Ubuntu commited on
Commit
acb7132
1 Parent(s): f871962
.gitignore ADDED
@@ -0,0 +1 @@
 
 
1
+ checkpoint-*/
latest ADDED
@@ -0,0 +1 @@
 
 
1
+ global_step10308
pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a03a40bb73fea5f8ae700fe7ecf5cdfbe50eda65faf7973e9a98b9e7bf099873
3
+ size 13476956611
rng_state_0.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c888e6ff3a7cd8d3cf49d7b58f132af388c0174b6f6fe01890af418823bf9d44
3
+ size 17563
rng_state_1.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c888e6ff3a7cd8d3cf49d7b58f132af388c0174b6f6fe01890af418823bf9d44
3
+ size 17563
rng_state_2.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c888e6ff3a7cd8d3cf49d7b58f132af388c0174b6f6fe01890af418823bf9d44
3
+ size 17563
rng_state_3.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c888e6ff3a7cd8d3cf49d7b58f132af388c0174b6f6fe01890af418823bf9d44
3
+ size 17563
trainer_state.json CHANGED
@@ -30930,15 +30930,6 @@
30930
  "learning_rate": 4.4648783368850255e-10,
30931
  "loss": 0.072,
30932
  "step": 10308
30933
- },
30934
- {
30935
- "epoch": 3.0,
30936
- "step": 10308,
30937
- "total_flos": 218874711244800.0,
30938
- "train_loss": 0.0,
30939
- "train_runtime": 2.854,
30940
- "train_samples_per_second": 57786.49,
30941
- "train_steps_per_second": 3611.787
30942
  }
30943
  ],
30944
  "max_steps": 10308,
 
30930
  "learning_rate": 4.4648783368850255e-10,
30931
  "loss": 0.072,
30932
  "step": 10308
 
 
 
 
 
 
 
 
 
30933
  }
30934
  ],
30935
  "max_steps": 10308,
training_args.bin CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:9a109cb16366a014a2ebc4ea2ee43005adbd9306c4f57f86032765dffdd46b18
3
  size 4975
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b004cd5b8169b77d514ca278e1ac0ec9f94a5d2a71c89f7a87adcb1f042645e2
3
  size 4975
zero_to_fp32.py ADDED
@@ -0,0 +1,584 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python
2
+
3
+ # Copyright (c) Microsoft Corporation.
4
+ # SPDX-License-Identifier: Apache-2.0
5
+
6
+ # DeepSpeed Team
7
+
8
+ # This script extracts fp32 consolidated weights from a zero 2 and 3 DeepSpeed checkpoints. It gets
9
+ # copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
10
+ # the future. Once extracted, the weights don't require DeepSpeed and can be used in any
11
+ # application.
12
+ #
13
+ # example: python zero_to_fp32.py . pytorch_model.bin
14
+
15
+ import argparse
16
+ import torch
17
+ import glob
18
+ import math
19
+ import os
20
+ import re
21
+ from collections import OrderedDict
22
+ from dataclasses import dataclass
23
+
24
+ # while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
25
+ # DeepSpeed data structures it has to be available in the current python environment.
26
+ from deepspeed.utils import logger
27
+ from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
28
+ FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
29
+ FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
30
+
31
+
32
+ @dataclass
33
+ class zero_model_state:
34
+ buffers: dict()
35
+ param_shapes: dict()
36
+ shared_params: list
37
+ ds_version: int
38
+ frozen_param_shapes: dict()
39
+ frozen_param_fragments: dict()
40
+
41
+
42
+ debug = 0
43
+
44
+ # load to cpu
45
+ device = torch.device('cpu')
46
+
47
+
48
+ def atoi(text):
49
+ return int(text) if text.isdigit() else text
50
+
51
+
52
+ def natural_keys(text):
53
+ '''
54
+ alist.sort(key=natural_keys) sorts in human order
55
+ http://nedbatchelder.com/blog/200712/human_sorting.html
56
+ (See Toothy's implementation in the comments)
57
+ '''
58
+ return [atoi(c) for c in re.split(r'(\d+)', text)]
59
+
60
+
61
+ def get_model_state_file(checkpoint_dir, zero_stage):
62
+ if not os.path.isdir(checkpoint_dir):
63
+ raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
64
+
65
+ # there should be only one file
66
+ if zero_stage == 2:
67
+ file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
68
+ elif zero_stage == 3:
69
+ file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
70
+
71
+ if not os.path.exists(file):
72
+ raise FileNotFoundError(f"can't find model states file at '{file}'")
73
+
74
+ return file
75
+
76
+
77
+ def get_checkpoint_files(checkpoint_dir, glob_pattern):
78
+ # XXX: need to test that this simple glob rule works for multi-node setup too
79
+ ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
80
+
81
+ if len(ckpt_files) == 0:
82
+ raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
83
+
84
+ return ckpt_files
85
+
86
+
87
+ def get_optim_files(checkpoint_dir):
88
+ return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
89
+
90
+
91
+ def get_model_state_files(checkpoint_dir):
92
+ return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
93
+
94
+
95
+ def parse_model_states(files):
96
+ zero_model_states = []
97
+ for file in files:
98
+ state_dict = torch.load(file, map_location=device)
99
+
100
+ if BUFFER_NAMES not in state_dict:
101
+ raise ValueError(f"{file} is not a model state checkpoint")
102
+ buffer_names = state_dict[BUFFER_NAMES]
103
+ if debug:
104
+ print("Found buffers:", buffer_names)
105
+
106
+ # recover just the buffers while restoring them to fp32 if they were saved in fp16
107
+ buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
108
+ param_shapes = state_dict[PARAM_SHAPES]
109
+
110
+ # collect parameters that are included in param_shapes
111
+ param_names = []
112
+ for s in param_shapes:
113
+ for name in s.keys():
114
+ param_names.append(name)
115
+
116
+ # update with frozen parameters
117
+ frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
118
+ if frozen_param_shapes is not None:
119
+ if debug:
120
+ print(f"Found frozen_param_shapes: {frozen_param_shapes}")
121
+ param_names += list(frozen_param_shapes.keys())
122
+
123
+ # record shared parameters so that they can be recovered based on partners
124
+ # this is because such parameters holding reference only are not saved by optimizer
125
+ shared_params = []
126
+ for param in state_dict["module"]:
127
+ if param not in [*param_names, *buffer_names]:
128
+ for share_param in state_dict["module"]:
129
+ if (state_dict["module"][share_param].data_ptr() == state_dict["module"][param].data_ptr()
130
+ and share_param != param):
131
+ shared_params.append([param, share_param])
132
+ break
133
+
134
+ ds_version = state_dict.get(DS_VERSION, None)
135
+
136
+ frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
137
+
138
+ z_model_state = zero_model_state(buffers=buffers,
139
+ param_shapes=param_shapes,
140
+ shared_params=shared_params,
141
+ ds_version=ds_version,
142
+ frozen_param_shapes=frozen_param_shapes,
143
+ frozen_param_fragments=frozen_param_fragments)
144
+ zero_model_states.append(z_model_state)
145
+
146
+ return zero_model_states
147
+
148
+
149
+ def parse_optim_states(files, ds_checkpoint_dir):
150
+
151
+ total_files = len(files)
152
+ state_dicts = []
153
+ for f in files:
154
+ state_dicts.append(torch.load(f, map_location=device))
155
+
156
+ if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
157
+ raise ValueError(f"{files[0]} is not a zero checkpoint")
158
+ zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
159
+ world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
160
+
161
+ # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
162
+ # parameters can be different from data parallelism for non-expert parameters. So we can just
163
+ # use the max of the partition_count to get the dp world_size.
164
+
165
+ if type(world_size) is list:
166
+ world_size = max(world_size)
167
+
168
+ if world_size != total_files:
169
+ raise ValueError(
170
+ f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
171
+ "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
172
+ )
173
+
174
+ # the groups are named differently in each stage
175
+ if zero_stage == 2:
176
+ fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
177
+ elif zero_stage == 3:
178
+ fp32_groups_key = FP32_FLAT_GROUPS
179
+ else:
180
+ raise ValueError(f"unknown zero stage {zero_stage}")
181
+
182
+ if zero_stage == 2:
183
+ fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
184
+ elif zero_stage == 3:
185
+ # if there is more than one param group, there will be multiple flattened tensors - one
186
+ # flattened tensor per group - for simplicity merge them into a single tensor
187
+ #
188
+ # XXX: could make the script more memory efficient for when there are multiple groups - it
189
+ # will require matching the sub-lists of param_shapes for each param group flattened tensor
190
+
191
+ fp32_flat_groups = [
192
+ torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
193
+ ]
194
+
195
+ return zero_stage, world_size, fp32_flat_groups
196
+
197
+
198
+ def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
199
+ """
200
+ Returns fp32 state_dict reconstructed from ds checkpoint
201
+
202
+ Args:
203
+ - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
204
+
205
+ """
206
+ print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
207
+
208
+ optim_files = get_optim_files(ds_checkpoint_dir)
209
+ zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
210
+ print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
211
+
212
+ model_files = get_model_state_files(ds_checkpoint_dir)
213
+
214
+ zero_model_states = parse_model_states(model_files)
215
+ print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
216
+
217
+ if zero_stage == 2:
218
+ return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states)
219
+ elif zero_stage == 3:
220
+ return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
221
+
222
+
223
+ def _zero2_merge_frozen_params(state_dict, zero_model_states):
224
+ if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
225
+ return
226
+
227
+ frozen_param_shapes = zero_model_states[0].frozen_param_shapes
228
+ frozen_param_fragments = zero_model_states[0].frozen_param_fragments
229
+
230
+ if debug:
231
+ num_elem = sum(s.numel() for s in frozen_param_shapes.values())
232
+ print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
233
+
234
+ wanted_params = len(frozen_param_shapes)
235
+ wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
236
+ avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
237
+ print(f'Frozen params: Have {avail_numel} numels to process.')
238
+ print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
239
+
240
+ total_params = 0
241
+ total_numel = 0
242
+ for name, shape in frozen_param_shapes.items():
243
+ total_params += 1
244
+ unpartitioned_numel = shape.numel()
245
+ total_numel += unpartitioned_numel
246
+
247
+ state_dict[name] = frozen_param_fragments[name]
248
+
249
+ if debug:
250
+ print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
251
+
252
+ print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
253
+
254
+
255
+ def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
256
+ param_shapes = zero_model_states[0].param_shapes
257
+
258
+ # Reconstruction protocol:
259
+ #
260
+ # XXX: document this
261
+
262
+ if debug:
263
+ for i in range(world_size):
264
+ for j in range(len(fp32_flat_groups[0])):
265
+ print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
266
+
267
+ # XXX: memory usage doubles here (zero2)
268
+ num_param_groups = len(fp32_flat_groups[0])
269
+ merged_single_partition_of_fp32_groups = []
270
+ for i in range(num_param_groups):
271
+ merged_partitions = [sd[i] for sd in fp32_flat_groups]
272
+ full_single_fp32_vector = torch.cat(merged_partitions, 0)
273
+ merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
274
+ avail_numel = sum(
275
+ [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
276
+
277
+ if debug:
278
+ wanted_params = sum([len(shapes) for shapes in param_shapes])
279
+ wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
280
+ # not asserting if there is a mismatch due to possible padding
281
+ print(f"Have {avail_numel} numels to process.")
282
+ print(f"Need {wanted_numel} numels in {wanted_params} params.")
283
+
284
+ # params
285
+ # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
286
+ # out-of-core computing solution
287
+ total_numel = 0
288
+ total_params = 0
289
+ for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
290
+ offset = 0
291
+ avail_numel = full_single_fp32_vector.numel()
292
+ for name, shape in shapes.items():
293
+
294
+ unpartitioned_numel = shape.numel()
295
+ total_numel += unpartitioned_numel
296
+ total_params += 1
297
+
298
+ if debug:
299
+ print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
300
+ state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
301
+ offset += unpartitioned_numel
302
+
303
+ # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
304
+ # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
305
+ # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
306
+ # live optimizer object, so we are checking that the numbers are within the right range
307
+ align_to = 2 * world_size
308
+
309
+ def zero2_align(x):
310
+ return align_to * math.ceil(x / align_to)
311
+
312
+ if debug:
313
+ print(f"original offset={offset}, avail_numel={avail_numel}")
314
+
315
+ offset = zero2_align(offset)
316
+ avail_numel = zero2_align(avail_numel)
317
+
318
+ if debug:
319
+ print(f"aligned offset={offset}, avail_numel={avail_numel}")
320
+
321
+ # Sanity check
322
+ if offset != avail_numel:
323
+ raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
324
+
325
+ print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
326
+
327
+
328
+ def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):
329
+ state_dict = OrderedDict()
330
+
331
+ # buffers
332
+ buffers = zero_model_states[0].buffers
333
+ state_dict.update(buffers)
334
+ if debug:
335
+ print(f"added {len(buffers)} buffers")
336
+
337
+ _zero2_merge_frozen_params(state_dict, zero_model_states)
338
+
339
+ _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
340
+
341
+ # recover shared parameters
342
+ for pair in zero_model_states[0].shared_params:
343
+ state_dict[pair[0]] = state_dict[pair[1]]
344
+
345
+ return state_dict
346
+
347
+
348
+ def zero3_partitioned_param_info(unpartitioned_numel, world_size):
349
+ remainder = unpartitioned_numel % world_size
350
+ padding_numel = (world_size - remainder) if remainder else 0
351
+ partitioned_numel = math.ceil(unpartitioned_numel / world_size)
352
+ return partitioned_numel, padding_numel
353
+
354
+
355
+ def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
356
+ if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
357
+ return
358
+
359
+ if debug:
360
+ for i in range(world_size):
361
+ num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
362
+ print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
363
+
364
+ frozen_param_shapes = zero_model_states[0].frozen_param_shapes
365
+ wanted_params = len(frozen_param_shapes)
366
+ wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
367
+ avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
368
+ print(f'Frozen params: Have {avail_numel} numels to process.')
369
+ print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
370
+
371
+ total_params = 0
372
+ total_numel = 0
373
+ for name, shape in zero_model_states[0].frozen_param_shapes.items():
374
+ total_params += 1
375
+ unpartitioned_numel = shape.numel()
376
+ total_numel += unpartitioned_numel
377
+
378
+ param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
379
+ state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
380
+
381
+ partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
382
+
383
+ if debug:
384
+ print(
385
+ f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
386
+ )
387
+
388
+ print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
389
+
390
+
391
+ def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
392
+ param_shapes = zero_model_states[0].param_shapes
393
+ avail_numel = fp32_flat_groups[0].numel() * world_size
394
+ # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
395
+ # param, re-consolidating each param, while dealing with padding if any
396
+
397
+ # merge list of dicts, preserving order
398
+ param_shapes = {k: v for d in param_shapes for k, v in d.items()}
399
+
400
+ if debug:
401
+ for i in range(world_size):
402
+ print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
403
+
404
+ wanted_params = len(param_shapes)
405
+ wanted_numel = sum(shape.numel() for shape in param_shapes.values())
406
+ # not asserting if there is a mismatch due to possible padding
407
+ avail_numel = fp32_flat_groups[0].numel() * world_size
408
+ print(f"Trainable params: Have {avail_numel} numels to process.")
409
+ print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
410
+
411
+ # params
412
+ # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
413
+ # out-of-core computing solution
414
+ offset = 0
415
+ total_numel = 0
416
+ total_params = 0
417
+ for name, shape in param_shapes.items():
418
+
419
+ unpartitioned_numel = shape.numel()
420
+ total_numel += unpartitioned_numel
421
+ total_params += 1
422
+
423
+ partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
424
+
425
+ if debug:
426
+ print(
427
+ f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
428
+ )
429
+
430
+ # XXX: memory usage doubles here
431
+ state_dict[name] = torch.cat(
432
+ tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
433
+ 0).narrow(0, 0, unpartitioned_numel).view(shape)
434
+ offset += partitioned_numel
435
+
436
+ offset *= world_size
437
+
438
+ # Sanity check
439
+ if offset != avail_numel:
440
+ raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
441
+
442
+ print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
443
+
444
+
445
+ def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):
446
+ state_dict = OrderedDict()
447
+
448
+ # buffers
449
+ buffers = zero_model_states[0].buffers
450
+ state_dict.update(buffers)
451
+ if debug:
452
+ print(f"added {len(buffers)} buffers")
453
+
454
+ _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
455
+
456
+ _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
457
+
458
+ # recover shared parameters
459
+ for pair in zero_model_states[0].shared_params:
460
+ state_dict[pair[0]] = state_dict[pair[1]]
461
+
462
+ return state_dict
463
+
464
+
465
+ def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
466
+ """
467
+ Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
468
+ ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
469
+ via a model hub.
470
+
471
+ Args:
472
+ - ``checkpoint_dir``: path to the desired checkpoint folder
473
+ - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
474
+
475
+ Returns:
476
+ - pytorch ``state_dict``
477
+
478
+ Note: this approach may not work if your application doesn't have sufficient free CPU memory and
479
+ you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
480
+ the checkpoint.
481
+
482
+ A typical usage might be ::
483
+
484
+ from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
485
+ # do the training and checkpoint saving
486
+ state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
487
+ model = model.cpu() # move to cpu
488
+ model.load_state_dict(state_dict)
489
+ # submit to model hub or save the model to share with others
490
+
491
+ In this example the ``model`` will no longer be usable in the deepspeed context of the same
492
+ application. i.e. you will need to re-initialize the deepspeed engine, since
493
+ ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
494
+
495
+ If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
496
+
497
+ """
498
+ if tag is None:
499
+ latest_path = os.path.join(checkpoint_dir, 'latest')
500
+ if os.path.isfile(latest_path):
501
+ with open(latest_path, 'r') as fd:
502
+ tag = fd.read().strip()
503
+ else:
504
+ raise ValueError(f"Unable to find 'latest' file at {latest_path}")
505
+
506
+ ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
507
+
508
+ if not os.path.isdir(ds_checkpoint_dir):
509
+ raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
510
+
511
+ return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
512
+
513
+
514
+ def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
515
+ """
516
+ Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
517
+ loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
518
+
519
+ Args:
520
+ - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
521
+ - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
522
+ - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
523
+ """
524
+
525
+ state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
526
+ print(f"Saving fp32 state dict to {output_file}")
527
+ torch.save(state_dict, output_file)
528
+
529
+
530
+ def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
531
+ """
532
+ 1. Put the provided model to cpu
533
+ 2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
534
+ 3. Load it into the provided model
535
+
536
+ Args:
537
+ - ``model``: the model object to update
538
+ - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
539
+ - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
540
+
541
+ Returns:
542
+ - ``model`: modified model
543
+
544
+ Make sure you have plenty of CPU memory available before you call this function. If you don't
545
+ have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
546
+ conveniently placed for you in the checkpoint folder.
547
+
548
+ A typical usage might be ::
549
+
550
+ from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
551
+ model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
552
+ # submit to model hub or save the model to share with others
553
+
554
+ Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
555
+ of the same application. i.e. you will need to re-initialize the deepspeed engine, since
556
+ ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
557
+
558
+ """
559
+ logger.info(f"Extracting fp32 weights")
560
+ state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
561
+
562
+ logger.info(f"Overwriting model with fp32 weights")
563
+ model = model.cpu()
564
+ model.load_state_dict(state_dict, strict=False)
565
+
566
+ return model
567
+
568
+
569
+ if __name__ == "__main__":
570
+
571
+ parser = argparse.ArgumentParser()
572
+ parser.add_argument("checkpoint_dir",
573
+ type=str,
574
+ help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
575
+ parser.add_argument(
576
+ "output_file",
577
+ type=str,
578
+ help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
579
+ parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
580
+ args = parser.parse_args()
581
+
582
+ debug = args.debug
583
+
584
+ convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file)