# Copyright (c) Microsoft Corporation. # SPDX-License-Identifier: Apache-2.0 # DeepSpeed Team from collections import OrderedDict import torch import sys from torch._utils import _flatten_dense_tensors, _unflatten_dense_tensors from deepspeed import comm as dist from deepspeed.runtime.constants import PIPE_REPLICATED from deepspeed.runtime.base_optimizer import ZeROOptimizer from packaging import version as pkg_version from deepspeed.git_version_info import version from deepspeed.runtime.utils import (get_global_norm_of_tensors, clip_tensors_by_global_norm, DummyOptim, align_dense_tensors, all_gather_dp_groups, is_model_parallel_parameter, see_memory_usage, graph_process, get_norm_with_moe_layers) from deepspeed.utils import link_hp_params, lazy_init_hp_params_optimizer_state, fragment_address, groups from deepspeed.moe.utils import is_moe_param, is_moe_param_group from deepspeed.utils.bwc import bwc_tensor_model_parallel_rank from deepspeed.checkpoint import enable_universal_checkpoint from deepspeed.checkpoint.constants import (DS_VERSION, PARTITION_COUNT, BASE_OPTIMIZER_STATE, SINGLE_PARTITION_OF_FP32_GROUPS, CLIP_GRAD, GROUP_PADDINGS, PARAM_SLICE_MAPPINGS) setattr(sys.modules[__name__], 'fragment_address', fragment_address) class BF16_Optimizer(ZeROOptimizer): def __init__(self, init_optimizer, param_names, mpu=None, clip_grad=0.0, norm_type=2, allgather_bucket_size=5000000000, dp_process_group=None, timers=None, grad_acc_dtype=None, graph_harvesting=False, immediate_grad_update=False, has_moe_layers=False): super().__init__() see_memory_usage('begin bf16_optimizer', force=True) self.timers = timers self.optimizer = init_optimizer self.param_names = param_names self.using_real_optimizer = not isinstance(self.optimizer, DummyOptim) assert grad_acc_dtype in [torch.float32, torch.bfloat16 ], f"BF16Optimizer: Unsupported gradient accumulation data type: {grad_acc_dtype}" self.grad_acc_dtype = grad_acc_dtype self.immediate_grad_update = immediate_grad_update self.clip_grad = clip_grad self.norm_type = norm_type self.mpu = mpu self.allgather_bucket_size = int(allgather_bucket_size) self.dp_process_group = dp_process_group self.dp_rank = dist.get_rank(group=self.dp_process_group) self.has_moe_layers = has_moe_layers self.non_expert_gradients = [] self.real_dp_process_group = [dp_process_group for i in range(len(self.optimizer.param_groups))] if self.has_moe_layers: self._configure_moe_settings() # Use torch (un)flatten ops self.flatten = _flatten_dense_tensors self.unflatten = _unflatten_dense_tensors #align nccl all-gather send buffers to 4-bye boundary self.nccl_start_alignment_factor = 2 # 4-byte alignment/sizeof(fp16) = 2 # Build BF16/FP32 groups self.bf16_groups = [] self.bf16_groups_flat = [] self.bf16_partitioned_groups = [] self.fp32_groups_flat_partition = [] # Maintain different fp32 gradients views for convenience self.fp32_groups_gradients = [] self.fp32_groups_gradient_dict = {} self.fp32_groups_gradients_flat = [] self.fp32_groups_actual_gradients_flat = [] self.fp32_groups_gradient_flat_partition = [] self.fp32_groups_has_gradients = [] self.group_paddings = [] self.graph_harvesting = graph_harvesting if self.using_real_optimizer: self._setup_for_real_optimizer() see_memory_usage('end bf16_optimizer', force=True) def _configure_moe_settings(self): assert any( [is_moe_param_group(group) for group in self.optimizer.param_groups] ), "The model has moe layers, but None of the param groups are marked as MoE. Create a param group with 'moe' key set to True before creating optimizer" for i, group in enumerate(self.optimizer.param_groups): if is_moe_param_group(group): assert all([is_moe_param(param) for param in group['params']]), "All params in MoE group must be MoE params" self.real_dp_process_group[i] = groups._get_expert_data_parallel_group(group['name']) self.expert_gradients = {} if self.has_moe_layers: for key in groups._get_expert_data_parallel_group_dict().keys(): self.expert_gradients[key] = [] def _setup_for_real_optimizer(self): self.partition_count = [dist.get_world_size(group=pg) for pg in self.real_dp_process_group] for i, param_group in enumerate(self.optimizer.param_groups): real_dp_world_size = dist.get_world_size(group=self.real_dp_process_group[i]) see_memory_usage(f'before initializing group {i}', force=True) partition_id = dist.get_rank(group=self.real_dp_process_group[i]) # grab the original list trainable_parameters = [param for param in param_group['params'] if param.requires_grad] self.bf16_groups.append(trainable_parameters) # create flat bf16 params self.bf16_groups_flat.append( self._flatten_dense_tensors_aligned(self.bf16_groups[i], self.nccl_start_alignment_factor * real_dp_world_size)) # Make bf16 params point to flat tensor storage self._update_storage_to_flattened_tensor(tensor_list=self.bf16_groups[i], flat_tensor=self.bf16_groups_flat[i]) # divide flat weights into equal sized partitions partition_size = self.bf16_groups_flat[i].numel() // real_dp_world_size bf16_dp_partitions = [ self.bf16_groups_flat[i].narrow(0, dp_index * partition_size, partition_size) for dp_index in range(real_dp_world_size) ] self.bf16_partitioned_groups.append(bf16_dp_partitions) # create fp32 params partition self.fp32_groups_flat_partition.append(bf16_dp_partitions[partition_id].clone().float().detach()) self.fp32_groups_flat_partition[i].requires_grad = True num_elem_list = [t.numel() for t in self.bf16_groups[i]] # create fp32 gradients fp32_flat_buffer = torch.zeros_like(self.bf16_groups_flat[i], dtype=self.grad_acc_dtype) self.fp32_groups_gradients_flat.append(fp32_flat_buffer) if self.has_moe_layers and is_moe_param_group(param_group): self.expert_gradients[param_group['name']].append(fp32_flat_buffer) else: self.non_expert_gradients.append(fp32_flat_buffer) # track individual fp32 gradients for entire model fp32_gradients = self._split_flat_tensor(flat_tensor=self.fp32_groups_gradients_flat[i], num_elem_list=num_elem_list) self.fp32_groups_gradients.append(fp32_gradients) self.fp32_groups_gradient_dict[i] = fp32_gradients # flat tensor corresponding to actual fp32 gradients (i.e., minus alignment padding) length_without_padding = sum(num_elem_list) self.fp32_groups_actual_gradients_flat.append( torch.narrow(self.fp32_groups_gradients_flat[i], 0, 0, length_without_padding)) # flat tensor corresponding to gradient partition self.fp32_groups_gradient_flat_partition.append( torch.narrow(self.fp32_groups_gradients_flat[i], 0, partition_id * partition_size, partition_size)) # track fp32 gradient updates self.fp32_groups_has_gradients.append([False] * len(self.bf16_groups[i])) # Record padding required for alignment if partition_id == dist.get_world_size(group=self.real_dp_process_group[i]) - 1: padding = self.bf16_groups_flat[i].numel() - length_without_padding else: padding = 0 self.group_paddings.append(padding) # update optimizer param groups to reference fp32 params partition param_group['params'] = [self.fp32_groups_flat_partition[i]] see_memory_usage(f'after initializing group {i}', force=True) see_memory_usage('before initialize_optimizer', force=True) self.initialize_optimizer_states() see_memory_usage('end initialize_optimizer', force=True) if self.immediate_grad_update: self.create_grad_acc_hooks() # Need optimizer states initialized before linking lp to optimizer state self._link_all_hp_params() self._hp_optimizer_states_linked = False self._enable_universal_checkpoint() self._param_slice_mappings = self._create_param_mapping() def _enable_universal_checkpoint(self): for lp_param_group in self.bf16_groups: enable_universal_checkpoint(param_list=lp_param_group) def _create_param_mapping(self): param_mapping = [] for i, _ in enumerate(self.optimizer.param_groups): param_mapping_per_group = OrderedDict() for lp in self.bf16_groups[i]: if lp._hp_mapping is not None: lp_name = self.param_names[lp] param_mapping_per_group[lp_name] = lp._hp_mapping.get_hp_fragment_address() param_mapping.append(param_mapping_per_group) return param_mapping def _link_all_hp_params(self): for i, _ in enumerate(self.optimizer.param_groups): real_dp_world_size = dist.get_world_size(group=self.real_dp_process_group[i]) # Link bf16 and fp32 params in partition partition_id = dist.get_rank(group=self.real_dp_process_group[i]) partition_size = self.bf16_groups_flat[i].numel() // real_dp_world_size flat_hp_partition = self.fp32_groups_flat_partition[i] link_hp_params(lp_param_list=self.bf16_groups[i], flat_hp_partition=flat_hp_partition, gradient_dict=self.fp32_groups_gradient_dict, offload_gradient_dict=None, use_offload=False, param_group_index=i, partition_start=partition_id * partition_size, partition_size=partition_size, dp_group=self.real_dp_process_group[i]) def _lazy_init_hp_params_optimizer_state(self): if not self._hp_optimizer_states_linked: for i, _ in enumerate(self.optimizer.param_groups): lazy_init_hp_params_optimizer_state(self.bf16_groups[i], self.fp32_groups_flat_partition[i], self.optimizer.state) self._hp_optimizer_states_linked = True def initialize_optimizer_states(self): """Take an optimizer step with zero-valued gradients to allocate internal optimizer state. This helps prevent memory fragmentation by allocating optimizer state at the beginning of training instead of after activations have been allocated. """ for param_partition, grad_partition in zip(self.fp32_groups_flat_partition, self.fp32_groups_gradient_flat_partition): # In case of grad acc dtype different than FP32, need to cast to high precision. param_partition.grad = grad_partition.to( param_partition.dtype) if grad_partition.dtype != param_partition.dtype else grad_partition if self.grad_acc_dtype is not torch.float32: for param_partition in self.fp32_groups_flat_partition: param_partition.grad = None self.clear_hp_grads() def _split_flat_tensor(self, flat_tensor, num_elem_list): assert sum(num_elem_list) <= flat_tensor.numel() tensor_list = [] offset = 0 for num_elem in num_elem_list: dense_tensor = torch.narrow(flat_tensor, 0, offset, num_elem) tensor_list.append(dense_tensor) offset += num_elem return tensor_list def _update_storage_to_flattened_tensor(self, tensor_list, flat_tensor): updated_params = self.unflatten(flat_tensor, tensor_list) for p, q in zip(tensor_list, updated_params): p.data = q.data def _flatten_dense_tensors_aligned(self, tensor_list, alignment): return self.flatten(align_dense_tensors(tensor_list, alignment)) @torch.no_grad() def step(self, closure=None): if closure is not None: raise NotImplementedError(f'{self.__class__} does not support closure.') non_expert_grads_for_norm, expert_grads_for_norm = self.get_grads_for_norm() non_expert_groups_norm = get_global_norm_of_tensors(input_tensors=non_expert_grads_for_norm, mpu=self.mpu, norm_type=self.norm_type, use_graph=self.graph_harvesting) all_groups_norm = non_expert_groups_norm if self.has_moe_layers: all_groups_norm = get_norm_with_moe_layers(non_expert_groups_norm, mpu=self.mpu, expert_tensors=expert_grads_for_norm, norm_type=self.norm_type) self._global_grad_norm = all_groups_norm assert all_groups_norm > 0. if self.clip_grad > 0.: clip_tensors_by_global_norm(input_tensors=self.get_grads_for_norm(for_clipping=True), max_norm=self.clip_grad, global_norm=all_groups_norm, mpu=self.mpu, use_graph=self.graph_harvesting) self.optimizer.step() # We need to link optimizer state after the first step() call self._lazy_init_hp_params_optimizer_state() self.update_lp_params() self.clear_hp_grads() def backward(self, loss, update_hp_grads=True, clear_lp_grads=False, **bwd_kwargs): """Perform a backward pass and copy the low-precision gradients to the high-precision copy. We copy/accumulate to the high-precision grads now to prevent accumulating in the bf16 grads after successive backward() calls (i.e., grad accumulation steps > 1) The low-precision grads are deallocated during this procedure. """ self.clear_lp_grads() loss.backward(**bwd_kwargs) if update_hp_grads: self.update_hp_grads(clear_lp_grads=clear_lp_grads) @torch.no_grad() def _update_hp_grad(self, lp, group_idx, param_idx, clear_lp_grads): if lp.grad is None: return hp_grad = self.fp32_groups_gradients[group_idx][param_idx] assert hp_grad is not None, \ f'high precision param has no gradient, lp param_id = {id(lp)} group_info = [{group_idx}][{param_idx}]' hp_grad.data.add_(lp.grad.data.to(hp_grad.dtype).view(hp_grad.shape)) lp._hp_grad = hp_grad self.fp32_groups_has_gradients[group_idx][param_idx] = True # clear gradients if clear_lp_grads: lp.grad.zero_() @torch.no_grad() def _update_hp_grads_func(self, clear_lp_grads=False): for i, group in enumerate(self.bf16_groups): for j, lp in enumerate(group): self._update_hp_grad(lp, i, j, clear_lp_grads) @torch.no_grad() def update_hp_grads(self, clear_lp_grads=False): if self.immediate_grad_update: return if self.graph_harvesting: graph_process(False, self._update_hp_grads_func, clear_lp_grads) else: self._update_hp_grads_func(clear_lp_grads) #cpu op for i, group in enumerate(self.bf16_groups): for j, lp in enumerate(group): if lp.grad is None: continue self.fp32_groups_has_gradients[i][j] = True @torch.no_grad() def get_grads_for_reduction(self): if self.has_moe_layers: return self.non_expert_gradients, self.expert_gradients return self.non_expert_gradients, {} @torch.no_grad() def get_grads_for_norm(self, for_clipping=False): """ Returns: tuple[list[Tensor], dict[ep_name, List[Tensor]] | list: If for_clipping, return all gradients. Otherwise, separate and return dict of expert_grad and list of non_expert_grad """ # (grads, expert_group_name) expert_grads_for_norm = {} # grads non_expert_grads_for_norm = [] all_grads_for_clip = [] tensor_mp_rank = bwc_tensor_model_parallel_rank(mpu=self.mpu) assert len(self.bf16_groups) == len(self.optimizer.param_groups) for i, group in enumerate(self.bf16_groups): for j, lp in enumerate(group): if not for_clipping: if hasattr(lp, PIPE_REPLICATED) and lp.ds_pipe_replicated: continue # skip duplicated parameters. perform norm only on cards with tp_rank=0. # non-duplicated parameters include: # - Parameters with tp: Use allreducesum of mp_group. # - Moe Parameters with ep: Use allreducesum of ep_group. if not (tensor_mp_rank == 0 or is_model_parallel_parameter(lp) or is_moe_param(lp)): continue if not self.fp32_groups_has_gradients[i][j]: continue if not for_clipping: param_group = self.optimizer.param_groups[i] if self.has_moe_layers and is_moe_param_group(param_group): if param_group['name'] not in expert_grads_for_norm: expert_grads_for_norm[param_group['name']] = [] expert_grads_for_norm[param_group['name']].append(self.fp32_groups_gradients[i][j]) else: non_expert_grads_for_norm.append(self.fp32_groups_gradients[i][j]) else: all_grads_for_clip.append(self.fp32_groups_gradients[i][j]) if not for_clipping: return non_expert_grads_for_norm, expert_grads_for_norm return all_grads_for_clip @torch.no_grad() def update_lp_params(self): for i, (bf16_partitions, fp32_partition) in enumerate(zip(self.bf16_partitioned_groups, self.fp32_groups_flat_partition)): partition_id = dist.get_rank(group=self.real_dp_process_group[i]) bf16_partitions[partition_id].data.copy_(fp32_partition.data) # print_rank_0(f'update_lp_params {i=} {partition_id=}', force=True) # if i == 0: # print_rank_0(f'{fp32_partition[:10]=}', force=True) all_gather_dp_groups(groups_flat=self.bf16_groups_flat, partitioned_param_groups=self.bf16_partitioned_groups, dp_process_group=self.real_dp_process_group, start_alignment_factor=self.nccl_start_alignment_factor, allgather_bucket_size=self.allgather_bucket_size) def clear_hp_grads(self): for flat_gradients in self.fp32_groups_gradients_flat: flat_gradients.zero_() for i, group in enumerate(self.fp32_groups_gradients): self.fp32_groups_has_gradients[i] = [False] * len(group) def clear_lp_grads(self): # using zero_() fixed memory address for graph replay set_to_none = False if self.graph_harvesting else True zero_grads_list = [] for group in self.bf16_groups: for param in group: if set_to_none: param.grad = None elif param.grad is not None: if param.grad.grad_fn is not None: param.grad.detach_() zero_grads_list.append(param.grad) if not set_to_none and len(zero_grads_list) > 0: torch._foreach_zero_(zero_grads_list) def state_dict(self): state_dict = {} state_dict[CLIP_GRAD] = self.clip_grad state_dict[BASE_OPTIMIZER_STATE] = self.optimizer.state_dict() state_dict[SINGLE_PARTITION_OF_FP32_GROUPS] = self.fp32_groups_flat_partition state_dict[GROUP_PADDINGS] = self.group_paddings state_dict[PARTITION_COUNT] = self.partition_count state_dict[DS_VERSION] = version state_dict[PARAM_SLICE_MAPPINGS] = self._param_slice_mappings return state_dict # Restore base optimizer fp32 weights bfloat16 weights def _restore_from_bit16_weights(self): for i, group in enumerate(self.bf16_groups): partition_id = dist.get_rank(group=self.real_dp_process_group[i]) for bf16_partitions, fp32_partition in zip(self.bf16_partitioned_groups, self.fp32_groups_flat_partition): fp32_partition.data.copy_(bf16_partitions[partition_id].data) def refresh_fp32_params(self): self._restore_from_bit16_weights() def load_state_dict(self, state_dict_list, checkpoint_folder, load_optimizer_states=True, load_from_fp32_weights=False, load_serial=None): if checkpoint_folder: self._load_universal_checkpoint(checkpoint_folder, load_optimizer_states, load_from_fp32_weights) else: self._load_legacy_checkpoint(state_dict_list, load_optimizer_states, load_from_fp32_weights) def _load_legacy_checkpoint(self, state_dict_list, load_optimizer_states=True, load_from_fp32_weights=False): dp_rank = dist.get_rank(group=self.dp_process_group) current_rank_sd = state_dict_list[dp_rank] ckpt_version = current_rank_sd.get(DS_VERSION, False) assert ckpt_version, f"Empty ds_version in checkpoint, not clear how to proceed" ckpt_version = pkg_version.parse(ckpt_version) self.clip_grad = current_rank_sd.get(CLIP_GRAD, self.clip_grad) if load_optimizer_states: print(f"_load_legacy_checkpoint current_rank_sd[BASE_OPTIMIZER_STATE]") self.optimizer.load_state_dict(current_rank_sd[BASE_OPTIMIZER_STATE]) if load_from_fp32_weights: for current, saved in zip(self.fp32_groups_flat_partition, current_rank_sd[SINGLE_PARTITION_OF_FP32_GROUPS]): src_tensor = _get_padded_tensor(saved, current.numel()) current.data.copy_(src_tensor.data) if load_optimizer_states: self._link_all_hp_params() def _load_universal_checkpoint(self, checkpoint_folder, load_optimizer_states, load_from_fp32_weights): self.load_hp_checkpoint_state_from_checkpoint_dir("bf16_groups", checkpoint_folder) def _load_global_state(self, sd): pass @property def param_groups(self): """Forward the wrapped optimizer's parameters.""" return self.optimizer.param_groups def accumulate_hp_grads_and_remove_lp(self, lp_param, group_idx, param_idx): assert self.immediate_grad_update self._update_hp_grad(lp_param, group_idx, param_idx, clear_lp_grads=True) def create_grad_acc_hooks(self): self.grad_accs = [] for i, param_group in enumerate(self.bf16_groups): for j, param in enumerate(param_group): if param.requires_grad: def wrapper(param, i, j): param_tmp = param.expand_as(param) grad_acc = param_tmp.grad_fn.next_functions[0][0] def accumulate_hp_grads_and_remove_lp(*notneeded): self.accumulate_hp_grads_and_remove_lp(param, i, j) grad_acc.register_hook(accumulate_hp_grads_and_remove_lp) self.grad_accs.append(grad_acc) wrapper(param, i, j) def _get_padded_tensor(src_tensor, size): if src_tensor.numel() >= size: return src_tensor padded_tensor = torch.zeros(size, dtype=src_tensor.dtype, device=src_tensor.device) slice_tensor = torch.narrow(padded_tensor, 0, 0, src_tensor.numel()) slice_tensor.data.copy_(src_tensor.data) return padded_tensor