options = [estimate_for_budget(0.0), estimate_for_budget(1.0)] if options[0][1:] != options[1][1:]: bisects = [(options[0], options[1])] while bisects: lhs, rhs = bisects.pop() if rhs[0] - lhs[0] < 1e-3: options.append(lhs) options.append(rhs) continue mid = estimate_for_budget((lhs[0] + rhs[0]) / 2) if mid[1:] != lhs[1:]: bisects.append((lhs, mid)) if mid[1:] != rhs[1:]: bisects.append((mid, rhs)) options.sort() import matplotlib.pyplot as plt x_values = [item[2] for item in options] y_values = [item[1] for item in options] # Plotting the values with updated axis labels and chart title plt.figure(figsize=(10, 6)) plt.plot(x_values, y_values, marker="o") # Adding labels for each point for i, txt in enumerate(x_values): plt.annotate( f"{txt:.4f}", (txt, y_values[i]), textcoords="offset points", xytext=(0, 10), ha="center", ) plt.xlabel("Memory Budget") plt.ylabel("Runtime of Recomputed Components") plt.title("Pareto Frontier of Memory Budget vs. Recomputation Runtime") plt.grid(True) fig = plt.gcf() plt.show() fig_dir = os.getcwd() if config.memory_budget_pareto_dir is not None: fig_dir = config.memory_budget_pareto_dir os.makedirs(fig_dir, exist_ok=True) rank_suffix = "" if torch.distributed.is_available() and torch.distributed.is_initialized(): rank_suffix = f"_rank_{torch.distributed.get_rank()}" fig_name = os.path.join( fig_dir, f"memory_budget_pareto{rank_suffix}_{get_aot_graph_name()}.svg" ) fig.savefig(fig_name) log.warning("Generated Pareto frontier curve at %s", fig_name) # todo(chilli): Estimated doesn't align exactly with actual - actual is # usually less memory than estimated. i'm guessing (actually quite # unsure about this) that's because estimated is just only including # tensors we actually banned from recompute, but there may be other # tensors that we choose to save. return get_saved_values_knapsack( memory_budget=memory_budget, node_info=node_info, joint_graph=joint_graph )[0] def min_cut_rematerialization_partition( joint_module: fx.GraphModule, _joint_inputs, compiler="inductor", *, num_fwd_outputs, ) -> tuple[fx.GraphModule, fx.GraphModule]: """ Partitions the joint graph such that the backward recomputes the forward. Recomputing helps in trading off memory bandwidth with computation. To create the fwd and bwd graph, we copy the joint graph, manually set the outputs to just original forward or backward outputs. And then we run the resulting graphs through dead code elimination. .. warning:: This API is experimental and likely to change. Args: joint_module(fx.GraphModule): The joint forward and backward graph. This is the result of AOT Autograd tracing. _joint_inputs: The inputs to the joint graph. This is unused. compiler: This option determines the default set of recomputable ops. Currently, there are two options: ``nvfuser`` and ``inductor``. recomputable_ops: This is an optional set of recomputable ops. If this is not None, then this set of ops will be used instead of the default set of ops. num_fwd_outputs: The number of outputs from the forward graph. Returns: Returns the generated forward and backward Fx graph modules. """ joint_module.graph.eliminate_dead_code() joint_module.recompile() fx_g = joint_module.graph # add the CSE pass if config.cse: cse_graph = fx_graph_cse(fx_g) joint_module.graph = cse_graph joint_graph = joint_module.graph graph_has_recomputable_ops = has_recomputable_ops(joint_module) graph_has_recomputable_rng_ops = has_recomputable_rng_ops(joint_module) if graph_has_recomputable_ops: joint_module = cleanup_recompute_tags(joint_module) def classify_nodes(joint_module): name_to_node = get_name_to_node(joint_module.graph) required_bw_nodes: OrderedSet[fx.Node] = OrderedSet() for node in joint_module.graph.nodes: if node.op == "placeholder" and "tangents" in node.target: required_bw_nodes.add(node) elif _must_be_in_backward(node): required_bw_nodes.add(node) if node in required_bw_nodes: required_bw_nodes.update(node.users) primal_inputs = list(filter(_is_primal, joint_module.graph.nodes)) fwd_seed_offset_inputs = list( filter(_is_fwd_seed_offset, joint_module.graph.nodes) ) inputs = primal_inputs + fwd_seed_offset_inputs fwd_outputs, bwd_outputs = _extract_fwd_bwd_outputs( joint_module, num_fwd_outputs=num_fwd_outputs ) required_bw_nodes.update( o for o in bwd_outputs if o is not None and o.op != "output" ) forward_only_graph = _extract_graph_with_inputs_outputs( joint_module.graph, inputs, fwd_outputs, "forward" ) required_fw_nodes: OrderedSet[fx.Node] = OrderedSet( name_to_node[node.name] for node in forward_only_graph.nodes if node.op != "output" ) unclaimed_nodes: OrderedSet[fx.Node] = OrderedSet( node for node in joint_module.graph.nodes if node not in required_fw_nodes and node not in required_bw_nodes ) fw_cnt = 0 fw_order = {} for node in joint_module.graph.nodes: if node in required_fw_nodes: fw_order[node] = fw_cnt fw_cnt += 1 return NodeInfo( inputs, required_fw_nodes, required_bw_nodes, unclaimed_nodes, fw_order ) node_info = classify_nodes(joint_module) # networkx blows up on graphs with no required backward nodes # Since there's nothing to partition anyway, and the default partitioner can "handle" # this case, send our graph over to the default partitioner. if len(node_info.required_bw_nodes) == 0: return default_partition( joint_module, _joint_inputs, num_fwd_outputs=num_fwd_outputs ) for node in reversed(joint_module.graph.nodes): if node.op == "output": node.dist_from_bw = int(1e9) elif not node_info.is_required_fw(node): node.dist_from_bw = 0 else: node.dist_from_bw = int(1e9) for user in node.users: node.dist_from_bw = min(node.dist_from_bw, user.dist_from_bw + 1) memory_budget = config.activation_memory_budget for node in joint_graph.nodes: if isinstance(node.meta.get("memory_budget", None), float): memory_budget = node.meta["memory_budget"] break saved_values = choose_saved_values_set( joint_graph, node_info, memory_budget=memory_budget, ) # save_for_backward on tensors and stashes symints in autograd .ctx saved_sym_nodes = list(filter(is_sym_node, saved_values)) saved_values = list(filter(lambda n: not is_sym_node(n), saved_values)) # NB: saved_sym_nodes will be mutated to reflect the actual saved symbols fw_module, bw_module = _extract_fwd_bwd_modules( joint_module, saved_values, saved_sym_nodes=saved_sym_nodes, num_fwd_outputs=num_fwd_outputs, ) if graph_has_recomputable_ops: if graph_has_recomputable_rng_ops: fw_module, bw_module = functionalize_rng_ops( joint_module, fw_module, bw_module, len(saved_sym_nodes) ) bw_module = reordering_to_mimic_autograd_engine(bw_module) if AOT_PARTITIONER_DEBUG: # Calculate sorted sizes of saved values sorted_sizes = sorted([(_size_of(i), str(i)) for i in saved_values]) # Log total theoretical activations stored total_activations_size_gb = sum(_size_of(i) for i in saved_values) / 1e9 log.info("Theoretical Activations Stored: %.2f GB", total_activations_size_gb) # Log theoretical per activation storage sizes log.info("Theoretical Per Activation Storage Sizes: %s", sorted_sizes) fw_module_nodes = OrderedSet( node.name for node in fw_module.graph.nodes if node.op == "call_function" ) bw_module_nodes = OrderedSet( node.name for node in bw_module.graph.nodes if node.op == "call_function" ) remat_nodes = fw_module_nodes & bw_module_nodes counts: dict[str, int] = defaultdict(int) for node in fw_module.graph.nodes: if node.name in remat_nodes and hasattr(node.target, "_overloadpacket"): counts[str(node.target._overloadpacket)] += 1 log.info( "# remat/fw/bw: %d/%d/%d", len(remat_nodes), len(fw_module_nodes), len(bw_module_nodes), ) rematerialized_ops = sorted(counts.items(), key=lambda x: x[1], reverse=True) log.info("Count of Ops Rematerialized: %s", rematerialized_ops) return fw_module, bw_module def draw_graph( traced: torch.fx.GraphModule, fname: str, figname: str = "fx_graph", clear_meta: bool = True, prog: Optional[Union[str, list[str]]] = None, parse_stack_trace: bool = False, dot_graph_shape: Optional[str] = None, ) -> None: if clear_meta: new_graph = copy.deepcopy(traced.graph) traced = fx.GraphModule(traced, new_graph) for node in traced.graph.nodes: node.meta = {} base, ext = os.path.splitext(fname) if not ext: ext = "." + config.torch_compile_graph_format log.info("Writing FX graph to file: %s%s", base, ext) g = graph_drawer.FxGraphDrawer( traced, figname, parse_stack_trace=parse_stack_trace, dot_graph_shape=dot_graph_shape, ) x = g.get_main_dot_graph() write_method = getattr(x, "write_" + ext.lstrip(".")) fname = f"{base}{ext}" if prog is None: write_method(fname) else: write_method(fname, prog=prog)