# SPDX-License-Identifier: Apache-2.0 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project """ Expert parallelism load balancer (EPLB) metrics and states. # Glossary - **Logical Expert**: An expert that is part of the model's logical structure. It holds a set of weights and is replicated across multiple physical experts. - **Redundant Expert**: To achieve load balancing, for some popular logical experts, we create additional copies of the expert weights. During inference, each of these copies can be routed to by the same set of tokens. - **Physical Expert**: An expert that is instantiated on a specific device. It is a replica of a logical expert and can be rearranged across devices. I.e., one logical expert may have multiple sets of weights initialized on different devices, and each of these sets is a physical expert. - **Local Physical Expert**: A physical expert that is instantiated on the current device. For example: DeepSeek-R1 has 256 logical experts, so each MoE layer has 256 sets of linear layer weights in the model parameters. If we add 32 redundant experts, DeepSeek-R1 will have 256 + 32 = 288 physical experts in total. And when deploying, we'll have 288 sets of linear layer weights for each MoE layer. If we have 32 EP ranks, then each GPU will hold 288 / 32 = 9 local physical experts. """ import time from collections.abc import Sequence from dataclasses import dataclass import torch from torch.distributed import ProcessGroup, all_reduce from vllm.config import ModelConfig, ParallelConfig from vllm.distributed.parallel_state import ( get_ep_group, get_node_count, in_the_same_node_as, ) from vllm.distributed.utils import StatelessProcessGroup from vllm.logger import init_logger from vllm.model_executor.models.interfaces import MixtureOfExperts from .rebalance_algo import rebalance_experts from .rebalance_execute import rearrange_expert_weights_inplace logger = init_logger(__name__) @dataclass class EplbModelState: """EPLB metrics.""" physical_to_logical_map: torch.Tensor """ Mapping from physical experts to logical experts. Shape: (num_moe_layers, num_physical_experts) # Example For a 2-layer MoE model with 6 physical experts and 4 logical experts on 3 EP ranks, the mapping could look like this: ``` [[0, 1, 2, 3, 0, 1], [0, 2, 0, 1, 0, 3]] ``` """ logical_to_physical_map: torch.Tensor """ Mapping from logical experts to physical experts. This is a sparse matrix, where -1 indicates no mapping. Shape: (num_moe_layers, num_logical_experts, num_redundant_experts + 1) # Example For a 2-layer MoE model with 6 physical experts and 4 logical experts on 3 EP ranks, the mapping could look like this: ``` [[[0, 4, -1], [1, 5, -1], [2, -1, -1], [3, -1, -1]], [[0, 2, 4], [3, -1, -1], [1, -1, -1], [5, -1, -1]]] ``` """ logical_replica_count: torch.Tensor """ Number of replicas for each logical expert. This is exactly the non-`-1` count in the `logical_to_physical_map`. Shape: (num_moe_layers, num_logical_experts) # Example For a 2-layer MoE model with 6 physical experts and 4 logical experts on 3 EP ranks, the count could look like this: ``` [[2, 2, 1, 1], [3, 1, 1, 1]] """ expert_load_pass: torch.Tensor """ Expert load during this forward pass. We use the token count each expert processes as the load. Shape: (num_moe_layers, num_physical_experts) """ expert_load_window: torch.Tensor """ A sliding window of expert load. Shape: (window_size, num_moe_layers, num_physical_experts) NOTE: The expert_load_view now records load for all physical experts rather than just local experts. This ensures consistent load statistics across different dispatch methods (naive all-to-all, DeepEP, pplx-kernels). The recorded load will be multiplied by dp_size when using naive all-to-all due to each DP rank contributing the same token set to the calculation. See: https://github.com/vllm-project/vllm/pull/22167#pullrequestreview-3086143856 """ model_name: str model: MixtureOfExperts class EplbState: """ EplbState of each expert parallel model. Key is the model config hash. """ def __init__(self, parallel_config: ParallelConfig, device: torch.device): self.parallel_config = parallel_config self.device = device self.model_states: dict[str, EplbModelState] = {} """ Current step in the sliding window. Different from `expert_rearrangement_step`, each EP rank may have its own `expert_load_window_step`. """ self.expert_load_window_step: int = 0 """ Size of the expert load sliding window. This is a constant and is taken from the config. """ self.expert_load_window_size: int = 0 """ Steps after last rearrangement. Will trigger a rearrangement if it exceeds the threshold. NOTE: Keep in mind that all EP ranks need to have the same `expert_rearrangement_step` value to ensure synchronization. Otherwise, the rearrangement will hang at collective communication calls. """ self.expert_rearrangement_step: int = 0 """ Interval for expert rearrangement steps. This is a constant and is taken from the config. """ self.expert_rearrangement_step_interval: int = 0 @staticmethod def build_initial_global_physical_to_logical_map( num_routed_experts: int, num_redundant_experts: int, ) -> Sequence[int]: """ Build an initial expert arrangement using the following structure: [original routed experts, redundant experts] Returns: physical_to_logical_map (Sequence[int]): A list of integers, where each integer is the index of the logical expert that the corresponding physical expert maps to. """ global_physical_to_logical_map = list(range(num_routed_experts)) global_physical_to_logical_map += [ i % num_routed_experts for i in range(num_redundant_experts) ] return global_physical_to_logical_map def validate_ep_configuration(self, new_model: MixtureOfExperts): """ Validate that the expert parallel configuration of the new model is the same as the existing models. """ if len(self.model_states) > 0: model = next(iter(self.model_states.values())).model if ( model.num_routed_experts != new_model.num_routed_experts or model.num_redundant_experts != new_model.num_redundant_experts or model.num_physical_experts != new_model.num_physical_experts or model.num_logical_experts != new_model.num_logical_experts or model.num_expert_groups != new_model.num_expert_groups ): raise RuntimeError( "Model: {} " "with config {} " "{} {} {} {} " "mismatch with new model {} " "with config {} " "{} {} {} {}".format( type(model), model.num_routed_experts, model.num_redundant_experts, model.num_physical_experts, model.num_logical_experts, model.num_expert_groups, type(new_model), new_model.num_routed_experts, new_model.num_redundant_experts, new_model.num_physical_experts, new_model.num_logical_experts, new_model.num_expert_groups, ) ) def add_model( self, model: MixtureOfExperts, model_config: ModelConfig, global_expert_load: torch.Tensor | None = None, old_global_expert_indices: torch.Tensor | None = None, rank_mapping: dict[int, int] | None = None, ): """ Build the initial EPLB state. """ self.validate_ep_configuration(model) physical_to_logical_map_list = ( EplbState.build_initial_global_physical_to_logical_map( model.num_routed_experts, model.num_redundant_experts, ) ) physical_to_logical_map = torch.tensor( physical_to_logical_map_list, device=self.device, ) # Assuming 8 GPUs per node, this supports up to # (1023 + 1) / 8 = 128 nodes for now. # TODO(rui): make this configurable MAX_EXPERT_REDUNDANCY = 1023 assert model.num_redundant_experts <= MAX_EXPERT_REDUNDANCY, ( f"num_redundant_experts {model.num_redundant_experts} " f"must be less than or equal to {MAX_EXPERT_REDUNDANCY}" ) max_slots_per_logical_expert = MAX_EXPERT_REDUNDANCY + 1 logical_to_physical_map = torch.full( (model.num_logical_experts, max_slots_per_logical_expert), -1, device=self.device, ) logical_replica_count = torch.zeros( (model.num_logical_experts,), device=self.device, dtype=torch.long, ) for i in range(model.num_physical_experts): logical_idx = physical_to_logical_map[i] logical_to_physical_map[logical_idx, logical_replica_count[logical_idx]] = i logical_replica_count[logical_idx] += 1 # Duplicate initial mapping for all layers physical_to_logical_map = ( physical_to_logical_map.unsqueeze(0) .expand( model.num_moe_layers, -1, ) .contiguous() ) logical_to_physical_map = ( logical_to_physical_map.unsqueeze(0) .expand( model.num_moe_layers, -1, -1, ) .contiguous() ) logical_replica_count = ( logical_replica_count.unsqueeze(0) .expand( model.num_moe_layers, -1, ) .contiguous() ) expert_load_pass = torch.zeros( (model.num_moe_layers, model.num_physical_experts), dtype=torch.int32, device=self.device, ) self.expert_load_window_size = self.parallel_config.eplb_config.window_size expert_load_window = torch.zeros( ( self.expert_load_window_size, model.num_moe_layers, model.num_physical_experts, ), dtype=torch.int32, device=self.device, ) # Set the initial progress of rearrangement to 3/4 eplb_step_interval = self.parallel_config.eplb_config.step_interval self.expert_rearrangement_step = max( 0, eplb_step_interval - eplb_step_interval // 4 ) self.expert_rearrangement_step_interval = eplb_step_interval if global_expert_load is not None: ep_group = get_ep_group().device_group assert global_expert_load.shape == ( model.num_moe_layers, model.num_logical_experts, ) assert global_expert_load.dtype == torch.int64 num_replicas = model.num_physical_experts num_groups = model.num_expert_groups num_nodes = get_node_count() num_gpus = ep_group.size() if num_gpus % num_nodes != 0: num_nodes = 1 logger.warning_once( f"num_gpus % num_nodes != 0, " "not using hierarchical rearrangement algorithm.\n" f"{num_gpus=}, {num_nodes=}" ) # Get new expert mappings ( new_physical_to_logical_map, new_logical_to_physical_map, new_logical_replica_count, ) = rebalance_experts( global_expert_load, num_replicas, num_groups, num_nodes, num_gpus, ) max_physical_slots = new_logical_to_physical_map.shape[-1] assert max_physical_slots <= logical_to_physical_map.shape[-1] new_logical_to_physical_map = torch.nn.functional.pad( new_logical_to_physical_map, (0, logical_to_physical_map.shape[-1] - max_physical_slots), value=-1, ) physical_to_logical_map = new_physical_to_logical_map.to(self.device) logical_to_physical_map.copy_(new_logical_to_physical_map) logical_replica_count.copy_(new_logical_replica_count) model.set_eplb_state( expert_load_pass, logical_to_physical_map, logical_replica_count, ) if global_expert_load is not None: rearrange_expert_weights_inplace( old_global_expert_indices, new_physical_to_logical_map, model.expert_weights, ep_group, False, rank_mapping, ) self.expert_rearrangement_step = 0 self.model_states[model_config.compute_hash()] = EplbModelState( physical_to_logical_map, logical_to_physical_map, logical_replica_count, expert_load_pass, expert_load_window, model_config.model, model, ) def step( self, is_dummy: bool = False, is_profile: bool = False, log_stats: bool = False, ) -> None: """ Step the EPLB state. Args: is_dummy (bool): If `True`, this is a dummy step and the load metrics recorded in this forward pass will not count. Defaults to `False`. is_profile (bool): If `True`, perform a dummy rearrangement with maximum communication cost. This is used in `profile_run` to reserve enough memory for the communication buffer. log_stats (bool): If `True`, log the expert load metrics. # Stats The metrics are all summed up across layers. - `avg_tokens`: The average load across ranks. - `max_tokens`: The maximum load across ranks. - `balancedness`: The ratio of average load to maximum load. """ if is_profile: self.rearrange(is_profile=True) return if is_dummy: # Do not record load metrics for dummy steps for eplb_model_state in self.model_states.values(): eplb_model_state.expert_load_pass.zero_() if log_stats: # Sync the expert load pass for each model (main and drafter). # expert_load_pass: (num_moe_layers, num_physical_experts) expert_load_pass_list = self._sync_load_pass() ep_group = get_ep_group().device_group for expert_load_pass, eplb_model_state in zip( expert_load_pass_list, self.model_states.values() ): # num_tokens_per_rank: (num_moe_layers, num_ranks) num_tokens_per_rank = ( expert_load_pass.reshape( expert_load_pass.shape[0], ep_group.size(), -1 ) .sum(dim=-1) .float() ) # Compute balancedness ratio: # for each layer: # (mean load across ranks) / (max load across ranks) avg_tokens_tensor = num_tokens_per_rank.mean(dim=0).sum(dim=0) max_tokens_tensor = num_tokens_per_rank.max(dim=0).values.sum(dim=0) # Just to make type checker happy tokens_tensors: list[float] = torch.stack( [avg_tokens_tensor, max_tokens_tensor] ).tolist() avg_tokens, max_tokens = tokens_tensors balancedness = avg_tokens / max_tokens if max_tokens > 0 else 0.0 if ep_group.rank() == 0: logger.info( "EPLB step: %d for model %s: avg_tokens=%.2f, " "max_tokens=%d, balancedness=%.4f", self.expert_rearrangement_step, eplb_model_state.model_name, avg_tokens, max_tokens, balancedness, ) # Update the expert load sliding window if not is_dummy: for eplb_model_state in self.model_states.values(): eplb_model_state.expert_load_window[self.expert_load_window_step] = ( eplb_model_state.expert_load_pass.clone() ) eplb_model_state.expert_load_pass.zero_() self.expert_load_window_step += 1 if self.expert_load_window_step >= self.expert_load_window_size: self.expert_load_window_step = 0 # Step the expert rearrangement step # Note that even if this is a dummy step, we still increment the # rearrangement step and perform rearrangement to ensure all ranks are # performing collective communication. self.expert_rearrangement_step += 1 if self.expert_rearrangement_step >= self.expert_rearrangement_step_interval: self.expert_rearrangement_step = 0 self.rearrange() def rearrange( self, is_profile: bool = False, execute_shuffle: bool = True, global_expert_loads: list[torch.Tensor] | None = None, rank_mapping: dict[int, int] | None = None, ) -> torch.Tensor | None: """ Rearrange the experts according to the current load. Args: is_profile (bool): If `True`, perform a dummy rearrangement. This is used in `profile_run` to reserve enough memory, no memory movement will be performed. Default is False. execute_shuffle (bool): If `True`, execute the shuffle in elastic expert parallel (EEP). Default is True. global_expert_loads (list[torch.Tensor] | None): The global expert loads when scaling is done in EEP. List of expert loads for the main and drafter (when spec decode is used) models. rank_mapping (dict[int, int] | None): The rank mapping when scaling is done in EEP. """ ep_group = get_ep_group().device_group ep_rank = ep_group.rank() time_start = None is_main_rank = ep_rank == 0 if is_main_rank: torch.cuda.synchronize() time_start = time.perf_counter() logger.info("Rearranging experts %s...", "(profile)" if is_profile else "") if global_expert_loads is None: # Map the physical expert load to global logical experts global_expert_load_windows = [] if not execute_shuffle: num_models = torch.tensor( [len(self.model_states)], dtype=torch.int32, device="cpu" ) torch.distributed.broadcast( num_models, group=get_ep_group().cpu_group, group_src=0 ) for eplb_model_state in self.model_states.values(): logical_expert_load_window = torch.zeros( self.expert_load_window_size, eplb_model_state.model.num_moe_layers, eplb_model_state.model.num_logical_experts, dtype=eplb_model_state.expert_load_window.dtype, device=eplb_model_state.expert_load_window.device, ) logical_expert_load_window.scatter_add_( dim=-1, index=eplb_model_state.physical_to_logical_map.unsqueeze(0) .expand_as(eplb_model_state.expert_load_window) .long(), src=eplb_model_state.expert_load_window, ) if not execute_shuffle: metadata = torch.tensor( [ eplb_model_state.model.num_moe_layers, eplb_model_state.model.num_logical_experts, eplb_model_state.physical_to_logical_map.shape[1], ], dtype=torch.int32, device="cpu", ) torch.distributed.broadcast( metadata, group=get_ep_group().cpu_group, group_src=0 ) global_expert_load_window = logical_expert_load_window.sum(dim=0) global_expert_load_windows.append(global_expert_load_window) # Perform all-reduce to get the expert load across all ranks for each model global_expert_load_windows = self._allreduce_list( global_expert_load_windows ) if not execute_shuffle: for eplb_model_state, global_expert_load_window in zip( self.model_states.values(), global_expert_load_windows ): # (num_moe_layers, old_num_physical_experts) old_global_expert_indices = eplb_model_state.physical_to_logical_map torch.distributed.broadcast( old_global_expert_indices, group=ep_group, group_src=0 ) if not execute_shuffle: return global_expert_load_windows else: assert execute_shuffle global_expert_load_windows = global_expert_loads # TODO(bowen): Treat differently for prefill and decode nodes eplb_model_state = next(iter(self.model_states.values())) model = eplb_model_state.model num_replicas = model.num_physical_experts num_groups = model.num_expert_groups if rank_mapping is not None and len(rank_mapping) == ep_group.size(): # NOTE(yongji): scale down, we need to rebalance the experts on # remaining GPUs, transfer the experts while we haven't shutdown # the GPUs to be released. cpu_group = get_ep_group().cpu_group num_nodes = _node_count_with_rank_mapping(cpu_group, rank_mapping) num_gpus = sum(new_rank != -1 for new_rank in rank_mapping.values()) num_replicas = ( num_replicas // ep_group.size() * num_gpus ) # handle num replicas change else: num_nodes = get_node_count() num_gpus = ep_group.size() if num_gpus % num_nodes != 0: self.num_nodes = 1 logger.warning_once( f"num_gpus % num_nodes != 0, " "not using hierarchical rearrangement algorithm.\n" f"{num_gpus=}, {num_nodes=}" ) for eplb_model_state, global_expert_load_window in zip( self.model_states.values(), global_expert_load_windows ): # Get new expert mappings for the model ( new_physical_to_logical_map, new_logical_to_physical_map, new_logical_replica_count, ) = rebalance_experts( global_expert_load_window, num_replicas, num_groups, num_nodes, num_gpus, ) # Update expert weights rearrange_expert_weights_inplace( eplb_model_state.physical_to_logical_map, new_physical_to_logical_map, eplb_model_state.model.expert_weights, ep_group, is_profile, rank_mapping, ) if not is_profile: if ( eplb_model_state.physical_to_logical_map.shape[1] != new_physical_to_logical_map.shape[1] ): eplb_model_state.physical_to_logical_map = ( new_physical_to_logical_map.to( eplb_model_state.physical_to_logical_map.device ) ) else: eplb_model_state.physical_to_logical_map.copy_( new_physical_to_logical_map ) max_physical_slots = new_logical_to_physical_map.shape[-1] assert ( max_physical_slots <= eplb_model_state.logical_to_physical_map.shape[-1] ) new_logical_to_physical_map = torch.nn.functional.pad( new_logical_to_physical_map, ( 0, eplb_model_state.logical_to_physical_map.shape[-1] - max_physical_slots, ), value=-1, ) eplb_model_state.logical_to_physical_map.copy_( new_logical_to_physical_map ) eplb_model_state.logical_replica_count.copy_(new_logical_replica_count) if is_main_rank: assert time_start is not None torch.cuda.synchronize() time_end = time.perf_counter() logger.info( "Rearranged experts%sin %.2f seconds.", " (profile) " if is_profile else " ", time_end - time_start, ) return None @staticmethod def recv_state() -> tuple[list[torch.Tensor], list[torch.Tensor]]: """ Receive the expert load and old placement from the master rank. """ ep_group = get_ep_group() num_models = torch.empty(1, dtype=torch.int32, device="cpu") torch.distributed.broadcast(num_models, group=ep_group.cpu_group, group_src=0) num_models = num_models.item() global_expert_loads = [] old_global_expert_indices_per_model = [] for _ in range(num_models): metadata = torch.empty(3, dtype=torch.int32, device="cpu") torch.distributed.broadcast(metadata, group=ep_group.cpu_group, group_src=0) num_moe_layers, num_logical_experts, num_old_physical_experts = ( metadata.tolist() ) global_expert_load = torch.zeros( (num_moe_layers, num_logical_experts), dtype=torch.int64, device=ep_group.device, ) all_reduce(global_expert_load, group=ep_group.device_group) old_global_expert_indices = torch.empty( (num_moe_layers, num_old_physical_experts), dtype=torch.int64, device=ep_group.device, ) torch.distributed.broadcast( old_global_expert_indices, group=ep_group.device_group, group_src=0, ) global_expert_loads.append(global_expert_load) old_global_expert_indices_per_model.append(old_global_expert_indices) return global_expert_loads, old_global_expert_indices_per_model @classmethod def get_eep_state( cls, parallel_config: ParallelConfig ) -> tuple[ list[torch.Tensor] | None, list[torch.Tensor] | None, dict[int, int] | None, ]: num_local_physical_experts = torch.empty(1, dtype=torch.int32, device="cpu") torch.distributed.broadcast( num_local_physical_experts, group=get_ep_group().cpu_group, group_src=0, ) num_local_physical_experts = int(num_local_physical_experts.item()) new_ep_size = get_ep_group().world_size global_expert_loads, old_global_expert_indices_per_model = ( EplbState.recv_state() ) # EP configuration for all models has to be the same so as eplb config num_logical_experts = global_expert_loads[0].shape[1] parallel_config.eplb_config.num_redundant_experts = ( num_local_physical_experts * new_ep_size - num_logical_experts ) assert ( old_global_expert_indices_per_model[0].shape[1] % num_local_physical_experts == 0 ) old_ep_size = ( old_global_expert_indices_per_model[0].shape[1] // num_local_physical_experts ) rank_mapping = {old_ep_rank: old_ep_rank for old_ep_rank in range(old_ep_size)} return ( global_expert_loads, old_global_expert_indices_per_model, rank_mapping, ) def _allreduce_list(self, tensor_list: list[torch.Tensor]) -> list[torch.Tensor]: """ All-reduce a list of tensors. """ if len(tensor_list) == 1: all_reduce(tensor_list[0], group=get_ep_group().device_group) return tensor_list assert all(t.dim() == 2 for t in tensor_list), "All tensors must be 2D." assert all(t.shape[1] == tensor_list[0].shape[1] for t in tensor_list), ( "All tensors must have the same shape[1]." ) # Concatenate, all_reduce, then unpack to original shapes. # We assume all tensors are 2D and shape[1] (num_physical_experts) # is the same across all models. shapes = [t.shape for t in tensor_list] concat_tensor = torch.cat(tensor_list, dim=0) ep_group = get_ep_group().device_group all_reduce(concat_tensor, group=ep_group) all_reduce_list = [] offset = 0 for shape in shapes: all_reduce_list.append(concat_tensor[offset : offset + shape[0], :]) offset += shape[0] return all_reduce_list def _sync_load_pass(self) -> list[torch.Tensor]: """ Sync the expert load pass across all ranks for log stats. Doesn't update the expert load pass in eplb_model_state. """ load_pass_list = [] for eplb_model_state in self.model_states.values(): load_pass_list.append(eplb_model_state.expert_load_pass.clone()) return self._allreduce_list(load_pass_list) def _node_count_with_rank_mapping( pg: ProcessGroup | StatelessProcessGroup, rank_mapping: dict[int, int], ) -> int: if isinstance(pg, ProcessGroup): world_size = torch.distributed.get_world_size(group=pg) else: world_size = pg.world_size if world_size == 1: return 1 # Build node assignment map node_assignment = [0] * world_size # rank -> node_id next_node_id = 0 for current_rank in range(world_size): if node_assignment[current_rank] != 0: continue # Already assigned to a node assert current_rank in rank_mapping if rank_mapping[current_rank] == -1: continue # Pending shutdown # Assign current rank to a new node next_node_id += 1 node_assignment[current_rank] = next_node_id # Find all ranks on the same node as current_rank same_node_flags = in_the_same_node_as(pg, current_rank) for other_rank, is_same_node in enumerate(same_node_flags): if is_same_node and node_assignment[other_rank] == 0: node_assignment[other_rank] = next_node_id return next_node_id