File size: 9,260 Bytes
2ea2835 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 | from datetime import timedelta
from functools import partial
import inspect
import os
import torch
import torch.distributed as dist
from torch.distributed.fsdp import BackwardPrefetch, FullStateDictConfig, FullOptimStateDictConfig, 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 (unsharded) optimizer state on rank 0 (CPU).
Collective: must be called by every rank. Returns the full optim state
dict on rank 0 and an empty/partial dict on other ranks (rank0_only=True).
"""
save_policy = FullStateDictConfig(offload_to_cpu=True, rank0_only=True)
optim_policy = FullOptimStateDictConfig(offload_to_cpu=True, rank0_only=True)
with FSDP.state_dict_type(
model,
StateDictType.FULL_STATE_DICT,
state_dict_config=save_policy,
optim_state_dict_config=optim_policy,
):
osd = FSDP.optim_state_dict(model, optimizer)
return osd
def fsdp_load_optim_state_dict(model, optimizer, full_osd):
"""Load a full optimizer state dict (present on rank 0) back into ``optimizer``.
Collective: every rank must call this. ``full_osd`` need only be non-None on
rank 0 (FSDP broadcasts/shards it to the other ranks), which keeps host memory
bounded to a single copy of the optimizer state.
"""
save_policy = FullStateDictConfig(offload_to_cpu=True, rank0_only=True)
optim_policy = FullOptimStateDictConfig(offload_to_cpu=True, rank0_only=True)
with FSDP.state_dict_type(
model,
StateDictType.FULL_STATE_DICT,
state_dict_config=save_policy,
optim_state_dict_config=optim_policy,
):
sharded = FSDP.optim_state_dict_to_load(
model=model, optim=optimizer, optim_state_dict=full_osd
)
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, backward_prefetch="pre"):
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":
if transformer_module is None:
raise ValueError("transformer wrap strategy requires transformer_module")
transformer_classes = (
{transformer_module}
if isinstance(transformer_module, type)
else set(transformer_module)
)
auto_wrap_policy = partial(
transformer_auto_wrap_policy,
transformer_layer_cls=transformer_classes
)
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]
backward_prefetch_modes = {
"pre": BackwardPrefetch.BACKWARD_PRE,
"post": BackwardPrefetch.BACKWARD_POST,
}
if backward_prefetch not in backward_prefetch_modes:
raise ValueError(
"FSDP backward_prefetch must be 'pre' or 'post', got "
f"{backward_prefetch!r}")
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),
backward_prefetch=backward_prefetch_modes[backward_prefetch],
# Every rank constructs and loads the same pretrained/checkpoint state
# before wrapping; no rank-0 broadcast is implied here.
sync_module_states=False
)
return module
def move_adamw_moment_state_(optimizer, target_device):
"""Move only AdamW moments without changing optimizer values."""
if not isinstance(optimizer, torch.optim.AdamW):
raise TypeError("optimizer moment offload requires torch.optim.AdamW")
target = torch.device(target_device)
movable_keys = {"exp_avg", "exp_avg_sq", "max_exp_avg_sq"}
tensor_count = 0
total_bytes = 0
initialized_parameters = 0
for state in optimizer.state.values():
if not state:
continue
initialized_parameters += 1
unknown = set(state) - movable_keys - {"step"}
if unknown:
raise RuntimeError(
"AdamW state contains unsupported entries for exact offload: "
f"{sorted(unknown)}")
if "exp_avg" not in state or "exp_avg_sq" not in state:
raise RuntimeError(
"initialized AdamW state is missing moment buffers")
step = state.get("step")
if not torch.is_tensor(step) or step.device.type != "cpu":
raise RuntimeError(
"non-capturable AdamW step state must remain on CPU")
for key in movable_keys:
value = state.get(key)
if value is None:
continue
if not torch.is_tensor(value):
raise RuntimeError(f"AdamW {key} state is not a tensor")
if value.device != target:
value = value.to(device=target, non_blocking=False)
state[key] = value
if value.device != target:
raise RuntimeError(
f"AdamW {key} state did not move to {target}")
tensor_count += 1
total_bytes += value.numel() * value.element_size()
return {
"target_device": str(target),
"initialized_parameters": initialized_parameters,
"tensor_count": tensor_count,
"bytes": total_bytes,
}
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"])
# Bind CUDA before NCCL creates communicators.
torch.cuda.set_device(local_rank)
if ":" in host: # IPv6
init_method = f"tcp://[{host}]:{port}"
else: # IPv4
init_method = f"tcp://{host}:{port}"
init_kwargs = {}
if (
backend == "nccl"
and "device_id" in inspect.signature(
dist.init_process_group).parameters
):
init_kwargs["device_id"] = torch.device("cuda", local_rank)
dist.init_process_group(
rank=rank,
world_size=world_size,
backend=backend,
init_method=init_method,
timeout=timedelta(minutes=30),
**init_kwargs,
)
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
with FSDP.summon_full_params(fsdp_module, writeback=False):
for n, p in fsdp_module.module.named_parameters():
self.shadow[n] = p.detach().clone().float().cpu()
@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():
self.shadow[n].mul_(d).add_(p.detach().float().cpu(), 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))
|