# mypy: allow-untyped-defs # pyre-strict from __future__ import annotations import heapq import importlib import itertools import logging import operator import sys import time from collections import defaultdict from dataclasses import dataclass from typing import Any, Optional, TYPE_CHECKING, Union import torch from torch._logging import trace_structured from torch.multiprocessing.reductions import StorageWeakRef from torch.utils._ordered_set import OrderedSet from . import config, config_comms, ir from .dependencies import WeakDep if TYPE_CHECKING: from .ir import IRNode, Operation from .memory import ( estimate_peak_memory, estimate_peak_memory_allocfree, FreeableInputBuffer, get_freeable_input_buf, SNodeMemory, ) from .utils import ( contains_collective, contains_wait, find_recursive_deps_of_node, find_recursive_users_of_node, is_collective, is_fallback_op, is_wait, ) from .virtualized import V log = logging.getLogger(__name__) overlap_log = torch._logging.getArtifactLogger(__name__, "overlap") if TYPE_CHECKING: from torch._inductor.scheduler import BaseSchedulerNode def align_runtime_estimations_across_all_distributed_ranks( snodes: list[BaseSchedulerNode], ): runtime_estimations = {} for snode in snodes: runtime_estimations[snode] = snode.get_estimated_runtime() import torch.distributed as dist from torch.distributed.distributed_c10d import _get_default_group world_size = dist.get_world_size() pg = _get_default_group() gathered_runtime_estimations: list[list[float]] = [[] for _ in range(world_size)] dist.all_gather_object( gathered_runtime_estimations, list(runtime_estimations.values()), pg ) median_runtime_estimations = torch.median( torch.tensor(gathered_runtime_estimations), dim=0 ).values.tolist() for i in range(len(snodes)): snodes[i].override_estimated_runtime = median_runtime_estimations[i] def sink_waits(snodes: list[BaseSchedulerNode]) -> list[BaseSchedulerNode]: """ Greedily schedules waits as late as possible. """ return _schedule_for_comm( snodes, raise_comms=False, sink_waits=True, reorder_for_overlap=False ) def raise_comms(snodes: list[BaseSchedulerNode]) -> list[BaseSchedulerNode]: """ Greedily schedules comms as early as possible. """ return _schedule_for_comm( snodes, raise_comms=True, sink_waits=False, reorder_for_overlap=False ) def reorder_compute_for_overlap( snodes: list[BaseSchedulerNode], ) -> list[BaseSchedulerNode]: """ This achieves the following overall scheduling procedure: Step 1: Given that we've currently scheduled comm N, we now schedule all compute nodes that are required for comm N + 1 but do not depend on comm N, to run at the same time with comm N. Step 2: If all those compute nodes are sufficient to overlap comm N, we're done. Otherwise, we now need to look elsewhere to find compute that overlaps with comm N. We prioritize compute nodes that are needed sooner. Step 3: We schedule the compute nodes dependent on comm N and required for comm N + 1. Step 4: We schedule comm N + 1. Repeat this for subsequent comm nodes. """ return _schedule_for_comm( snodes, raise_comms=True, sink_waits=True, reorder_for_overlap=True ) def reorder_communication_preserving_peak_memory( snodes: list[BaseSchedulerNode], ) -> list[BaseSchedulerNode]: """ Reorders communication ops relative to computation ops to improve communication-compute overlapping and hide comm latency. Stops moving a particular op if it reaches a point that would have increased the peak memory footprint. Currently, follows these heuristics (subject to change or tune): - never reorders collectives relative to one another, for SPMD safety - has an option for per-collective prefetch limit, but does not enable it by default - limits the total number of reorder steps to some factor of the graph size to prevent worst-case quadratic performance Prerequisite: sink_comms_and_waits - ensure comm and wait nodes are scheduled as late as possible, respecting data dependencies. That allows reorder_communication_preserving_peak_memory to take a best case peak-memory snapshot, and then monotonically improve latency by moving collectives backward in time. Peak memory impact is computed in an iterative fashion. First, memory use at each timestep is computed, and global peak memory is computed as a max over timesteps. Then, when swapping any two adjacent nodes, only the curr-memory for the earlier of the nodes after the swap is affected. This enables checking step by step whether a swap is peak-memory-safe, and bailing out if not. Example: 0 n0 C0 1 n1 C0 + Allocs(n1) - Frees(n1) 2 n2 C0 + Allocs(n1) - Frees(n1) + Allocs(n2) - Frees(n2) 0 n0 C0 1 n2 C0 + Allocs(n2) - Frees(n2) <-- After moving n2 to Time 1, only time1 memory changes 2 n1 C0 + Allocs(n2) - Frees(n2) + Allocs(n1) - Frees(n1) """ reordered_snodes, node_stats = ( _reorder_communication_preserving_peak_memory_internal(snodes) ) return reordered_snodes @dataclass class ReorderInfo: """ Debug info describing how an individual snode was reordered """ limiting_factor: str = "None" moves: int = 0 grouped: int = 0 grouped_info: str = "" comm_time: float = -1.0 comp_time: float = -1.0 initial_exposed: float = -1.0 final_exposed: float = -1.0 overlap_info: str = "None" @property def improvement(self): return self.initial_exposed - self.final_exposed def is_gemm_like(node: Optional[Union[IRNode, Operation]]) -> bool: if node is None: return False if is_fallback_op( node, # type: ignore[arg-type] torch.ops.aten._scaled_dot_product_flash_attention.default, ): return True if ( python_kernel_name := getattr(node, "python_kernel_name", None) ) and "extern_kernels" in python_kernel_name: return True return False def contains_gemm_like(snode: BaseSchedulerNode) -> bool: from torch._inductor.scheduler import GroupedSchedulerNode if isinstance(snode, GroupedSchedulerNode): return any(contains_gemm_like(x) for x in snode.snodes) else: return is_gemm_like(snode.node) def _temp_group_visit_leaves(snode: BaseSchedulerNode, fn): from torch._inductor.scheduler import GroupedSchedulerNode if isinstance(snode, GroupedSchedulerNode) and snode.temp_grouping: for _snode in snode.snodes: fn(_snode) else: fn(snode) def wait_exposed_communication_time( snodes_to_wait: list[BaseSchedulerNode], runtimes: dict[BaseSchedulerNode, float] ) -> tuple[float, float, str]: """ Calculate exposed communication time for a wait operation by finding its corresponding collective and accumulating overlapping compute time between them. The Wait node must be the last in snodes_to_wait. Compute time between corresponding Collective and Wait is accumulated. If there is another pair of Collective and Wait inside, Only compute before first such Wait' is considered as overlapping. Multiple process groups are not modeled so far. """ wait_snode = snodes_to_wait[-1] assert is_wait(wait_snode.node) assert len(snodes_to_wait) > 1 idx = len(snodes_to_wait) - 2 comm_time = 0.0 comp_time = 0.0 overlap_info = "" waits_found = [] for i in range(idx, -1, -1): c = snodes_to_wait[i] if contains_wait(c): waits_found.append(c) if contains_collective(c): if is_corresponding_collective_wait(c, wait_snode): comm_time = runtimes[c] overlap_info += f"->C[{c.get_name()}]" break if not contains_async_collective(c): # Sync Collective comp_time = 0.0 continue else: for w in waits_found: if is_corresponding_collective_wait(c, w): # Similar to Sync Collective # If after our Collective exist another Collective-Wait, # All compute after it will not be overlapping comp_time = 0.0 continue comp_time_before = comp_time def accumulate_time(_snode: BaseSchedulerNode) -> None: nonlocal comp_time comp_time += runtimes[_snode] _temp_group_visit_leaves(c, accumulate_time) comp_time_after = comp_time overlap_info += f"+{c.get_name()}[{comp_time_after - comp_time_before}]" return comm_time, comp_time, overlap_info def coll_exposed_communication_time( snodes: list[BaseSchedulerNode], runtimes: dict[BaseSchedulerNode, float], ) -> tuple[float, float, str]: """ Calculate exposed communication time for a collective operation by finding its corresponding wait and accumulating compute time that can overlap with communication. The Collective node must be the first in snodes. Compute time between corresponding Collective and Wait is accumulated. If there is another pair of Collective and Wait inside, Only compute before first such Wait' is considered as overlapping. Multiple process groups are not modeled so far. """ collective_snode = snodes[0] comm_time = runtimes[collective_snode] comp_time = 0.0 collective_outs: OrderedSet[str] = OrderedSet( o.get_name() for o in collective_snode.get_outputs() ) overlap_info = "" collectives_found: list[BaseSchedulerNode] = [] for snode in snodes[1:]: # We may have some ops without Wait, # e.g. DTensor torch.ops._dtensor.shard_dim_alltoall unmet_deps = OrderedSet( d.name for d in snode.unmet_dependencies if not _is_fake_dep(d) ) if unmet_deps & collective_outs: overlap_info += f"->W[{snode.get_name()}]" break if contains_collective(snode): if not contains_async_collective(snode): break else: collectives_found.append(snode) continue if contains_wait(snode): has_wait_for_collectives_found = False for _coll in collectives_found: if is_corresponding_collective_wait(collective_snode, snode): has_wait_for_collectives_found = True break if has_wait_for_collectives_found: # Any compute after not overlapping original Collective break comp_time_before = comp_time def accumulate_time(_snode: BaseSchedulerNode) -> None: nonlocal comp_time comp_time += runtimes[_snode] _temp_group_visit_leaves(snode, accumulate_time) comp_time_after = comp_time overlap_info += f"+{snode.get_name()}[{comp_time_after - comp_time_before}]" return comm_time, comp_time, overlap_info def _group_name(snode, with_bufs=False) -> str: ret = "" for n in snode.snodes: if ret: ret += "_" ret += n.get_name() if with_bufs: ret += f"{list(snode.get_buffer_names())}" return ret def _is_fake_dep(d): return isinstance(d, WeakDep) and d.is_fake def _group_names(gns: list[BaseSchedulerNode]) -> str: return "~".join([gn.get_name() for gn in gns]) def _initialize_memory_tracking(snodes, graph_inputs, graph_outputs): """Initialize memory tracking data structures""" name_to_freeable_input_buf = get_freeable_input_buf(snodes, graph_inputs) peak_memory, snodes_curr_memory, snodes_allocfree, buf_to_snode_last_use = ( estimate_peak_memory_allocfree( snodes, name_to_freeable_input_buf, graph_outputs ) ) _curr_memory = dict(zip(snodes, snodes_curr_memory)) _curr_memory[None] = (0, 0) return ( peak_memory, _curr_memory, snodes_allocfree, buf_to_snode_last_use, name_to_freeable_input_buf, ) def _initialize_double_linked_list( snodes: list[BaseSchedulerNode], ) -> tuple[ dict[BaseSchedulerNode, Optional[BaseSchedulerNode]], dict[BaseSchedulerNode, Optional[BaseSchedulerNode]], BaseSchedulerNode, ]: """Create double-linked list structure from snodes""" _prev = {} _next = {} for i, snode in enumerate(snodes): _prev[snode] = snodes[i - 1] if i > 0 else None _next[snode] = snodes[i + 1] if i < len(snodes) - 1 else None _head = snodes[0] return _prev, _next, _head def is_corresponding_collective_wait( collective_snode: BaseSchedulerNode, wait_snode: BaseSchedulerNode ) -> bool: """ Check if a wait node corresponds to a given collective node by verifying if the wait depends on outputs from the collective. """ collective_outs = OrderedSet(o.get_name() for o in collective_snode.get_outputs()) unmet_deps = OrderedSet(d.name for d in wait_snode.unmet_dependencies) return bool(unmet_deps & collective_outs) def _op_runtime_estimate_mult(snode): # Apply multipliers for faster experimentation. # TODO(ivankobzarev): Remove after confirmation that runtime estimations are correct. if contains_collective(snode): return config_comms.reorder_sink_runtime_estimations_comm_mult return config_comms.reorder_sink_runtime_estimations_non_comm_mult def is_async_collective(snode): """ Filtering out ops that contain Collective and Wait inside and considered as Collectives. See contains_collective function. If the op contains Wait inside - consider as Synchronous compute. """ if python_kernel_name := getattr(snode.node, "python_kernel_name", None): if "torch.ops._dtensor.shard_dim_alltoall.default" in python_kernel_name: return False return True def contains_async_collective(snode): return contains_collective(snode, is_async_collective) def _group_nodes_from_linked_list( head: Optional[BaseSchedulerNode], tail: Optional[BaseSchedulerNode], next_dict: dict[BaseSchedulerNode, Optional[BaseSchedulerNode]], ) -> list[BaseSchedulerNode]: """ Traverse doubly-linked list from head to tail and return nodes as a list. Args: head: Starting node of the segment tail: Ending node of the segment (inclusive) next_dict: Dictionary mapping each node to its next node Returns: List of nodes from head to tail (inclusive) """ ret = [] n = head while True: if n is not None: ret.append(n) if n == tail: break n = next_dict[n] # type: ignore[index] return ret def _perform_double_linked_list_swap( candidate: BaseSchedulerNode, group_head: BaseSchedulerNode, group_tail: BaseSchedulerNode, prev_dict: dict[BaseSchedulerNode, Optional[BaseSchedulerNode]], next_dict: dict[BaseSchedulerNode, Optional[BaseSchedulerNode]], head: BaseSchedulerNode, ) -> BaseSchedulerNode: """ Swap positions of candidate and group in doubly-linked list. Transforms: candidate_prev -> candidate -> group_head...group_tail -> group_tail_next Into: candidate_prev -> group_head...group_tail -> candidate -> group_tail_next Args: candidate: Node to swap with group group_head: First node of group group_tail: Last node of group prev_dict: Dictionary mapping nodes to their previous nodes next_dict: Dictionary mapping nodes to their next nodes head: Current head of the linked list Returns: New head of the linked list (may change if candidate was the head) """ # 0: Update candidate's previous node candidate_prev = prev_dict[candidate] if candidate_prev: next_dict[candidate_prev] = group_head prev_dict[group_head] = candidate_prev # 2: Update group_tail's next node group_tail_next = next_dict[group_tail] if group_tail_next: prev_dict[group_tail_next] = candidate next_dict[candidate] = group_tail_next # 1: Link group_tail to candidate prev_dict[candidate] = group_tail next_dict[group_tail] = candidate # Update head if candidate was the head if head == candidate: return group_head return head def _calculate_potential_peak_memory_reorder( candidate: BaseSchedulerNode, gns: list[BaseSchedulerNode], group_tail: BaseSchedulerNode, group_peak_memory: int, candidate_delta_mem: int, candidate_allocfree: SNodeMemory, group_n_to_bufs_after_swap_dealloc_by_candidate: dict, curr_memory: dict, ) -> tuple[int, dict[BaseSchedulerNode, int]]: """ Calculate potential peak memory after swapping candidate with group (reorder version). Computes new memory levels for all affected nodes and returns the potential peak memory along with cached post-allocation memory values for each node. Args: candidate: Node being moved gns: Group nodes group_tail: Last node of group group_peak_memory: Current peak memory within the group candidate_delta_mem: Net memory change from candidate (alloc - free) candidate_allocfree: Candidate's allocation/free info group_n_to_bufs_after_swap_dealloc_by_candidate: Buffers whose deallocation moves to candidate curr_memory: Current memory state dict Returns: Tuple of (potential_peak_memory, post_alloc_update_dict) """ # Caching calculations of memory for group nodes and candidate, # to apply without recalculation after swap. _post_alloc_update: dict[BaseSchedulerNode, int] = {} potential_peak: int = 0 if not group_n_to_bufs_after_swap_dealloc_by_candidate: # Not accounting for buffers last use change potential_peak = max( group_peak_memory - candidate_delta_mem, curr_memory[group_tail][1] - candidate_delta_mem + candidate_allocfree.size_alloc, ) return potential_peak, _post_alloc_update # If candidate will be after group, the starting memory level of group nodes # changes to the -(candidate.size_alloc - candidate.size_free) mem_after_reorder_delta: int = -candidate_delta_mem for gn in gns: gn_post_alloc_mem = curr_memory[gn][0] + mem_after_reorder_delta _post_alloc_update[gn] = gn_post_alloc_mem potential_peak = max(potential_peak, gn_post_alloc_mem) bufs = group_n_to_bufs_after_swap_dealloc_by_candidate.get(gn) if bufs is not None: for buf in bufs: # Candidate will deallocate those buffers mem_after_reorder_delta += buf.mpi_buffer.size_free candidate_mem_post_alloc = ( curr_memory[group_tail][1] + mem_after_reorder_delta + candidate_allocfree.size_alloc ) _post_alloc_update[candidate] = candidate_mem_post_alloc potential_peak = max(potential_peak, candidate_mem_post_alloc) return potential_peak, _post_alloc_update def _update_memory_tracking_after_swap_reorder( candidate: BaseSchedulerNode, gns: list[BaseSchedulerNode], group_tail: BaseSchedulerNode, candidate_delta_mem: int, candidate_allocfree: SNodeMemory, group_n_to_bufs_after_swap_dealloc_by_candidate: dict, post_alloc_update: dict[BaseSchedulerNode, int], curr_memory: dict, buf_to_snode_last_use: dict, snodes_allocfree: dict, ) -> None: """ Update memory tracking structures after swap (reorder version). Updates curr_memory, buf_to_snode_last_use, and snodes_allocfree dictionaries to reflect the new memory state after swapping candidate with group. Args: candidate: Node that was moved gns: Group nodes group_tail: Last node of group candidate_delta_mem: Net memory change from candidate (alloc - free) candidate_allocfree: Candidate's allocation/free info group_n_to_bufs_after_swap_dealloc_by_candidate: Buffers whose deallocation moves to candidate post_alloc_update: Cached post-allocation memory values curr_memory: Current memory state dict (mutated) buf_to_snode_last_use: Buffer to last-use node mapping (mutated) snodes_allocfree: Node allocation/free info dict (mutated) """ if not group_n_to_bufs_after_swap_dealloc_by_candidate: for gn in gns: cm = curr_memory[gn] curr_memory[gn] = ( cm[0] - candidate_delta_mem, cm[1] - candidate_delta_mem, ) _candidate_post_alloc_mem = ( curr_memory[group_tail][1] + candidate_allocfree.size_alloc ) _candidate_post_free_mem = ( _candidate_post_alloc_mem - candidate_allocfree.size_free ) curr_memory[candidate] = ( _candidate_post_alloc_mem, _candidate_post_free_mem, ) return # Candidate becomes last use of some bufs for bufs in group_n_to_bufs_after_swap_dealloc_by_candidate.values(): for buf in bufs: buf_to_snode_last_use[buf] = candidate size_free_to_move_to_candidate_sum: int = 0 for n in gns: _gn_post_alloc_mem: int = post_alloc_update[n] size_free_to_move_to_candidate: int = sum( buf.mpi_buffer.size_free for buf in group_n_to_bufs_after_swap_dealloc_by_candidate[n] ) size_free_to_move_to_candidate_sum += size_free_to_move_to_candidate # group node does not deallocate this after swap snodes_allocfree[n].size_free -= size_free_to_move_to_candidate gn_post_free_mem: int = _gn_post_alloc_mem - snodes_allocfree[n].size_free curr_memory[n] = (_gn_post_alloc_mem, gn_post_free_mem) _candidate_post_alloc_mem = post_alloc_update[candidate] snodes_allocfree[candidate].size_free += size_free_to_move_to_candidate_sum candidate_post_free_mem = ( _candidate_post_alloc_mem - snodes_allocfree[candidate].size_free ) curr_memory[candidate] = ( _candidate_post_alloc_mem, candidate_post_free_mem, ) def _find_buffers_with_changed_last_use( candidate: BaseSchedulerNode, gns: list[BaseSchedulerNode], buf_to_snode_last_use: dict, ) -> dict[BaseSchedulerNode, list[Union[FreeableInputBuffer, Any]]]: """ Find buffers whose last use will change after swapping candidate with group. When we swap [candidate [group]] to [[group] candidate], some buffers that were last used by a group node will now be last used by candidate instead. This affects memory deallocation timing. Args: candidate: The node being moved gns: Group nodes being swapped with candidate buf_to_snode_last_use: Mapping of buffers to their current last-use nodes Returns: Dict mapping group nodes to buffers that will change their last-use node """ group_n_to_bufs_after_swap_dealloc_by_candidate: dict[ BaseSchedulerNode, list[Union[FreeableInputBuffer, Any]] ] = defaultdict(list) for ( buf, snode_last_use, ) in buf_to_snode_last_use.items(): succ_nodes = buf.mpi_buffer.succ_nodes if candidate not in succ_nodes: continue if not any(gn == snode_last_use for gn in gns): continue group_n_to_bufs_after_swap_dealloc_by_candidate[snode_last_use].append(buf) return group_n_to_bufs_after_swap_dealloc_by_candidate def _is_node_groupable_for_reorder( candidate: BaseSchedulerNode, ) -> tuple[bool, Optional[str]]: """ Check if a candidate node can be grouped with collective during reordering. This pass processes collectives left to right, so we avoid grouping with already-processed collectives based on configuration. Args: candidate: Node to check for groupability Returns: Tuple of (is_groupable, reason_if_not_groupable) """ # This pass processes collectives left to right, # Do not group with processed collectives. # Leaving config for experimentation in 2D if not config_comms.reorder_iterative_group_with_collectives: if contains_async_collective(candidate): return ( False, f"candidate contains_collective {candidate.get_name()}", ) if not config_comms.reorder_iterative_use_runtime_estimations: if contains_gemm_like(candidate): return False, "contains_gemm_like" return True, None def _format_and_log_reordering_stats( stats: dict[BaseSchedulerNode, ReorderInfo], head: BaseSchedulerNode, next_dict: dict[BaseSchedulerNode, Optional[BaseSchedulerNode]], original_snodes_num: int, peak_memory: int, name_to_freeable_input_buf: dict, graph_outputs: OrderedSet[str], ) -> list[BaseSchedulerNode]: """ Format reordering statistics, log them, and return final node list. Computes improvement metrics, creates a formatted table (using tabulate if available), validates the reordered node count, recalculates peak memory, and logs all information. Args: stats: Per-node reordering statistics head: Head of the reordered linked list next_dict: Linked list next pointers original_snodes_num: Original number of nodes (for validation) peak_memory: Initial peak memory before reordering name_to_freeable_input_buf: Buffer memory tracking info graph_outputs: Graph output names Returns: Final reordered list of scheduler nodes """ node_stats = stats improvement = {snode: node_stats[snode].improvement for snode in node_stats} total_improvement = sum([improvement[snode] for snode in improvement]) total_moves = sum([node_stats[snode].moves for snode in node_stats]) reorder_log_str = ( f"reorder_communication_preserving_peak_memory improved overlap by {total_improvement} ns" f" after {total_moves} reorders.\n" ) headers = [ "Collective node", "comm_time(us)", "comp_time(us)", "initial exposed(us)", "final exposed(us)", "improvement(us)", "limiting factor", "moves", "grouped", "grouped_info", "overlap_info", ] rows = [ [ node_summary(snode), node_info.comm_time / 1e3, node_info.comp_time / 1e3, node_info.initial_exposed / 1e3, node_info.final_exposed / 1e3, node_info.improvement / 1e3, node_info.limiting_factor, node_info.moves, node_info.grouped, node_info.grouped_info, node_info.overlap_info, ] for snode, node_info in node_stats.items() ] if importlib.util.find_spec("tabulate"): # pyrefly: ignore[import-error] from tabulate import tabulate reorder_log_str += tabulate( rows, headers=headers, ) else: reorder_log_str += ( "Please `pip install tabulate` to nicely render overlap stats.\n" ) reorder_log_str += str(headers) + "\n" reorder_log_str += "\n".join(map(str, rows)) new_snodes = _group_nodes_from_linked_list(head, None, next_dict) assert len(new_snodes) == original_snodes_num new_peak_memory, _, _, _ = estimate_peak_memory_allocfree( new_snodes, name_to_freeable_input_buf, graph_outputs ) reorder_log_str += f"\n peak_memory_before:{peak_memory}" reorder_log_str += f"\n peak_memory_after:{new_peak_memory}" overlap_log.info(reorder_log_str) trace_structured( "artifact", metadata_fn=lambda: { "name": "reorder_communication_preserving_peak_memory", "encoding": "string", }, payload_fn=lambda: reorder_log_str, ) return new_snodes def _reorder_communication_preserving_peak_memory_internal( snodes: list[BaseSchedulerNode], ) -> tuple[list[BaseSchedulerNode], dict[BaseSchedulerNode, ReorderInfo]]: """ Internal testing helper that also returns debug info. Returns: - reordered snodes list - dict {snode: ReorderInfo} """ has_collectives = False for snode in snodes: if contains_collective(snode): has_collectives = True break if not has_collectives: return snodes, {} from torch._inductor.scheduler import GroupedSchedulerNode original_snodes_num = len(snodes) # heuristic to avoid degenerating to quadratic time graph_inputs: OrderedSet[str] = OrderedSet(V.graph.graph_inputs.keys()) graph_outputs: OrderedSet[str] = OrderedSet(V.graph.get_output_names()) ( peak_memory, _curr_memory, snodes_allocfree, buf_to_snode_last_use, name_to_freeable_input_buf, ) = _initialize_memory_tracking(snodes, graph_inputs, graph_outputs) runtimes: dict[BaseSchedulerNode, float] = { snode: estimate_op_runtime(snode) * _op_runtime_estimate_mult(snode) for snode in snodes } # debug stats stats: dict[BaseSchedulerNode, ReorderInfo] = {} total_moves = 0 _prev, _next, _head = _initialize_double_linked_list(snodes) debug_num_collectives_to_reorder: Optional[int] = ( config_comms.reorder_iterative_debug_limit_to_reorder ) num_processed_collectives: int = 0 curr: Optional[BaseSchedulerNode] = _head debug_iterative_memory_recompute = ( config_comms.reorder_iterative_debug_memory_recompute ) iterative_recompute_error = False while curr is not None and _next[curr] is not None: _next_curr = _next[curr] if iterative_recompute_error: break # pyrefly: ignore [bad-argument-type] if not contains_async_collective(curr): curr = _next_curr continue if debug_num_collectives_to_reorder is not None and ( num_processed_collectives >= debug_num_collectives_to_reorder ): break num_processed_collectives += 1 info = stats[curr] = ReorderInfo() comm_time, comp_time, overlap_info = coll_exposed_communication_time( _group_nodes_from_linked_list(curr, None, _next), runtimes ) info.comm_time = comm_time info.comp_time = comp_time info.initial_exposed = info.final_exposed = comm_time - comp_time info.overlap_info = overlap_info candidate = _prev[curr] group_head = curr group_tail = curr group_waits = {} group_runtime = 0.0 group_peak_memory = _curr_memory[curr][0] # post_alloc memory while candidate is not None: if config_comms.reorder_iterative_use_runtime_estimations and ( info.final_exposed < -config_comms.reorder_iterative_extra_comm_comp_overlap * info.comm_time ): info.limiting_factor = "unexposed by runtime estimations" break if ( not config_comms.reorder_iterative_unsafe_collectives_reorder and contains_collective(candidate) ): info.limiting_factor = "collective ordering" break gns: list[BaseSchedulerNode] = _group_nodes_from_linked_list( group_head, group_tail, _next ) group = GroupedSchedulerNode( curr.scheduler, gns, temp_grouping=True, ) # We can have multiple deps with the same name. # As we ignore WeakDep(is_fake=True) => # filter them out first to avoid overwriting of real dep. data_deps = { d.name: d for d in group.unmet_dependencies if not _is_fake_dep(d) } candidate_outs = candidate.get_outputs() data_dep = None for o in candidate_outs: if d := data_deps.get(o.get_name(), None): data_dep = d break if data_dep is not None: is_groupable_result, grouping_reason = _is_node_groupable_for_reorder( candidate ) if is_groupable_result: group_head = candidate # pyrefly: ignore[unbound-name] if config_comms.reorder_iterative_use_runtime_estimations: if contains_wait(candidate): comm_time, comp_time, _ = wait_exposed_communication_time( _group_nodes_from_linked_list(_head, candidate, _next), runtimes, ) group_waits[candidate] = comm_time, comp_time if not contains_async_collective(candidate): group_runtime += runtimes[candidate] group_peak_memory = max( group_peak_memory, _curr_memory[candidate][0] ) info.grouped += 1 info.grouped_info = _group_names(gns) candidate = _prev[candidate] continue else: msg = ( f"data dependency {data_dep}(dep_names:{list(data_deps.keys())})" f"\n candidate:{candidate.get_name()}(outs:{[candidate.get_buffer_names()]})" f"dep on {_group_names(gns)}" f"\n non_group_reason:{grouping_reason}" ) info.limiting_factor = msg break # pyrefly: ignore[unbound-name] if config_comms.reorder_iterative_use_runtime_estimations: # Check if candidate has sync runtime if not contains_async_collective(candidate): c_runtime = runtimes[candidate] if c_runtime > 0 and len(group_waits) > 0: # pyrefly: ignore[no-matching-overload] exposed_before = max(0, info.comm_time - info.comp_time) # pyrefly: ignore[no-matching-overload] exposed_after = max( 0, info.comm_time - info.comp_time - c_runtime ) exposed_delta = exposed_after - exposed_before for gw_comm_time, gw_comp_time in group_waits.values(): gw_exposed_before = max(0, gw_comm_time - gw_comp_time) gw_exposed_after = max( 0, gw_comm_time - gw_comp_time + c_runtime ) exposed_delta += gw_exposed_after - gw_exposed_before if exposed_delta > 0: info.limiting_factor = ( f"candidate has compute {c_runtime}," f" group contains waits, total_exposed_delta {exposed_delta}" ) break else: # Update all group_colls comm_time, comp_time for gw, ( gw_comm_time, gw_comp_time, ) in group_waits.items(): group_waits[gw] = ( gw_comm_time, gw_comp_time - c_runtime, ) else: # Candidate is async_collective # Unsafe collectives reordering # Cj -> [...group_runtime..., Ci] -> Wj # Checking that we are not increasing exposed time of Cj if group_runtime > 0: comm_time, comp_time, _ = coll_exposed_communication_time( _group_nodes_from_linked_list(candidate, None, _next), runtimes, ) # pyrefly: ignore[no-matching-overload] exposed_before = max(0, comm_time - comp_time) # pyrefly: ignore[no-matching-overload] exposed_after = max(0, comm_time - comp_time + group_runtime) exposed_delta = exposed_after - exposed_before if exposed_delta > 0: info.limiting_factor = ( f"candidate {candidate.get_name()} is collective," f" group_runtime:{group_runtime}," f" exposed_delta:{exposed_delta} c_comm_time:{comm_time} c_comp_time:{comp_time}" ) break candidate_allocfree: SNodeMemory = snodes_allocfree[candidate] candidate_delta_mem: int = ( candidate_allocfree.size_alloc - candidate_allocfree.size_free ) # candidate and one of group nodes are successors of the same buffer # and last use of the buffer happen in group nodes. # This last use deallocates it. # If we swap [candidate [group]] to [[group] candidate], # candidate becomes the last use # and deallocated this buffer instead of group node. # we need to update size_free accordingly to group_node and candidate, # and recalculate post_alloc, post_free for them. # # Buf that changes its last use snode, # after swap will be deallocated only by candidate, # while before it was deallocated by group node. group_n_to_bufs_after_swap_dealloc_by_candidate = ( _find_buffers_with_changed_last_use( candidate, gns, buf_to_snode_last_use ) ) potential_peak, _post_alloc_update = ( _calculate_potential_peak_memory_reorder( candidate, gns, group_tail, group_peak_memory, candidate_delta_mem, candidate_allocfree, group_n_to_bufs_after_swap_dealloc_by_candidate, _curr_memory, ) ) if ( potential_peak - peak_memory # pyrefly: ignore[unbound-name] > peak_memory * config_comms.reorder_iterative_peak_memory_budget ): info.limiting_factor = ( f"peak memory new:{potential_peak} vs base:{peak_memory}" ) break info.moves += 1 total_moves += 1 _head = _perform_double_linked_list_swap( candidate, group_head, group_tail, _prev, _next, _head ) comm_time, comp_time, overlap_info = coll_exposed_communication_time( _group_nodes_from_linked_list(curr, None, _next), runtimes ) info.comm_time = comm_time info.comp_time = comp_time info.overlap_info = overlap_info info.final_exposed = comm_time - comp_time _update_memory_tracking_after_swap_reorder( candidate, gns, group_tail, candidate_delta_mem, candidate_allocfree, group_n_to_bufs_after_swap_dealloc_by_candidate, _post_alloc_update, _curr_memory, buf_to_snode_last_use, snodes_allocfree, ) if debug_iterative_memory_recompute: # Compare iteratively recomputed memory data # with full run of estimate_peak_memory from .comms_debug import _debug_iterative_memory_recompute iterative_recompute_error = _debug_iterative_memory_recompute( candidate, gns, _group_names(gns), _group_nodes_from_linked_list(_head, None, _next), name_to_freeable_input_buf, graph_outputs, peak_memory, _curr_memory, snodes_allocfree, "reorder_communication_preserving_peak_memory", group_n_to_bufs_after_swap_dealloc_by_candidate, ) if iterative_recompute_error: break candidate = _prev[group_head] curr = _next_curr new_snodes = _format_and_log_reordering_stats( stats, _head, _next, original_snodes_num, peak_memory, name_to_freeable_input_buf, graph_outputs, ) return new_snodes, stats def _schedule_for_comm( snodes: list[BaseSchedulerNode], raise_comms: bool, sink_waits: bool, reorder_for_overlap: bool, ) -> list[BaseSchedulerNode]: """ Schedule `snodes` for various comm optimization objectives. Args: snodes: the nodes to be scheduled. raise_comms: whether to greedily schedule collectives as early as possible sink_wait: whether to greedily schedule waits as late as possible reorder_compute_for_overlap: whether to reorder compute nodes to optimize for compute/communication overlapping. Returns: The new schedule order. Some notes on the synergy between different options: - `raise_comms` provides more overlapping oppurtunies for `reorder_compute_for_overlap`. - When both `raise_comms` and `sink_waits` is `True`, `raise_comms` is prioritized. """ # We assign each node a tuple of scores (score_0, score_1, score_2), # decreasing in importance, with a lower value indicating a higher ranking: # # - score_0: the lowest comm_idx among the comm nodes that the node blocks. # If a node doesn't block any comm nodes, its score_0 is set to # sys.maxsize. This score ensures that comm nodes get scheduled as early as # possible. # - score_1: 1 if the node is a wait node, 0 otherwise. This score ensures # that wait nodes are deferred as late as possible. # - score_2: the index of the node in the original topological order. This # score provides stability in case of ties. # # When only raise_comms is True, only score_0 and score_2 are considered. # When only sink_waits is True, only score_1 and score_2 are considered. # When neither is True, the original order is yielded. buf_name_to_snode = {} name_to_fused_node = {} scores_0, scores_1, scores_2 = {}, {}, {} for idx, snode in enumerate(snodes): for buf_name in snode.get_buffer_names(): buf_name_to_snode[buf_name] = snode for op_name in snode.get_operation_names(): name_to_fused_node[op_name] = snode name_to_fused_node[snode.get_name()] = snode node_name = snode.get_name() scores_0[node_name] = sys.maxsize scores_1[node_name] = 0 scores_2[node_name] = idx comm_idx = 0 for snode in snodes: if raise_comms and contains_collective(snode): scores_0[snode.get_name()] = comm_idx for ancestor in snode.ancestors: anc_fused_name = name_to_fused_node[ancestor].get_name() scores_0[anc_fused_name] = min(scores_0[anc_fused_name], comm_idx) comm_idx += 1 elif sink_waits and contains_wait(snode): scores_1[snode.get_name()] = 1 class Runnable: def __init__(self, snode) -> None: self.snode = snode name = next(iter(snode.get_operation_names())) fused_name = name_to_fused_node[name].get_name() self.score = ( scores_0[fused_name], scores_1[fused_name], scores_2[fused_name], ) def __lt__(self, other): return self.score < other.score unmet_deps: dict[BaseSchedulerNode, OrderedSet[str]] = { snode: OrderedSet(dep.name for dep in snode.unmet_dependencies) for snode in snodes } ready: list[Runnable] = [] buffer_users: dict[str, OrderedSet[BaseSchedulerNode]] = defaultdict(OrderedSet) snode_to_cost = {snode: estimate_op_runtime(snode) for snode in snodes} for snode, deps in unmet_deps.items(): if len(deps) == 0: heapq.heappush(ready, Runnable(snode)) for dep in deps: buffer_users[dep].add(snode) scheduled = [] def schedule(snode): """ Schedules `snode` and put all unblocked nodes onto the ready queue. """ scheduled.append(snode) for buf_name in snode.get_buffer_names(): for snode in buffer_users[buf_name]: unmet_deps[snode].remove(buf_name) if len(unmet_deps[snode]) == 0: heapq.heappush(ready, Runnable(snode)) def get_overlapping_candidate(): """ Return the next node in the ready queue that's neither a collective or a wait. """ candidates = [ x for x in ready if not contains_collective(x.snode) and not contains_wait(x.snode) ] if len(candidates) == 0: return None return min(candidates, key=lambda x: x.score) def schedule_collective_for_overlap(snode): """ Schedules collective node `snode`, along with one or more compute nodes to overlap with it. The strategy is described in the comment of `reorder_compute_for_overlap`. """ assert contains_collective(snode) schedule(snode) collective_cost = snode_to_cost[snode] while ( collective_cost > 0 and (candidate := get_overlapping_candidate()) is not None ): ready.remove(candidate) schedule(candidate.snode) collective_cost -= snode_to_cost[candidate.snode] heapq.heapify(ready) while ready: snode = heapq.heappop(ready).snode if reorder_for_overlap and contains_collective(snode): schedule_collective_for_overlap(snode) else: schedule(snode) for deps in unmet_deps.values(): assert len(deps) == 0, ( f"Detected unscheduled nodes. Nodes with unmet dependencies: {unmet_deps}" ) return scheduled def decide_global_ordering_of_comms( nodes: list[BaseSchedulerNode], name_to_buf, name_to_fused_node ) -> list[BaseSchedulerNode]: """ Decide global ordering of comms, by just enforcing the ordering that's in the input graph (might not be the same ordering as the eager mode program). TODO: Come up with a better approach """ if not torch.distributed.is_available(): return nodes comm_nodes = [n for n in nodes if contains_collective(n)] for i in range(1, len(comm_nodes)): # Enforce ordering by making previous comm a `WeakDep` dependency of the next comm mutating_buf = next(iter(comm_nodes[i].get_buffer_names())) for buf in comm_nodes[i - 1].get_buffer_names(): comm_nodes[i].add_fake_dep( WeakDep(buf, mutating_buf=mutating_buf, is_fake=True) ) return nodes @dataclass class SinkWaitInfo: grouped: int = 0 grouped_info: str = "" moves: int = 0 moves_info: str = "" limiting_factor: str = "None" comm_time: float = -1.0 comp_time: float = -1.0 initial_exposed: float = -1.0 final_exposed: float = -1.0 overlap_info: str = "None" @property def improvement(self): return self.initial_exposed - self.final_exposed def _is_node_groupable_for_sink_waits( candidate: BaseSchedulerNode, ) -> tuple[bool, Optional[str]]: """ Check if a candidate node can be grouped during sink_waits pass. Sink Waits traverses waits right to left, so we don't group with processed waits on the right or with async collectives. Args: candidate: Node to check for groupability Returns: Tuple of (is_groupable, reason_if_not_groupable) """ # Sink Waits traverse Waits right to left, # => we do not group with processed Waits on the right. if contains_wait(candidate): return False, f"candidate contains wait {candidate.get_name()}" if contains_async_collective(candidate): return ( False, f"candidate contains_async_collective {candidate.get_name()}", ) # pyrefly: ignore[unbound-name] if not config_comms.sink_iterative_use_runtime_estimations: # Heuristics pre-use_runtime_estimations: # TODO(ivankobzarev): Remove them after confirming, # that using runtime estimations always give better results. # We do not want to group with collectives to not reorder them forward. if contains_collective(candidate): return ( False, f"candidate contains collective {candidate.get_name()}", ) if contains_gemm_like(candidate): return ( False, f"candidate contains gemm_like {candidate.get_name()}", ) return True, None def _update_memory_tracking_after_swap_sink_waits( candidate: BaseSchedulerNode, gns: list[BaseSchedulerNode], candidate_delta_mem: int, candidate_allocfree: SNodeMemory, group_n_to_bufs_after_swap_dealloc_instead_of_candidate: dict, post_alloc_update: dict[BaseSchedulerNode, int], size_free_delta_update: dict[BaseSchedulerNode, int], curr_memory: dict, snodes_allocfree: dict, ) -> None: """ Update memory tracking structures after swap (sink_waits version). Updates curr_memory and snodes_allocfree dictionaries to reflect the new memory state after swapping candidate with group. Args: candidate: Node that was moved gns: Group nodes candidate_delta_mem: Net memory change from candidate (alloc - free) candidate_allocfree: Candidate's allocation/free info group_n_to_bufs_after_swap_dealloc_instead_of_candidate: Buffers whose deallocation moves from candidate to group post_alloc_update: Cached post-allocation memory values size_free_delta_update: Cached size-free delta values curr_memory: Current memory state dict (mutated) snodes_allocfree: Node allocation/free info dict (mutated) """ group_head = gns[0] pre_group_mem = curr_memory[group_head][0] - snodes_allocfree[group_head].size_alloc if not group_n_to_bufs_after_swap_dealloc_instead_of_candidate: candidate_post_alloc = pre_group_mem + candidate_allocfree.size_alloc curr_memory[candidate] = ( candidate_post_alloc, candidate_post_alloc - candidate_allocfree.size_free, ) for gn in gns: cm = curr_memory[gn] curr_memory[gn] = ( cm[0] + candidate_delta_mem, cm[1] + candidate_delta_mem, ) return for n in [candidate, *gns]: post_alloc = post_alloc_update[n] snodes_allocfree[n].size_free += size_free_delta_update.get(n, 0) curr_memory[n] = ( post_alloc, post_alloc - snodes_allocfree[n].size_free, ) def _calculate_potential_peak_memory_sink_waits( candidate: BaseSchedulerNode, gns: list[BaseSchedulerNode], group_head: BaseSchedulerNode, group_peak_memory: int, candidate_delta_mem: int, candidate_allocfree: SNodeMemory, group_n_to_bufs_after_swap_dealloc_instead_of_candidate: dict, curr_memory: dict, snodes_allocfree: dict, ) -> tuple[int, dict[BaseSchedulerNode, int], dict[BaseSchedulerNode, int]]: """ Calculate potential peak memory after swapping candidate with group (sink_waits version). Computes new memory levels for all affected nodes and returns the potential peak memory along with cached post-allocation and size-free delta values. Args: candidate: Node being moved gns: Group nodes group_head: First node of group group_peak_memory: Current peak memory within the group candidate_delta_mem: Net memory change from candidate (alloc - free) candidate_allocfree: Candidate's allocation/free info group_n_to_bufs_after_swap_dealloc_instead_of_candidate: Buffers whose deallocation moves from candidate to group curr_memory: Current memory state dict snodes_allocfree: Allocation/free info for all nodes Returns: Tuple of (potential_peak_memory, post_alloc_update_dict, size_free_delta_update_dict) """ pre_group_mem = curr_memory[group_head][0] - snodes_allocfree[group_head].size_alloc # Stash memory tracing updates to not recompute them after swap _post_alloc_update: dict[BaseSchedulerNode, int] = {} _size_free_delta_update: dict[BaseSchedulerNode, int] = {} potential_peak = 0 if not group_n_to_bufs_after_swap_dealloc_instead_of_candidate: # Not accounting for buffers liveliness change potential_peak = max( group_peak_memory + candidate_delta_mem, pre_group_mem + candidate_allocfree.size_alloc, ) return potential_peak, _post_alloc_update, _size_free_delta_update candidate_post_alloc = pre_group_mem + candidate_allocfree.size_alloc _post_alloc_update[candidate] = candidate_post_alloc potential_peak = candidate_post_alloc candidate_size_free_to_move = sum( buf.mpi_buffer.size_free # type: ignore[attr-defined] for buf in itertools.chain.from_iterable( group_n_to_bufs_after_swap_dealloc_instead_of_candidate.values() ) ) _size_free_delta_update[candidate] = -candidate_size_free_to_move delta_mem = candidate_delta_mem + candidate_size_free_to_move for gn in gns: gn_post_alloc = curr_memory[gn][0] + delta_mem _post_alloc_update[gn] = gn_post_alloc potential_peak = max(potential_peak, gn_post_alloc) gn_size_free_to_add = 0 if gn in group_n_to_bufs_after_swap_dealloc_instead_of_candidate: bufs = group_n_to_bufs_after_swap_dealloc_instead_of_candidate[gn] for buf in bufs: gn_size_free_to_add += buf.mpi_buffer.size_free _size_free_delta_update[gn] = gn_size_free_to_add delta_mem -= gn_size_free_to_add return potential_peak, _post_alloc_update, _size_free_delta_update def _perform_double_linked_list_swap_sink_waits( candidate: BaseSchedulerNode, group_head: BaseSchedulerNode, group_tail: BaseSchedulerNode, prev_dict: dict[BaseSchedulerNode, Optional[BaseSchedulerNode]], next_dict: dict[BaseSchedulerNode, Optional[BaseSchedulerNode]], head: BaseSchedulerNode, ) -> BaseSchedulerNode: """ Swap positions of candidate and group in doubly-linked list (sink_waits version). Transforms (moves candidate to the left): group_head_prev -> group_head...group_tail -> candidate -> candidate_next Into: group_head_prev -> candidate -> group_head...group_tail -> candidate_next Args: candidate: Node to swap with group group_head: First node of group group_tail: Last node of group prev_dict: Dictionary mapping nodes to their previous nodes next_dict: Dictionary mapping nodes to their next nodes head: Current head of the linked list Returns: New head of the linked list (may change if group_head was the head) """ # 0: Update group_head's previous node group_head_prev = prev_dict[group_head] if group_head_prev: next_dict[group_head_prev] = candidate prev_dict[candidate] = group_head_prev # 2: Update candidate's next node candidate_next = next_dict[candidate] if candidate_next: prev_dict[candidate_next] = group_tail next_dict[group_tail] = candidate_next # 1: Link candidate to group_head prev_dict[group_head] = candidate next_dict[candidate] = group_head # Update head if group_head was the head if group_head == head: return candidate return head def _format_and_log_sink_waits_stats( stats: dict[BaseSchedulerNode, SinkWaitInfo], head: BaseSchedulerNode, next_dict: dict[BaseSchedulerNode, Optional[BaseSchedulerNode]], original_snodes_num: int, peak_memory: int, name_to_freeable_input_buf: dict, graph_outputs: OrderedSet[str], ) -> list[BaseSchedulerNode]: """ Format sink_waits statistics, log them, and return final node list. Computes improvement metrics, creates a formatted table (using tabulate if available), validates the reordered node count, recalculates peak memory, and logs all information. Args: stats: Per-node sink_waits statistics head: Head of the reordered linked list next_dict: Linked list next pointers original_snodes_num: Original number of nodes (for validation) peak_memory: Initial peak memory before reordering name_to_freeable_input_buf: Buffer memory tracking info graph_outputs: Graph output names Returns: Final reordered list of scheduler nodes """ headers = [ "Wait node", "comm_time(us)", "comp_time(us)", "initial exposed(us)", "final exposed(us)", "improvement(us)", "limiting factor", "grouped", "grouped_info", "moves", "moves_info", "overlap_info", ] rows = [ [ node_summary(snode), info.comm_time / 1e3, info.comp_time / 1e3, info.initial_exposed / 1e3, info.final_exposed / 1e3, info.improvement / 1e3, info.limiting_factor, info.grouped, info.grouped_info, info.moves, info.moves_info, info.overlap_info, ] for snode, info in stats.items() ] log_str = "" if importlib.util.find_spec("tabulate"): # pyrefly: ignore[import-error] from tabulate import tabulate log_str += tabulate( rows, headers=headers, ) else: log_str += "Please `pip install tabulate` to nicely render overlap stats.\n" log_str += str(headers) + "\n" log_str += "\n".join(map(str, rows)) overlap_log.info(log_str) new_snodes = _group_nodes_from_linked_list(head, None, next_dict) assert len(new_snodes) == original_snodes_num new_peak_memory, _, _, _ = estimate_peak_memory_allocfree( new_snodes, name_to_freeable_input_buf, graph_outputs ) log_str += f"\n sink_waits_iterative peak_memory_before:{peak_memory}" log_str += f"\n sink_waits_iterative peak_memory_after:{new_peak_memory}" trace_structured( "artifact", metadata_fn=lambda: { "name": "sink_waits_iterative_info", "encoding": "string", }, payload_fn=lambda: log_str, ) return new_snodes def _find_buffers_with_changed_last_use_sink_waits( candidate: BaseSchedulerNode, gns: list[BaseSchedulerNode], buf_to_snode_last_use: dict, ) -> dict[BaseSchedulerNode, list[Union[FreeableInputBuffer, Any]]]: """ Find buffers whose last use will change after swapping in sink_waits pass. When we swap [group] candidate to candidate [group], some buffers that were last used by candidate will now be last used by a group node instead. This is the opposite direction from the reorder version. Args: candidate: The node being moved (currently last use) gns: Group nodes being swapped with candidate buf_to_snode_last_use: Mapping of buffers to their current last-use nodes Returns: Dict mapping group nodes to buffers that will change their last-use node """ group_n_to_bufs_after_swap_dealloc_instead_of_candidate: dict[ BaseSchedulerNode, list[Union[FreeableInputBuffer, Any]] ] = defaultdict(list) for ( buf, snode_last_use, ) in buf_to_snode_last_use.items(): succ_nodes = buf.mpi_buffer.succ_nodes if snode_last_use != candidate: # noqa: E711 continue # candidate is last use of buf last_succ_gn = None for gn in gns: if gn in succ_nodes: last_succ_gn = gn if last_succ_gn is None: continue # gn has successors of buf that after potential swap will become # last use of buf and start deallocating buf instead of candidate group_n_to_bufs_after_swap_dealloc_instead_of_candidate[last_succ_gn].append( buf ) return group_n_to_bufs_after_swap_dealloc_instead_of_candidate def _sink_waits_iterative_internal( snodes: list[BaseSchedulerNode], ) -> tuple[list[BaseSchedulerNode], dict[BaseSchedulerNode, SinkWaitInfo]]: from torch._inductor.scheduler import GroupedSchedulerNode original_snodes_num = len(snodes) if original_snodes_num == 0: return snodes, {} graph_inputs: OrderedSet[str] = OrderedSet(V.graph.graph_inputs.keys()) graph_outputs: OrderedSet[str] = OrderedSet(V.graph.get_output_names()) ( peak_memory, _curr_memory, snodes_allocfree, buf_to_snode_last_use, name_to_freeable_input_buf, ) = _initialize_memory_tracking(snodes, graph_inputs, graph_outputs) _prev, _next, _head = _initialize_double_linked_list(snodes) stats: dict[BaseSchedulerNode, SinkWaitInfo] = {} runtimes: dict[BaseSchedulerNode, float] = { snode: estimate_op_runtime(snode) * _op_runtime_estimate_mult(snode) for snode in snodes } curr: Optional[BaseSchedulerNode] = snodes[-1] processed_waits = OrderedSet() # type: ignore[var-annotated] debug_iterative_memory_recompute = ( config_comms.reorder_iterative_debug_memory_recompute ) debug_num_sink_waits_to_reorder: Optional[int] = ( config_comms.sink_waits_iterative_debug_limit_to_sink ) iterative_recompute_error = False while curr is not None and _prev[curr] is not None: _prev_curr = _prev[curr] if iterative_recompute_error: break if ( debug_num_sink_waits_to_reorder is not None and len(processed_waits) >= debug_num_sink_waits_to_reorder ): break # pyrefly: ignore [bad-argument-type] if not (contains_wait(curr) and curr not in processed_waits): curr = _prev_curr continue processed_waits.add(curr) info = stats[curr] = SinkWaitInfo() comm_time, comp_time, overlap_info = wait_exposed_communication_time( _group_nodes_from_linked_list(_head, curr, _next), runtimes ) info.initial_exposed = info.final_exposed = comm_time - comp_time info.comm_time = comm_time info.comp_time = comp_time info.overlap_info = overlap_info candidate = _next[curr] wait_snode = curr group_head = curr group_tail = curr group_colls = {} group_runtime = 0.0 group_peak_memory = _curr_memory[curr][0] while candidate is not None: if config_comms.sink_iterative_use_runtime_estimations and ( info.final_exposed < -config_comms.sink_iterative_extra_comm_comp_overlap * info.comm_time ): info.limiting_factor = "unexposed by runtime estimations" break gns: list[BaseSchedulerNode] = _group_nodes_from_linked_list( group_head, group_tail, _next ) group = GroupedSchedulerNode( wait_snode.scheduler, gns, temp_grouping=True, ) # We can have multiple deps with the same name. # As we ignore WeakDep(is_fake=True) => # filter them out first to avoid overwriting of real dep. data_deps = { d.name: d for d in candidate.unmet_dependencies if not _is_fake_dep(d) } group_outs = group.get_outputs() data_dep = None for o in group_outs: if d := data_deps.get(o.get_name(), None): data_dep = d break # Conservative sink wait, limiting by space before next collective. # The global strategy is that bucketing should create space. # For 2D we can experiment with allowing to sink Wait beyond non current group collective. # pyrefly: ignore[unbound-name] if not config_comms.sink_waits_iterative_swap_with_collectives: if contains_async_collective(candidate): info.limiting_factor = ( f"candidate contains_async_collective {candidate.get_name()}" ) break # 1. If we have data_dep - we can not swap => trying to group # 2. If swap candidate and current node both contain collectives => trying to group if data_dep is not None or ( both_contain_comms := ( contains_collective(group) and contains_collective(candidate) ) ): _is_groupable, groupable_reason = _is_node_groupable_for_sink_waits( candidate ) if _is_groupable: group_tail = candidate if ( # pyrefly: ignore[unbound-name] config_comms.sink_iterative_use_runtime_estimations and contains_collective(candidate) ): comm_time, comp_time, _ = coll_exposed_communication_time( _group_nodes_from_linked_list(candidate, None, _next), runtimes, ) group_colls[candidate] = (comm_time, comp_time) if not contains_async_collective(candidate): group_runtime += runtimes[candidate] group_peak_memory = max( group_peak_memory, _curr_memory[candidate][0] ) info.grouped += 1 info.grouped_info = _group_names(gns) candidate = _next[candidate] continue elif data_dep is None: if ( # pyrefly: ignore[unbound-name] not config_comms.sink_waits_iterative_unsafe_collectives_reorder and both_contain_comms ): info.limiting_factor = ( f"collective ordering {_group_names(gns)}" f"\n with candidate:{candidate.get_name()}" ) break else: info.limiting_factor = ( f"data dependency {data_dep}(dep_names:{list(data_deps.keys())})" f"\n candidate:{candidate.get_name()}(os:{[candidate.get_buffer_names()]})" f"\n dep on {_group_names(gns)}" f"\n outs:{[o.get_name() for o in group_outs]}" f"\n non_group_reason:{groupable_reason}" ) break # pyrefly: ignore[unbound-name] if config_comms.sink_iterative_use_runtime_estimations: if is_wait(candidate.node): # Corresponding collective is before the group, # Swap can increase exposed time of corresponding collective comm_time, comp_time, _ = wait_exposed_communication_time( _group_nodes_from_linked_list(_head, candidate, _next), runtimes ) # pyrefly: ignore[no-matching-overload] exposed_before = max(0, comm_time - comp_time) # pyrefly: ignore[no-matching-overload] exposed_after = max(0, comm_time - comp_time + group_runtime) # We do not know how much we can sink more after this swap, # Just comparing advantage at the moment for now. if exposed_after > exposed_before: info.limiting_factor = ( "candidate is wait," f" exposed_before:{exposed_before} vs exposed_after:{exposed_after}" ) break # Check if candidate has sync runtime if not contains_async_collective(candidate): # If candidate has sync runtime, # Waits of gorup_colls are on the right from group. # Swap can increase their exposed time. c_runtime = runtimes[candidate] if c_runtime > 0 and len(group_colls) > 0: # Advantage for current Wait to do the Swap # pyrefly: ignore[no-matching-overload] exposed_delta = max( 0, info.comm_time - info.comp_time, ) # pyrefly: ignore[no-matching-overload] -max(0, info.comm_time - info.comp_time - c_runtime) for gc_comm_time, gc_comp_time in group_colls.values(): exposed_delta += max(0, gc_comm_time - gc_comp_time) - max( 0, gc_comm_time - gc_comp_time + c_runtime ) if exposed_delta > 0: info.limiting_factor = ( f"candidate has compute {c_runtime}, group contains collectives," f" total_exposed_delta {exposed_delta}" ) break else: # Update all group_colls comm_time, comp_time for gc, ( gc_comm_time, gc_comp_time, ) in group_colls.items(): group_colls[gc] = ( gc_comm_time, gc_comp_time - c_runtime, ) candidate_allocfree: SNodeMemory = snodes_allocfree[candidate] candidate_delta_mem = ( candidate_allocfree.size_alloc - candidate_allocfree.size_free ) # [group] candidate -> candidate [group] # Check for buffers with successors in group and candidate last successor # # Buf that changes its last use snode, # It was deallocated by candidate, # but after swap it will be deallocated by group node. group_n_to_bufs_after_swap_dealloc_instead_of_candidate = ( _find_buffers_with_changed_last_use_sink_waits( candidate, gns, buf_to_snode_last_use ) ) potential_peak, _post_alloc_update, _size_free_delta_update = ( _calculate_potential_peak_memory_sink_waits( candidate, gns, group_head, group_peak_memory, candidate_delta_mem, candidate_allocfree, group_n_to_bufs_after_swap_dealloc_instead_of_candidate, _curr_memory, snodes_allocfree, ) ) if ( potential_peak - peak_memory # pyrefly: ignore[unbound-name] > peak_memory * config_comms.sink_iterative_peak_memory_budget ): info.limiting_factor = ( f"peak memory new:{potential_peak} vs base:{peak_memory}" ) break info.moves += 1 info.moves_info += f"+{candidate.get_name()}" _head = _perform_double_linked_list_swap_sink_waits( candidate, group_head, group_tail, _prev, _next, _head ) comm_time, comp_time, overlap_info = wait_exposed_communication_time( _group_nodes_from_linked_list(_head, curr, _next), runtimes ) info.comm_time = comm_time info.comp_time = comp_time info.final_exposed = comm_time - comp_time info.overlap_info = overlap_info _update_memory_tracking_after_swap_sink_waits( candidate, gns, candidate_delta_mem, candidate_allocfree, group_n_to_bufs_after_swap_dealloc_instead_of_candidate, _post_alloc_update, _size_free_delta_update, _curr_memory, snodes_allocfree, ) if debug_iterative_memory_recompute: from .comms_debug import _debug_iterative_memory_recompute iterative_recompute_error = _debug_iterative_memory_recompute( candidate, gns, _group_names(gns), _group_nodes_from_linked_list(_head, None, _next), name_to_freeable_input_buf, graph_outputs, peak_memory, _curr_memory, snodes_allocfree, "sink_waits_iterative", group_n_to_bufs_after_swap_dealloc_instead_of_candidate, ) if iterative_recompute_error: break candidate = _next[group_tail] curr = _prev_curr new_snodes = _format_and_log_sink_waits_stats( stats, _head, _next, original_snodes_num, peak_memory, name_to_freeable_input_buf, graph_outputs, ) return new_snodes, stats def sink_waits_iterative(snodes: list[BaseSchedulerNode]) -> list[BaseSchedulerNode]: """ Similarly to reorder_communication_preserving_peak_memory this pass will try to iteratively push Wait nodes later, recomputing estimated peak memory before each swap, and preventing peak memory regressions. Pass will be applied to every Wait node. If there are immediate dependencies with next node, pass will try to group them together and on the next step to swap the group with next candidate. If _inductor.config_comms.sink_iterative_use_runtime_estimations is set True, pass will stop reordering of Wait once corresponding Collective is unexposed, based on runtime estimations. inductor.config_comms.sink_iterative_peak_memory_budget allows to tune how much pass can regress initial peak memory. E.g.: sink_iterative_peak_memory_budget == 0.0 - No regression of initial peak memory is allowed sink_iterative_peak_memory_budget == 0.2 - Pass can improve comm-compute overlap, sacrificing 20% of initial peak memory value. inductor.config_comms.sink_iterative_extra_comm_comp_overlap config allows to more aggressively sink waits, stopping only when overlap_compute >= (1 + extra_comm_comp_overlap) * comm_time """ return _sink_waits_iterative_internal(snodes)[0] def estimate_op_runtime(snode: BaseSchedulerNode) -> float: """ Returns estimated op runtime in milliseconds (ms) """ if config.estimate_op_runtime == "default": runtime = snode.get_estimated_runtime() else: assert callable(config.estimate_op_runtime) runtime = config.estimate_op_runtime(snode) return runtime def node_summary(snode): snodes = snode.get_nodes() if len(snodes) == 1: detail = "" if isinstance(snode.node, (ir.ExternKernelOut, ir._CollectiveKernel)): outs_str = f"outs:{[o.get_name() for o in snode.get_outputs()]}" ins_str = f"ins:{[d.name for d in snode.unmet_dependencies]}" detail = f" {snode.get_name()} ({snode.node.python_kernel_name})\n {outs_str}({ins_str})" layouts = [child.node.get_output_spec() for child in snode.get_nodes()] out_tensor_info = ",".join( [ f" (size={layout.size}, stride={layout.stride})" if isinstance(layout, ir.Layout) else "" for layout in layouts ] ) try: node_name = snode.node.maybe_get_name() except AttributeError: # TODO: node_summary was written without FusedSchedulerNode in mind, generally needs to be hardened node_name = "" return f"{snode.node.__class__.__name__}{detail}{out_tensor_info} ({node_name} ({snode.get_estimated_runtime():.0f} ns)" # Flatten the summaries for Fused/Foreach/Grouped nodes summaries = [] for child_snode in snodes: summaries.append(node_summary(child_snode)) return f"{snode.__class__.__name__}: {', '.join(summaries)}" def visualize_overlap(order): # TODO - this function probably doesn't do a very good job estimating the runtime because it doesn't carefully model # streams and overlap. For now its mostly useful as a debug visualization. total_est_runtime: float = 0.0 cur_comm_node = None def step_log(step, msg): overlap_log.debug(f"{step:>6}: {msg}") # noqa: G004 for step, snode in enumerate(order): if cur_comm_node is None: if contains_collective(snode): total_est_runtime += estimate_op_runtime(snode) cur_comm_node = snode.node elif is_wait(snode.node): # raise AssertionError( # "Wait is not expected when there is no collective running" # ) pass else: # exposed compute op total_est_runtime += estimate_op_runtime(snode) step_log(step, f"{node_summary(snode)}") else: # cur_comm_node is not None if contains_collective(snode): total_est_runtime += estimate_op_runtime(snode) cur_comm_node = snode.node step_log(step, f"{node_summary(snode)}") # noqa: G004 elif is_wait(snode.node): # end of this comm op step_log(step, f"{node_summary(snode)}") cur_comm_node = None else: # overlapped compute op step_log(step, f"| {node_summary(snode)}") overlap_log.debug( f"Est. runtime (ms): {total_est_runtime / 1000 / 1000}" # noqa: G004 ) def reorder_compute_and_comm_for_overlap( snodes: list[BaseSchedulerNode], ) -> list[BaseSchedulerNode]: order = snodes graph_inputs: OrderedSet[str] = OrderedSet(V.graph.graph_inputs.keys()) graph_outputs: OrderedSet[str] = OrderedSet(V.graph.get_output_names()) for p in config.reorder_for_compute_comm_overlap_passes: if isinstance(p, str) and p in globals(): p = globals()[p] # it is a builtin pass assert callable(p), ( f"Invalid reorder_compute_and_comm_for_overlap pass: {p} is not callable" ) peak_memory, _ = estimate_peak_memory( snodes, get_freeable_input_buf(snodes, graph_inputs), graph_outputs ) if torch.distributed.get_rank() == 0: overlap_log.debug( f"==== Visualize overlap before reordering pass {p}, {peak_memory=} ====" # noqa: G004 ) try: visualize_overlap(order) except Exception as e: overlap_log.debug("", exc_info=e) t0 = time.time() order = p(order) # type: ignore[operator] t = time.time() - t0 if torch.distributed.get_rank() == 0: overlap_log.debug( f"==== Visualize overlap after reordering pass {p} (ran in {t} sec)====" # noqa: G004 ) try: visualize_overlap(order) except Exception as e: overlap_log.debug("", exc_info=e) peak_memory, _ = estimate_peak_memory( snodes, get_freeable_input_buf(snodes, graph_inputs), graph_outputs ) print(f"final {peak_memory=}") # pyrefly: ignore [bad-return] return order def remove_fsdp2_unsharded_param_graph_input_usage(graph: torch.fx.Graph): """ This FX graph pass replaces uses of FSDP2 unsharded params with their corresponding graph intermediates that were fsdp.copy_ into the unsharded params in the original graph. NOTE: Can only apply this pass to any of the FSDP2 unsharded params that have this pattern (or repetition of): `resize_(full) -> copy_ -> resize_(0)`. Because of this, for partial-graph case where `resize_(full) -> copy_` is in one graph and `resize_(0)` is in another graph, we can't remove these resize and copy ops and thus we will have worse performance there. In other words, "do we try to remove all the resize_(full) -> copy_ -> resize_(0) nodes for this unsharded param" is actually a per-unsharded-param decision, since for each unsharded param, we look at its resize sequence pattern (in `check_resize_pattern()`) to determine if its set of resize and copy nodes can be removed. """ node_list = list(graph.nodes) # Find all graph inputs and their resize counts graph_input_to_resized_to_full_node_idxes = defaultdict(list) graph_input_to_resized_to_0_node_idxes = defaultdict(list) for idx, node in enumerate(node_list): if ( node.op == "call_function" and node.target is torch.ops.inductor.resize_storage_bytes_.default ): assert node.args[0].op == "placeholder", f"""\ Resize can only operate on graph inputs, but got {node} which is resizing non-graph-input {node.args[0]} """ graph_input = node.args[0] new_size = node.args[1] if new_size > 0: graph_input_to_resized_to_full_node_idxes[graph_input].append(idx) else: graph_input_to_resized_to_0_node_idxes[graph_input].append(idx) def check_resize_pattern(graph_input): # Check the number of resize-to-full and resize-to-0 nodes are equal, # and that for each (resize-to-full, resize-to-0) pair, the resize-to-full node # always happens before the resize-to-0 node. # This is the precondition for being able to remove all the resize and copy nodes # for this specific unsharded param. resized_to_full_idxes = graph_input_to_resized_to_full_node_idxes.get( graph_input, [] ) resized_to_0_idxes = graph_input_to_resized_to_0_node_idxes.get(graph_input, []) if len(resized_to_full_idxes) != len(resized_to_0_idxes): log.warning( f""" Unequal number of resize-to-full and resize-to-0 nodes for graph input {graph_input}: {len(resized_to_full_idxes)} vs. {len(resized_to_0_idxes)}. Skipping `remove_fsdp2_unsharded_param_graph_input_usage` FX graph pass. """ # noqa: G004 ) return False # Check the sequence: (resize_to_full -> resize_to_0)+ for resize_to_full_idx, resize_to_0_idx in zip( resized_to_full_idxes, resized_to_0_idxes ): if resize_to_full_idx >= resize_to_0_idx: log.warning( f""" For graph input {graph_input}: resize-to-full node {node_list[resize_to_full_idx]} at index {resize_to_full_idx} happens after resize-to-0 node {node_list[resize_to_0_idx]} at index {resize_to_0_idx}. Skipping `remove_fsdp2_unsharded_param_graph_input_usage` FX graph pass for that unsharded param. """ # noqa: G004 ) return False return True # Find all eligible unsharded params and their corresponding graph intermediates. unsharded_param_to_fsdp_copy_node_idxes = defaultdict(list) for idx, node in enumerate(node_list): if node.op == "call_function" and node.target is torch.ops.fsdp.copy_.default: fsdp_copy_node = node unsharded_param = node.args[0] assert unsharded_param.op == "placeholder", f""" Assumed all FSDP2 `unsharded_param`s to be graph input, but it's not true! Offending node: {unsharded_param}. Graph: {graph} """ if check_resize_pattern(unsharded_param): unsharded_param_to_fsdp_copy_node_idxes[unsharded_param].append(idx) def is_allowed_mutation(node): return ( node.target is torch.ops.fsdp.copy_.default or node.target is torch.ops.inductor.resize_storage_bytes_.default ) def is_node_mutating_unsharded_param_or_its_alias(node, unsharded_params): # Check whether the node is mutating any of the unsharded params or their aliases. mutated_arg_idxes = ( [ i for i, x in enumerate(node.target._schema.arguments) if x.alias_info is not None and x.alias_info.is_write ] if isinstance(node.target, torch._ops.OpOverload) else [] ) mutated_node_arg_storages = OrderedSet( [ StorageWeakRef(node.args[i].meta["val"].untyped_storage()) for i in mutated_arg_idxes ] ) storages_of_unsharded_params = OrderedSet( [ StorageWeakRef(unsharded_param.meta["val"].untyped_storage()) for unsharded_param in unsharded_params ] ) return len(mutated_node_arg_storages & storages_of_unsharded_params) > 0 # Check no user mutation on any unsharded_param for node in node_list: if ( node.op == "call_function" and isinstance(node.target, torch._ops.OpOverload) and node.target._schema.is_mutable and not is_allowed_mutation(node) ): assert not is_node_mutating_unsharded_param_or_its_alias( node, unsharded_param_to_fsdp_copy_node_idxes.keys() ), f"""\ User mutation on FSDP2 unsharded param is not allowed when Traceable FSDP2 is used. Violating node: {node} """ # For each `fsdp.copy_(unsharded_param, Y)`, replace downstream usage of `unsharded_param` with `Y`. # # NOTE: Because of "layer reuse" use case, there could be multiple `fsdp.copy_` to the same `unsharded_param` graph input. # e.g. # ``` # fsdp_copy_1 = fsdp.copy_(unsharded_param_1, Y1) # ... (use of unsharded_param_1) -> Subgraph 1 # fsdp_copy_2 = fsdp.copy_(unsharded_param_1, Y2) # ... (use of unsharded_param_1) -> Subgraph 2 # fsdp_copy_3 = fsdp.copy_(unsharded_param_1, Y3) # ... (use of unsharded_param_1) -> Subgraph 3 # ``` # We must do the replacement only within each subgraph. for ( unsharded_param, fsdp_copy_node_idxes, ) in unsharded_param_to_fsdp_copy_node_idxes.items(): for i, fsdp_copy_node_idx in enumerate(fsdp_copy_node_idxes): fsdp_copy_node = node_list[fsdp_copy_node_idx] assert fsdp_copy_node.args[0] is unsharded_param _, replacement = fsdp_copy_node.args # subgraph_start_idx is exclusive subgraph_start_idx = fsdp_copy_node_idx + 1 # subgraph_end_idx is exclusive (also intentionally don't replace args in return op) subgraph_end_idx = ( fsdp_copy_node_idxes[i + 1] if i < len(fsdp_copy_node_idxes) - 1 else len(node_list) - 1 ) subgraph_nodes = node_list[subgraph_start_idx:subgraph_end_idx] assert not any( is_node_mutating_unsharded_param_or_its_alias(node, [unsharded_param]) for node in subgraph_nodes ), f"""\ Assumed no ops mutating unsharded param {unsharded_param} in subgraph {subgraph_nodes}, but it's not true! Graph: {graph} """ for node in subgraph_nodes: if ( node.op == "call_function" and unsharded_param in node.args and node.target != torch.ops.inductor.resize_storage_bytes_.default ): # TODO(yf225): implement replacement in kwargs new_args = tuple( replacement if arg is unsharded_param else arg for arg in node.args ) node.args = new_args # Delete `fsdp.copy_(unsharded_param, Y)` nodes for fsdp_copy_node_idxes in unsharded_param_to_fsdp_copy_node_idxes.values(): for fsdp_copy_node_idx in fsdp_copy_node_idxes: fsdp_copy_node = node_list[fsdp_copy_node_idx] graph.erase_node(fsdp_copy_node) # Delete `resize_(unsharded_param, ...)` nodes for node in node_list: if ( node.op == "call_function" and node.target is torch.ops.inductor.resize_storage_bytes_.default and node.args[0] in unsharded_param_to_fsdp_copy_node_idxes ): graph.erase_node(node) def reinplace_fsdp_all_gather(graph: torch.fx.Graph) -> None: try: import torch.distributed.fsdp._fully_shard._fsdp_collectives assert torch.distributed.is_available() # Assert existence of these ops assert ( torch.ops._c10d_functional.all_gather_into_tensor and torch.ops._c10d_functional.all_gather_into_tensor_out ) except (ImportError, AttributeError, AssertionError): return from .pattern_matcher import ( CallFunction, KeywordArg, Match, PatternMatcherPass, register_graph_pattern, ) """ all_gather_copy_in = torch.ops.fsdp.all_gather_copy_in.default(...); getitem = all_gather_copy_in[0]; (getitem_1 = all_gather_copy_in[1];) # optional all_gather_into_tensor = torch.ops._c10d_functional.all_gather_into_tensor.default(getitem, ...); -> all_gather_copy_in = torch.ops.fsdp.all_gather_copy_in.default(...); getitem = all_gather_copy_in[0]; getitem_1 = all_gather_copy_in[1]; all_gather_into_tensor = torch.ops._c10d_functional.all_gather_into_tensor_out.default(getitem, ..., out=getitem_1); """ def remove_unused_getitem(g): # Remove `getitem_X = all_gather_copy_in[1]` which is never used. node_list = list(g.nodes) for n in node_list: if ( n.target is operator.getitem and n.args[0].target is torch.ops.fsdp.all_gather_copy_in.default and n.args[1] == 1 ): g.erase_node(n) graph_pass = PatternMatcherPass() @register_graph_pattern( CallFunction( torch.ops._c10d_functional.all_gather_into_tensor.default, CallFunction( operator.getitem, CallFunction( torch.ops.fsdp.all_gather_copy_in.default, KeywordArg("all_gather_inputs"), KeywordArg("all_gather_output"), KeywordArg("inp_split_sizes"), KeywordArg("all_gather_input_numel"), KeywordArg("rank"), ), KeywordArg("item_idx"), ), KeywordArg("group_size"), KeywordArg("group_name"), ), # pyrefly: ignore [bad-argument-type] pass_dict=graph_pass, extra_check=lambda match: match.kwargs["item_idx"] == 0, ) def reinplace_all_gather(match: Match, *args, **kwargs): def repl( *args, ): copy_in_args = args[:-2] group_size = args[-2] group_name = args[-1] all_gather_copy_in = torch.ops.fsdp.all_gather_copy_in.default( *copy_in_args ) getitem = all_gather_copy_in[0] getitem_1 = all_gather_copy_in[1] all_gather_into_tensor = ( torch.ops._c10d_functional.all_gather_into_tensor_out.default( getitem, group_size, group_name, out=getitem_1 ) ) return all_gather_into_tensor match.replace_by_example( # pyrefly: ignore [bad-argument-type] repl, [ kwargs["all_gather_inputs"], kwargs["all_gather_output"], kwargs["inp_split_sizes"], kwargs["all_gather_input_numel"], kwargs["rank"], kwargs["group_size"], kwargs["group_name"], ], ) remove_unused_getitem(graph) graph_pass.apply(graph) # type: ignore[arg-type] def get_op_idx(snode): assert not isinstance( snode, ( torch._inductor.scheduler.FusedSchedulerNode, torch._inductor.scheduler.GroupedSchedulerNode, ), ) return int(snode.get_name()[2:]) def enforce_comm_ordering_for_fsdp( snodes: list[torch._inductor.scheduler.BaseSchedulerNode], name_to_buf: dict[str, torch._inductor.scheduler.SchedulerBuffer], name_to_fused_node: dict[str, BaseSchedulerNode], ) -> list[torch._inductor.scheduler.BaseSchedulerNode]: from . import scheduler new_order: list[BaseSchedulerNode] = [] scheduled = OrderedSet[Any]() ag_exists = False rs_exists = False ag_grouped_node_to_wait_grouped_node = {} rs_grouped_node_to_wait_grouped_node = {} snode_name_to_final_snode = {} def _create_group_node(snodes_to_group): group_node = scheduler.GroupedSchedulerNode.create(snodes_to_group) for snode in snodes_to_group: snode_name_to_final_snode[snode.get_name()] = group_node snode_name_to_final_snode[group_node.get_name()] = group_node return group_node # Create grouped nodes for specific sets of ops for snode in snodes: # Case 1: Handle AllGather if is_collective( snode.node, op=torch.ops._c10d_functional.all_gather_into_tensor_out.default ) and any( is_fallback_op( name_to_fused_node[x].node, torch.ops.fsdp.all_gather_copy_in.default ) for x in snode.ancestors ): ag_exists = True ag_snode = snode ag_related_snode_set: OrderedSet[scheduler.BaseSchedulerNode] = OrderedSet() # Find the "cast + copy_in + getitem + all_gather" code block find_recursive_deps_of_node( ag_snode, ag_related_snode_set, name_to_buf, name_to_fused_node, ) # Find the "all_gather + all_gather_wait_tensor + copy_out" code block allowed_ops = OrderedSet( [ torch.ops._c10d_functional.all_gather_into_tensor_out.default, torch.ops._c10d_functional.wait_tensor.default, torch.ops.fsdp.split_with_sizes_copy.default, ] ) find_recursive_users_of_node( ag_snode, ag_related_snode_set, name_to_buf, name_to_fused_node, criteria_cb=lambda x: not ( isinstance(x, scheduler.NopKernelSchedulerNode) or ( isinstance(x, scheduler.ExternKernelSchedulerNode) and x.node.op_overload in allowed_ops # type: ignore[union-attr] ) ), ) # sort nodes by original operation order ag_related_snodes = sorted( ag_related_snode_set, key=lambda x: get_op_idx(x) ) # In the "reuse layer" case, some ops in the 2nd all-gather code block could also # depend on ops in the 1st all-gather code block, and we don't want to group them together. end_idx_of_current_ag_block = len(ag_related_snodes) copy_out_count = 0 for i in range(len(ag_related_snodes)): cur_snode = ag_related_snodes[i] if is_fallback_op( cur_snode.node, torch.ops.fsdp.split_with_sizes_copy.default ): copy_out_count += 1 if copy_out_count > 1: end_idx_of_current_ag_block = i break ag_related_snodes = ag_related_snodes[:end_idx_of_current_ag_block] # Group "cast + copy_in + getitem + all_gather" into one GroupedSchedulerNode wait_node_idx = None for i in range(len(ag_related_snodes) - 1): if isinstance(ag_related_snodes[i + 1].node, ir._WaitKernel): wait_node_idx = i + 1 break assert wait_node_idx is not None ag_group_node = _create_group_node(ag_related_snodes[:wait_node_idx]) # Group "all_gather_wait_tensor + copy_out" into one GroupedSchedulerNode ag_wait_group_node = _create_group_node(ag_related_snodes[wait_node_idx:]) ag_grouped_node_to_wait_grouped_node[ag_group_node] = ag_wait_group_node # Case 2: Handle ReduceScatter elif is_fallback_op(snode.node, torch.ops.fsdp.chunk_cat.default): rs_exists = True rs_snode = snode # Find the "reduce_scatter copy-in + reduce_scatter comm + reduce_scatter wait" code block rs_related_snode_set: OrderedSet[scheduler.BaseSchedulerNode] = OrderedSet() find_recursive_users_of_node( rs_snode, rs_related_snode_set, name_to_buf, name_to_fused_node, ) # sort nodes by original operation order rs_related_snodes = sorted( rs_related_snode_set, key=lambda x: get_op_idx(x) ) # Group "reduce_scatter copy-in + reduce_scatter comm" into one GroupedSchedulerNode wait_node_idx = None for i in range(len(rs_related_snodes) - 1): if isinstance(rs_related_snodes[i + 1].node, ir._WaitKernel): wait_node_idx = i + 1 break assert wait_node_idx is not None rs_group_node = _create_group_node(rs_related_snodes[:wait_node_idx]) # Group "reduce_scatter wait + related output nodes" into one GroupedSchedulerNode rs_wait_group_node = _create_group_node(rs_related_snodes[wait_node_idx:]) rs_grouped_node_to_wait_grouped_node[rs_group_node] = rs_wait_group_node assert len(snode_name_to_final_snode) > 0 if ag_exists: assert len(ag_grouped_node_to_wait_grouped_node) > 0 if rs_exists: assert len(rs_grouped_node_to_wait_grouped_node) > 0 # Build the new node schedule, taking GroupedSchedulerNode into account for snode in snodes: if snode.get_name() in snode_name_to_final_snode: snode = snode_name_to_final_snode[snode.get_name()] if snode in scheduled: continue new_order.append(snode) scheduled.add(snode) # Enforce AllGather ordering: previous AllGather's "wait then copy_out" group node must run # before next AllGather's "copy_in then AG" group node prev_ag_wait = None for ag_group_node, wait_group_node in ag_grouped_node_to_wait_grouped_node.items(): if prev_ag_wait is not None: mutating_buf = next(iter(ag_group_node.get_buffer_names())) for o in prev_ag_wait.get_outputs(): ag_group_node.add_fake_dep( WeakDep(o.get_name(), mutating_buf=mutating_buf, is_fake=True) ) prev_ag_wait = wait_group_node # Enforce ReduceScatter ordering: previous ReduceScatter's "wait" group node must run # before next ReduceScatter's "copy_in then RS" group node prev_rs_wait = None for rs_group_node, wait_group_node in rs_grouped_node_to_wait_grouped_node.items(): if prev_rs_wait is not None: mutating_buf = next(iter(rs_group_node.get_buffer_names())) for o in prev_rs_wait.get_outputs(): rs_group_node.add_fake_dep( WeakDep(o.get_name(), mutating_buf=mutating_buf, is_fake=True) ) prev_rs_wait = wait_group_node return new_order # type: ignore[return-value]