julep commited on
Commit
3cbc0ff
1 Parent(s): 1b218f9

feat: Add trained model

Browse files

Signed-off-by: Julep AI <developers@julep.ai>

config.json ADDED
@@ -0,0 +1,32 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "prakharz/DIAL-T0",
3
+ "architectures": [
4
+ "T5ForConditionalGeneration"
5
+ ],
6
+ "d_ff": 5120,
7
+ "d_kv": 64,
8
+ "d_model": 2048,
9
+ "decoder_start_token_id": 0,
10
+ "dense_act_fn": "gelu_new",
11
+ "dropout_rate": 0.1,
12
+ "eos_token_id": 1,
13
+ "feed_forward_proj": "gated-gelu",
14
+ "gradient_checkpointing": false,
15
+ "initializer_factor": 1.0,
16
+ "is_encoder_decoder": true,
17
+ "is_gated_act": true,
18
+ "layer_norm_epsilon": 1e-06,
19
+ "model_type": "t5",
20
+ "num_decoder_layers": 24,
21
+ "num_heads": 32,
22
+ "num_layers": 24,
23
+ "output_past": true,
24
+ "pad_token_id": 0,
25
+ "relative_attention_max_distance": 128,
26
+ "relative_attention_num_buckets": 32,
27
+ "tie_word_embeddings": false,
28
+ "torch_dtype": "bfloat16",
29
+ "transformers_version": "4.25.1",
30
+ "use_cache": false,
31
+ "vocab_size": 32100
32
+ }
global_step1600/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cb2bf10ab79a734abf48b4526e85617b23cb1c5b7111ba3ad379dbac908bad78
3
+ size 8548942587
global_step1600/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a5a8b042a08e277e9607ae79d179bc99de7f521b92e835ee4da73b2c46aeb054
3
+ size 8548943675
global_step1600/bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b94e350dee7dc475cd000bdcb40cb1defaa8bb627d8fd31733e726003f4393a5
3
+ size 8548944827
global_step1600/bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ce62491926e9ac6b1c1cc2b7dad2e51995e342c008162e81bff5549bc55054f7
3
+ size 8548943867
global_step1600/mp_rank_00_model_states.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4a5eefa82e897593644630ca4a35533b7d933cb8a29d64ca442390b9cb02df76
3
+ size 5699481819
latest ADDED
@@ -0,0 +1 @@
 
 
1
+ global_step1600
pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:19fb7c9a42c8115a0b2518631b5bbb17dea18f382fc7c1533c4c534df2229d15
3
+ size 5699366451
rng_state_0.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c41b81c8227fa28b27fd97d2ad7721930a51b65a4edde92cf2f4fab8c936ef99
3
+ size 14583
rng_state_1.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:191c7a58927c9f14ad9603e587b64634de332f567bbc5010a7f0d63bda38db1b
3
+ size 14583
rng_state_2.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:030bb0a6f5114351f55cfc1c6138a702b0c7aabf96b28f202d0656ade264bc2d
3
+ size 14583
rng_state_3.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:26bd79bd6efb5789c1f2bbb43bb094cf34184e47e809cc4029a946738f0927d9
3
+ size 14583
special_tokens_map.json ADDED
@@ -0,0 +1,107 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "additional_special_tokens": [
3
+ "<extra_id_0>",
4
+ "<extra_id_1>",
5
+ "<extra_id_2>",
6
+ "<extra_id_3>",
7
+ "<extra_id_4>",
8
+ "<extra_id_5>",
9
+ "<extra_id_6>",
10
+ "<extra_id_7>",
11
+ "<extra_id_8>",
12
+ "<extra_id_9>",
13
+ "<extra_id_10>",
14
+ "<extra_id_11>",
15
+ "<extra_id_12>",
16
+ "<extra_id_13>",
17
+ "<extra_id_14>",
18
+ "<extra_id_15>",
19
+ "<extra_id_16>",
20
+ "<extra_id_17>",
21
+ "<extra_id_18>",
22
+ "<extra_id_19>",
23
+ "<extra_id_20>",
24
+ "<extra_id_21>",
25
+ "<extra_id_22>",
26
+ "<extra_id_23>",
27
+ "<extra_id_24>",
28
+ "<extra_id_25>",
29
+ "<extra_id_26>",
30
+ "<extra_id_27>",
31
+ "<extra_id_28>",
32
+ "<extra_id_29>",
33
+ "<extra_id_30>",
34
+ "<extra_id_31>",
35
+ "<extra_id_32>",
36
+ "<extra_id_33>",
37
+ "<extra_id_34>",
38
+ "<extra_id_35>",
39
+ "<extra_id_36>",
40
+ "<extra_id_37>",
41
+ "<extra_id_38>",
42
+ "<extra_id_39>",
43
+ "<extra_id_40>",
44
+ "<extra_id_41>",
45
+ "<extra_id_42>",
46
+ "<extra_id_43>",
47
+ "<extra_id_44>",
48
+ "<extra_id_45>",
49
+ "<extra_id_46>",
50
+ "<extra_id_47>",
51
+ "<extra_id_48>",
52
+ "<extra_id_49>",
53
+ "<extra_id_50>",
54
+ "<extra_id_51>",
55
+ "<extra_id_52>",
56
+ "<extra_id_53>",
57
+ "<extra_id_54>",
58
+ "<extra_id_55>",
59
+ "<extra_id_56>",
60
+ "<extra_id_57>",
61
+ "<extra_id_58>",
62
+ "<extra_id_59>",
63
+ "<extra_id_60>",
64
+ "<extra_id_61>",
65
+ "<extra_id_62>",
66
+ "<extra_id_63>",
67
+ "<extra_id_64>",
68
+ "<extra_id_65>",
69
+ "<extra_id_66>",
70
+ "<extra_id_67>",
71
+ "<extra_id_68>",
72
+ "<extra_id_69>",
73
+ "<extra_id_70>",
74
+ "<extra_id_71>",
75
+ "<extra_id_72>",
76
+ "<extra_id_73>",
77
+ "<extra_id_74>",
78
+ "<extra_id_75>",
79
+ "<extra_id_76>",
80
+ "<extra_id_77>",
81
+ "<extra_id_78>",
82
+ "<extra_id_79>",
83
+ "<extra_id_80>",
84
+ "<extra_id_81>",
85
+ "<extra_id_82>",
86
+ "<extra_id_83>",
87
+ "<extra_id_84>",
88
+ "<extra_id_85>",
89
+ "<extra_id_86>",
90
+ "<extra_id_87>",
91
+ "<extra_id_88>",
92
+ "<extra_id_89>",
93
+ "<extra_id_90>",
94
+ "<extra_id_91>",
95
+ "<extra_id_92>",
96
+ "<extra_id_93>",
97
+ "<extra_id_94>",
98
+ "<extra_id_95>",
99
+ "<extra_id_96>",
100
+ "<extra_id_97>",
101
+ "<extra_id_98>",
102
+ "<extra_id_99>"
103
+ ],
104
+ "eos_token": "</s>",
105
+ "pad_token": "<pad>",
106
+ "unk_token": "<unk>"
107
+ }
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer_config.json ADDED
@@ -0,0 +1,113 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "additional_special_tokens": [
3
+ "<extra_id_0>",
4
+ "<extra_id_1>",
5
+ "<extra_id_2>",
6
+ "<extra_id_3>",
7
+ "<extra_id_4>",
8
+ "<extra_id_5>",
9
+ "<extra_id_6>",
10
+ "<extra_id_7>",
11
+ "<extra_id_8>",
12
+ "<extra_id_9>",
13
+ "<extra_id_10>",
14
+ "<extra_id_11>",
15
+ "<extra_id_12>",
16
+ "<extra_id_13>",
17
+ "<extra_id_14>",
18
+ "<extra_id_15>",
19
+ "<extra_id_16>",
20
+ "<extra_id_17>",
21
+ "<extra_id_18>",
22
+ "<extra_id_19>",
23
+ "<extra_id_20>",
24
+ "<extra_id_21>",
25
+ "<extra_id_22>",
26
+ "<extra_id_23>",
27
+ "<extra_id_24>",
28
+ "<extra_id_25>",
29
+ "<extra_id_26>",
30
+ "<extra_id_27>",
31
+ "<extra_id_28>",
32
+ "<extra_id_29>",
33
+ "<extra_id_30>",
34
+ "<extra_id_31>",
35
+ "<extra_id_32>",
36
+ "<extra_id_33>",
37
+ "<extra_id_34>",
38
+ "<extra_id_35>",
39
+ "<extra_id_36>",
40
+ "<extra_id_37>",
41
+ "<extra_id_38>",
42
+ "<extra_id_39>",
43
+ "<extra_id_40>",
44
+ "<extra_id_41>",
45
+ "<extra_id_42>",
46
+ "<extra_id_43>",
47
+ "<extra_id_44>",
48
+ "<extra_id_45>",
49
+ "<extra_id_46>",
50
+ "<extra_id_47>",
51
+ "<extra_id_48>",
52
+ "<extra_id_49>",
53
+ "<extra_id_50>",
54
+ "<extra_id_51>",
55
+ "<extra_id_52>",
56
+ "<extra_id_53>",
57
+ "<extra_id_54>",
58
+ "<extra_id_55>",
59
+ "<extra_id_56>",
60
+ "<extra_id_57>",
61
+ "<extra_id_58>",
62
+ "<extra_id_59>",
63
+ "<extra_id_60>",
64
+ "<extra_id_61>",
65
+ "<extra_id_62>",
66
+ "<extra_id_63>",
67
+ "<extra_id_64>",
68
+ "<extra_id_65>",
69
+ "<extra_id_66>",
70
+ "<extra_id_67>",
71
+ "<extra_id_68>",
72
+ "<extra_id_69>",
73
+ "<extra_id_70>",
74
+ "<extra_id_71>",
75
+ "<extra_id_72>",
76
+ "<extra_id_73>",
77
+ "<extra_id_74>",
78
+ "<extra_id_75>",
79
+ "<extra_id_76>",
80
+ "<extra_id_77>",
81
+ "<extra_id_78>",
82
+ "<extra_id_79>",
83
+ "<extra_id_80>",
84
+ "<extra_id_81>",
85
+ "<extra_id_82>",
86
+ "<extra_id_83>",
87
+ "<extra_id_84>",
88
+ "<extra_id_85>",
89
+ "<extra_id_86>",
90
+ "<extra_id_87>",
91
+ "<extra_id_88>",
92
+ "<extra_id_89>",
93
+ "<extra_id_90>",
94
+ "<extra_id_91>",
95
+ "<extra_id_92>",
96
+ "<extra_id_93>",
97
+ "<extra_id_94>",
98
+ "<extra_id_95>",
99
+ "<extra_id_96>",
100
+ "<extra_id_97>",
101
+ "<extra_id_98>",
102
+ "<extra_id_99>"
103
+ ],
104
+ "eos_token": "</s>",
105
+ "extra_ids": 100,
106
+ "model_max_length": 512,
107
+ "name_or_path": "prakharz/DIAL-T0",
108
+ "pad_token": "<pad>",
109
+ "sp_model_kwargs": {},
110
+ "special_tokens_map_file": "/home/ubuntu/.cache/huggingface/transformers/303fbee39a17e96552ac07e02b70ba62ff0ad760609687e3a1b92b4ad2dff58c.c94798918c92ded6aeef2d2f0e666d2cc4145eca1aa6e1336fde07f2e13e2f46",
111
+ "tokenizer_class": "T5Tokenizer",
112
+ "unk_token": "<unk>"
113
+ }
trainer_state.json ADDED
@@ -0,0 +1,444 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "best_metric": 85.9274,
3
+ "best_model_checkpoint": "./output/checkpoint-1600",
4
+ "epoch": 1.463719512195122,
5
+ "global_step": 1600,
6
+ "is_hyper_param_search": false,
7
+ "is_local_process_zero": true,
8
+ "is_world_process_zero": true,
9
+ "log_history": [
10
+ {
11
+ "epoch": 0.02,
12
+ "learning_rate": 4.9496336996337e-05,
13
+ "loss": 0.7035,
14
+ "step": 25
15
+ },
16
+ {
17
+ "epoch": 0.05,
18
+ "learning_rate": 4.8923992673992674e-05,
19
+ "loss": 0.3973,
20
+ "step": 50
21
+ },
22
+ {
23
+ "epoch": 0.07,
24
+ "learning_rate": 4.8351648351648355e-05,
25
+ "loss": 0.3092,
26
+ "step": 75
27
+ },
28
+ {
29
+ "epoch": 0.09,
30
+ "learning_rate": 4.7779304029304036e-05,
31
+ "loss": 0.2813,
32
+ "step": 100
33
+ },
34
+ {
35
+ "epoch": 0.11,
36
+ "learning_rate": 4.720695970695971e-05,
37
+ "loss": 0.2603,
38
+ "step": 125
39
+ },
40
+ {
41
+ "epoch": 0.14,
42
+ "learning_rate": 4.6634615384615384e-05,
43
+ "loss": 0.2371,
44
+ "step": 150
45
+ },
46
+ {
47
+ "epoch": 0.16,
48
+ "learning_rate": 4.606227106227106e-05,
49
+ "loss": 0.2334,
50
+ "step": 175
51
+ },
52
+ {
53
+ "epoch": 0.18,
54
+ "learning_rate": 4.548992673992674e-05,
55
+ "loss": 0.2296,
56
+ "step": 200
57
+ },
58
+ {
59
+ "epoch": 0.21,
60
+ "learning_rate": 4.491758241758242e-05,
61
+ "loss": 0.2219,
62
+ "step": 225
63
+ },
64
+ {
65
+ "epoch": 0.23,
66
+ "learning_rate": 4.4345238095238095e-05,
67
+ "loss": 0.2078,
68
+ "step": 250
69
+ },
70
+ {
71
+ "epoch": 0.25,
72
+ "learning_rate": 4.3772893772893776e-05,
73
+ "loss": 0.2123,
74
+ "step": 275
75
+ },
76
+ {
77
+ "epoch": 0.27,
78
+ "learning_rate": 4.320054945054946e-05,
79
+ "loss": 0.2063,
80
+ "step": 300
81
+ },
82
+ {
83
+ "epoch": 0.3,
84
+ "learning_rate": 4.262820512820513e-05,
85
+ "loss": 0.2002,
86
+ "step": 325
87
+ },
88
+ {
89
+ "epoch": 0.32,
90
+ "learning_rate": 4.2055860805860806e-05,
91
+ "loss": 0.2102,
92
+ "step": 350
93
+ },
94
+ {
95
+ "epoch": 0.34,
96
+ "learning_rate": 4.148351648351649e-05,
97
+ "loss": 0.2003,
98
+ "step": 375
99
+ },
100
+ {
101
+ "epoch": 0.37,
102
+ "learning_rate": 4.091117216117216e-05,
103
+ "loss": 0.2019,
104
+ "step": 400
105
+ },
106
+ {
107
+ "epoch": 0.39,
108
+ "learning_rate": 4.033882783882784e-05,
109
+ "loss": 0.1944,
110
+ "step": 425
111
+ },
112
+ {
113
+ "epoch": 0.41,
114
+ "learning_rate": 3.9766483516483516e-05,
115
+ "loss": 0.1935,
116
+ "step": 450
117
+ },
118
+ {
119
+ "epoch": 0.43,
120
+ "learning_rate": 3.91941391941392e-05,
121
+ "loss": 0.1974,
122
+ "step": 475
123
+ },
124
+ {
125
+ "epoch": 0.46,
126
+ "learning_rate": 3.862179487179488e-05,
127
+ "loss": 0.1922,
128
+ "step": 500
129
+ },
130
+ {
131
+ "epoch": 0.48,
132
+ "learning_rate": 3.804945054945055e-05,
133
+ "loss": 0.188,
134
+ "step": 525
135
+ },
136
+ {
137
+ "epoch": 0.5,
138
+ "learning_rate": 3.747710622710623e-05,
139
+ "loss": 0.1945,
140
+ "step": 550
141
+ },
142
+ {
143
+ "epoch": 0.53,
144
+ "learning_rate": 3.690476190476191e-05,
145
+ "loss": 0.1897,
146
+ "step": 575
147
+ },
148
+ {
149
+ "epoch": 0.55,
150
+ "learning_rate": 3.633241758241758e-05,
151
+ "loss": 0.1797,
152
+ "step": 600
153
+ },
154
+ {
155
+ "epoch": 0.57,
156
+ "learning_rate": 3.576007326007326e-05,
157
+ "loss": 0.1752,
158
+ "step": 625
159
+ },
160
+ {
161
+ "epoch": 0.59,
162
+ "learning_rate": 3.5187728937728944e-05,
163
+ "loss": 0.1797,
164
+ "step": 650
165
+ },
166
+ {
167
+ "epoch": 0.62,
168
+ "learning_rate": 3.461538461538462e-05,
169
+ "loss": 0.1785,
170
+ "step": 675
171
+ },
172
+ {
173
+ "epoch": 0.64,
174
+ "learning_rate": 3.404304029304029e-05,
175
+ "loss": 0.185,
176
+ "step": 700
177
+ },
178
+ {
179
+ "epoch": 0.66,
180
+ "learning_rate": 3.347069597069597e-05,
181
+ "loss": 0.1751,
182
+ "step": 725
183
+ },
184
+ {
185
+ "epoch": 0.69,
186
+ "learning_rate": 3.289835164835165e-05,
187
+ "loss": 0.1799,
188
+ "step": 750
189
+ },
190
+ {
191
+ "epoch": 0.71,
192
+ "learning_rate": 3.232600732600733e-05,
193
+ "loss": 0.1801,
194
+ "step": 775
195
+ },
196
+ {
197
+ "epoch": 0.73,
198
+ "learning_rate": 3.1753663003663003e-05,
199
+ "loss": 0.1748,
200
+ "step": 800
201
+ },
202
+ {
203
+ "epoch": 0.73,
204
+ "eval_Bleu_1": 0.8552,
205
+ "eval_Bleu_2": 0.0009,
206
+ "eval_Bleu_3": 0.0001,
207
+ "eval_Bleu_4": 0.0,
208
+ "eval_ROUGE_L": 0.8552,
209
+ "eval_accuracy": 85.52,
210
+ "eval_f1": 85.52,
211
+ "eval_gen_len": 2.2084,
212
+ "eval_loss": 0.17816422879695892,
213
+ "eval_precision": 85.52,
214
+ "eval_recall": 85.52,
215
+ "eval_rouge1": 85.6587,
216
+ "eval_rouge2": 3.7183,
217
+ "eval_rougeL": 85.6528,
218
+ "eval_rougeLsum": 85.6348,
219
+ "eval_runtime": 1004.3102,
220
+ "eval_samples_per_second": 34.705,
221
+ "eval_steps_per_second": 4.338,
222
+ "step": 800
223
+ },
224
+ {
225
+ "epoch": 0.75,
226
+ "learning_rate": 3.1181318681318685e-05,
227
+ "loss": 0.1698,
228
+ "step": 825
229
+ },
230
+ {
231
+ "epoch": 0.78,
232
+ "learning_rate": 3.0608974358974366e-05,
233
+ "loss": 0.1763,
234
+ "step": 850
235
+ },
236
+ {
237
+ "epoch": 0.8,
238
+ "learning_rate": 3.0036630036630036e-05,
239
+ "loss": 0.1655,
240
+ "step": 875
241
+ },
242
+ {
243
+ "epoch": 0.82,
244
+ "learning_rate": 2.9464285714285718e-05,
245
+ "loss": 0.1703,
246
+ "step": 900
247
+ },
248
+ {
249
+ "epoch": 0.85,
250
+ "learning_rate": 2.8891941391941392e-05,
251
+ "loss": 0.174,
252
+ "step": 925
253
+ },
254
+ {
255
+ "epoch": 0.87,
256
+ "learning_rate": 2.831959706959707e-05,
257
+ "loss": 0.1723,
258
+ "step": 950
259
+ },
260
+ {
261
+ "epoch": 0.89,
262
+ "learning_rate": 2.774725274725275e-05,
263
+ "loss": 0.1739,
264
+ "step": 975
265
+ },
266
+ {
267
+ "epoch": 0.91,
268
+ "learning_rate": 2.7174908424908425e-05,
269
+ "loss": 0.1673,
270
+ "step": 1000
271
+ },
272
+ {
273
+ "epoch": 0.94,
274
+ "learning_rate": 2.6602564102564102e-05,
275
+ "loss": 0.1647,
276
+ "step": 1025
277
+ },
278
+ {
279
+ "epoch": 0.96,
280
+ "learning_rate": 2.6030219780219783e-05,
281
+ "loss": 0.1739,
282
+ "step": 1050
283
+ },
284
+ {
285
+ "epoch": 0.98,
286
+ "learning_rate": 2.5457875457875458e-05,
287
+ "loss": 0.1703,
288
+ "step": 1075
289
+ },
290
+ {
291
+ "epoch": 1.01,
292
+ "learning_rate": 2.488553113553114e-05,
293
+ "loss": 0.1727,
294
+ "step": 1100
295
+ },
296
+ {
297
+ "epoch": 1.03,
298
+ "learning_rate": 2.4313186813186813e-05,
299
+ "loss": 0.1603,
300
+ "step": 1125
301
+ },
302
+ {
303
+ "epoch": 1.05,
304
+ "learning_rate": 2.374084249084249e-05,
305
+ "loss": 0.1511,
306
+ "step": 1150
307
+ },
308
+ {
309
+ "epoch": 1.07,
310
+ "learning_rate": 2.3168498168498172e-05,
311
+ "loss": 0.1544,
312
+ "step": 1175
313
+ },
314
+ {
315
+ "epoch": 1.1,
316
+ "learning_rate": 2.2596153846153846e-05,
317
+ "loss": 0.1517,
318
+ "step": 1200
319
+ },
320
+ {
321
+ "epoch": 1.12,
322
+ "learning_rate": 2.2023809523809524e-05,
323
+ "loss": 0.1556,
324
+ "step": 1225
325
+ },
326
+ {
327
+ "epoch": 1.14,
328
+ "learning_rate": 2.14514652014652e-05,
329
+ "loss": 0.1688,
330
+ "step": 1250
331
+ },
332
+ {
333
+ "epoch": 1.17,
334
+ "learning_rate": 2.0879120879120882e-05,
335
+ "loss": 0.1658,
336
+ "step": 1275
337
+ },
338
+ {
339
+ "epoch": 1.19,
340
+ "learning_rate": 2.0306776556776557e-05,
341
+ "loss": 0.1587,
342
+ "step": 1300
343
+ },
344
+ {
345
+ "epoch": 1.21,
346
+ "learning_rate": 1.9734432234432234e-05,
347
+ "loss": 0.1664,
348
+ "step": 1325
349
+ },
350
+ {
351
+ "epoch": 1.24,
352
+ "learning_rate": 1.9162087912087915e-05,
353
+ "loss": 0.1558,
354
+ "step": 1350
355
+ },
356
+ {
357
+ "epoch": 1.26,
358
+ "learning_rate": 1.858974358974359e-05,
359
+ "loss": 0.1591,
360
+ "step": 1375
361
+ },
362
+ {
363
+ "epoch": 1.28,
364
+ "learning_rate": 1.8017399267399267e-05,
365
+ "loss": 0.1558,
366
+ "step": 1400
367
+ },
368
+ {
369
+ "epoch": 1.3,
370
+ "learning_rate": 1.7445054945054945e-05,
371
+ "loss": 0.1583,
372
+ "step": 1425
373
+ },
374
+ {
375
+ "epoch": 1.33,
376
+ "learning_rate": 1.6872710622710626e-05,
377
+ "loss": 0.155,
378
+ "step": 1450
379
+ },
380
+ {
381
+ "epoch": 1.35,
382
+ "learning_rate": 1.63003663003663e-05,
383
+ "loss": 0.1644,
384
+ "step": 1475
385
+ },
386
+ {
387
+ "epoch": 1.37,
388
+ "learning_rate": 1.5728021978021978e-05,
389
+ "loss": 0.1574,
390
+ "step": 1500
391
+ },
392
+ {
393
+ "epoch": 1.4,
394
+ "learning_rate": 1.5155677655677656e-05,
395
+ "loss": 0.1555,
396
+ "step": 1525
397
+ },
398
+ {
399
+ "epoch": 1.42,
400
+ "learning_rate": 1.4583333333333335e-05,
401
+ "loss": 0.1542,
402
+ "step": 1550
403
+ },
404
+ {
405
+ "epoch": 1.44,
406
+ "learning_rate": 1.4010989010989013e-05,
407
+ "loss": 0.1537,
408
+ "step": 1575
409
+ },
410
+ {
411
+ "epoch": 1.46,
412
+ "learning_rate": 1.3438644688644688e-05,
413
+ "loss": 0.1493,
414
+ "step": 1600
415
+ },
416
+ {
417
+ "epoch": 1.46,
418
+ "eval_Bleu_1": 0.8593,
419
+ "eval_Bleu_2": 0.0009,
420
+ "eval_Bleu_3": 0.0001,
421
+ "eval_Bleu_4": 0.0,
422
+ "eval_ROUGE_L": 0.8593,
423
+ "eval_accuracy": 85.9274,
424
+ "eval_f1": 85.9274,
425
+ "eval_gen_len": 2.2038,
426
+ "eval_loss": 0.1681133359670639,
427
+ "eval_precision": 85.9274,
428
+ "eval_recall": 85.9274,
429
+ "eval_rouge1": 86.0601,
430
+ "eval_rouge2": 3.5662,
431
+ "eval_rougeL": 86.0539,
432
+ "eval_rougeLsum": 86.0496,
433
+ "eval_runtime": 1001.4875,
434
+ "eval_samples_per_second": 34.803,
435
+ "eval_steps_per_second": 4.351,
436
+ "step": 1600
437
+ }
438
+ ],
439
+ "max_steps": 2186,
440
+ "num_train_epochs": 2,
441
+ "total_flos": 2.7769320388994007e+18,
442
+ "trial_name": null,
443
+ "trial_params": null
444
+ }
training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2850b9e779255d3efce03dff445e52ea503bbb829fd1b72d24779361b420c4a9
3
+ size 4603
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)