File size: 6,111 Bytes
59630ba | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 | from datetime import timedelta
from functools import partial
import os
import torch
import torch.distributed as dist
from torch.distributed.fsdp import FullOptimStateDictConfig, FullStateDictConfig, FullyShardedDataParallel as FSDP, MixedPrecision, ShardingStrategy, StateDictType
from torch.distributed.fsdp.api import CPUOffload
from torch.distributed.fsdp.wrap import size_based_auto_wrap_policy, transformer_auto_wrap_policy
def fsdp_state_dict(model):
fsdp_fullstate_save_policy = FullStateDictConfig(
offload_to_cpu=True, rank0_only=True
)
with FSDP.state_dict_type(
model, StateDictType.FULL_STATE_DICT, fsdp_fullstate_save_policy
):
checkpoint = model.state_dict()
return checkpoint
def fsdp_optim_state_dict(model, optimizer):
"""Gather the full optimizer state on rank 0 (CPU). Other ranks return an empty dict."""
state_dict_config = FullStateDictConfig(offload_to_cpu=True, rank0_only=True)
optim_state_dict_config = FullOptimStateDictConfig(offload_to_cpu=True, rank0_only=True)
with FSDP.state_dict_type(
model,
StateDictType.FULL_STATE_DICT,
state_dict_config=state_dict_config,
optim_state_dict_config=optim_state_dict_config,
):
return FSDP.optim_state_dict(model, optimizer)
def load_fsdp_optim_state_dict(model, optimizer, full_optim_state_dict):
"""Scatter a full optimizer state (valid on rank 0) into the local FSDP shards."""
with FSDP.state_dict_type(model, StateDictType.FULL_STATE_DICT):
sharded = FSDP.optim_state_dict_to_load(
model=model,
optim=optimizer,
optim_state_dict=full_optim_state_dict,
)
optimizer.load_state_dict(sharded)
def fsdp_wrap(module, sharding_strategy="full", mixed_precision=False, wrap_strategy="size", min_num_params=int(5e7), transformer_module=None, cpu_offload=False):
if mixed_precision:
mixed_precision_policy = MixedPrecision(
param_dtype=torch.bfloat16,
reduce_dtype=torch.float32,
buffer_dtype=torch.float32,
cast_forward_inputs=False
)
else:
mixed_precision_policy = None
if wrap_strategy == "transformer":
auto_wrap_policy = partial(
transformer_auto_wrap_policy,
transformer_layer_cls=transformer_module
)
elif wrap_strategy == "size":
auto_wrap_policy = partial(
size_based_auto_wrap_policy,
min_num_params=min_num_params
)
else:
raise ValueError(f"Invalid wrap strategy: {wrap_strategy}")
os.environ["NCCL_CROSS_NIC"] = "1"
sharding_strategy = {
"full": ShardingStrategy.FULL_SHARD,
"hybrid_full": ShardingStrategy.HYBRID_SHARD,
"hybrid_zero2": ShardingStrategy._HYBRID_SHARD_ZERO2,
"no_shard": ShardingStrategy.NO_SHARD,
}[sharding_strategy]
module = FSDP(
module,
auto_wrap_policy=auto_wrap_policy,
sharding_strategy=sharding_strategy,
mixed_precision=mixed_precision_policy,
device_id=torch.cuda.current_device(),
limit_all_gathers=True,
use_orig_params=True,
cpu_offload=CPUOffload(offload_params=cpu_offload),
sync_module_states=False # Load ckpt on rank 0 and sync to other ranks
)
return module
def barrier():
if dist.is_initialized():
dist.barrier()
def launch_distributed_job(backend: str = "nccl"):
rank = int(os.environ["RANK"])
local_rank = int(os.environ["LOCAL_RANK"])
world_size = int(os.environ["WORLD_SIZE"])
host = os.environ["MASTER_ADDR"]
port = int(os.environ["MASTER_PORT"])
if ":" in host: # IPv6
init_method = f"tcp://[{host}]:{port}"
else: # IPv4
init_method = f"tcp://{host}:{port}"
dist.init_process_group(rank=rank, world_size=world_size, backend=backend,
init_method=init_method, timeout=timedelta(minutes=30))
torch.cuda.set_device(local_rank)
class EMA_FSDP:
def __init__(self, fsdp_module: torch.nn.Module, decay: float = 0.999):
self.decay = decay
self.shadow = {}
self._init_shadow(fsdp_module)
@torch.no_grad()
def _init_shadow(self, fsdp_module):
from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
# Anchor shadow to the local CUDA device. summon_full_params can return
# CPU views of params under some FSDP/resume code paths, so don't
# implicitly inherit p's device.
target_device = torch.device("cuda", torch.cuda.current_device())
with FSDP.summon_full_params(fsdp_module, writeback=False):
for n, p in fsdp_module.module.named_parameters():
self.shadow[n] = p.detach().to(target_device).float().clone()
@torch.no_grad()
def update(self, fsdp_module):
d = self.decay
from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
with FSDP.summon_full_params(fsdp_module, writeback=False):
for n, p in fsdp_module.module.named_parameters():
shadow = self.shadow[n]
# Always align the source tensor to shadow's device to avoid
# cuda/cpu mismatches when FSDP exposes params on CPU.
shadow.mul_(d).add_(
p.detach().to(shadow.device).float(), alpha=1. - d
)
# Optional helpers ---------------------------------------------------
def state_dict(self):
return self.shadow # picklable
def load_state_dict(self, sd):
self.shadow = {k: v.clone() for k, v in sd.items()}
def copy_to(self, fsdp_module):
# load EMA weights into an (unwrapped) copy of the generator
from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
with FSDP.summon_full_params(fsdp_module, writeback=True):
for n, p in fsdp_module.module.named_parameters():
if n in self.shadow:
p.data.copy_(self.shadow[n].to(p.dtype, device=p.device))
|