added_tokens.json ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ {
2
+ "<CLS>": 41070,
3
+ "<EOD>": 41072,
4
+ "<MASK>": 41073,
5
+ "<PAD>": 41074,
6
+ "<SEP>": 41071
7
+ }
config.json ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "/project/lt900048-ai24tn/models/openthaigpt/openthaigpt-1.0.0-7b-chat",
3
+ "architectures": [
4
+ "LlamaForCausalLM"
5
+ ],
6
+ "attention_bias": false,
7
+ "attention_dropout": 0.0,
8
+ "bos_token_id": 1,
9
+ "eos_token_id": 2,
10
+ "hidden_act": "silu",
11
+ "hidden_size": 4096,
12
+ "initializer_range": 0.02,
13
+ "intermediate_size": 11008,
14
+ "max_position_embeddings": 4096,
15
+ "model_type": "llama",
16
+ "num_attention_heads": 32,
17
+ "num_hidden_layers": 32,
18
+ "num_key_value_heads": 32,
19
+ "pretraining_tp": 1,
20
+ "rms_norm_eps": 1e-05,
21
+ "rope_scaling": null,
22
+ "rope_theta": 10000.0,
23
+ "tie_word_embeddings": false,
24
+ "torch_dtype": "bfloat16",
25
+ "transformers_version": "4.40.1",
26
+ "use_cache": false,
27
+ "vocab_size": 41088
28
+ }
latest ADDED
@@ -0,0 +1 @@
 
 
1
+ global_step285
special_tokens_map.json ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "additional_special_tokens": [
3
+ "<unk>",
4
+ "<s>",
5
+ "</s>"
6
+ ],
7
+ "bos_token": {
8
+ "content": "<s>",
9
+ "lstrip": false,
10
+ "normalized": true,
11
+ "rstrip": false,
12
+ "single_word": false
13
+ },
14
+ "eos_token": {
15
+ "content": "</s>",
16
+ "lstrip": false,
17
+ "normalized": true,
18
+ "rstrip": false,
19
+ "single_word": false
20
+ },
21
+ "pad_token": "</s>",
22
+ "unk_token": {
23
+ "content": "<unk>",
24
+ "lstrip": false,
25
+ "normalized": true,
26
+ "rstrip": false,
27
+ "single_word": false
28
+ }
29
+ }
tokenizer.model ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:02df43dcae8c7b5b122d45f642e42c96577cdd09fd949c6996051886c72ab002
3
+ size 717508
tokenizer_config.json ADDED
@@ -0,0 +1,88 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_bos_token": true,
3
+ "add_eos_token": false,
4
+ "add_prefix_space": true,
5
+ "added_tokens_decoder": {
6
+ "0": {
7
+ "content": "<unk>",
8
+ "lstrip": false,
9
+ "normalized": true,
10
+ "rstrip": false,
11
+ "single_word": false,
12
+ "special": true
13
+ },
14
+ "1": {
15
+ "content": "<s>",
16
+ "lstrip": false,
17
+ "normalized": true,
18
+ "rstrip": false,
19
+ "single_word": false,
20
+ "special": true
21
+ },
22
+ "2": {
23
+ "content": "</s>",
24
+ "lstrip": false,
25
+ "normalized": true,
26
+ "rstrip": false,
27
+ "single_word": false,
28
+ "special": true
29
+ },
30
+ "41070": {
31
+ "content": "<CLS>",
32
+ "lstrip": false,
33
+ "normalized": false,
34
+ "rstrip": false,
35
+ "single_word": false,
36
+ "special": true
37
+ },
38
+ "41071": {
39
+ "content": "<SEP>",
40
+ "lstrip": false,
41
+ "normalized": false,
42
+ "rstrip": false,
43
+ "single_word": false,
44
+ "special": true
45
+ },
46
+ "41072": {
47
+ "content": "<EOD>",
48
+ "lstrip": false,
49
+ "normalized": false,
50
+ "rstrip": false,
51
+ "single_word": false,
52
+ "special": true
53
+ },
54
+ "41073": {
55
+ "content": "<MASK>",
56
+ "lstrip": false,
57
+ "normalized": false,
58
+ "rstrip": false,
59
+ "single_word": false,
60
+ "special": true
61
+ },
62
+ "41074": {
63
+ "content": "<PAD>",
64
+ "lstrip": false,
65
+ "normalized": false,
66
+ "rstrip": false,
67
+ "single_word": false,
68
+ "special": true
69
+ }
70
+ },
71
+ "additional_special_tokens": [
72
+ "<unk>",
73
+ "<s>",
74
+ "</s>"
75
+ ],
76
+ "bos_token": "<s>",
77
+ "clean_up_tokenization_spaces": false,
78
+ "eos_token": "</s>",
79
+ "legacy": true,
80
+ "model_max_length": 2048,
81
+ "pad_token": "</s>",
82
+ "padding_side": "right",
83
+ "sp_model_kwargs": {},
84
+ "spaces_between_special_tokens": false,
85
+ "tokenizer_class": "LlamaTokenizer",
86
+ "unk_token": "<unk>",
87
+ "use_default_system_prompt": true
88
+ }
trainer_state.json ADDED
@@ -0,0 +1,2025 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "best_metric": null,
3
+ "best_model_checkpoint": null,
4
+ "epoch": 1.0,
5
+ "eval_steps": 500,
6
+ "global_step": 285,
7
+ "is_hyper_param_search": false,
8
+ "is_local_process_zero": true,
9
+ "is_world_process_zero": true,
10
+ "log_history": [
11
+ {
12
+ "epoch": 0.0035087719298245615,
13
+ "grad_norm": 18.65219621501764,
14
+ "learning_rate": 0.0,
15
+ "loss": 1.9653,
16
+ "step": 1
17
+ },
18
+ {
19
+ "epoch": 0.007017543859649123,
20
+ "grad_norm": 5.5068522642682325,
21
+ "learning_rate": 2.5237190142858296e-05,
22
+ "loss": 1.4912,
23
+ "step": 2
24
+ },
25
+ {
26
+ "epoch": 0.010526315789473684,
27
+ "grad_norm": 5.300187955783763,
28
+ "learning_rate": 4e-05,
29
+ "loss": 1.488,
30
+ "step": 3
31
+ },
32
+ {
33
+ "epoch": 0.014035087719298246,
34
+ "grad_norm": 3.607547540937298,
35
+ "learning_rate": 5.047438028571659e-05,
36
+ "loss": 1.3376,
37
+ "step": 4
38
+ },
39
+ {
40
+ "epoch": 0.017543859649122806,
41
+ "grad_norm": 2.8856871756952254,
42
+ "learning_rate": 5.859894082871708e-05,
43
+ "loss": 1.2775,
44
+ "step": 5
45
+ },
46
+ {
47
+ "epoch": 0.021052631578947368,
48
+ "grad_norm": 2.1943773684803283,
49
+ "learning_rate": 6.52371901428583e-05,
50
+ "loss": 1.1621,
51
+ "step": 6
52
+ },
53
+ {
54
+ "epoch": 0.02456140350877193,
55
+ "grad_norm": 1.6148482128986932,
56
+ "learning_rate": 7.08497499664569e-05,
57
+ "loss": 1.1393,
58
+ "step": 7
59
+ },
60
+ {
61
+ "epoch": 0.028070175438596492,
62
+ "grad_norm": 1.4378600160323183,
63
+ "learning_rate": 7.571157042857488e-05,
64
+ "loss": 1.1342,
65
+ "step": 8
66
+ },
67
+ {
68
+ "epoch": 0.031578947368421054,
69
+ "grad_norm": 1.3979572631886956,
70
+ "learning_rate": 8e-05,
71
+ "loss": 1.1196,
72
+ "step": 9
73
+ },
74
+ {
75
+ "epoch": 0.03508771929824561,
76
+ "grad_norm": 1.3666278253029815,
77
+ "learning_rate": 8e-05,
78
+ "loss": 1.1479,
79
+ "step": 10
80
+ },
81
+ {
82
+ "epoch": 0.03859649122807018,
83
+ "grad_norm": 1.2483549253543214,
84
+ "learning_rate": 7.971014492753623e-05,
85
+ "loss": 1.1125,
86
+ "step": 11
87
+ },
88
+ {
89
+ "epoch": 0.042105263157894736,
90
+ "grad_norm": 1.347256910957925,
91
+ "learning_rate": 7.942028985507246e-05,
92
+ "loss": 1.1588,
93
+ "step": 12
94
+ },
95
+ {
96
+ "epoch": 0.0456140350877193,
97
+ "grad_norm": 1.3748876632936469,
98
+ "learning_rate": 7.91304347826087e-05,
99
+ "loss": 1.0846,
100
+ "step": 13
101
+ },
102
+ {
103
+ "epoch": 0.04912280701754386,
104
+ "grad_norm": 1.329787006205099,
105
+ "learning_rate": 7.884057971014494e-05,
106
+ "loss": 1.1839,
107
+ "step": 14
108
+ },
109
+ {
110
+ "epoch": 0.05263157894736842,
111
+ "grad_norm": 1.1855019705035863,
112
+ "learning_rate": 7.855072463768117e-05,
113
+ "loss": 1.1546,
114
+ "step": 15
115
+ },
116
+ {
117
+ "epoch": 0.056140350877192984,
118
+ "grad_norm": 1.2020228064951217,
119
+ "learning_rate": 7.82608695652174e-05,
120
+ "loss": 1.0999,
121
+ "step": 16
122
+ },
123
+ {
124
+ "epoch": 0.05964912280701754,
125
+ "grad_norm": 1.2879067395522685,
126
+ "learning_rate": 7.797101449275363e-05,
127
+ "loss": 1.1466,
128
+ "step": 17
129
+ },
130
+ {
131
+ "epoch": 0.06315789473684211,
132
+ "grad_norm": 1.256590992120541,
133
+ "learning_rate": 7.768115942028987e-05,
134
+ "loss": 1.1593,
135
+ "step": 18
136
+ },
137
+ {
138
+ "epoch": 0.06666666666666667,
139
+ "grad_norm": 1.1654027316591642,
140
+ "learning_rate": 7.73913043478261e-05,
141
+ "loss": 1.1806,
142
+ "step": 19
143
+ },
144
+ {
145
+ "epoch": 0.07017543859649122,
146
+ "grad_norm": 1.2007307897010868,
147
+ "learning_rate": 7.710144927536232e-05,
148
+ "loss": 1.1416,
149
+ "step": 20
150
+ },
151
+ {
152
+ "epoch": 0.07368421052631578,
153
+ "grad_norm": 1.1499545393774593,
154
+ "learning_rate": 7.681159420289856e-05,
155
+ "loss": 1.1203,
156
+ "step": 21
157
+ },
158
+ {
159
+ "epoch": 0.07719298245614035,
160
+ "grad_norm": 1.0810536737354983,
161
+ "learning_rate": 7.652173913043479e-05,
162
+ "loss": 1.0875,
163
+ "step": 22
164
+ },
165
+ {
166
+ "epoch": 0.08070175438596491,
167
+ "grad_norm": 1.249217128847186,
168
+ "learning_rate": 7.623188405797102e-05,
169
+ "loss": 1.1878,
170
+ "step": 23
171
+ },
172
+ {
173
+ "epoch": 0.08421052631578947,
174
+ "grad_norm": 1.171401194067982,
175
+ "learning_rate": 7.594202898550726e-05,
176
+ "loss": 1.1416,
177
+ "step": 24
178
+ },
179
+ {
180
+ "epoch": 0.08771929824561403,
181
+ "grad_norm": 1.1864422136018933,
182
+ "learning_rate": 7.565217391304349e-05,
183
+ "loss": 1.1559,
184
+ "step": 25
185
+ },
186
+ {
187
+ "epoch": 0.0912280701754386,
188
+ "grad_norm": 1.1084100899911409,
189
+ "learning_rate": 7.536231884057971e-05,
190
+ "loss": 1.127,
191
+ "step": 26
192
+ },
193
+ {
194
+ "epoch": 0.09473684210526316,
195
+ "grad_norm": 1.0848431056969332,
196
+ "learning_rate": 7.507246376811594e-05,
197
+ "loss": 1.1621,
198
+ "step": 27
199
+ },
200
+ {
201
+ "epoch": 0.09824561403508772,
202
+ "grad_norm": 1.160465065160364,
203
+ "learning_rate": 7.478260869565218e-05,
204
+ "loss": 1.1086,
205
+ "step": 28
206
+ },
207
+ {
208
+ "epoch": 0.10175438596491228,
209
+ "grad_norm": 1.0888824123502896,
210
+ "learning_rate": 7.449275362318841e-05,
211
+ "loss": 1.1402,
212
+ "step": 29
213
+ },
214
+ {
215
+ "epoch": 0.10526315789473684,
216
+ "grad_norm": 1.136530946229479,
217
+ "learning_rate": 7.420289855072465e-05,
218
+ "loss": 1.1533,
219
+ "step": 30
220
+ },
221
+ {
222
+ "epoch": 0.10877192982456141,
223
+ "grad_norm": 1.1193975180437097,
224
+ "learning_rate": 7.391304347826088e-05,
225
+ "loss": 1.1194,
226
+ "step": 31
227
+ },
228
+ {
229
+ "epoch": 0.11228070175438597,
230
+ "grad_norm": 1.032391835314362,
231
+ "learning_rate": 7.36231884057971e-05,
232
+ "loss": 1.086,
233
+ "step": 32
234
+ },
235
+ {
236
+ "epoch": 0.11578947368421053,
237
+ "grad_norm": 1.1734719012857107,
238
+ "learning_rate": 7.333333333333333e-05,
239
+ "loss": 1.1515,
240
+ "step": 33
241
+ },
242
+ {
243
+ "epoch": 0.11929824561403508,
244
+ "grad_norm": 1.1223429839275594,
245
+ "learning_rate": 7.304347826086957e-05,
246
+ "loss": 1.1067,
247
+ "step": 34
248
+ },
249
+ {
250
+ "epoch": 0.12280701754385964,
251
+ "grad_norm": 1.0919749212433656,
252
+ "learning_rate": 7.27536231884058e-05,
253
+ "loss": 1.1298,
254
+ "step": 35
255
+ },
256
+ {
257
+ "epoch": 0.12631578947368421,
258
+ "grad_norm": 1.061313889855867,
259
+ "learning_rate": 7.246376811594204e-05,
260
+ "loss": 1.1323,
261
+ "step": 36
262
+ },
263
+ {
264
+ "epoch": 0.12982456140350876,
265
+ "grad_norm": 0.9886502393238654,
266
+ "learning_rate": 7.217391304347827e-05,
267
+ "loss": 1.1358,
268
+ "step": 37
269
+ },
270
+ {
271
+ "epoch": 0.13333333333333333,
272
+ "grad_norm": 1.090149070089068,
273
+ "learning_rate": 7.18840579710145e-05,
274
+ "loss": 1.1083,
275
+ "step": 38
276
+ },
277
+ {
278
+ "epoch": 0.1368421052631579,
279
+ "grad_norm": 1.2141824886927202,
280
+ "learning_rate": 7.159420289855072e-05,
281
+ "loss": 1.1659,
282
+ "step": 39
283
+ },
284
+ {
285
+ "epoch": 0.14035087719298245,
286
+ "grad_norm": 1.2804981481033841,
287
+ "learning_rate": 7.130434782608696e-05,
288
+ "loss": 1.1769,
289
+ "step": 40
290
+ },
291
+ {
292
+ "epoch": 0.14385964912280702,
293
+ "grad_norm": 1.085391641580103,
294
+ "learning_rate": 7.101449275362319e-05,
295
+ "loss": 1.1307,
296
+ "step": 41
297
+ },
298
+ {
299
+ "epoch": 0.14736842105263157,
300
+ "grad_norm": 1.0432511632089205,
301
+ "learning_rate": 7.072463768115943e-05,
302
+ "loss": 1.1152,
303
+ "step": 42
304
+ },
305
+ {
306
+ "epoch": 0.15087719298245614,
307
+ "grad_norm": 1.0805636921488617,
308
+ "learning_rate": 7.043478260869566e-05,
309
+ "loss": 1.1131,
310
+ "step": 43
311
+ },
312
+ {
313
+ "epoch": 0.1543859649122807,
314
+ "grad_norm": 1.0831222641710494,
315
+ "learning_rate": 7.014492753623189e-05,
316
+ "loss": 1.1585,
317
+ "step": 44
318
+ },
319
+ {
320
+ "epoch": 0.15789473684210525,
321
+ "grad_norm": 0.9919690920164701,
322
+ "learning_rate": 6.985507246376811e-05,
323
+ "loss": 1.1235,
324
+ "step": 45
325
+ },
326
+ {
327
+ "epoch": 0.16140350877192983,
328
+ "grad_norm": 0.9498489248223517,
329
+ "learning_rate": 6.956521739130436e-05,
330
+ "loss": 1.0856,
331
+ "step": 46
332
+ },
333
+ {
334
+ "epoch": 0.1649122807017544,
335
+ "grad_norm": 1.0276715714307674,
336
+ "learning_rate": 6.927536231884058e-05,
337
+ "loss": 1.1378,
338
+ "step": 47
339
+ },
340
+ {
341
+ "epoch": 0.16842105263157894,
342
+ "grad_norm": 1.094458012231917,
343
+ "learning_rate": 6.898550724637682e-05,
344
+ "loss": 1.117,
345
+ "step": 48
346
+ },
347
+ {
348
+ "epoch": 0.17192982456140352,
349
+ "grad_norm": 1.089782170199694,
350
+ "learning_rate": 6.869565217391305e-05,
351
+ "loss": 1.1471,
352
+ "step": 49
353
+ },
354
+ {
355
+ "epoch": 0.17543859649122806,
356
+ "grad_norm": 1.0908602591310133,
357
+ "learning_rate": 6.840579710144928e-05,
358
+ "loss": 1.1711,
359
+ "step": 50
360
+ },
361
+ {
362
+ "epoch": 0.17894736842105263,
363
+ "grad_norm": 1.126099914721245,
364
+ "learning_rate": 6.811594202898552e-05,
365
+ "loss": 1.1437,
366
+ "step": 51
367
+ },
368
+ {
369
+ "epoch": 0.1824561403508772,
370
+ "grad_norm": 1.1004469394026717,
371
+ "learning_rate": 6.782608695652175e-05,
372
+ "loss": 1.1502,
373
+ "step": 52
374
+ },
375
+ {
376
+ "epoch": 0.18596491228070175,
377
+ "grad_norm": 1.0690976698679064,
378
+ "learning_rate": 6.753623188405797e-05,
379
+ "loss": 1.1436,
380
+ "step": 53
381
+ },
382
+ {
383
+ "epoch": 0.18947368421052632,
384
+ "grad_norm": 1.0893372270742567,
385
+ "learning_rate": 6.72463768115942e-05,
386
+ "loss": 1.1028,
387
+ "step": 54
388
+ },
389
+ {
390
+ "epoch": 0.19298245614035087,
391
+ "grad_norm": 1.038406028318434,
392
+ "learning_rate": 6.695652173913044e-05,
393
+ "loss": 1.1601,
394
+ "step": 55
395
+ },
396
+ {
397
+ "epoch": 0.19649122807017544,
398
+ "grad_norm": 0.9960149874705063,
399
+ "learning_rate": 6.666666666666667e-05,
400
+ "loss": 1.102,
401
+ "step": 56
402
+ },
403
+ {
404
+ "epoch": 0.2,
405
+ "grad_norm": 1.073973228867499,
406
+ "learning_rate": 6.637681159420291e-05,
407
+ "loss": 1.1213,
408
+ "step": 57
409
+ },
410
+ {
411
+ "epoch": 0.20350877192982456,
412
+ "grad_norm": 1.1952354445407958,
413
+ "learning_rate": 6.608695652173914e-05,
414
+ "loss": 1.1046,
415
+ "step": 58
416
+ },
417
+ {
418
+ "epoch": 0.20701754385964913,
419
+ "grad_norm": 1.0175331854073473,
420
+ "learning_rate": 6.579710144927537e-05,
421
+ "loss": 1.1192,
422
+ "step": 59
423
+ },
424
+ {
425
+ "epoch": 0.21052631578947367,
426
+ "grad_norm": 1.053567224648277,
427
+ "learning_rate": 6.550724637681159e-05,
428
+ "loss": 1.1514,
429
+ "step": 60
430
+ },
431
+ {
432
+ "epoch": 0.21403508771929824,
433
+ "grad_norm": 1.1549538449027954,
434
+ "learning_rate": 6.521739130434783e-05,
435
+ "loss": 1.161,
436
+ "step": 61
437
+ },
438
+ {
439
+ "epoch": 0.21754385964912282,
440
+ "grad_norm": 0.9980361090477836,
441
+ "learning_rate": 6.492753623188406e-05,
442
+ "loss": 1.0715,
443
+ "step": 62
444
+ },
445
+ {
446
+ "epoch": 0.22105263157894736,
447
+ "grad_norm": 1.0664047714210805,
448
+ "learning_rate": 6.46376811594203e-05,
449
+ "loss": 1.172,
450
+ "step": 63
451
+ },
452
+ {
453
+ "epoch": 0.22456140350877193,
454
+ "grad_norm": 1.0338684936573976,
455
+ "learning_rate": 6.434782608695653e-05,
456
+ "loss": 1.0547,
457
+ "step": 64
458
+ },
459
+ {
460
+ "epoch": 0.22807017543859648,
461
+ "grad_norm": 1.0164199027592216,
462
+ "learning_rate": 6.405797101449276e-05,
463
+ "loss": 1.0821,
464
+ "step": 65
465
+ },
466
+ {
467
+ "epoch": 0.23157894736842105,
468
+ "grad_norm": 1.08287922650492,
469
+ "learning_rate": 6.376811594202898e-05,
470
+ "loss": 1.112,
471
+ "step": 66
472
+ },
473
+ {
474
+ "epoch": 0.23508771929824562,
475
+ "grad_norm": 1.012260870534961,
476
+ "learning_rate": 6.347826086956523e-05,
477
+ "loss": 1.1339,
478
+ "step": 67
479
+ },
480
+ {
481
+ "epoch": 0.23859649122807017,
482
+ "grad_norm": 0.9284957156873009,
483
+ "learning_rate": 6.318840579710145e-05,
484
+ "loss": 1.0655,
485
+ "step": 68
486
+ },
487
+ {
488
+ "epoch": 0.24210526315789474,
489
+ "grad_norm": 0.9699566883456087,
490
+ "learning_rate": 6.28985507246377e-05,
491
+ "loss": 1.0447,
492
+ "step": 69
493
+ },
494
+ {
495
+ "epoch": 0.24561403508771928,
496
+ "grad_norm": 0.9953128820241032,
497
+ "learning_rate": 6.260869565217392e-05,
498
+ "loss": 1.0979,
499
+ "step": 70
500
+ },
501
+ {
502
+ "epoch": 0.24912280701754386,
503
+ "grad_norm": 1.2255506094662556,
504
+ "learning_rate": 6.231884057971015e-05,
505
+ "loss": 1.1266,
506
+ "step": 71
507
+ },
508
+ {
509
+ "epoch": 0.25263157894736843,
510
+ "grad_norm": 1.1053469341884643,
511
+ "learning_rate": 6.202898550724638e-05,
512
+ "loss": 1.1083,
513
+ "step": 72
514
+ },
515
+ {
516
+ "epoch": 0.256140350877193,
517
+ "grad_norm": 1.0471395283209621,
518
+ "learning_rate": 6.173913043478262e-05,
519
+ "loss": 1.1081,
520
+ "step": 73
521
+ },
522
+ {
523
+ "epoch": 0.2596491228070175,
524
+ "grad_norm": 1.1372933010569457,
525
+ "learning_rate": 6.144927536231884e-05,
526
+ "loss": 1.1191,
527
+ "step": 74
528
+ },
529
+ {
530
+ "epoch": 0.2631578947368421,
531
+ "grad_norm": 1.094007360015913,
532
+ "learning_rate": 6.115942028985507e-05,
533
+ "loss": 1.1498,
534
+ "step": 75
535
+ },
536
+ {
537
+ "epoch": 0.26666666666666666,
538
+ "grad_norm": 1.037731837007179,
539
+ "learning_rate": 6.086956521739131e-05,
540
+ "loss": 1.078,
541
+ "step": 76
542
+ },
543
+ {
544
+ "epoch": 0.27017543859649124,
545
+ "grad_norm": 0.9878690341929469,
546
+ "learning_rate": 6.057971014492755e-05,
547
+ "loss": 1.0925,
548
+ "step": 77
549
+ },
550
+ {
551
+ "epoch": 0.2736842105263158,
552
+ "grad_norm": 1.0251650017115257,
553
+ "learning_rate": 6.0289855072463774e-05,
554
+ "loss": 1.0992,
555
+ "step": 78
556
+ },
557
+ {
558
+ "epoch": 0.2771929824561403,
559
+ "grad_norm": 0.9677079627985357,
560
+ "learning_rate": 6.000000000000001e-05,
561
+ "loss": 1.0228,
562
+ "step": 79
563
+ },
564
+ {
565
+ "epoch": 0.2807017543859649,
566
+ "grad_norm": 1.03084630118753,
567
+ "learning_rate": 5.9710144927536236e-05,
568
+ "loss": 1.1026,
569
+ "step": 80
570
+ },
571
+ {
572
+ "epoch": 0.28421052631578947,
573
+ "grad_norm": 1.149025536043054,
574
+ "learning_rate": 5.942028985507246e-05,
575
+ "loss": 1.0971,
576
+ "step": 81
577
+ },
578
+ {
579
+ "epoch": 0.28771929824561404,
580
+ "grad_norm": 0.8977367483124615,
581
+ "learning_rate": 5.91304347826087e-05,
582
+ "loss": 1.0665,
583
+ "step": 82
584
+ },
585
+ {
586
+ "epoch": 0.2912280701754386,
587
+ "grad_norm": 0.982319885338846,
588
+ "learning_rate": 5.884057971014494e-05,
589
+ "loss": 1.0987,
590
+ "step": 83
591
+ },
592
+ {
593
+ "epoch": 0.29473684210526313,
594
+ "grad_norm": 0.8863683819755959,
595
+ "learning_rate": 5.8550724637681166e-05,
596
+ "loss": 1.0603,
597
+ "step": 84
598
+ },
599
+ {
600
+ "epoch": 0.2982456140350877,
601
+ "grad_norm": 1.0566465114428347,
602
+ "learning_rate": 5.82608695652174e-05,
603
+ "loss": 1.0961,
604
+ "step": 85
605
+ },
606
+ {
607
+ "epoch": 0.3017543859649123,
608
+ "grad_norm": 0.9391730832092359,
609
+ "learning_rate": 5.797101449275363e-05,
610
+ "loss": 1.1184,
611
+ "step": 86
612
+ },
613
+ {
614
+ "epoch": 0.30526315789473685,
615
+ "grad_norm": 0.9725750087741657,
616
+ "learning_rate": 5.768115942028986e-05,
617
+ "loss": 1.1131,
618
+ "step": 87
619
+ },
620
+ {
621
+ "epoch": 0.3087719298245614,
622
+ "grad_norm": 0.9134535098625444,
623
+ "learning_rate": 5.739130434782609e-05,
624
+ "loss": 1.0635,
625
+ "step": 88
626
+ },
627
+ {
628
+ "epoch": 0.312280701754386,
629
+ "grad_norm": 0.9210153237928186,
630
+ "learning_rate": 5.710144927536232e-05,
631
+ "loss": 1.0208,
632
+ "step": 89
633
+ },
634
+ {
635
+ "epoch": 0.3157894736842105,
636
+ "grad_norm": 1.073076083662152,
637
+ "learning_rate": 5.681159420289856e-05,
638
+ "loss": 1.1314,
639
+ "step": 90
640
+ },
641
+ {
642
+ "epoch": 0.3192982456140351,
643
+ "grad_norm": 0.9246736723369094,
644
+ "learning_rate": 5.652173913043479e-05,
645
+ "loss": 1.0458,
646
+ "step": 91
647
+ },
648
+ {
649
+ "epoch": 0.32280701754385965,
650
+ "grad_norm": 0.9415979605613132,
651
+ "learning_rate": 5.623188405797102e-05,
652
+ "loss": 1.0432,
653
+ "step": 92
654
+ },
655
+ {
656
+ "epoch": 0.3263157894736842,
657
+ "grad_norm": 1.0029090729561863,
658
+ "learning_rate": 5.594202898550725e-05,
659
+ "loss": 1.1045,
660
+ "step": 93
661
+ },
662
+ {
663
+ "epoch": 0.3298245614035088,
664
+ "grad_norm": 1.0012689311514984,
665
+ "learning_rate": 5.565217391304348e-05,
666
+ "loss": 1.0923,
667
+ "step": 94
668
+ },
669
+ {
670
+ "epoch": 0.3333333333333333,
671
+ "grad_norm": 0.9253904632856298,
672
+ "learning_rate": 5.5362318840579714e-05,
673
+ "loss": 1.028,
674
+ "step": 95
675
+ },
676
+ {
677
+ "epoch": 0.3368421052631579,
678
+ "grad_norm": 1.006794875176782,
679
+ "learning_rate": 5.507246376811594e-05,
680
+ "loss": 1.0772,
681
+ "step": 96
682
+ },
683
+ {
684
+ "epoch": 0.34035087719298246,
685
+ "grad_norm": 0.9652341181001821,
686
+ "learning_rate": 5.478260869565218e-05,
687
+ "loss": 1.0544,
688
+ "step": 97
689
+ },
690
+ {
691
+ "epoch": 0.34385964912280703,
692
+ "grad_norm": 0.9633053495992878,
693
+ "learning_rate": 5.449275362318841e-05,
694
+ "loss": 1.0369,
695
+ "step": 98
696
+ },
697
+ {
698
+ "epoch": 0.3473684210526316,
699
+ "grad_norm": 0.889193134564978,
700
+ "learning_rate": 5.4202898550724644e-05,
701
+ "loss": 1.0571,
702
+ "step": 99
703
+ },
704
+ {
705
+ "epoch": 0.3508771929824561,
706
+ "grad_norm": 1.0688458576983824,
707
+ "learning_rate": 5.391304347826087e-05,
708
+ "loss": 1.0875,
709
+ "step": 100
710
+ },
711
+ {
712
+ "epoch": 0.3543859649122807,
713
+ "grad_norm": 0.9618220162761959,
714
+ "learning_rate": 5.3623188405797106e-05,
715
+ "loss": 1.0989,
716
+ "step": 101
717
+ },
718
+ {
719
+ "epoch": 0.35789473684210527,
720
+ "grad_norm": 1.0116504967458333,
721
+ "learning_rate": 5.333333333333333e-05,
722
+ "loss": 1.1362,
723
+ "step": 102
724
+ },
725
+ {
726
+ "epoch": 0.36140350877192984,
727
+ "grad_norm": 0.9446835004736188,
728
+ "learning_rate": 5.304347826086957e-05,
729
+ "loss": 1.0549,
730
+ "step": 103
731
+ },
732
+ {
733
+ "epoch": 0.3649122807017544,
734
+ "grad_norm": 1.2634297228252125,
735
+ "learning_rate": 5.27536231884058e-05,
736
+ "loss": 1.0865,
737
+ "step": 104
738
+ },
739
+ {
740
+ "epoch": 0.3684210526315789,
741
+ "grad_norm": 1.0070752681998931,
742
+ "learning_rate": 5.2463768115942036e-05,
743
+ "loss": 1.071,
744
+ "step": 105
745
+ },
746
+ {
747
+ "epoch": 0.3719298245614035,
748
+ "grad_norm": 1.0889370997100492,
749
+ "learning_rate": 5.217391304347826e-05,
750
+ "loss": 1.1063,
751
+ "step": 106
752
+ },
753
+ {
754
+ "epoch": 0.37543859649122807,
755
+ "grad_norm": 0.9667935705662773,
756
+ "learning_rate": 5.18840579710145e-05,
757
+ "loss": 1.0666,
758
+ "step": 107
759
+ },
760
+ {
761
+ "epoch": 0.37894736842105264,
762
+ "grad_norm": 0.9817859943047047,
763
+ "learning_rate": 5.1594202898550725e-05,
764
+ "loss": 1.1025,
765
+ "step": 108
766
+ },
767
+ {
768
+ "epoch": 0.3824561403508772,
769
+ "grad_norm": 0.8710066297645505,
770
+ "learning_rate": 5.130434782608696e-05,
771
+ "loss": 1.1057,
772
+ "step": 109
773
+ },
774
+ {
775
+ "epoch": 0.38596491228070173,
776
+ "grad_norm": 1.0647772098482442,
777
+ "learning_rate": 5.1014492753623186e-05,
778
+ "loss": 1.1315,
779
+ "step": 110
780
+ },
781
+ {
782
+ "epoch": 0.3894736842105263,
783
+ "grad_norm": 0.9399046787831713,
784
+ "learning_rate": 5.072463768115943e-05,
785
+ "loss": 1.0572,
786
+ "step": 111
787
+ },
788
+ {
789
+ "epoch": 0.3929824561403509,
790
+ "grad_norm": 0.8768124884597763,
791
+ "learning_rate": 5.043478260869566e-05,
792
+ "loss": 1.0401,
793
+ "step": 112
794
+ },
795
+ {
796
+ "epoch": 0.39649122807017545,
797
+ "grad_norm": 1.009086025516714,
798
+ "learning_rate": 5.014492753623189e-05,
799
+ "loss": 1.0604,
800
+ "step": 113
801
+ },
802
+ {
803
+ "epoch": 0.4,
804
+ "grad_norm": 0.9419034237826691,
805
+ "learning_rate": 4.9855072463768116e-05,
806
+ "loss": 1.0858,
807
+ "step": 114
808
+ },
809
+ {
810
+ "epoch": 0.40350877192982454,
811
+ "grad_norm": 0.9876303358742554,
812
+ "learning_rate": 4.956521739130435e-05,
813
+ "loss": 1.083,
814
+ "step": 115
815
+ },
816
+ {
817
+ "epoch": 0.4070175438596491,
818
+ "grad_norm": 0.934675922910989,
819
+ "learning_rate": 4.927536231884058e-05,
820
+ "loss": 1.0749,
821
+ "step": 116
822
+ },
823
+ {
824
+ "epoch": 0.4105263157894737,
825
+ "grad_norm": 0.9813286336072813,
826
+ "learning_rate": 4.898550724637682e-05,
827
+ "loss": 1.0721,
828
+ "step": 117
829
+ },
830
+ {
831
+ "epoch": 0.41403508771929826,
832
+ "grad_norm": 1.0168504955660564,
833
+ "learning_rate": 4.869565217391305e-05,
834
+ "loss": 1.0276,
835
+ "step": 118
836
+ },
837
+ {
838
+ "epoch": 0.41754385964912283,
839
+ "grad_norm": 0.9087837180593036,
840
+ "learning_rate": 4.840579710144928e-05,
841
+ "loss": 1.0104,
842
+ "step": 119
843
+ },
844
+ {
845
+ "epoch": 0.42105263157894735,
846
+ "grad_norm": 0.8834303398491904,
847
+ "learning_rate": 4.8115942028985514e-05,
848
+ "loss": 1.032,
849
+ "step": 120
850
+ },
851
+ {
852
+ "epoch": 0.4245614035087719,
853
+ "grad_norm": 0.9525280076015293,
854
+ "learning_rate": 4.782608695652174e-05,
855
+ "loss": 1.0257,
856
+ "step": 121
857
+ },
858
+ {
859
+ "epoch": 0.4280701754385965,
860
+ "grad_norm": 0.9493713737636021,
861
+ "learning_rate": 4.7536231884057976e-05,
862
+ "loss": 1.0273,
863
+ "step": 122
864
+ },
865
+ {
866
+ "epoch": 0.43157894736842106,
867
+ "grad_norm": 0.9098239721172197,
868
+ "learning_rate": 4.72463768115942e-05,
869
+ "loss": 0.981,
870
+ "step": 123
871
+ },
872
+ {
873
+ "epoch": 0.43508771929824563,
874
+ "grad_norm": 0.9664557596159875,
875
+ "learning_rate": 4.6956521739130444e-05,
876
+ "loss": 1.1167,
877
+ "step": 124
878
+ },
879
+ {
880
+ "epoch": 0.43859649122807015,
881
+ "grad_norm": 0.8663800542676098,
882
+ "learning_rate": 4.666666666666667e-05,
883
+ "loss": 1.0566,
884
+ "step": 125
885
+ },
886
+ {
887
+ "epoch": 0.4421052631578947,
888
+ "grad_norm": 0.8587329676076015,
889
+ "learning_rate": 4.6376811594202906e-05,
890
+ "loss": 1.0457,
891
+ "step": 126
892
+ },
893
+ {
894
+ "epoch": 0.4456140350877193,
895
+ "grad_norm": 1.0139380452518938,
896
+ "learning_rate": 4.608695652173913e-05,
897
+ "loss": 1.0871,
898
+ "step": 127
899
+ },
900
+ {
901
+ "epoch": 0.44912280701754387,
902
+ "grad_norm": 0.87218312389074,
903
+ "learning_rate": 4.579710144927537e-05,
904
+ "loss": 0.9934,
905
+ "step": 128
906
+ },
907
+ {
908
+ "epoch": 0.45263157894736844,
909
+ "grad_norm": 0.945265788938204,
910
+ "learning_rate": 4.5507246376811595e-05,
911
+ "loss": 1.0598,
912
+ "step": 129
913
+ },
914
+ {
915
+ "epoch": 0.45614035087719296,
916
+ "grad_norm": 1.0447712876512982,
917
+ "learning_rate": 4.521739130434783e-05,
918
+ "loss": 1.0208,
919
+ "step": 130
920
+ },
921
+ {
922
+ "epoch": 0.45964912280701753,
923
+ "grad_norm": 0.8863649966336605,
924
+ "learning_rate": 4.492753623188406e-05,
925
+ "loss": 0.9916,
926
+ "step": 131
927
+ },
928
+ {
929
+ "epoch": 0.4631578947368421,
930
+ "grad_norm": 0.9389072020523048,
931
+ "learning_rate": 4.46376811594203e-05,
932
+ "loss": 1.039,
933
+ "step": 132
934
+ },
935
+ {
936
+ "epoch": 0.4666666666666667,
937
+ "grad_norm": 1.00128465703121,
938
+ "learning_rate": 4.4347826086956525e-05,
939
+ "loss": 1.007,
940
+ "step": 133
941
+ },
942
+ {
943
+ "epoch": 0.47017543859649125,
944
+ "grad_norm": 0.9218948102233011,
945
+ "learning_rate": 4.405797101449276e-05,
946
+ "loss": 1.0489,
947
+ "step": 134
948
+ },
949
+ {
950
+ "epoch": 0.47368421052631576,
951
+ "grad_norm": 0.8980509099966125,
952
+ "learning_rate": 4.3768115942028986e-05,
953
+ "loss": 1.0539,
954
+ "step": 135
955
+ },
956
+ {
957
+ "epoch": 0.47719298245614034,
958
+ "grad_norm": 0.8922781997772711,
959
+ "learning_rate": 4.347826086956522e-05,
960
+ "loss": 1.0653,
961
+ "step": 136
962
+ },
963
+ {
964
+ "epoch": 0.4807017543859649,
965
+ "grad_norm": 0.9229578265481162,
966
+ "learning_rate": 4.318840579710145e-05,
967
+ "loss": 1.0359,
968
+ "step": 137
969
+ },
970
+ {
971
+ "epoch": 0.4842105263157895,
972
+ "grad_norm": 0.9864176248898312,
973
+ "learning_rate": 4.289855072463769e-05,
974
+ "loss": 1.0711,
975
+ "step": 138
976
+ },
977
+ {
978
+ "epoch": 0.48771929824561405,
979
+ "grad_norm": 0.9040178292405336,
980
+ "learning_rate": 4.2608695652173916e-05,
981
+ "loss": 1.0615,
982
+ "step": 139
983
+ },
984
+ {
985
+ "epoch": 0.49122807017543857,
986
+ "grad_norm": 0.9065042000785789,
987
+ "learning_rate": 4.231884057971015e-05,
988
+ "loss": 1.0267,
989
+ "step": 140
990
+ },
991
+ {
992
+ "epoch": 0.49473684210526314,
993
+ "grad_norm": 0.9289938532156966,
994
+ "learning_rate": 4.202898550724638e-05,
995
+ "loss": 1.0269,
996
+ "step": 141
997
+ },
998
+ {
999
+ "epoch": 0.4982456140350877,
1000
+ "grad_norm": 0.8703398745310128,
1001
+ "learning_rate": 4.173913043478261e-05,
1002
+ "loss": 0.9952,
1003
+ "step": 142
1004
+ },
1005
+ {
1006
+ "epoch": 0.5017543859649123,
1007
+ "grad_norm": 0.9754555079004449,
1008
+ "learning_rate": 4.144927536231884e-05,
1009
+ "loss": 1.0744,
1010
+ "step": 143
1011
+ },
1012
+ {
1013
+ "epoch": 0.5052631578947369,
1014
+ "grad_norm": 1.0698316819851512,
1015
+ "learning_rate": 4.115942028985507e-05,
1016
+ "loss": 1.0179,
1017
+ "step": 144
1018
+ },
1019
+ {
1020
+ "epoch": 0.5087719298245614,
1021
+ "grad_norm": 0.9380968706540932,
1022
+ "learning_rate": 4.0869565217391314e-05,
1023
+ "loss": 0.9932,
1024
+ "step": 145
1025
+ },
1026
+ {
1027
+ "epoch": 0.512280701754386,
1028
+ "grad_norm": 0.9380079488737321,
1029
+ "learning_rate": 4.057971014492754e-05,
1030
+ "loss": 1.0655,
1031
+ "step": 146
1032
+ },
1033
+ {
1034
+ "epoch": 0.5157894736842106,
1035
+ "grad_norm": 0.9764065323167558,
1036
+ "learning_rate": 4.0289855072463776e-05,
1037
+ "loss": 1.0096,
1038
+ "step": 147
1039
+ },
1040
+ {
1041
+ "epoch": 0.519298245614035,
1042
+ "grad_norm": 0.917590332443622,
1043
+ "learning_rate": 4e-05,
1044
+ "loss": 1.0376,
1045
+ "step": 148
1046
+ },
1047
+ {
1048
+ "epoch": 0.5228070175438596,
1049
+ "grad_norm": 0.9466091457378278,
1050
+ "learning_rate": 3.971014492753623e-05,
1051
+ "loss": 1.0722,
1052
+ "step": 149
1053
+ },
1054
+ {
1055
+ "epoch": 0.5263157894736842,
1056
+ "grad_norm": 0.9998149019870339,
1057
+ "learning_rate": 3.942028985507247e-05,
1058
+ "loss": 1.0747,
1059
+ "step": 150
1060
+ },
1061
+ {
1062
+ "epoch": 0.5298245614035088,
1063
+ "grad_norm": 0.9592124249246545,
1064
+ "learning_rate": 3.91304347826087e-05,
1065
+ "loss": 0.9772,
1066
+ "step": 151
1067
+ },
1068
+ {
1069
+ "epoch": 0.5333333333333333,
1070
+ "grad_norm": 0.8702377525213307,
1071
+ "learning_rate": 3.884057971014493e-05,
1072
+ "loss": 1.0416,
1073
+ "step": 152
1074
+ },
1075
+ {
1076
+ "epoch": 0.5368421052631579,
1077
+ "grad_norm": 0.9408998854641563,
1078
+ "learning_rate": 3.855072463768116e-05,
1079
+ "loss": 1.053,
1080
+ "step": 153
1081
+ },
1082
+ {
1083
+ "epoch": 0.5403508771929825,
1084
+ "grad_norm": 0.913519043609596,
1085
+ "learning_rate": 3.8260869565217395e-05,
1086
+ "loss": 1.0653,
1087
+ "step": 154
1088
+ },
1089
+ {
1090
+ "epoch": 0.543859649122807,
1091
+ "grad_norm": 1.0325941411560875,
1092
+ "learning_rate": 3.797101449275363e-05,
1093
+ "loss": 1.0248,
1094
+ "step": 155
1095
+ },
1096
+ {
1097
+ "epoch": 0.5473684210526316,
1098
+ "grad_norm": 1.0088857759548826,
1099
+ "learning_rate": 3.7681159420289856e-05,
1100
+ "loss": 1.0984,
1101
+ "step": 156
1102
+ },
1103
+ {
1104
+ "epoch": 0.5508771929824562,
1105
+ "grad_norm": 0.9428277439596274,
1106
+ "learning_rate": 3.739130434782609e-05,
1107
+ "loss": 1.0176,
1108
+ "step": 157
1109
+ },
1110
+ {
1111
+ "epoch": 0.5543859649122806,
1112
+ "grad_norm": 0.9286332315925182,
1113
+ "learning_rate": 3.7101449275362325e-05,
1114
+ "loss": 1.0141,
1115
+ "step": 158
1116
+ },
1117
+ {
1118
+ "epoch": 0.5578947368421052,
1119
+ "grad_norm": 0.8869336003769103,
1120
+ "learning_rate": 3.681159420289855e-05,
1121
+ "loss": 1.0075,
1122
+ "step": 159
1123
+ },
1124
+ {
1125
+ "epoch": 0.5614035087719298,
1126
+ "grad_norm": 0.8749772917176272,
1127
+ "learning_rate": 3.6521739130434786e-05,
1128
+ "loss": 0.9922,
1129
+ "step": 160
1130
+ },
1131
+ {
1132
+ "epoch": 0.5649122807017544,
1133
+ "grad_norm": 0.9177592855322187,
1134
+ "learning_rate": 3.623188405797102e-05,
1135
+ "loss": 1.047,
1136
+ "step": 161
1137
+ },
1138
+ {
1139
+ "epoch": 0.5684210526315789,
1140
+ "grad_norm": 0.9681450836764032,
1141
+ "learning_rate": 3.594202898550725e-05,
1142
+ "loss": 0.9947,
1143
+ "step": 162
1144
+ },
1145
+ {
1146
+ "epoch": 0.5719298245614035,
1147
+ "grad_norm": 0.8944949579910934,
1148
+ "learning_rate": 3.565217391304348e-05,
1149
+ "loss": 1.0414,
1150
+ "step": 163
1151
+ },
1152
+ {
1153
+ "epoch": 0.5754385964912281,
1154
+ "grad_norm": 0.823418699987978,
1155
+ "learning_rate": 3.5362318840579716e-05,
1156
+ "loss": 0.9991,
1157
+ "step": 164
1158
+ },
1159
+ {
1160
+ "epoch": 0.5789473684210527,
1161
+ "grad_norm": 0.9010716175104698,
1162
+ "learning_rate": 3.5072463768115943e-05,
1163
+ "loss": 0.957,
1164
+ "step": 165
1165
+ },
1166
+ {
1167
+ "epoch": 0.5824561403508772,
1168
+ "grad_norm": 0.8553233244911422,
1169
+ "learning_rate": 3.478260869565218e-05,
1170
+ "loss": 0.9758,
1171
+ "step": 166
1172
+ },
1173
+ {
1174
+ "epoch": 0.5859649122807018,
1175
+ "grad_norm": 0.8955709532913586,
1176
+ "learning_rate": 3.449275362318841e-05,
1177
+ "loss": 1.0323,
1178
+ "step": 167
1179
+ },
1180
+ {
1181
+ "epoch": 0.5894736842105263,
1182
+ "grad_norm": 0.9632249383426715,
1183
+ "learning_rate": 3.420289855072464e-05,
1184
+ "loss": 1.0214,
1185
+ "step": 168
1186
+ },
1187
+ {
1188
+ "epoch": 0.5929824561403508,
1189
+ "grad_norm": 1.0070882268479324,
1190
+ "learning_rate": 3.391304347826087e-05,
1191
+ "loss": 1.0443,
1192
+ "step": 169
1193
+ },
1194
+ {
1195
+ "epoch": 0.5964912280701754,
1196
+ "grad_norm": 0.9089812750457459,
1197
+ "learning_rate": 3.36231884057971e-05,
1198
+ "loss": 0.9986,
1199
+ "step": 170
1200
+ },
1201
+ {
1202
+ "epoch": 0.6,
1203
+ "grad_norm": 0.9687903432716775,
1204
+ "learning_rate": 3.3333333333333335e-05,
1205
+ "loss": 1.0469,
1206
+ "step": 171
1207
+ },
1208
+ {
1209
+ "epoch": 0.6035087719298246,
1210
+ "grad_norm": 0.9065456102432492,
1211
+ "learning_rate": 3.304347826086957e-05,
1212
+ "loss": 1.0736,
1213
+ "step": 172
1214
+ },
1215
+ {
1216
+ "epoch": 0.6070175438596491,
1217
+ "grad_norm": 0.8191700226087758,
1218
+ "learning_rate": 3.2753623188405796e-05,
1219
+ "loss": 0.9516,
1220
+ "step": 173
1221
+ },
1222
+ {
1223
+ "epoch": 0.6105263157894737,
1224
+ "grad_norm": 0.910478607503674,
1225
+ "learning_rate": 3.246376811594203e-05,
1226
+ "loss": 1.0594,
1227
+ "step": 174
1228
+ },
1229
+ {
1230
+ "epoch": 0.6140350877192983,
1231
+ "grad_norm": 0.9847223979786262,
1232
+ "learning_rate": 3.2173913043478265e-05,
1233
+ "loss": 1.0592,
1234
+ "step": 175
1235
+ },
1236
+ {
1237
+ "epoch": 0.6175438596491228,
1238
+ "grad_norm": 0.9377862201207361,
1239
+ "learning_rate": 3.188405797101449e-05,
1240
+ "loss": 1.0477,
1241
+ "step": 176
1242
+ },
1243
+ {
1244
+ "epoch": 0.6210526315789474,
1245
+ "grad_norm": 0.9077059344710785,
1246
+ "learning_rate": 3.1594202898550726e-05,
1247
+ "loss": 1.0394,
1248
+ "step": 177
1249
+ },
1250
+ {
1251
+ "epoch": 0.624561403508772,
1252
+ "grad_norm": 1.038075786423051,
1253
+ "learning_rate": 3.130434782608696e-05,
1254
+ "loss": 1.0548,
1255
+ "step": 178
1256
+ },
1257
+ {
1258
+ "epoch": 0.6280701754385964,
1259
+ "grad_norm": 0.9160209790965669,
1260
+ "learning_rate": 3.101449275362319e-05,
1261
+ "loss": 0.9976,
1262
+ "step": 179
1263
+ },
1264
+ {
1265
+ "epoch": 0.631578947368421,
1266
+ "grad_norm": 0.8109415841452218,
1267
+ "learning_rate": 3.072463768115942e-05,
1268
+ "loss": 1.0091,
1269
+ "step": 180
1270
+ },
1271
+ {
1272
+ "epoch": 0.6350877192982456,
1273
+ "grad_norm": 0.8492567440061373,
1274
+ "learning_rate": 3.0434782608695656e-05,
1275
+ "loss": 1.0038,
1276
+ "step": 181
1277
+ },
1278
+ {
1279
+ "epoch": 0.6385964912280702,
1280
+ "grad_norm": 0.9589169451188458,
1281
+ "learning_rate": 3.0144927536231887e-05,
1282
+ "loss": 1.0207,
1283
+ "step": 182
1284
+ },
1285
+ {
1286
+ "epoch": 0.6421052631578947,
1287
+ "grad_norm": 0.8848409951191845,
1288
+ "learning_rate": 2.9855072463768118e-05,
1289
+ "loss": 0.9767,
1290
+ "step": 183
1291
+ },
1292
+ {
1293
+ "epoch": 0.6456140350877193,
1294
+ "grad_norm": 0.8955318088948372,
1295
+ "learning_rate": 2.956521739130435e-05,
1296
+ "loss": 1.0078,
1297
+ "step": 184
1298
+ },
1299
+ {
1300
+ "epoch": 0.6491228070175439,
1301
+ "grad_norm": 0.8900555696345945,
1302
+ "learning_rate": 2.9275362318840583e-05,
1303
+ "loss": 1.058,
1304
+ "step": 185
1305
+ },
1306
+ {
1307
+ "epoch": 0.6526315789473685,
1308
+ "grad_norm": 0.9579354937026371,
1309
+ "learning_rate": 2.8985507246376814e-05,
1310
+ "loss": 1.0052,
1311
+ "step": 186
1312
+ },
1313
+ {
1314
+ "epoch": 0.656140350877193,
1315
+ "grad_norm": 0.8773237300221255,
1316
+ "learning_rate": 2.8695652173913044e-05,
1317
+ "loss": 1.0132,
1318
+ "step": 187
1319
+ },
1320
+ {
1321
+ "epoch": 0.6596491228070176,
1322
+ "grad_norm": 0.9290671958630004,
1323
+ "learning_rate": 2.840579710144928e-05,
1324
+ "loss": 0.9978,
1325
+ "step": 188
1326
+ },
1327
+ {
1328
+ "epoch": 0.6631578947368421,
1329
+ "grad_norm": 0.9098091343168585,
1330
+ "learning_rate": 2.811594202898551e-05,
1331
+ "loss": 0.986,
1332
+ "step": 189
1333
+ },
1334
+ {
1335
+ "epoch": 0.6666666666666666,
1336
+ "grad_norm": 0.9494694332915156,
1337
+ "learning_rate": 2.782608695652174e-05,
1338
+ "loss": 1.0194,
1339
+ "step": 190
1340
+ },
1341
+ {
1342
+ "epoch": 0.6701754385964912,
1343
+ "grad_norm": 0.8801338941113726,
1344
+ "learning_rate": 2.753623188405797e-05,
1345
+ "loss": 0.969,
1346
+ "step": 191
1347
+ },
1348
+ {
1349
+ "epoch": 0.6736842105263158,
1350
+ "grad_norm": 0.9510753232451794,
1351
+ "learning_rate": 2.7246376811594205e-05,
1352
+ "loss": 0.9742,
1353
+ "step": 192
1354
+ },
1355
+ {
1356
+ "epoch": 0.6771929824561403,
1357
+ "grad_norm": 0.9589458183246781,
1358
+ "learning_rate": 2.6956521739130436e-05,
1359
+ "loss": 1.0525,
1360
+ "step": 193
1361
+ },
1362
+ {
1363
+ "epoch": 0.6807017543859649,
1364
+ "grad_norm": 0.879997612951922,
1365
+ "learning_rate": 2.6666666666666667e-05,
1366
+ "loss": 0.9824,
1367
+ "step": 194
1368
+ },
1369
+ {
1370
+ "epoch": 0.6842105263157895,
1371
+ "grad_norm": 0.8389910470867041,
1372
+ "learning_rate": 2.63768115942029e-05,
1373
+ "loss": 1.0056,
1374
+ "step": 195
1375
+ },
1376
+ {
1377
+ "epoch": 0.6877192982456141,
1378
+ "grad_norm": 0.9163566396967311,
1379
+ "learning_rate": 2.608695652173913e-05,
1380
+ "loss": 0.9976,
1381
+ "step": 196
1382
+ },
1383
+ {
1384
+ "epoch": 0.6912280701754386,
1385
+ "grad_norm": 0.9772320516342279,
1386
+ "learning_rate": 2.5797101449275362e-05,
1387
+ "loss": 1.0338,
1388
+ "step": 197
1389
+ },
1390
+ {
1391
+ "epoch": 0.6947368421052632,
1392
+ "grad_norm": 0.960770516269451,
1393
+ "learning_rate": 2.5507246376811593e-05,
1394
+ "loss": 1.0611,
1395
+ "step": 198
1396
+ },
1397
+ {
1398
+ "epoch": 0.6982456140350877,
1399
+ "grad_norm": 0.8890098864104716,
1400
+ "learning_rate": 2.521739130434783e-05,
1401
+ "loss": 1.0168,
1402
+ "step": 199
1403
+ },
1404
+ {
1405
+ "epoch": 0.7017543859649122,
1406
+ "grad_norm": 0.8658933316106927,
1407
+ "learning_rate": 2.4927536231884058e-05,
1408
+ "loss": 1.011,
1409
+ "step": 200
1410
+ },
1411
+ {
1412
+ "epoch": 0.7052631578947368,
1413
+ "grad_norm": 0.8576704263418472,
1414
+ "learning_rate": 2.463768115942029e-05,
1415
+ "loss": 0.9807,
1416
+ "step": 201
1417
+ },
1418
+ {
1419
+ "epoch": 0.7087719298245614,
1420
+ "grad_norm": 0.9186773959277461,
1421
+ "learning_rate": 2.4347826086956526e-05,
1422
+ "loss": 0.9284,
1423
+ "step": 202
1424
+ },
1425
+ {
1426
+ "epoch": 0.712280701754386,
1427
+ "grad_norm": 0.858587966202896,
1428
+ "learning_rate": 2.4057971014492757e-05,
1429
+ "loss": 0.9795,
1430
+ "step": 203
1431
+ },
1432
+ {
1433
+ "epoch": 0.7157894736842105,
1434
+ "grad_norm": 0.9422658689165894,
1435
+ "learning_rate": 2.3768115942028988e-05,
1436
+ "loss": 0.9764,
1437
+ "step": 204
1438
+ },
1439
+ {
1440
+ "epoch": 0.7192982456140351,
1441
+ "grad_norm": 1.0137944232610083,
1442
+ "learning_rate": 2.3478260869565222e-05,
1443
+ "loss": 1.0627,
1444
+ "step": 205
1445
+ },
1446
+ {
1447
+ "epoch": 0.7228070175438597,
1448
+ "grad_norm": 0.9853442185067721,
1449
+ "learning_rate": 2.3188405797101453e-05,
1450
+ "loss": 1.0289,
1451
+ "step": 206
1452
+ },
1453
+ {
1454
+ "epoch": 0.7263157894736842,
1455
+ "grad_norm": 0.8923047177872453,
1456
+ "learning_rate": 2.2898550724637684e-05,
1457
+ "loss": 1.0172,
1458
+ "step": 207
1459
+ },
1460
+ {
1461
+ "epoch": 0.7298245614035088,
1462
+ "grad_norm": 0.8439161677593868,
1463
+ "learning_rate": 2.2608695652173914e-05,
1464
+ "loss": 1.0484,
1465
+ "step": 208
1466
+ },
1467
+ {
1468
+ "epoch": 0.7333333333333333,
1469
+ "grad_norm": 0.8201410283081917,
1470
+ "learning_rate": 2.231884057971015e-05,
1471
+ "loss": 1.0168,
1472
+ "step": 209
1473
+ },
1474
+ {
1475
+ "epoch": 0.7368421052631579,
1476
+ "grad_norm": 0.8442899557106462,
1477
+ "learning_rate": 2.202898550724638e-05,
1478
+ "loss": 0.9728,
1479
+ "step": 210
1480
+ },
1481
+ {
1482
+ "epoch": 0.7403508771929824,
1483
+ "grad_norm": 0.9098630031184369,
1484
+ "learning_rate": 2.173913043478261e-05,
1485
+ "loss": 1.0304,
1486
+ "step": 211
1487
+ },
1488
+ {
1489
+ "epoch": 0.743859649122807,
1490
+ "grad_norm": 0.8217923688708064,
1491
+ "learning_rate": 2.1449275362318844e-05,
1492
+ "loss": 1.0334,
1493
+ "step": 212
1494
+ },
1495
+ {
1496
+ "epoch": 0.7473684210526316,
1497
+ "grad_norm": 0.8420848196609938,
1498
+ "learning_rate": 2.1159420289855075e-05,
1499
+ "loss": 0.9665,
1500
+ "step": 213
1501
+ },
1502
+ {
1503
+ "epoch": 0.7508771929824561,
1504
+ "grad_norm": 0.8810436312516507,
1505
+ "learning_rate": 2.0869565217391306e-05,
1506
+ "loss": 0.9792,
1507
+ "step": 214
1508
+ },
1509
+ {
1510
+ "epoch": 0.7543859649122807,
1511
+ "grad_norm": 0.9749166994556522,
1512
+ "learning_rate": 2.0579710144927537e-05,
1513
+ "loss": 1.0163,
1514
+ "step": 215
1515
+ },
1516
+ {
1517
+ "epoch": 0.7578947368421053,
1518
+ "grad_norm": 0.9619307294480671,
1519
+ "learning_rate": 2.028985507246377e-05,
1520
+ "loss": 1.0276,
1521
+ "step": 216
1522
+ },
1523
+ {
1524
+ "epoch": 0.7614035087719299,
1525
+ "grad_norm": 0.8475320304538961,
1526
+ "learning_rate": 2e-05,
1527
+ "loss": 1.0175,
1528
+ "step": 217
1529
+ },
1530
+ {
1531
+ "epoch": 0.7649122807017544,
1532
+ "grad_norm": 0.9215056645276478,
1533
+ "learning_rate": 1.9710144927536236e-05,
1534
+ "loss": 1.0349,
1535
+ "step": 218
1536
+ },
1537
+ {
1538
+ "epoch": 0.7684210526315789,
1539
+ "grad_norm": 0.8318566015217065,
1540
+ "learning_rate": 1.9420289855072467e-05,
1541
+ "loss": 0.9736,
1542
+ "step": 219
1543
+ },
1544
+ {
1545
+ "epoch": 0.7719298245614035,
1546
+ "grad_norm": 0.8841348427961992,
1547
+ "learning_rate": 1.9130434782608697e-05,
1548
+ "loss": 1.0028,
1549
+ "step": 220
1550
+ },
1551
+ {
1552
+ "epoch": 0.775438596491228,
1553
+ "grad_norm": 0.8726663458778539,
1554
+ "learning_rate": 1.8840579710144928e-05,
1555
+ "loss": 0.9837,
1556
+ "step": 221
1557
+ },
1558
+ {
1559
+ "epoch": 0.7789473684210526,
1560
+ "grad_norm": 0.8933308297868434,
1561
+ "learning_rate": 1.8550724637681162e-05,
1562
+ "loss": 0.9875,
1563
+ "step": 222
1564
+ },
1565
+ {
1566
+ "epoch": 0.7824561403508772,
1567
+ "grad_norm": 0.8473190274519702,
1568
+ "learning_rate": 1.8260869565217393e-05,
1569
+ "loss": 0.9822,
1570
+ "step": 223
1571
+ },
1572
+ {
1573
+ "epoch": 0.7859649122807018,
1574
+ "grad_norm": 0.8786630261494585,
1575
+ "learning_rate": 1.7971014492753624e-05,
1576
+ "loss": 0.9785,
1577
+ "step": 224
1578
+ },
1579
+ {
1580
+ "epoch": 0.7894736842105263,
1581
+ "grad_norm": 0.8769937018569535,
1582
+ "learning_rate": 1.7681159420289858e-05,
1583
+ "loss": 0.9863,
1584
+ "step": 225
1585
+ },
1586
+ {
1587
+ "epoch": 0.7929824561403509,
1588
+ "grad_norm": 0.9418240902774399,
1589
+ "learning_rate": 1.739130434782609e-05,
1590
+ "loss": 0.997,
1591
+ "step": 226
1592
+ },
1593
+ {
1594
+ "epoch": 0.7964912280701755,
1595
+ "grad_norm": 0.9360134126214504,
1596
+ "learning_rate": 1.710144927536232e-05,
1597
+ "loss": 0.9863,
1598
+ "step": 227
1599
+ },
1600
+ {
1601
+ "epoch": 0.8,
1602
+ "grad_norm": 0.999458819703,
1603
+ "learning_rate": 1.681159420289855e-05,
1604
+ "loss": 0.9498,
1605
+ "step": 228
1606
+ },
1607
+ {
1608
+ "epoch": 0.8035087719298246,
1609
+ "grad_norm": 0.8238131544325801,
1610
+ "learning_rate": 1.6521739130434785e-05,
1611
+ "loss": 0.9609,
1612
+ "step": 229
1613
+ },
1614
+ {
1615
+ "epoch": 0.8070175438596491,
1616
+ "grad_norm": 0.9110168722541372,
1617
+ "learning_rate": 1.6231884057971015e-05,
1618
+ "loss": 0.9726,
1619
+ "step": 230
1620
+ },
1621
+ {
1622
+ "epoch": 0.8105263157894737,
1623
+ "grad_norm": 0.915280887700176,
1624
+ "learning_rate": 1.5942028985507246e-05,
1625
+ "loss": 0.9687,
1626
+ "step": 231
1627
+ },
1628
+ {
1629
+ "epoch": 0.8140350877192982,
1630
+ "grad_norm": 0.8997906447720105,
1631
+ "learning_rate": 1.565217391304348e-05,
1632
+ "loss": 1.0236,
1633
+ "step": 232
1634
+ },
1635
+ {
1636
+ "epoch": 0.8175438596491228,
1637
+ "grad_norm": 0.8927454629454755,
1638
+ "learning_rate": 1.536231884057971e-05,
1639
+ "loss": 0.96,
1640
+ "step": 233
1641
+ },
1642
+ {
1643
+ "epoch": 0.8210526315789474,
1644
+ "grad_norm": 0.9150163082832309,
1645
+ "learning_rate": 1.5072463768115944e-05,
1646
+ "loss": 0.9843,
1647
+ "step": 234
1648
+ },
1649
+ {
1650
+ "epoch": 0.8245614035087719,
1651
+ "grad_norm": 0.9497695522124332,
1652
+ "learning_rate": 1.4782608695652174e-05,
1653
+ "loss": 1.0254,
1654
+ "step": 235
1655
+ },
1656
+ {
1657
+ "epoch": 0.8280701754385965,
1658
+ "grad_norm": 0.8134079237524361,
1659
+ "learning_rate": 1.4492753623188407e-05,
1660
+ "loss": 0.9874,
1661
+ "step": 236
1662
+ },
1663
+ {
1664
+ "epoch": 0.8315789473684211,
1665
+ "grad_norm": 0.854532647109886,
1666
+ "learning_rate": 1.420289855072464e-05,
1667
+ "loss": 0.9697,
1668
+ "step": 237
1669
+ },
1670
+ {
1671
+ "epoch": 0.8350877192982457,
1672
+ "grad_norm": 0.902401593136911,
1673
+ "learning_rate": 1.391304347826087e-05,
1674
+ "loss": 1.0083,
1675
+ "step": 238
1676
+ },
1677
+ {
1678
+ "epoch": 0.8385964912280702,
1679
+ "grad_norm": 0.8365313967364223,
1680
+ "learning_rate": 1.3623188405797103e-05,
1681
+ "loss": 0.9479,
1682
+ "step": 239
1683
+ },
1684
+ {
1685
+ "epoch": 0.8421052631578947,
1686
+ "grad_norm": 0.8385782446580461,
1687
+ "learning_rate": 1.3333333333333333e-05,
1688
+ "loss": 0.998,
1689
+ "step": 240
1690
+ },
1691
+ {
1692
+ "epoch": 0.8456140350877193,
1693
+ "grad_norm": 0.9130856203558422,
1694
+ "learning_rate": 1.3043478260869566e-05,
1695
+ "loss": 0.9934,
1696
+ "step": 241
1697
+ },
1698
+ {
1699
+ "epoch": 0.8491228070175438,
1700
+ "grad_norm": 0.8261442785562587,
1701
+ "learning_rate": 1.2753623188405797e-05,
1702
+ "loss": 0.9274,
1703
+ "step": 242
1704
+ },
1705
+ {
1706
+ "epoch": 0.8526315789473684,
1707
+ "grad_norm": 0.8560538840518993,
1708
+ "learning_rate": 1.2463768115942029e-05,
1709
+ "loss": 0.9511,
1710
+ "step": 243
1711
+ },
1712
+ {
1713
+ "epoch": 0.856140350877193,
1714
+ "grad_norm": 0.8699104552871371,
1715
+ "learning_rate": 1.2173913043478263e-05,
1716
+ "loss": 0.958,
1717
+ "step": 244
1718
+ },
1719
+ {
1720
+ "epoch": 0.8596491228070176,
1721
+ "grad_norm": 0.9107920631411749,
1722
+ "learning_rate": 1.1884057971014494e-05,
1723
+ "loss": 1.0177,
1724
+ "step": 245
1725
+ },
1726
+ {
1727
+ "epoch": 0.8631578947368421,
1728
+ "grad_norm": 0.8776802277406592,
1729
+ "learning_rate": 1.1594202898550726e-05,
1730
+ "loss": 0.9261,
1731
+ "step": 246
1732
+ },
1733
+ {
1734
+ "epoch": 0.8666666666666667,
1735
+ "grad_norm": 0.8999713914824392,
1736
+ "learning_rate": 1.1304347826086957e-05,
1737
+ "loss": 0.9973,
1738
+ "step": 247
1739
+ },
1740
+ {
1741
+ "epoch": 0.8701754385964913,
1742
+ "grad_norm": 0.8835866950552989,
1743
+ "learning_rate": 1.101449275362319e-05,
1744
+ "loss": 1.0028,
1745
+ "step": 248
1746
+ },
1747
+ {
1748
+ "epoch": 0.8736842105263158,
1749
+ "grad_norm": 0.9243563863707843,
1750
+ "learning_rate": 1.0724637681159422e-05,
1751
+ "loss": 0.9971,
1752
+ "step": 249
1753
+ },
1754
+ {
1755
+ "epoch": 0.8771929824561403,
1756
+ "grad_norm": 0.9287603572835252,
1757
+ "learning_rate": 1.0434782608695653e-05,
1758
+ "loss": 0.9902,
1759
+ "step": 250
1760
+ },
1761
+ {
1762
+ "epoch": 0.8807017543859649,
1763
+ "grad_norm": 0.8670790287211418,
1764
+ "learning_rate": 1.0144927536231885e-05,
1765
+ "loss": 1.0039,
1766
+ "step": 251
1767
+ },
1768
+ {
1769
+ "epoch": 0.8842105263157894,
1770
+ "grad_norm": 0.8663791969051364,
1771
+ "learning_rate": 9.855072463768118e-06,
1772
+ "loss": 0.9817,
1773
+ "step": 252
1774
+ },
1775
+ {
1776
+ "epoch": 0.887719298245614,
1777
+ "grad_norm": 0.8631015068740149,
1778
+ "learning_rate": 9.565217391304349e-06,
1779
+ "loss": 1.0173,
1780
+ "step": 253
1781
+ },
1782
+ {
1783
+ "epoch": 0.8912280701754386,
1784
+ "grad_norm": 0.8829365467183184,
1785
+ "learning_rate": 9.275362318840581e-06,
1786
+ "loss": 0.9578,
1787
+ "step": 254
1788
+ },
1789
+ {
1790
+ "epoch": 0.8947368421052632,
1791
+ "grad_norm": 0.9001038483145964,
1792
+ "learning_rate": 8.985507246376812e-06,
1793
+ "loss": 1.0258,
1794
+ "step": 255
1795
+ },
1796
+ {
1797
+ "epoch": 0.8982456140350877,
1798
+ "grad_norm": 0.8834099073060356,
1799
+ "learning_rate": 8.695652173913044e-06,
1800
+ "loss": 0.9726,
1801
+ "step": 256
1802
+ },
1803
+ {
1804
+ "epoch": 0.9017543859649123,
1805
+ "grad_norm": 0.8221574328830703,
1806
+ "learning_rate": 8.405797101449275e-06,
1807
+ "loss": 0.9318,
1808
+ "step": 257
1809
+ },
1810
+ {
1811
+ "epoch": 0.9052631578947369,
1812
+ "grad_norm": 0.8603145991699224,
1813
+ "learning_rate": 8.115942028985508e-06,
1814
+ "loss": 0.9663,
1815
+ "step": 258
1816
+ },
1817
+ {
1818
+ "epoch": 0.9087719298245615,
1819
+ "grad_norm": 0.8557514339905812,
1820
+ "learning_rate": 7.82608695652174e-06,
1821
+ "loss": 0.9968,
1822
+ "step": 259
1823
+ },
1824
+ {
1825
+ "epoch": 0.9122807017543859,
1826
+ "grad_norm": 0.8263665060539463,
1827
+ "learning_rate": 7.536231884057972e-06,
1828
+ "loss": 1.0022,
1829
+ "step": 260
1830
+ },
1831
+ {
1832
+ "epoch": 0.9157894736842105,
1833
+ "grad_norm": 0.9157072836324722,
1834
+ "learning_rate": 7.246376811594203e-06,
1835
+ "loss": 0.9387,
1836
+ "step": 261
1837
+ },
1838
+ {
1839
+ "epoch": 0.9192982456140351,
1840
+ "grad_norm": 0.9233305451677062,
1841
+ "learning_rate": 6.956521739130435e-06,
1842
+ "loss": 0.9934,
1843
+ "step": 262
1844
+ },
1845
+ {
1846
+ "epoch": 0.9228070175438596,
1847
+ "grad_norm": 0.903604456186341,
1848
+ "learning_rate": 6.666666666666667e-06,
1849
+ "loss": 0.9502,
1850
+ "step": 263
1851
+ },
1852
+ {
1853
+ "epoch": 0.9263157894736842,
1854
+ "grad_norm": 0.9285076793745337,
1855
+ "learning_rate": 6.376811594202898e-06,
1856
+ "loss": 0.9645,
1857
+ "step": 264
1858
+ },
1859
+ {
1860
+ "epoch": 0.9298245614035088,
1861
+ "grad_norm": 0.888909609341721,
1862
+ "learning_rate": 6.086956521739132e-06,
1863
+ "loss": 0.946,
1864
+ "step": 265
1865
+ },
1866
+ {
1867
+ "epoch": 0.9333333333333333,
1868
+ "grad_norm": 0.9438455431857006,
1869
+ "learning_rate": 5.797101449275363e-06,
1870
+ "loss": 1.0153,
1871
+ "step": 266
1872
+ },
1873
+ {
1874
+ "epoch": 0.9368421052631579,
1875
+ "grad_norm": 0.8551557910334344,
1876
+ "learning_rate": 5.507246376811595e-06,
1877
+ "loss": 0.958,
1878
+ "step": 267
1879
+ },
1880
+ {
1881
+ "epoch": 0.9403508771929825,
1882
+ "grad_norm": 0.8234220234488231,
1883
+ "learning_rate": 5.2173913043478265e-06,
1884
+ "loss": 0.96,
1885
+ "step": 268
1886
+ },
1887
+ {
1888
+ "epoch": 0.9438596491228071,
1889
+ "grad_norm": 0.9003874584417242,
1890
+ "learning_rate": 4.927536231884059e-06,
1891
+ "loss": 0.9757,
1892
+ "step": 269
1893
+ },
1894
+ {
1895
+ "epoch": 0.9473684210526315,
1896
+ "grad_norm": 0.89021747241037,
1897
+ "learning_rate": 4.637681159420291e-06,
1898
+ "loss": 0.9511,
1899
+ "step": 270
1900
+ },
1901
+ {
1902
+ "epoch": 0.9508771929824561,
1903
+ "grad_norm": 0.850549521859951,
1904
+ "learning_rate": 4.347826086956522e-06,
1905
+ "loss": 0.9619,
1906
+ "step": 271
1907
+ },
1908
+ {
1909
+ "epoch": 0.9543859649122807,
1910
+ "grad_norm": 0.8477206031920781,
1911
+ "learning_rate": 4.057971014492754e-06,
1912
+ "loss": 0.9169,
1913
+ "step": 272
1914
+ },
1915
+ {
1916
+ "epoch": 0.9578947368421052,
1917
+ "grad_norm": 0.9199733593357698,
1918
+ "learning_rate": 3.768115942028986e-06,
1919
+ "loss": 1.0299,
1920
+ "step": 273
1921
+ },
1922
+ {
1923
+ "epoch": 0.9614035087719298,
1924
+ "grad_norm": 0.8038841148302012,
1925
+ "learning_rate": 3.4782608695652175e-06,
1926
+ "loss": 0.9584,
1927
+ "step": 274
1928
+ },
1929
+ {
1930
+ "epoch": 0.9649122807017544,
1931
+ "grad_norm": 0.8488075839645941,
1932
+ "learning_rate": 3.188405797101449e-06,
1933
+ "loss": 0.9868,
1934
+ "step": 275
1935
+ },
1936
+ {
1937
+ "epoch": 0.968421052631579,
1938
+ "grad_norm": 0.8355215697229909,
1939
+ "learning_rate": 2.8985507246376816e-06,
1940
+ "loss": 1.009,
1941
+ "step": 276
1942
+ },
1943
+ {
1944
+ "epoch": 0.9719298245614035,
1945
+ "grad_norm": 0.9354156162463755,
1946
+ "learning_rate": 2.6086956521739132e-06,
1947
+ "loss": 0.9914,
1948
+ "step": 277
1949
+ },
1950
+ {
1951
+ "epoch": 0.9754385964912281,
1952
+ "grad_norm": 0.8504047654484729,
1953
+ "learning_rate": 2.3188405797101453e-06,
1954
+ "loss": 1.0137,
1955
+ "step": 278
1956
+ },
1957
+ {
1958
+ "epoch": 0.9789473684210527,
1959
+ "grad_norm": 0.9106574593705012,
1960
+ "learning_rate": 2.028985507246377e-06,
1961
+ "loss": 0.9963,
1962
+ "step": 279
1963
+ },
1964
+ {
1965
+ "epoch": 0.9824561403508771,
1966
+ "grad_norm": 0.8356909679028329,
1967
+ "learning_rate": 1.7391304347826088e-06,
1968
+ "loss": 1.0176,
1969
+ "step": 280
1970
+ },
1971
+ {
1972
+ "epoch": 0.9859649122807017,
1973
+ "grad_norm": 0.8904235043433295,
1974
+ "learning_rate": 1.4492753623188408e-06,
1975
+ "loss": 0.9684,
1976
+ "step": 281
1977
+ },
1978
+ {
1979
+ "epoch": 0.9894736842105263,
1980
+ "grad_norm": 0.8663330662929399,
1981
+ "learning_rate": 1.1594202898550726e-06,
1982
+ "loss": 0.9852,
1983
+ "step": 282
1984
+ },
1985
+ {
1986
+ "epoch": 0.9929824561403509,
1987
+ "grad_norm": 0.8453724734461978,
1988
+ "learning_rate": 8.695652173913044e-07,
1989
+ "loss": 0.9599,
1990
+ "step": 283
1991
+ },
1992
+ {
1993
+ "epoch": 0.9964912280701754,
1994
+ "grad_norm": 0.8659094349837689,
1995
+ "learning_rate": 5.797101449275363e-07,
1996
+ "loss": 1.0253,
1997
+ "step": 284
1998
+ },
1999
+ {
2000
+ "epoch": 1.0,
2001
+ "grad_norm": 0.8132420261672163,
2002
+ "learning_rate": 2.8985507246376816e-07,
2003
+ "loss": 0.8957,
2004
+ "step": 285
2005
+ },
2006
+ {
2007
+ "epoch": 1.0,
2008
+ "step": 285,
2009
+ "total_flos": 224781354270720.0,
2010
+ "train_loss": 1.0555618083267881,
2011
+ "train_runtime": 2779.4288,
2012
+ "train_samples_per_second": 52.454,
2013
+ "train_steps_per_second": 0.103
2014
+ }
2015
+ ],
2016
+ "logging_steps": 1.0,
2017
+ "max_steps": 285,
2018
+ "num_input_tokens_seen": 0,
2019
+ "num_train_epochs": 1,
2020
+ "save_steps": 700,
2021
+ "total_flos": 224781354270720.0,
2022
+ "train_batch_size": 8,
2023
+ "trial_name": null,
2024
+ "trial_params": null
2025
+ }
training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1dca7fb36802253051575e9de9c44be43e98f54a18c8c686f622a8da95a7669a
3
+ size 7096
zero_to_fp32.py ADDED
@@ -0,0 +1,604 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python
2
+
3
+ # Copyright (c) Microsoft Corporation.
4
+ # SPDX-License-Identifier: Apache-2.0
5
+
6
+ # DeepSpeed Team
7
+
8
+ # This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
9
+ # copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
10
+ # the future. Once extracted, the weights don't require DeepSpeed and can be used in any
11
+ # application.
12
+ #
13
+ # example: python zero_to_fp32.py . pytorch_model.bin
14
+
15
+ import argparse
16
+ import torch
17
+ import glob
18
+ import math
19
+ import os
20
+ import re
21
+ from collections import OrderedDict
22
+ from dataclasses import dataclass
23
+
24
+ # while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
25
+ # DeepSpeed data structures it has to be available in the current python environment.
26
+ from deepspeed.utils import logger
27
+ from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
28
+ FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
29
+ FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
30
+
31
+
32
+ @dataclass
33
+ class zero_model_state:
34
+ buffers: dict()
35
+ param_shapes: dict()
36
+ shared_params: list
37
+ ds_version: int
38
+ frozen_param_shapes: dict()
39
+ frozen_param_fragments: dict()
40
+
41
+
42
+ debug = 0
43
+
44
+ # load to cpu
45
+ device = torch.device('cpu')
46
+
47
+
48
+ def atoi(text):
49
+ return int(text) if text.isdigit() else text
50
+
51
+
52
+ def natural_keys(text):
53
+ '''
54
+ alist.sort(key=natural_keys) sorts in human order
55
+ http://nedbatchelder.com/blog/200712/human_sorting.html
56
+ (See Toothy's implementation in the comments)
57
+ '''
58
+ return [atoi(c) for c in re.split(r'(\d+)', text)]
59
+
60
+
61
+ def get_model_state_file(checkpoint_dir, zero_stage):
62
+ if not os.path.isdir(checkpoint_dir):
63
+ raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
64
+
65
+ # there should be only one file
66
+ if zero_stage <= 2:
67
+ file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
68
+ elif zero_stage == 3:
69
+ file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
70
+
71
+ if not os.path.exists(file):
72
+ raise FileNotFoundError(f"can't find model states file at '{file}'")
73
+
74
+ return file
75
+
76
+
77
+ def get_checkpoint_files(checkpoint_dir, glob_pattern):
78
+ # XXX: need to test that this simple glob rule works for multi-node setup too
79
+ ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
80
+
81
+ if len(ckpt_files) == 0:
82
+ raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
83
+
84
+ return ckpt_files
85
+
86
+
87
+ def get_optim_files(checkpoint_dir):
88
+ return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
89
+
90
+
91
+ def get_model_state_files(checkpoint_dir):
92
+ return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
93
+
94
+
95
+ def parse_model_states(files):
96
+ zero_model_states = []
97
+ for file in files:
98
+ state_dict = torch.load(file, map_location=device)
99
+
100
+ if BUFFER_NAMES not in state_dict:
101
+ raise ValueError(f"{file} is not a model state checkpoint")
102
+ buffer_names = state_dict[BUFFER_NAMES]
103
+ if debug:
104
+ print("Found buffers:", buffer_names)
105
+
106
+ # recover just the buffers while restoring them to fp32 if they were saved in fp16
107
+ buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
108
+ param_shapes = state_dict[PARAM_SHAPES]
109
+
110
+ # collect parameters that are included in param_shapes
111
+ param_names = []
112
+ for s in param_shapes:
113
+ for name in s.keys():
114
+ param_names.append(name)
115
+
116
+ # update with frozen parameters
117
+ frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
118
+ if frozen_param_shapes is not None:
119
+ if debug:
120
+ print(f"Found frozen_param_shapes: {frozen_param_shapes}")
121
+ param_names += list(frozen_param_shapes.keys())
122
+
123
+ # handle shared params
124
+ shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
125
+
126
+ ds_version = state_dict.get(DS_VERSION, None)
127
+
128
+ frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
129
+
130
+ z_model_state = zero_model_state(buffers=buffers,
131
+ param_shapes=param_shapes,
132
+ shared_params=shared_params,
133
+ ds_version=ds_version,
134
+ frozen_param_shapes=frozen_param_shapes,
135
+ frozen_param_fragments=frozen_param_fragments)
136
+ zero_model_states.append(z_model_state)
137
+
138
+ return zero_model_states
139
+
140
+
141
+ def parse_optim_states(files, ds_checkpoint_dir):
142
+
143
+ total_files = len(files)
144
+ state_dicts = []
145
+ for f in files:
146
+ state_dict = torch.load(f, map_location=device)
147
+ # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
148
+ # and also handle the case where it was already removed by another helper script
149
+ state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
150
+ state_dicts.append(state_dict)
151
+
152
+ if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
153
+ raise ValueError(f"{files[0]} is not a zero checkpoint")
154
+ zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
155
+ world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
156
+
157
+ # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
158
+ # parameters can be different from data parallelism for non-expert parameters. So we can just
159
+ # use the max of the partition_count to get the dp world_size.
160
+
161
+ if type(world_size) is list:
162
+ world_size = max(world_size)
163
+
164
+ if world_size != total_files:
165
+ raise ValueError(
166
+ f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
167
+ "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
168
+ )
169
+
170
+ # the groups are named differently in each stage
171
+ if zero_stage <= 2:
172
+ fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
173
+ elif zero_stage == 3:
174
+ fp32_groups_key = FP32_FLAT_GROUPS
175
+ else:
176
+ raise ValueError(f"unknown zero stage {zero_stage}")
177
+
178
+ if zero_stage <= 2:
179
+ fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
180
+ elif zero_stage == 3:
181
+ # if there is more than one param group, there will be multiple flattened tensors - one
182
+ # flattened tensor per group - for simplicity merge them into a single tensor
183
+ #
184
+ # XXX: could make the script more memory efficient for when there are multiple groups - it
185
+ # will require matching the sub-lists of param_shapes for each param group flattened tensor
186
+
187
+ fp32_flat_groups = [
188
+ torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
189
+ ]
190
+
191
+ return zero_stage, world_size, fp32_flat_groups
192
+
193
+
194
+ def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
195
+ """
196
+ Returns fp32 state_dict reconstructed from ds checkpoint
197
+
198
+ Args:
199
+ - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
200
+
201
+ """
202
+ print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
203
+
204
+ optim_files = get_optim_files(ds_checkpoint_dir)
205
+ zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
206
+ print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
207
+
208
+ model_files = get_model_state_files(ds_checkpoint_dir)
209
+
210
+ zero_model_states = parse_model_states(model_files)
211
+ print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
212
+
213
+ if zero_stage <= 2:
214
+ return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
215
+ exclude_frozen_parameters)
216
+ elif zero_stage == 3:
217
+ return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
218
+ exclude_frozen_parameters)
219
+
220
+
221
+ def _zero2_merge_frozen_params(state_dict, zero_model_states):
222
+ if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
223
+ return
224
+
225
+ frozen_param_shapes = zero_model_states[0].frozen_param_shapes
226
+ frozen_param_fragments = zero_model_states[0].frozen_param_fragments
227
+
228
+ if debug:
229
+ num_elem = sum(s.numel() for s in frozen_param_shapes.values())
230
+ print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
231
+
232
+ wanted_params = len(frozen_param_shapes)
233
+ wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
234
+ avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
235
+ print(f'Frozen params: Have {avail_numel} numels to process.')
236
+ print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
237
+
238
+ total_params = 0
239
+ total_numel = 0
240
+ for name, shape in frozen_param_shapes.items():
241
+ total_params += 1
242
+ unpartitioned_numel = shape.numel()
243
+ total_numel += unpartitioned_numel
244
+
245
+ state_dict[name] = frozen_param_fragments[name]
246
+
247
+ if debug:
248
+ print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
249
+
250
+ print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
251
+
252
+
253
+ def _has_callable(obj, fn):
254
+ attr = getattr(obj, fn, None)
255
+ return callable(attr)
256
+
257
+
258
+ def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
259
+ param_shapes = zero_model_states[0].param_shapes
260
+
261
+ # Reconstruction protocol:
262
+ #
263
+ # XXX: document this
264
+
265
+ if debug:
266
+ for i in range(world_size):
267
+ for j in range(len(fp32_flat_groups[0])):
268
+ print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
269
+
270
+ # XXX: memory usage doubles here (zero2)
271
+ num_param_groups = len(fp32_flat_groups[0])
272
+ merged_single_partition_of_fp32_groups = []
273
+ for i in range(num_param_groups):
274
+ merged_partitions = [sd[i] for sd in fp32_flat_groups]
275
+ full_single_fp32_vector = torch.cat(merged_partitions, 0)
276
+ merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
277
+ avail_numel = sum(
278
+ [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
279
+
280
+ if debug:
281
+ wanted_params = sum([len(shapes) for shapes in param_shapes])
282
+ wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
283
+ # not asserting if there is a mismatch due to possible padding
284
+ print(f"Have {avail_numel} numels to process.")
285
+ print(f"Need {wanted_numel} numels in {wanted_params} params.")
286
+
287
+ # params
288
+ # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
289
+ # out-of-core computing solution
290
+ total_numel = 0
291
+ total_params = 0
292
+ for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
293
+ offset = 0
294
+ avail_numel = full_single_fp32_vector.numel()
295
+ for name, shape in shapes.items():
296
+
297
+ unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
298
+ total_numel += unpartitioned_numel
299
+ total_params += 1
300
+
301
+ if debug:
302
+ print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
303
+ state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
304
+ offset += unpartitioned_numel
305
+
306
+ # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
307
+ # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
308
+ # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
309
+ # live optimizer object, so we are checking that the numbers are within the right range
310
+ align_to = 2 * world_size
311
+
312
+ def zero2_align(x):
313
+ return align_to * math.ceil(x / align_to)
314
+
315
+ if debug:
316
+ print(f"original offset={offset}, avail_numel={avail_numel}")
317
+
318
+ offset = zero2_align(offset)
319
+ avail_numel = zero2_align(avail_numel)
320
+
321
+ if debug:
322
+ print(f"aligned offset={offset}, avail_numel={avail_numel}")
323
+
324
+ # Sanity check
325
+ if offset != avail_numel:
326
+ raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
327
+
328
+ print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
329
+
330
+
331
+ def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
332
+ exclude_frozen_parameters):
333
+ state_dict = OrderedDict()
334
+
335
+ # buffers
336
+ buffers = zero_model_states[0].buffers
337
+ state_dict.update(buffers)
338
+ if debug:
339
+ print(f"added {len(buffers)} buffers")
340
+
341
+ if not exclude_frozen_parameters:
342
+ _zero2_merge_frozen_params(state_dict, zero_model_states)
343
+
344
+ _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
345
+
346
+ # recover shared parameters
347
+ for pair in zero_model_states[0].shared_params:
348
+ if pair[1] in state_dict:
349
+ state_dict[pair[0]] = state_dict[pair[1]]
350
+
351
+ return state_dict
352
+
353
+
354
+ def zero3_partitioned_param_info(unpartitioned_numel, world_size):
355
+ remainder = unpartitioned_numel % world_size
356
+ padding_numel = (world_size - remainder) if remainder else 0
357
+ partitioned_numel = math.ceil(unpartitioned_numel / world_size)
358
+ return partitioned_numel, padding_numel
359
+
360
+
361
+ def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
362
+ if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
363
+ return
364
+
365
+ if debug:
366
+ for i in range(world_size):
367
+ num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
368
+ print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
369
+
370
+ frozen_param_shapes = zero_model_states[0].frozen_param_shapes
371
+ wanted_params = len(frozen_param_shapes)
372
+ wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
373
+ avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
374
+ print(f'Frozen params: Have {avail_numel} numels to process.')
375
+ print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
376
+
377
+ total_params = 0
378
+ total_numel = 0
379
+ for name, shape in zero_model_states[0].frozen_param_shapes.items():
380
+ total_params += 1
381
+ unpartitioned_numel = shape.numel()
382
+ total_numel += unpartitioned_numel
383
+
384
+ param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
385
+ state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
386
+
387
+ partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
388
+
389
+ if debug:
390
+ print(
391
+ f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
392
+ )
393
+
394
+ print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
395
+
396
+
397
+ def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
398
+ param_shapes = zero_model_states[0].param_shapes
399
+ avail_numel = fp32_flat_groups[0].numel() * world_size
400
+ # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
401
+ # param, re-consolidating each param, while dealing with padding if any
402
+
403
+ # merge list of dicts, preserving order
404
+ param_shapes = {k: v for d in param_shapes for k, v in d.items()}
405
+
406
+ if debug:
407
+ for i in range(world_size):
408
+ print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
409
+
410
+ wanted_params = len(param_shapes)
411
+ wanted_numel = sum(shape.numel() for shape in param_shapes.values())
412
+ # not asserting if there is a mismatch due to possible padding
413
+ avail_numel = fp32_flat_groups[0].numel() * world_size
414
+ print(f"Trainable params: Have {avail_numel} numels to process.")
415
+ print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
416
+
417
+ # params
418
+ # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
419
+ # out-of-core computing solution
420
+ offset = 0
421
+ total_numel = 0
422
+ total_params = 0
423
+ for name, shape in param_shapes.items():
424
+
425
+ unpartitioned_numel = shape.numel()
426
+ total_numel += unpartitioned_numel
427
+ total_params += 1
428
+
429
+ partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
430
+
431
+ if debug:
432
+ print(
433
+ f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
434
+ )
435
+
436
+ # XXX: memory usage doubles here
437
+ state_dict[name] = torch.cat(
438
+ tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
439
+ 0).narrow(0, 0, unpartitioned_numel).view(shape)
440
+ offset += partitioned_numel
441
+
442
+ offset *= world_size
443
+
444
+ # Sanity check
445
+ if offset != avail_numel:
446
+ raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
447
+
448
+ print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
449
+
450
+
451
+ def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
452
+ exclude_frozen_parameters):
453
+ state_dict = OrderedDict()
454
+
455
+ # buffers
456
+ buffers = zero_model_states[0].buffers
457
+ state_dict.update(buffers)
458
+ if debug:
459
+ print(f"added {len(buffers)} buffers")
460
+
461
+ if not exclude_frozen_parameters:
462
+ _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
463
+
464
+ _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
465
+
466
+ # recover shared parameters
467
+ for pair in zero_model_states[0].shared_params:
468
+ if pair[1] in state_dict:
469
+ state_dict[pair[0]] = state_dict[pair[1]]
470
+
471
+ return state_dict
472
+
473
+
474
+ def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None, exclude_frozen_parameters=False):
475
+ """
476
+ Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
477
+ ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
478
+ via a model hub.
479
+
480
+ Args:
481
+ - ``checkpoint_dir``: path to the desired checkpoint folder
482
+ - ``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``
483
+ - ``exclude_frozen_parameters``: exclude frozen parameters
484
+
485
+ Returns:
486
+ - pytorch ``state_dict``
487
+
488
+ Note: this approach may not work if your application doesn't have sufficient free CPU memory and
489
+ you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
490
+ the checkpoint.
491
+
492
+ A typical usage might be ::
493
+
494
+ from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
495
+ # do the training and checkpoint saving
496
+ state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
497
+ model = model.cpu() # move to cpu
498
+ model.load_state_dict(state_dict)
499
+ # submit to model hub or save the model to share with others
500
+
501
+ In this example the ``model`` will no longer be usable in the deepspeed context of the same
502
+ application. i.e. you will need to re-initialize the deepspeed engine, since
503
+ ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
504
+
505
+ If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
506
+
507
+ """
508
+ if tag is None:
509
+ latest_path = os.path.join(checkpoint_dir, 'latest')
510
+ if os.path.isfile(latest_path):
511
+ with open(latest_path, 'r') as fd:
512
+ tag = fd.read().strip()
513
+ else:
514
+ raise ValueError(f"Unable to find 'latest' file at {latest_path}")
515
+
516
+ ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
517
+
518
+ if not os.path.isdir(ds_checkpoint_dir):
519
+ raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
520
+
521
+ return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
522
+
523
+
524
+ def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None, exclude_frozen_parameters=False):
525
+ """
526
+ Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
527
+ loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
528
+
529
+ Args:
530
+ - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
531
+ - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
532
+ - ``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``
533
+ - ``exclude_frozen_parameters``: exclude frozen parameters
534
+ """
535
+
536
+ state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag, exclude_frozen_parameters)
537
+ print(f"Saving fp32 state dict to {output_file}")
538
+ torch.save(state_dict, output_file)
539
+
540
+
541
+ def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
542
+ """
543
+ 1. Put the provided model to cpu
544
+ 2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
545
+ 3. Load it into the provided model
546
+
547
+ Args:
548
+ - ``model``: the model object to update
549
+ - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
550
+ - ``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``
551
+
552
+ Returns:
553
+ - ``model`: modified model
554
+
555
+ Make sure you have plenty of CPU memory available before you call this function. If you don't
556
+ have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
557
+ conveniently placed for you in the checkpoint folder.
558
+
559
+ A typical usage might be ::
560
+
561
+ from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
562
+ model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
563
+ # submit to model hub or save the model to share with others
564
+
565
+ Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
566
+ of the same application. i.e. you will need to re-initialize the deepspeed engine, since
567
+ ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
568
+
569
+ """
570
+ logger.info(f"Extracting fp32 weights")
571
+ state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
572
+
573
+ logger.info(f"Overwriting model with fp32 weights")
574
+ model = model.cpu()
575
+ model.load_state_dict(state_dict, strict=False)
576
+
577
+ return model
578
+
579
+
580
+ if __name__ == "__main__":
581
+
582
+ parser = argparse.ArgumentParser()
583
+ parser.add_argument("checkpoint_dir",
584
+ type=str,
585
+ help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
586
+ parser.add_argument(
587
+ "output_file",
588
+ type=str,
589
+ help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
590
+ parser.add_argument("-t",
591
+ "--tag",
592
+ type=str,
593
+ default=None,
594
+ help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
595
+ parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
596
+ parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
597
+ args = parser.parse_args()
598
+
599
+ debug = args.debug
600
+
601
+ convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
602
+ args.output_file,
603
+ tag=args.tag,
604
+ exclude_frozen_parameters=args.exclude_frozen_parameters)