Bohdan Ivashchenko commited on
Commit
063933c
1 Parent(s): 8e347c0

upload weights

Browse files
config.json ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "EleutherAI/gpt-j-6B",
3
+ "activation_function": "gelu_new",
4
+ "architectures": [
5
+ "GPTJForCausalLM"
6
+ ],
7
+ "attn_pdrop": 0.0,
8
+ "bos_token_id": 50256,
9
+ "embd_pdrop": 0.0,
10
+ "eos_token_id": 50256,
11
+ "initializer_range": 0.02,
12
+ "layer_norm_epsilon": 1e-05,
13
+ "model_type": "gptj",
14
+ "n_embd": 4096,
15
+ "n_head": 16,
16
+ "n_inner": null,
17
+ "n_layer": 28,
18
+ "n_positions": 2048,
19
+ "resid_pdrop": 0.0,
20
+ "rotary": true,
21
+ "rotary_dim": 64,
22
+ "scale_attn_weights": true,
23
+ "summary_activation": null,
24
+ "summary_first_dropout": 0.1,
25
+ "summary_proj_to_labels": true,
26
+ "summary_type": "cls_index",
27
+ "summary_use_proj": true,
28
+ "task_specific_params": {
29
+ "text-generation": {
30
+ "do_sample": true,
31
+ "max_length": 50,
32
+ "temperature": 1.0
33
+ }
34
+ },
35
+ "tie_word_embeddings": false,
36
+ "tokenizer_class": "GPT2Tokenizer",
37
+ "torch_dtype": "float16",
38
+ "transformers_version": "4.28.1",
39
+ "use_cache": false,
40
+ "vocab_size": 50257
41
+ }
generation_config.json ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "bos_token_id": 50256,
4
+ "eos_token_id": 50256,
5
+ "transformers_version": "4.28.1",
6
+ "use_cache": false
7
+ }
latest ADDED
@@ -0,0 +1 @@
 
 
1
+ global_step680
merges.txt ADDED
The diff for this file is too large to render. See raw diff
 
pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b81fe246fe829ef1903ad6d268793f05ac9bf3fd03cef393ca5a0eb22cc89e52
3
+ size 12216972905
rng_state_0.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b0b47786a021fed7e4aacc8d852ba10a39c20411f6f2cf6d8ffc637f14955a74
3
+ size 14583
rng_state_1.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a9ebaeb0d5485091d948a19f87ccdc8f7dc709b1c2f9d7386c9ee19f298e132d
3
+ size 14583
rng_state_2.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:766323763276f8db8b322269cb86e0e257714803820af7a0f4297c2b0b635806
3
+ size 14583
rng_state_3.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bbddaf920d7cbb89da457891cc70e8694977c94948058fb63294c63f16f5b5ba
3
+ size 14583
special_tokens_map.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token": {
3
+ "content": "<|endoftext|>",
4
+ "lstrip": false,
5
+ "normalized": true,
6
+ "rstrip": false,
7
+ "single_word": false
8
+ },
9
+ "eos_token": {
10
+ "content": "<|endoftext|>",
11
+ "lstrip": false,
12
+ "normalized": true,
13
+ "rstrip": false,
14
+ "single_word": false
15
+ },
16
+ "pad_token": "<|endoftext|>",
17
+ "unk_token": {
18
+ "content": "<|endoftext|>",
19
+ "lstrip": false,
20
+ "normalized": true,
21
+ "rstrip": false,
22
+ "single_word": false
23
+ }
24
+ }
tokenizer_config.json ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_bos_token": false,
3
+ "add_prefix_space": false,
4
+ "bos_token": {
5
+ "__type": "AddedToken",
6
+ "content": "<|endoftext|>",
7
+ "lstrip": false,
8
+ "normalized": true,
9
+ "rstrip": false,
10
+ "single_word": false
11
+ },
12
+ "clean_up_tokenization_spaces": true,
13
+ "eos_token": {
14
+ "__type": "AddedToken",
15
+ "content": "<|endoftext|>",
16
+ "lstrip": false,
17
+ "normalized": true,
18
+ "rstrip": false,
19
+ "single_word": false
20
+ },
21
+ "errors": "replace",
22
+ "model_max_length": 2048,
23
+ "pad_token": null,
24
+ "tokenizer_class": "GPT2Tokenizer",
25
+ "unk_token": {
26
+ "__type": "AddedToken",
27
+ "content": "<|endoftext|>",
28
+ "lstrip": false,
29
+ "normalized": true,
30
+ "rstrip": false,
31
+ "single_word": false
32
+ }
33
+ }
trainer_state.json ADDED
@@ -0,0 +1,294 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "best_metric": 0.057281494140625,
3
+ "best_model_checkpoint": "/opt/ml/checkpoints/checkpoint-680",
4
+ "epoch": 0.26143790849673204,
5
+ "global_step": 680,
6
+ "is_hyper_param_search": false,
7
+ "is_local_process_zero": true,
8
+ "is_world_process_zero": true,
9
+ "log_history": [
10
+ {
11
+ "epoch": 0.01,
12
+ "eval_loss": 0.07366943359375,
13
+ "eval_runtime": 1988.0122,
14
+ "eval_samples_per_second": 5.232,
15
+ "eval_steps_per_second": 0.164,
16
+ "step": 20
17
+ },
18
+ {
19
+ "epoch": 0.02,
20
+ "eval_loss": 0.06402587890625,
21
+ "eval_runtime": 1988.5161,
22
+ "eval_samples_per_second": 5.231,
23
+ "eval_steps_per_second": 0.164,
24
+ "step": 40
25
+ },
26
+ {
27
+ "epoch": 0.02,
28
+ "eval_loss": 0.060882568359375,
29
+ "eval_runtime": 1988.0139,
30
+ "eval_samples_per_second": 5.232,
31
+ "eval_steps_per_second": 0.164,
32
+ "step": 60
33
+ },
34
+ {
35
+ "epoch": 0.03,
36
+ "eval_loss": 0.05999755859375,
37
+ "eval_runtime": 1986.9046,
38
+ "eval_samples_per_second": 5.235,
39
+ "eval_steps_per_second": 0.164,
40
+ "step": 80
41
+ },
42
+ {
43
+ "epoch": 0.04,
44
+ "eval_loss": 0.059539794921875,
45
+ "eval_runtime": 1986.5359,
46
+ "eval_samples_per_second": 5.236,
47
+ "eval_steps_per_second": 0.164,
48
+ "step": 100
49
+ },
50
+ {
51
+ "epoch": 0.05,
52
+ "eval_loss": 0.05914306640625,
53
+ "eval_runtime": 1986.2404,
54
+ "eval_samples_per_second": 5.237,
55
+ "eval_steps_per_second": 0.164,
56
+ "step": 120
57
+ },
58
+ {
59
+ "epoch": 0.05,
60
+ "eval_loss": 0.0589599609375,
61
+ "eval_runtime": 1985.8515,
62
+ "eval_samples_per_second": 5.238,
63
+ "eval_steps_per_second": 0.164,
64
+ "step": 140
65
+ },
66
+ {
67
+ "epoch": 0.06,
68
+ "eval_loss": 0.058807373046875,
69
+ "eval_runtime": 1985.7582,
70
+ "eval_samples_per_second": 5.238,
71
+ "eval_steps_per_second": 0.164,
72
+ "step": 160
73
+ },
74
+ {
75
+ "epoch": 0.07,
76
+ "eval_loss": 0.05853271484375,
77
+ "eval_runtime": 1986.2944,
78
+ "eval_samples_per_second": 5.236,
79
+ "eval_steps_per_second": 0.164,
80
+ "step": 180
81
+ },
82
+ {
83
+ "epoch": 0.08,
84
+ "eval_loss": 0.05841064453125,
85
+ "eval_runtime": 1986.0105,
86
+ "eval_samples_per_second": 5.237,
87
+ "eval_steps_per_second": 0.164,
88
+ "step": 200
89
+ },
90
+ {
91
+ "epoch": 0.08,
92
+ "eval_loss": 0.058349609375,
93
+ "eval_runtime": 1986.7151,
94
+ "eval_samples_per_second": 5.235,
95
+ "eval_steps_per_second": 0.164,
96
+ "step": 220
97
+ },
98
+ {
99
+ "epoch": 0.09,
100
+ "eval_loss": 0.058197021484375,
101
+ "eval_runtime": 1985.7618,
102
+ "eval_samples_per_second": 5.238,
103
+ "eval_steps_per_second": 0.164,
104
+ "step": 240
105
+ },
106
+ {
107
+ "epoch": 0.1,
108
+ "eval_loss": 0.058135986328125,
109
+ "eval_runtime": 1985.9202,
110
+ "eval_samples_per_second": 5.237,
111
+ "eval_steps_per_second": 0.164,
112
+ "step": 260
113
+ },
114
+ {
115
+ "epoch": 0.11,
116
+ "eval_loss": 0.058135986328125,
117
+ "eval_runtime": 1986.6184,
118
+ "eval_samples_per_second": 5.236,
119
+ "eval_steps_per_second": 0.164,
120
+ "step": 280
121
+ },
122
+ {
123
+ "epoch": 0.12,
124
+ "eval_loss": 0.0579833984375,
125
+ "eval_runtime": 1986.6502,
126
+ "eval_samples_per_second": 5.235,
127
+ "eval_steps_per_second": 0.164,
128
+ "step": 300
129
+ },
130
+ {
131
+ "epoch": 0.12,
132
+ "eval_loss": 0.0579833984375,
133
+ "eval_runtime": 1985.5739,
134
+ "eval_samples_per_second": 5.238,
135
+ "eval_steps_per_second": 0.164,
136
+ "step": 320
137
+ },
138
+ {
139
+ "epoch": 0.13,
140
+ "eval_loss": 0.057861328125,
141
+ "eval_runtime": 1985.7642,
142
+ "eval_samples_per_second": 5.238,
143
+ "eval_steps_per_second": 0.164,
144
+ "step": 340
145
+ },
146
+ {
147
+ "epoch": 0.14,
148
+ "eval_loss": 0.0577392578125,
149
+ "eval_runtime": 1985.2011,
150
+ "eval_samples_per_second": 5.239,
151
+ "eval_steps_per_second": 0.164,
152
+ "step": 360
153
+ },
154
+ {
155
+ "epoch": 0.15,
156
+ "eval_loss": 0.057769775390625,
157
+ "eval_runtime": 1985.6821,
158
+ "eval_samples_per_second": 5.238,
159
+ "eval_steps_per_second": 0.164,
160
+ "step": 380
161
+ },
162
+ {
163
+ "epoch": 0.15,
164
+ "eval_loss": 0.057647705078125,
165
+ "eval_runtime": 1986.1381,
166
+ "eval_samples_per_second": 5.237,
167
+ "eval_steps_per_second": 0.164,
168
+ "step": 400
169
+ },
170
+ {
171
+ "epoch": 0.16,
172
+ "eval_loss": 0.05767822265625,
173
+ "eval_runtime": 1985.6719,
174
+ "eval_samples_per_second": 5.238,
175
+ "eval_steps_per_second": 0.164,
176
+ "step": 420
177
+ },
178
+ {
179
+ "epoch": 0.17,
180
+ "eval_loss": 0.057647705078125,
181
+ "eval_runtime": 1985.643,
182
+ "eval_samples_per_second": 5.238,
183
+ "eval_steps_per_second": 0.164,
184
+ "step": 440
185
+ },
186
+ {
187
+ "epoch": 0.18,
188
+ "eval_loss": 0.057525634765625,
189
+ "eval_runtime": 1985.6414,
190
+ "eval_samples_per_second": 5.238,
191
+ "eval_steps_per_second": 0.164,
192
+ "step": 460
193
+ },
194
+ {
195
+ "epoch": 0.18,
196
+ "eval_loss": 0.05755615234375,
197
+ "eval_runtime": 1985.6316,
198
+ "eval_samples_per_second": 5.238,
199
+ "eval_steps_per_second": 0.164,
200
+ "step": 480
201
+ },
202
+ {
203
+ "epoch": 0.19,
204
+ "learning_rate": 5e-06,
205
+ "loss": 0.2414,
206
+ "step": 500
207
+ },
208
+ {
209
+ "epoch": 0.19,
210
+ "eval_loss": 0.057525634765625,
211
+ "eval_runtime": 1986.8786,
212
+ "eval_samples_per_second": 5.235,
213
+ "eval_steps_per_second": 0.164,
214
+ "step": 500
215
+ },
216
+ {
217
+ "epoch": 0.2,
218
+ "eval_loss": 0.0574951171875,
219
+ "eval_runtime": 1987.3346,
220
+ "eval_samples_per_second": 5.234,
221
+ "eval_steps_per_second": 0.164,
222
+ "step": 520
223
+ },
224
+ {
225
+ "epoch": 0.21,
226
+ "eval_loss": 0.057464599609375,
227
+ "eval_runtime": 1986.2938,
228
+ "eval_samples_per_second": 5.236,
229
+ "eval_steps_per_second": 0.164,
230
+ "step": 540
231
+ },
232
+ {
233
+ "epoch": 0.22,
234
+ "eval_loss": 0.0574951171875,
235
+ "eval_runtime": 1986.0284,
236
+ "eval_samples_per_second": 5.237,
237
+ "eval_steps_per_second": 0.164,
238
+ "step": 560
239
+ },
240
+ {
241
+ "epoch": 0.22,
242
+ "eval_loss": 0.057403564453125,
243
+ "eval_runtime": 1987.0081,
244
+ "eval_samples_per_second": 5.235,
245
+ "eval_steps_per_second": 0.164,
246
+ "step": 580
247
+ },
248
+ {
249
+ "epoch": 0.23,
250
+ "eval_loss": 0.057403564453125,
251
+ "eval_runtime": 1987.2441,
252
+ "eval_samples_per_second": 5.234,
253
+ "eval_steps_per_second": 0.164,
254
+ "step": 600
255
+ },
256
+ {
257
+ "epoch": 0.24,
258
+ "eval_loss": 0.057403564453125,
259
+ "eval_runtime": 1988.9229,
260
+ "eval_samples_per_second": 5.229,
261
+ "eval_steps_per_second": 0.164,
262
+ "step": 620
263
+ },
264
+ {
265
+ "epoch": 0.25,
266
+ "eval_loss": 0.057403564453125,
267
+ "eval_runtime": 1988.9574,
268
+ "eval_samples_per_second": 5.229,
269
+ "eval_steps_per_second": 0.164,
270
+ "step": 640
271
+ },
272
+ {
273
+ "epoch": 0.25,
274
+ "eval_loss": 0.057342529296875,
275
+ "eval_runtime": 1989.2299,
276
+ "eval_samples_per_second": 5.229,
277
+ "eval_steps_per_second": 0.164,
278
+ "step": 660
279
+ },
280
+ {
281
+ "epoch": 0.26,
282
+ "eval_loss": 0.057281494140625,
283
+ "eval_runtime": 1989.7704,
284
+ "eval_samples_per_second": 5.227,
285
+ "eval_steps_per_second": 0.164,
286
+ "step": 680
287
+ }
288
+ ],
289
+ "max_steps": 31212,
290
+ "num_train_epochs": 12,
291
+ "total_flos": 108266998726656.0,
292
+ "trial_name": null,
293
+ "trial_params": null
294
+ }
training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7a9191e8a5e0bb4e333a1882c36df2f9a5b47697d3ecd47a8f9a1ffca648d116
3
+ size 4923
vocab.json ADDED
The diff for this file is too large to render. See raw diff
 
zero_to_fp32.py ADDED
@@ -0,0 +1,483 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python
2
+ '''Copyright The Microsoft DeepSpeed Team'''
3
+
4
+ # This script extracts fp32 consolidated weights from a zero 2 and 3 DeepSpeed checkpoints. It gets
5
+ # copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
6
+ # the future. Once extracted, the weights don't require DeepSpeed and can be used in any
7
+ # application.
8
+ #
9
+ # example: python zero_to_fp32.py . pytorch_model.bin
10
+
11
+ import argparse
12
+ import torch
13
+ import glob
14
+ import math
15
+ import os
16
+ import re
17
+ from collections import OrderedDict
18
+
19
+ # while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
20
+ # DeepSpeed data structures it has to be available in the current python environment.
21
+ from deepspeed.utils import logger
22
+ from deepspeed.checkpoint.constants import (DS_VERSION,
23
+ OPTIMIZER_STATE_DICT,
24
+ SINGLE_PARTITION_OF_FP32_GROUPS,
25
+ FP32_FLAT_GROUPS,
26
+ ZERO_STAGE,
27
+ PARTITION_COUNT,
28
+ PARAM_SHAPES,
29
+ BUFFER_NAMES)
30
+
31
+ debug = 0
32
+
33
+ # load to cpu
34
+ device = torch.device('cpu')
35
+
36
+
37
+ def atoi(text):
38
+ return int(text) if text.isdigit() else text
39
+
40
+
41
+ def natural_keys(text):
42
+ '''
43
+ alist.sort(key=natural_keys) sorts in human order
44
+ http://nedbatchelder.com/blog/200712/human_sorting.html
45
+ (See Toothy's implementation in the comments)
46
+ '''
47
+ return [atoi(c) for c in re.split(r'(\d+)', text)]
48
+
49
+
50
+ def get_model_state_file(checkpoint_dir, zero_stage):
51
+ if not os.path.isdir(checkpoint_dir):
52
+ raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
53
+
54
+ # there should be only one file
55
+ if zero_stage == 2:
56
+ file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
57
+ elif zero_stage == 3:
58
+ file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
59
+
60
+ if not os.path.exists(file):
61
+ raise FileNotFoundError(f"can't find model states file at '{file}'")
62
+
63
+ return file
64
+
65
+
66
+ def get_optim_files(checkpoint_dir):
67
+ # XXX: need to test that this simple glob rule works for multi-node setup too
68
+ optim_files = sorted(glob.glob(os.path.join(checkpoint_dir,
69
+ "*_optim_states.pt")),
70
+ key=natural_keys)
71
+
72
+ if len(optim_files) == 0:
73
+ raise FileNotFoundError(
74
+ f"can't find '*_optim_states.pt' files in directory '{checkpoint_dir}'")
75
+
76
+ return optim_files
77
+
78
+
79
+ def parse_model_state(file):
80
+ state_dict = torch.load(file, map_location=device)
81
+
82
+ if BUFFER_NAMES not in state_dict:
83
+ raise ValueError(f"{file} is not a model state checkpoint")
84
+ buffer_names = state_dict[BUFFER_NAMES]
85
+ if debug:
86
+ print("Found buffers:", buffer_names)
87
+
88
+ # recover just the buffers while restoring them to fp32 if they were saved in fp16
89
+ buffers = {
90
+ k: v.float()
91
+ for k,
92
+ v in state_dict["module"].items() if k in buffer_names
93
+ }
94
+ param_shapes = state_dict[PARAM_SHAPES]
95
+
96
+ ds_version = state_dict.get(DS_VERSION, None)
97
+
98
+ return buffers, param_shapes, ds_version
99
+
100
+
101
+ def parse_optim_states(files, ds_checkpoint_dir):
102
+
103
+ total_files = len(files)
104
+ state_dicts = []
105
+ for f in files:
106
+ state_dicts.append(torch.load(f, map_location=device))
107
+
108
+ if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
109
+ raise ValueError(f"{files[0]} is not a zero checkpoint")
110
+ zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
111
+ world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
112
+
113
+ # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
114
+ # parameters can be different from data parallelism for non-expert parameters. So we can just
115
+ # use the max of the partition_count to get the dp world_size.
116
+
117
+ if type(world_size) is list:
118
+ world_size = max(world_size)
119
+
120
+ if world_size != total_files:
121
+ raise ValueError(
122
+ f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
123
+ "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
124
+ )
125
+
126
+ # the groups are named differently in each stage
127
+ if zero_stage == 2:
128
+ fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
129
+ elif zero_stage == 3:
130
+ fp32_groups_key = FP32_FLAT_GROUPS
131
+ else:
132
+ raise ValueError(f"unknown zero stage {zero_stage}")
133
+
134
+ if zero_stage == 2:
135
+ fp32_flat_groups = [
136
+ state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key]
137
+ for i in range(len(state_dicts))
138
+ ]
139
+ elif zero_stage == 3:
140
+ # if there is more than one param group, there will be multiple flattened tensors - one
141
+ # flattened tensor per group - for simplicity merge them into a single tensor
142
+ #
143
+ # XXX: could make the script more memory efficient for when there are multiple groups - it
144
+ # will require matching the sub-lists of param_shapes for each param group flattened tensor
145
+
146
+ fp32_flat_groups = [
147
+ torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key],
148
+ 0) for i in range(len(state_dicts))
149
+ ]
150
+
151
+ return zero_stage, world_size, fp32_flat_groups
152
+
153
+
154
+ def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
155
+ """
156
+ Returns fp32 state_dict reconstructed from ds checkpoint
157
+
158
+ Args:
159
+ - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
160
+
161
+ """
162
+ print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
163
+
164
+ optim_files = get_optim_files(ds_checkpoint_dir)
165
+ zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
166
+ print(
167
+ f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
168
+
169
+ model_file = get_model_state_file(ds_checkpoint_dir, zero_stage)
170
+ buffers, param_shapes, ds_version = parse_model_state(model_file)
171
+ print(f'Parsing checkpoint created by deepspeed=={ds_version}')
172
+
173
+ if zero_stage == 2:
174
+ return _get_fp32_state_dict_from_zero2_checkpoint(world_size,
175
+ param_shapes,
176
+ fp32_flat_groups,
177
+ buffers)
178
+ elif zero_stage == 3:
179
+ return _get_fp32_state_dict_from_zero3_checkpoint(world_size,
180
+ param_shapes,
181
+ fp32_flat_groups,
182
+ buffers)
183
+
184
+
185
+ def _get_fp32_state_dict_from_zero2_checkpoint(world_size,
186
+ param_shapes,
187
+ fp32_flat_groups,
188
+ buffers):
189
+
190
+ # Reconstruction protocol:
191
+ #
192
+ # XXX: document this
193
+
194
+ if debug:
195
+ for i in range(world_size):
196
+ for j in range(len(fp32_flat_groups[0])):
197
+ print(
198
+ f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
199
+
200
+ # XXX: memory usage doubles here (zero2)
201
+ num_param_groups = len(fp32_flat_groups[0])
202
+ merged_single_partition_of_fp32_groups = []
203
+ for i in range(num_param_groups):
204
+ merged_partitions = [sd[i] for sd in fp32_flat_groups]
205
+ full_single_fp32_vector = torch.cat(merged_partitions, 0)
206
+ merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
207
+ avail_numel = sum([
208
+ full_single_fp32_vector.numel()
209
+ for full_single_fp32_vector in merged_single_partition_of_fp32_groups
210
+ ])
211
+
212
+ if debug:
213
+ wanted_params = sum([len(shapes) for shapes in param_shapes])
214
+ wanted_numel = sum(
215
+ [sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
216
+ # not asserting if there is a mismatch due to possible padding
217
+ print(f"Have {avail_numel} numels to process.")
218
+ print(f"Need {wanted_numel} numels in {wanted_params} params.")
219
+
220
+ state_dict = OrderedDict()
221
+
222
+ # buffers
223
+ state_dict.update(buffers)
224
+ if debug:
225
+ print(f"added {len(buffers)} buffers")
226
+
227
+ # params
228
+ # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
229
+ # out-of-core computing solution
230
+ total_numel = 0
231
+ total_params = 0
232
+ for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
233
+ offset = 0
234
+ avail_numel = full_single_fp32_vector.numel()
235
+ for name, shape in shapes.items():
236
+
237
+ unpartitioned_numel = shape.numel()
238
+ total_numel += unpartitioned_numel
239
+ total_params += 1
240
+
241
+ if debug:
242
+ print(
243
+ f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} "
244
+ )
245
+ state_dict[name] = full_single_fp32_vector.narrow(
246
+ 0,
247
+ offset,
248
+ unpartitioned_numel).view(shape)
249
+ offset += unpartitioned_numel
250
+
251
+ # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
252
+ # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
253
+ # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
254
+ # live optimizer object, so we are checking that the numbers are within the right range
255
+ align_to = 2 * world_size
256
+
257
+ def zero2_align(x):
258
+ return align_to * math.ceil(x / align_to)
259
+
260
+ if debug:
261
+ print(f"original offset={offset}, avail_numel={avail_numel}")
262
+
263
+ offset = zero2_align(offset)
264
+ avail_numel = zero2_align(avail_numel)
265
+
266
+ if debug:
267
+ print(f"aligned offset={offset}, avail_numel={avail_numel}")
268
+
269
+ # Sanity check
270
+ if offset != avail_numel:
271
+ raise ValueError(
272
+ f"consumed {offset} numels out of {avail_numel} - something is wrong")
273
+
274
+ print(
275
+ f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements"
276
+ )
277
+
278
+ return state_dict
279
+
280
+
281
+ def zero3_partitioned_param_info(unpartitioned_numel, world_size):
282
+ remainder = unpartitioned_numel % world_size
283
+ padding_numel = (world_size - remainder) if remainder else 0
284
+ partitioned_numel = math.ceil(unpartitioned_numel / world_size)
285
+ return partitioned_numel, padding_numel
286
+
287
+
288
+ def _get_fp32_state_dict_from_zero3_checkpoint(world_size,
289
+ param_shapes,
290
+ fp32_flat_groups,
291
+ buffers):
292
+
293
+ # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
294
+ # param, re-consolidating each param, while dealing with padding if any
295
+
296
+ avail_numel = fp32_flat_groups[0].numel() * world_size
297
+ # merge list of dicts, preserving order
298
+ param_shapes = {k: v for d in param_shapes for k, v in d.items()}
299
+
300
+ if debug:
301
+ for i in range(world_size):
302
+ print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
303
+
304
+ wanted_params = len(param_shapes)
305
+ wanted_numel = sum(shape.numel() for shape in param_shapes.values())
306
+ # not asserting if there is a mismatch due to possible padding
307
+ print(f"Have {avail_numel} numels to process.")
308
+ print(f"Need {wanted_numel} numels in {wanted_params} params.")
309
+
310
+ state_dict = OrderedDict()
311
+
312
+ # buffers
313
+ state_dict.update(buffers)
314
+ if debug:
315
+ print(f"added {len(buffers)} buffers")
316
+
317
+ # params
318
+ # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
319
+ # out-of-core computing solution
320
+ offset = 0
321
+ total_numel = 0
322
+ total_params = 0
323
+ for name, shape in param_shapes.items():
324
+
325
+ unpartitioned_numel = shape.numel()
326
+ total_numel += unpartitioned_numel
327
+ total_params += 1
328
+
329
+ partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
330
+
331
+ if debug:
332
+ print(
333
+ f"{total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
334
+ )
335
+
336
+ # XXX: memory usage doubles here
337
+ state_dict[name] = torch.cat(
338
+ tuple(fp32_flat_groups[i].narrow(0,
339
+ offset,
340
+ partitioned_numel)
341
+ for i in range(world_size)),
342
+ 0).narrow(0,
343
+ 0,
344
+ unpartitioned_numel).view(shape)
345
+ offset += partitioned_numel
346
+
347
+ offset *= world_size
348
+
349
+ # Sanity check
350
+ if offset != avail_numel:
351
+ raise ValueError(
352
+ f"consumed {offset} numels out of {avail_numel} - something is wrong")
353
+
354
+ print(
355
+ f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements"
356
+ )
357
+
358
+ return state_dict
359
+
360
+
361
+ def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
362
+ """
363
+ Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
364
+ ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
365
+ via a model hub.
366
+
367
+ Args:
368
+ - ``checkpoint_dir``: path to the desired checkpoint folder
369
+ - ``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``
370
+
371
+ Returns:
372
+ - pytorch ``state_dict``
373
+
374
+ Note: this approach may not work if your application doesn't have sufficient free CPU memory and
375
+ you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
376
+ the checkpoint.
377
+
378
+ A typical usage might be ::
379
+
380
+ from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
381
+ # do the training and checkpoint saving
382
+ state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
383
+ model = model.cpu() # move to cpu
384
+ model.load_state_dict(state_dict)
385
+ # submit to model hub or save the model to share with others
386
+
387
+ In this example the ``model`` will no longer be usable in the deepspeed context of the same
388
+ application. i.e. you will need to re-initialize the deepspeed engine, since
389
+ ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
390
+
391
+ If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
392
+
393
+ """
394
+ if tag is None:
395
+ latest_path = os.path.join(checkpoint_dir, 'latest')
396
+ if os.path.isfile(latest_path):
397
+ with open(latest_path, 'r') as fd:
398
+ tag = fd.read().strip()
399
+ else:
400
+ raise ValueError(f"Unable to find 'latest' file at {latest_path}")
401
+
402
+ ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
403
+
404
+ if not os.path.isdir(ds_checkpoint_dir):
405
+ raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
406
+
407
+ return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
408
+
409
+
410
+ def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
411
+ """
412
+ Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
413
+ loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
414
+
415
+ Args:
416
+ - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
417
+ - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
418
+ - ``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``
419
+ """
420
+
421
+ state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
422
+ print(f"Saving fp32 state dict to {output_file}")
423
+ torch.save(state_dict, output_file)
424
+
425
+
426
+ def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
427
+ """
428
+ 1. Put the provided model to cpu
429
+ 2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
430
+ 3. Load it into the provided model
431
+
432
+ Args:
433
+ - ``model``: the model object to update
434
+ - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
435
+ - ``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``
436
+
437
+ Returns:
438
+ - ``model`: modified model
439
+
440
+ Make sure you have plenty of CPU memory available before you call this function. If you don't
441
+ have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
442
+ conveniently placed for you in the checkpoint folder.
443
+
444
+ A typical usage might be ::
445
+
446
+ from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
447
+ model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
448
+ # submit to model hub or save the model to share with others
449
+
450
+ Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
451
+ of the same application. i.e. you will need to re-initialize the deepspeed engine, since
452
+ ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
453
+
454
+ """
455
+ logger.info(f"Extracting fp32 weights")
456
+ state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
457
+
458
+ logger.info(f"Overwriting model with fp32 weights")
459
+ model = model.cpu()
460
+ model.load_state_dict(state_dict, strict=False)
461
+
462
+ return model
463
+
464
+
465
+ if __name__ == "__main__":
466
+
467
+ parser = argparse.ArgumentParser()
468
+ parser.add_argument(
469
+ "checkpoint_dir",
470
+ type=str,
471
+ help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
472
+ parser.add_argument(
473
+ "output_file",
474
+ type=str,
475
+ help=
476
+ "path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)"
477
+ )
478
+ parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
479
+ args = parser.parse_args()
480
+
481
+ debug = args.debug
482
+
483
+ convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file)