File size: 11,168 Bytes
ee3e701 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 |
#!/usr/bin/env python
# -*- encoding: utf-8 -*-
from typing import List
from torch import Tensor
from torch._utils import _flatten_dense_tensors, _unflatten_dense_tensors
from internlm.core.context import ParallelMode
from internlm.core.context import global_context as gpc
class BaseStore:
"""
Base Store
"""
def __init__(self, dp_parallel_mode=ParallelMode.DATA):
self._world_size = gpc.get_world_size(dp_parallel_mode)
self._local_rank = gpc.get_local_rank(dp_parallel_mode)
@property
def world_size(self):
return self._world_size
@property
def local_rank(self):
return self._local_rank
class BucketStore(BaseStore):
"""
Bucket Store
"""
def __init__(self, dp_parallel_mode):
super().__init__(dp_parallel_mode)
self._grads = dict()
self._params = dict()
self._num_elements_in_bucket = dict()
self.reset()
def num_elements_in_bucket(self, reduce_rank: int = None):
return self._num_elements_in_bucket[reduce_rank]
def add_num_elements_in_bucket(self, num_elements, reduce_rank: int = None):
self._num_elements_in_bucket[reduce_rank] += num_elements
def add_grad(self, tensor, reduce_rank: int = None):
self._grads[reduce_rank].append(tensor)
def add_param(self, tensor, reduce_rank: int = None):
self._params[reduce_rank].append(tensor)
def reset(self):
keys = [None] + list(range(self._world_size))
self._grads = {rank: [] for rank in keys}
self._params = {rank: [] for rank in keys}
self._num_elements_in_bucket = {rank: 0 for rank in keys}
def reset_by_rank(self, reduce_rank=None):
self._grads[reduce_rank] = []
self._params[reduce_rank] = []
self._num_elements_in_bucket[reduce_rank] = 0
def get_grad(self, reduce_rank: int = None):
return self._grads[reduce_rank]
def get_param(self, reduce_rank: int = None):
return self._params[reduce_rank]
class GradientStore(BaseStore):
"""
Gradient Store
"""
def __init__(self, *args):
super().__init__(*args)
# bookkeeping data structures
self._averaged_gradients = dict()
# for backward reduction hooks
self._grad_acc_objs = []
def add_accumulate_grad_object(self, obj):
"""
Keep :class:`AccumulateGrad` objects. If these objects are not kept, reduction hooks may not
be attached successfully.
:param obj: An object of :class:`AccumulateGrad` class
:type obj: :class:`AccumulateGrad`
"""
self._grad_acc_objs.append(obj)
def get_averaged_gradients_by_group(self, group_id: int) -> List[Tensor]:
"""
Return average gradients of a parameter group
:param group_id: The index of parameter group
:type group_id: int
:return: Return the list of averaged gradients of a parameter group. Each element is a gradient,
not a parameter.
:rtype: List[torch.Tensor]
"""
return self._averaged_gradients[group_id]
def add_average_gradient_by_group(self, group_id: int, tensor: Tensor) -> None:
"""
Append an average gradient to the list of averaged gradients of a parameter group
:param group_id: The index of a parameter group
:param tensor: A :class:`torch.Tensor` object
:type group_id: int
:type tensor: torch.Tensor
"""
if group_id in self._averaged_gradients:
self._averaged_gradients[group_id].append(tensor)
else:
self._averaged_gradients[group_id] = [tensor]
def reset_average_gradients_by_group(self, group_id: int) -> None:
"""
Reset the bookkeeping data structure for averaged gradients to an empty list
:param group_id: The index of a parameter group
:type group_id: int
"""
self._averaged_gradients[group_id] = []
class ParameterStore(BaseStore):
"""
Parameter Store
"""
def __init__(self, dp_paralle_mode):
super().__init__(dp_paralle_mode)
# param partitioning data structures
self._fp16_param_to_rank = dict()
self._rank_groupid_to_fp16_param_list = dict()
self._rank_group_id_to_flat_fp16_param = dict()
# param reduction data structures
self._is_param_reduced = dict()
self._reduced_param = []
self._former_bucket_reduced_param = {}
self._last_bucket_reduced_param = {}
self._former_bucket_reduced_grad = {}
self._last_bucket_reduced_grad = {}
def set_param_to_rank(self, tensor: Tensor, rank: int) -> None:
"""
Set the mapping between parameter to rank, each parameter should be owned by a rank.
:param tensor: A :class:`torch.Tensor` object
:type tensor: torch.Tensor
:param rank: The rank of which the process is responsible for updating the parameter
:type rank: int
"""
self._fp16_param_to_rank[tensor] = rank
def get_param_rank(self, tensor: Tensor) -> int:
"""
Gives the rank which the parameter belongs to
:param tensor: A :class:`torch.Tensor` object
:type tensor: torch.Tensor
"""
return self._fp16_param_to_rank[tensor]
def belongs_to_current_rank(self, tensor) -> bool:
"""
Check whether a parameter is supposed to be updated by the process of the current rank
:param tensor: A :class:`torch.Tensor` object
:type tensor: torch.Tensor
:return: True if the parameter should be updated by the current rank. Otherwise false.
:rtype: bool
"""
tensor_rank = self._fp16_param_to_rank[tensor]
return tensor_rank == self._local_rank
def add_fp16_param_list_by_rank_group(self, rank, group_id, tensor_list) -> None:
if rank not in self._rank_groupid_to_fp16_param_list:
self._rank_groupid_to_fp16_param_list[rank] = dict()
if group_id not in self._rank_groupid_to_fp16_param_list[rank]:
self._rank_groupid_to_fp16_param_list[rank][group_id] = []
self._rank_groupid_to_fp16_param_list[rank][group_id].extend(tensor_list)
def get_fp16_params_by_rank_group(self, rank, group_id) -> List[Tensor]:
return self._rank_groupid_to_fp16_param_list[rank][group_id]
def add_flat_fp16_param_by_rank_group(self, rank, group_id, tensor) -> None:
if rank not in self._rank_group_id_to_flat_fp16_param:
self._rank_group_id_to_flat_fp16_param[rank] = dict()
self._rank_group_id_to_flat_fp16_param[rank][group_id] = tensor
def get_flat_fp16_param_by_rank_group(self, rank, group_id) -> Tensor:
return self._rank_group_id_to_flat_fp16_param[rank][group_id]
def is_param_reduced(self, tensor):
return self._is_param_reduced[tensor]
def set_param_reduction_state(self, tensor, state):
self._is_param_reduced[tensor] = state
def get_param_reduction_states(self):
return self._is_param_reduced
def reset_previous_reduced_params(self):
self._reduced_param = []
def add_previous_reduced_param(self, tensor):
self._reduced_param.append(tensor)
def add_reduced_param_for_compute_norm(self, param, last_bucket=False):
group_id = getattr(param, "group_id")
if last_bucket:
if group_id not in self._last_bucket_reduced_param:
self._last_bucket_reduced_param[group_id] = []
self._last_bucket_reduced_grad[group_id] = []
self._last_bucket_reduced_param[group_id].append(param)
self._last_bucket_reduced_grad[group_id].append(param.grad)
else:
if group_id not in self._former_bucket_reduced_param:
self._former_bucket_reduced_param[group_id] = []
self._former_bucket_reduced_grad[group_id] = []
self._former_bucket_reduced_param[group_id].append(param)
self._former_bucket_reduced_grad[group_id].append(param.grad)
def get_reduced_param_for_compute_norm(self, group_id=0, last_bucket=False):
if not last_bucket:
if group_id not in self._former_bucket_reduced_param:
return [], []
return (
self._former_bucket_reduced_param[group_id],
self._former_bucket_reduced_grad[group_id],
)
else:
if group_id not in self._last_bucket_reduced_param:
return [], []
return (
self._last_bucket_reduced_param[group_id],
self._last_bucket_reduced_grad[group_id],
)
def reset_reduced_data_for_compute_norm(self):
self._former_bucket_reduced_param = {}
self._last_bucket_reduced_param = {}
self._former_bucket_reduced_grad = {}
self._last_bucket_reduced_grad = {}
def clear_grads_of_previous_reduced_params(self):
if len(self._reduced_param) > 0:
for param in self._reduced_param:
param.grad = None
self.reset_previous_reduced_params()
class TensorBucket:
"""
Tensor Bucket
"""
def __init__(self, size):
self._max_size = size
self._current_size = 0
self._bucket = []
self._flat_tensor = None
self._unflatten_and_copy_flag = False
self.commu_handle = None
@property
def max_size(self):
return self._max_size
@property
def current_size(self):
return self._current_size
def is_full_or_oversized(self):
return self._current_size >= self._max_size
def is_empty(self):
return len(self._bucket) == 0
def set_unflatten_and_copy_flag(self, flag):
self._unflatten_and_copy_flag = flag
def get_unflatten_and_copy_flag(self):
return self._unflatten_and_copy_flag
def get_flat_tensor(self):
return self._flat_tensor
def add_to_bucket(self, tensor, allow_oversize=False):
tensor_size = tensor.numel()
if not allow_oversize and self.will_exceed_max_size(tensor_size):
msg = f"The param bucket max size {self._max_size} is exceeded" + f"by tensor (size {tensor_size})"
raise RuntimeError(msg)
self._bucket.append(tensor)
self._current_size += tensor_size
def will_exceed_max_size(self, tensor_size):
expected_size = self._current_size + tensor_size
return expected_size > self._max_size
def get_bucket(self):
return self._bucket
def empty(self):
self._bucket = []
self._size = 0
self._flat_tensor = None
self.commu_handle = None
def flatten(self):
self._flat_tensor = _flatten_dense_tensors(self._bucket)
def unflatten_and_copy(self):
if self._unflatten_and_copy_flag:
unflattened_tensor_list = _unflatten_dense_tensors(self._flat_tensor, self._bucket)
for old, new in zip(self._bucket, unflattened_tensor_list):
old.copy_(new)
|