shailja commited on
Commit
c3538ca
1 Parent(s): b01a7b3

Update from st4920

Browse files
added_tokens.json ADDED
@@ -0,0 +1,40 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "\t\t": 50294,
3
+ "\t\t\t": 50293,
4
+ "\t\t\t\t": 50292,
5
+ "\t\t\t\t\t": 50291,
6
+ "\t\t\t\t\t\t": 50290,
7
+ "\t\t\t\t\t\t\t": 50289,
8
+ "\t\t\t\t\t\t\t\t": 50288,
9
+ "\t\t\t\t\t\t\t\t\t": 50287,
10
+ " ": 50286,
11
+ " ": 50285,
12
+ " ": 50284,
13
+ " ": 50283,
14
+ " ": 50282,
15
+ " ": 50281,
16
+ " ": 50280,
17
+ " ": 50279,
18
+ " ": 50278,
19
+ " ": 50277,
20
+ " ": 50276,
21
+ " ": 50275,
22
+ " ": 50274,
23
+ " ": 50273,
24
+ " ": 50272,
25
+ " ": 50271,
26
+ " ": 50270,
27
+ " ": 50269,
28
+ " ": 50268,
29
+ " ": 50267,
30
+ " ": 50266,
31
+ " ": 50265,
32
+ " ": 50264,
33
+ " ": 50263,
34
+ " ": 50262,
35
+ " ": 50261,
36
+ " ": 50260,
37
+ " ": 50259,
38
+ " ": 50258,
39
+ " ": 50257
40
+ }
config.json ADDED
@@ -0,0 +1,42 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "Salesforce/codegen-6B-multi",
3
+ "activation_function": "gelu_new",
4
+ "architectures": [
5
+ "CodeGenForCausalLM"
6
+ ],
7
+ "attn_pdrop": 0.0,
8
+ "bos_token_id": 1,
9
+ "embd_pdrop": 0.0,
10
+ "eos_token_id": 50256,
11
+ "gradient_checkpointing": false,
12
+ "initializer_range": 0.02,
13
+ "layer_norm_epsilon": 1e-05,
14
+ "model_type": "codegen",
15
+ "n_ctx": 2048,
16
+ "n_embd": 4096,
17
+ "n_head": 16,
18
+ "n_inner": null,
19
+ "n_layer": 33,
20
+ "n_positions": 2048,
21
+ "resid_pdrop": 0.0,
22
+ "rotary_dim": 64,
23
+ "scale_attn_weights": true,
24
+ "summary_activation": null,
25
+ "summary_first_dropout": 0.1,
26
+ "summary_proj_to_labels": true,
27
+ "summary_type": "cls_index",
28
+ "summary_use_proj": true,
29
+ "task_specific_params": {
30
+ "text-generation": {
31
+ "do_sample": true,
32
+ "max_length": 50,
33
+ "temperature": 1.0
34
+ }
35
+ },
36
+ "tie_word_embeddings": false,
37
+ "tokenizer_class": "GPT2Tokenizer",
38
+ "torch_dtype": "float16",
39
+ "transformers_version": "4.25.0.dev0",
40
+ "use_cache": true,
41
+ "vocab_size": 50295
42
+ }
latest ADDED
@@ -0,0 +1 @@
 
 
1
+ global_step10000
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 sha256:0540847db12ad7b4eb6bc04510005157b5f4fd6b8580da3ce4cb853d27a27718
3
+ size 28780940193
rng_state_0.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e1140f09963bcd2999177e311c63c2ac4e6ec21122cfc6b62349bf8084a6293a
3
+ size 14503
rng_state_1.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e1140f09963bcd2999177e311c63c2ac4e6ec21122cfc6b62349bf8084a6293a
3
+ size 14503
rng_state_2.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e1140f09963bcd2999177e311c63c2ac4e6ec21122cfc6b62349bf8084a6293a
3
+ size 14503
special_tokens_map.json ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ {
2
+ "bos_token": "<|endoftext|>",
3
+ "eos_token": "<|endoftext|>",
4
+ "unk_token": "<|endoftext|>"
5
+ }
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer_config.json ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_prefix_space": false,
3
+ "bos_token": "<|endoftext|>",
4
+ "eos_token": "<|endoftext|>",
5
+ "model_max_length": 2048,
6
+ "name_or_path": "Salesforce/codegen-16B-multi",
7
+ "special_tokens_map_file": null,
8
+ "tokenizer_class": "CodeGenTokenizer",
9
+ "unk_token": "<|endoftext|>"
10
+ }
trainer_state.json ADDED
@@ -0,0 +1,136 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "best_metric": null,
3
+ "best_model_checkpoint": null,
4
+ "epoch": 0.962822997027284,
5
+ "global_step": 10000,
6
+ "is_hyper_param_search": false,
7
+ "is_local_process_zero": true,
8
+ "is_world_process_zero": true,
9
+ "log_history": [
10
+ {
11
+ "epoch": 0.05,
12
+ "learning_rate": 1.904275808936826e-05,
13
+ "loss": 0.3678,
14
+ "step": 500
15
+ },
16
+ {
17
+ "epoch": 0.1,
18
+ "learning_rate": 1.807973805855162e-05,
19
+ "loss": 0.3141,
20
+ "step": 1000
21
+ },
22
+ {
23
+ "epoch": 0.14,
24
+ "learning_rate": 1.7116718027734977e-05,
25
+ "loss": 0.2928,
26
+ "step": 1500
27
+ },
28
+ {
29
+ "epoch": 0.19,
30
+ "learning_rate": 1.6153697996918337e-05,
31
+ "loss": 0.2788,
32
+ "step": 2000
33
+ },
34
+ {
35
+ "epoch": 0.24,
36
+ "learning_rate": 1.5190677966101697e-05,
37
+ "loss": 0.2664,
38
+ "step": 2500
39
+ },
40
+ {
41
+ "epoch": 0.29,
42
+ "learning_rate": 1.4227657935285056e-05,
43
+ "loss": 0.2551,
44
+ "step": 3000
45
+ },
46
+ {
47
+ "epoch": 0.34,
48
+ "learning_rate": 1.3264637904468413e-05,
49
+ "loss": 0.2465,
50
+ "step": 3500
51
+ },
52
+ {
53
+ "epoch": 0.39,
54
+ "learning_rate": 1.2301617873651772e-05,
55
+ "loss": 0.2373,
56
+ "step": 4000
57
+ },
58
+ {
59
+ "epoch": 0.43,
60
+ "learning_rate": 1.1338597842835134e-05,
61
+ "loss": 0.2305,
62
+ "step": 4500
63
+ },
64
+ {
65
+ "epoch": 0.48,
66
+ "learning_rate": 1.037557781201849e-05,
67
+ "loss": 0.2229,
68
+ "step": 5000
69
+ },
70
+ {
71
+ "epoch": 0.53,
72
+ "learning_rate": 9.416409861325117e-06,
73
+ "loss": 0.2154,
74
+ "step": 5500
75
+ },
76
+ {
77
+ "epoch": 0.58,
78
+ "learning_rate": 8.455315870570109e-06,
79
+ "loss": 0.2104,
80
+ "step": 6000
81
+ },
82
+ {
83
+ "epoch": 0.63,
84
+ "learning_rate": 7.492295839753467e-06,
85
+ "loss": 0.2043,
86
+ "step": 6500
87
+ },
88
+ {
89
+ "epoch": 0.67,
90
+ "learning_rate": 6.529275808936827e-06,
91
+ "loss": 0.1985,
92
+ "step": 7000
93
+ },
94
+ {
95
+ "epoch": 0.72,
96
+ "learning_rate": 5.568181818181818e-06,
97
+ "loss": 0.1953,
98
+ "step": 7500
99
+ },
100
+ {
101
+ "epoch": 0.77,
102
+ "learning_rate": 4.605161787365178e-06,
103
+ "loss": 0.1903,
104
+ "step": 8000
105
+ },
106
+ {
107
+ "epoch": 0.82,
108
+ "learning_rate": 3.645993836671803e-06,
109
+ "loss": 0.186,
110
+ "step": 8500
111
+ },
112
+ {
113
+ "epoch": 0.87,
114
+ "learning_rate": 2.682973805855162e-06,
115
+ "loss": 0.1837,
116
+ "step": 9000
117
+ },
118
+ {
119
+ "epoch": 0.91,
120
+ "learning_rate": 1.7238058551617876e-06,
121
+ "loss": 0.1803,
122
+ "step": 9500
123
+ },
124
+ {
125
+ "epoch": 0.96,
126
+ "learning_rate": 7.607858243451465e-07,
127
+ "loss": 0.1775,
128
+ "step": 10000
129
+ }
130
+ ],
131
+ "max_steps": 10386,
132
+ "num_train_epochs": 1,
133
+ "total_flos": 1.1258147859595264e+16,
134
+ "trial_name": null,
135
+ "trial_params": null
136
+ }
training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c412aba4d0cebed53bcb26182922fdbfa5f1c2b36fd568a2b8fb10b5ef8985a5
3
+ size 4719
vocab.json ADDED
The diff for this file is too large to render. See raw diff
 
zero_to_fp32.py ADDED
@@ -0,0 +1,482 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python
2
+
3
+ # This script extracts fp32 consolidated weights from a zero 2 and 3 DeepSpeed checkpoints. It gets
4
+ # copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
5
+ # the future. Once extracted, the weights don't require DeepSpeed and can be used in any
6
+ # application.
7
+ #
8
+ # example: python zero_to_fp32.py . pytorch_model.bin
9
+
10
+ import argparse
11
+ import torch
12
+ import glob
13
+ import math
14
+ import os
15
+ import re
16
+ from collections import OrderedDict
17
+
18
+ # while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
19
+ # DeepSpeed data structures it has to be available in the current python environment.
20
+ from deepspeed.utils import logger
21
+ from deepspeed.checkpoint.constants import (DS_VERSION,
22
+ OPTIMIZER_STATE_DICT,
23
+ SINGLE_PARTITION_OF_FP32_GROUPS,
24
+ FP32_FLAT_GROUPS,
25
+ ZERO_STAGE,
26
+ PARTITION_COUNT,
27
+ PARAM_SHAPES,
28
+ BUFFER_NAMES)
29
+
30
+ debug = 0
31
+
32
+ # load to cpu
33
+ device = torch.device('cpu')
34
+
35
+
36
+ def atoi(text):
37
+ return int(text) if text.isdigit() else text
38
+
39
+
40
+ def natural_keys(text):
41
+ '''
42
+ alist.sort(key=natural_keys) sorts in human order
43
+ http://nedbatchelder.com/blog/200712/human_sorting.html
44
+ (See Toothy's implementation in the comments)
45
+ '''
46
+ return [atoi(c) for c in re.split(r'(\d+)', text)]
47
+
48
+
49
+ def get_model_state_file(checkpoint_dir, zero_stage):
50
+ if not os.path.isdir(checkpoint_dir):
51
+ raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
52
+
53
+ # there should be only one file
54
+ if zero_stage == 2:
55
+ file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
56
+ elif zero_stage == 3:
57
+ file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
58
+
59
+ if not os.path.exists(file):
60
+ raise FileNotFoundError(f"can't find model states file at '{file}'")
61
+
62
+ return file
63
+
64
+
65
+ def get_optim_files(checkpoint_dir):
66
+ # XXX: need to test that this simple glob rule works for multi-node setup too
67
+ optim_files = sorted(glob.glob(os.path.join(checkpoint_dir,
68
+ "*_optim_states.pt")),
69
+ key=natural_keys)
70
+
71
+ if len(optim_files) == 0:
72
+ raise FileNotFoundError(
73
+ f"can't find '*_optim_states.pt' files in directory '{checkpoint_dir}'")
74
+
75
+ return optim_files
76
+
77
+
78
+ def parse_model_state(file):
79
+ state_dict = torch.load(file, map_location=device)
80
+
81
+ if BUFFER_NAMES not in state_dict:
82
+ raise ValueError(f"{file} is not a model state checkpoint")
83
+ buffer_names = state_dict[BUFFER_NAMES]
84
+ if debug:
85
+ print("Found buffers:", buffer_names)
86
+
87
+ # recover just the buffers while restoring them to fp32 if they were saved in fp16
88
+ buffers = {
89
+ k: v.float()
90
+ for k,
91
+ v in state_dict["module"].items() if k in buffer_names
92
+ }
93
+ param_shapes = state_dict[PARAM_SHAPES]
94
+
95
+ ds_version = state_dict.get(DS_VERSION, None)
96
+
97
+ return buffers, param_shapes, ds_version
98
+
99
+
100
+ def parse_optim_states(files, ds_checkpoint_dir):
101
+
102
+ total_files = len(files)
103
+ state_dicts = []
104
+ for f in files:
105
+ state_dicts.append(torch.load(f, map_location=device))
106
+
107
+ if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
108
+ raise ValueError(f"{files[0]} is not a zero checkpoint")
109
+ zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
110
+ world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
111
+
112
+ # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
113
+ # parameters can be different from data parallelism for non-expert parameters. So we can just
114
+ # use the max of the partition_count to get the dp world_size.
115
+
116
+ if type(world_size) is list:
117
+ world_size = max(world_size)
118
+
119
+ if world_size != total_files:
120
+ raise ValueError(
121
+ f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
122
+ "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
123
+ )
124
+
125
+ # the groups are named differently in each stage
126
+ if zero_stage == 2:
127
+ fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
128
+ elif zero_stage == 3:
129
+ fp32_groups_key = FP32_FLAT_GROUPS
130
+ else:
131
+ raise ValueError(f"unknown zero stage {zero_stage}")
132
+
133
+ if zero_stage == 2:
134
+ fp32_flat_groups = [
135
+ state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key]
136
+ for i in range(len(state_dicts))
137
+ ]
138
+ elif zero_stage == 3:
139
+ # if there is more than one param group, there will be multiple flattened tensors - one
140
+ # flattened tensor per group - for simplicity merge them into a single tensor
141
+ #
142
+ # XXX: could make the script more memory efficient for when there are multiple groups - it
143
+ # will require matching the sub-lists of param_shapes for each param group flattened tensor
144
+
145
+ fp32_flat_groups = [
146
+ torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key],
147
+ 0) for i in range(len(state_dicts))
148
+ ]
149
+
150
+ return zero_stage, world_size, fp32_flat_groups
151
+
152
+
153
+ def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
154
+ """
155
+ Returns fp32 state_dict reconstructed from ds checkpoint
156
+
157
+ Args:
158
+ - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
159
+
160
+ """
161
+ print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
162
+
163
+ optim_files = get_optim_files(ds_checkpoint_dir)
164
+ zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
165
+ print(
166
+ f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
167
+
168
+ model_file = get_model_state_file(ds_checkpoint_dir, zero_stage)
169
+ buffers, param_shapes, ds_version = parse_model_state(model_file)
170
+ print(f'Parsing checkpoint created by deepspeed=={ds_version}')
171
+
172
+ if zero_stage == 2:
173
+ return _get_fp32_state_dict_from_zero2_checkpoint(world_size,
174
+ param_shapes,
175
+ fp32_flat_groups,
176
+ buffers)
177
+ elif zero_stage == 3:
178
+ return _get_fp32_state_dict_from_zero3_checkpoint(world_size,
179
+ param_shapes,
180
+ fp32_flat_groups,
181
+ buffers)
182
+
183
+
184
+ def _get_fp32_state_dict_from_zero2_checkpoint(world_size,
185
+ param_shapes,
186
+ fp32_flat_groups,
187
+ buffers):
188
+
189
+ # Reconstruction protocol:
190
+ #
191
+ # XXX: document this
192
+
193
+ if debug:
194
+ for i in range(world_size):
195
+ for j in range(len(fp32_flat_groups[0])):
196
+ print(
197
+ f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
198
+
199
+ # XXX: memory usage doubles here (zero2)
200
+ num_param_groups = len(fp32_flat_groups[0])
201
+ merged_single_partition_of_fp32_groups = []
202
+ for i in range(num_param_groups):
203
+ merged_partitions = [sd[i] for sd in fp32_flat_groups]
204
+ full_single_fp32_vector = torch.cat(merged_partitions, 0)
205
+ merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
206
+ avail_numel = sum([
207
+ full_single_fp32_vector.numel()
208
+ for full_single_fp32_vector in merged_single_partition_of_fp32_groups
209
+ ])
210
+
211
+ if debug:
212
+ wanted_params = sum([len(shapes) for shapes in param_shapes])
213
+ wanted_numel = sum(
214
+ [sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
215
+ # not asserting if there is a mismatch due to possible padding
216
+ print(f"Have {avail_numel} numels to process.")
217
+ print(f"Need {wanted_numel} numels in {wanted_params} params.")
218
+
219
+ state_dict = OrderedDict()
220
+
221
+ # buffers
222
+ state_dict.update(buffers)
223
+ if debug:
224
+ print(f"added {len(buffers)} buffers")
225
+
226
+ # params
227
+ # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
228
+ # out-of-core computing solution
229
+ total_numel = 0
230
+ total_params = 0
231
+ for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
232
+ offset = 0
233
+ avail_numel = full_single_fp32_vector.numel()
234
+ for name, shape in shapes.items():
235
+
236
+ unpartitioned_numel = shape.numel()
237
+ total_numel += unpartitioned_numel
238
+ total_params += 1
239
+
240
+ if debug:
241
+ print(
242
+ f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} "
243
+ )
244
+ state_dict[name] = full_single_fp32_vector.narrow(
245
+ 0,
246
+ offset,
247
+ unpartitioned_numel).view(shape)
248
+ offset += unpartitioned_numel
249
+
250
+ # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
251
+ # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
252
+ # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
253
+ # live optimizer object, so we are checking that the numbers are within the right range
254
+ align_to = 2 * world_size
255
+
256
+ def zero2_align(x):
257
+ return align_to * math.ceil(x / align_to)
258
+
259
+ if debug:
260
+ print(f"original offset={offset}, avail_numel={avail_numel}")
261
+
262
+ offset = zero2_align(offset)
263
+ avail_numel = zero2_align(avail_numel)
264
+
265
+ if debug:
266
+ print(f"aligned offset={offset}, avail_numel={avail_numel}")
267
+
268
+ # Sanity check
269
+ if offset != avail_numel:
270
+ raise ValueError(
271
+ f"consumed {offset} numels out of {avail_numel} - something is wrong")
272
+
273
+ print(
274
+ f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements"
275
+ )
276
+
277
+ return state_dict
278
+
279
+
280
+ def zero3_partitioned_param_info(unpartitioned_numel, world_size):
281
+ remainder = unpartitioned_numel % world_size
282
+ padding_numel = (world_size - remainder) if remainder else 0
283
+ partitioned_numel = math.ceil(unpartitioned_numel / world_size)
284
+ return partitioned_numel, padding_numel
285
+
286
+
287
+ def _get_fp32_state_dict_from_zero3_checkpoint(world_size,
288
+ param_shapes,
289
+ fp32_flat_groups,
290
+ buffers):
291
+
292
+ # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
293
+ # param, re-consolidating each param, while dealing with padding if any
294
+
295
+ avail_numel = fp32_flat_groups[0].numel() * world_size
296
+ # merge list of dicts, preserving order
297
+ param_shapes = {k: v for d in param_shapes for k, v in d.items()}
298
+
299
+ if debug:
300
+ for i in range(world_size):
301
+ print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
302
+
303
+ wanted_params = len(param_shapes)
304
+ wanted_numel = sum(shape.numel() for shape in param_shapes.values())
305
+ # not asserting if there is a mismatch due to possible padding
306
+ print(f"Have {avail_numel} numels to process.")
307
+ print(f"Need {wanted_numel} numels in {wanted_params} params.")
308
+
309
+ state_dict = OrderedDict()
310
+
311
+ # buffers
312
+ state_dict.update(buffers)
313
+ if debug:
314
+ print(f"added {len(buffers)} buffers")
315
+
316
+ # params
317
+ # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
318
+ # out-of-core computing solution
319
+ offset = 0
320
+ total_numel = 0
321
+ total_params = 0
322
+ for name, shape in param_shapes.items():
323
+
324
+ unpartitioned_numel = shape.numel()
325
+ total_numel += unpartitioned_numel
326
+ total_params += 1
327
+
328
+ partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
329
+
330
+ if debug:
331
+ print(
332
+ f"{total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
333
+ )
334
+
335
+ # XXX: memory usage doubles here
336
+ state_dict[name] = torch.cat(
337
+ tuple(fp32_flat_groups[i].narrow(0,
338
+ offset,
339
+ partitioned_numel)
340
+ for i in range(world_size)),
341
+ 0).narrow(0,
342
+ 0,
343
+ unpartitioned_numel).view(shape)
344
+ offset += partitioned_numel
345
+
346
+ offset *= world_size
347
+
348
+ # Sanity check
349
+ if offset != avail_numel:
350
+ raise ValueError(
351
+ f"consumed {offset} numels out of {avail_numel} - something is wrong")
352
+
353
+ print(
354
+ f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements"
355
+ )
356
+
357
+ return state_dict
358
+
359
+
360
+ def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
361
+ """
362
+ Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
363
+ ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
364
+ via a model hub.
365
+
366
+ Args:
367
+ - ``checkpoint_dir``: path to the desired checkpoint folder
368
+ - ``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``
369
+
370
+ Returns:
371
+ - pytorch ``state_dict``
372
+
373
+ Note: this approach may not work if your application doesn't have sufficient free CPU memory and
374
+ you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
375
+ the checkpoint.
376
+
377
+ A typical usage might be ::
378
+
379
+ from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
380
+ # do the training and checkpoint saving
381
+ state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
382
+ model = model.cpu() # move to cpu
383
+ model.load_state_dict(state_dict)
384
+ # submit to model hub or save the model to share with others
385
+
386
+ In this example the ``model`` will no longer be usable in the deepspeed context of the same
387
+ application. i.e. you will need to re-initialize the deepspeed engine, since
388
+ ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
389
+
390
+ If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
391
+
392
+ """
393
+ if tag is None:
394
+ latest_path = os.path.join(checkpoint_dir, 'latest')
395
+ if os.path.isfile(latest_path):
396
+ with open(latest_path, 'r') as fd:
397
+ tag = fd.read().strip()
398
+ else:
399
+ raise ValueError(f"Unable to find 'latest' file at {latest_path}")
400
+
401
+ ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
402
+
403
+ if not os.path.isdir(ds_checkpoint_dir):
404
+ raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
405
+
406
+ return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
407
+
408
+
409
+ def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
410
+ """
411
+ Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
412
+ loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
413
+
414
+ Args:
415
+ - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
416
+ - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
417
+ - ``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``
418
+ """
419
+
420
+ state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
421
+ print(f"Saving fp32 state dict to {output_file}")
422
+ torch.save(state_dict, output_file)
423
+
424
+
425
+ def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
426
+ """
427
+ 1. Put the provided model to cpu
428
+ 2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
429
+ 3. Load it into the provided model
430
+
431
+ Args:
432
+ - ``model``: the model object to update
433
+ - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
434
+ - ``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``
435
+
436
+ Returns:
437
+ - ``model`: modified model
438
+
439
+ Make sure you have plenty of CPU memory available before you call this function. If you don't
440
+ have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
441
+ conveniently placed for you in the checkpoint folder.
442
+
443
+ A typical usage might be ::
444
+
445
+ from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
446
+ model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
447
+ # submit to model hub or save the model to share with others
448
+
449
+ Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
450
+ of the same application. i.e. you will need to re-initialize the deepspeed engine, since
451
+ ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
452
+
453
+ """
454
+ logger.info(f"Extracting fp32 weights")
455
+ state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
456
+
457
+ logger.info(f"Overwriting model with fp32 weights")
458
+ model = model.cpu()
459
+ model.load_state_dict(state_dict, strict=False)
460
+
461
+ return model
462
+
463
+
464
+ if __name__ == "__main__":
465
+
466
+ parser = argparse.ArgumentParser()
467
+ parser.add_argument(
468
+ "checkpoint_dir",
469
+ type=str,
470
+ help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
471
+ parser.add_argument(
472
+ "output_file",
473
+ type=str,
474
+ help=
475
+ "path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)"
476
+ )
477
+ parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
478
+ args = parser.parse_args()
479
+
480
+ debug = args.debug
481
+
482
+ convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file)