arianhosseini commited on
Commit
8d63b93
1 Parent(s): b088211

Training in progress, step 400, checkpoint

Browse files
checkpoint-400/config.json ADDED
@@ -0,0 +1,32 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "EleutherAI/pythia-2.8b",
3
+ "architectures": [
4
+ "GPTNeoForMultipleChoice"
5
+ ],
6
+ "attention_bias": true,
7
+ "attention_dropout": 0.0,
8
+ "bos_token_id": 0,
9
+ "classifier_dropout": 0.1,
10
+ "eos_token_id": 0,
11
+ "hidden_act": "gelu",
12
+ "hidden_dropout": 0.0,
13
+ "hidden_size": 2560,
14
+ "initializer_range": 0.02,
15
+ "intermediate_size": 10240,
16
+ "layer_norm_eps": 1e-05,
17
+ "max_length": 1024,
18
+ "max_position_embeddings": 2048,
19
+ "model_type": "gpt_neox",
20
+ "num_attention_heads": 32,
21
+ "num_hidden_layers": 32,
22
+ "pad_token_id": 0,
23
+ "rope_scaling": null,
24
+ "rotary_emb_base": 10000,
25
+ "rotary_pct": 0.25,
26
+ "tie_word_embeddings": false,
27
+ "torch_dtype": "float16",
28
+ "transformers_version": "4.41.0",
29
+ "use_cache": true,
30
+ "use_parallel_residual": true,
31
+ "vocab_size": 50304
32
+ }
checkpoint-400/global_step400/mp_rank_00_model_states.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ea95116d051d73cf5f427087957b0f9c60257bfe4dbbc34634ec1eb61beb536d
3
+ size 5292979768
checkpoint-400/global_step400/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 sha256:869f526b5be820f293af34ea00dcba18de01cae947d7c27579f9c75033651667
3
+ size 15878620944
checkpoint-400/latest ADDED
@@ -0,0 +1 @@
 
 
1
+ global_step400
checkpoint-400/model-00001-of-00002.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4ac6368dcde9d02e930c04ebd844ba5099576bafd8fb3fd893390e655ea95d35
3
+ size 4978208880
checkpoint-400/model-00002-of-00002.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3ded594c2887b79dde8178fbdcdc2753585954d49812f4648dc089dee6bba4d2
3
+ size 314703498
checkpoint-400/model.safetensors.index.json ADDED
@@ -0,0 +1,396 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "metadata": {
3
+ "total_size": 5292866562
4
+ },
5
+ "weight_map": {
6
+ "classifier.bias": "model-00002-of-00002.safetensors",
7
+ "classifier.weight": "model-00002-of-00002.safetensors",
8
+ "gpt_neox.embed_in.weight": "model-00001-of-00002.safetensors",
9
+ "gpt_neox.final_layer_norm.bias": "model-00002-of-00002.safetensors",
10
+ "gpt_neox.final_layer_norm.weight": "model-00002-of-00002.safetensors",
11
+ "gpt_neox.layers.0.attention.dense.bias": "model-00001-of-00002.safetensors",
12
+ "gpt_neox.layers.0.attention.dense.weight": "model-00001-of-00002.safetensors",
13
+ "gpt_neox.layers.0.attention.query_key_value.bias": "model-00001-of-00002.safetensors",
14
+ "gpt_neox.layers.0.attention.query_key_value.weight": "model-00001-of-00002.safetensors",
15
+ "gpt_neox.layers.0.input_layernorm.bias": "model-00001-of-00002.safetensors",
16
+ "gpt_neox.layers.0.input_layernorm.weight": "model-00001-of-00002.safetensors",
17
+ "gpt_neox.layers.0.mlp.dense_4h_to_h.bias": "model-00001-of-00002.safetensors",
18
+ "gpt_neox.layers.0.mlp.dense_4h_to_h.weight": "model-00001-of-00002.safetensors",
19
+ "gpt_neox.layers.0.mlp.dense_h_to_4h.bias": "model-00001-of-00002.safetensors",
20
+ "gpt_neox.layers.0.mlp.dense_h_to_4h.weight": "model-00001-of-00002.safetensors",
21
+ "gpt_neox.layers.0.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
22
+ "gpt_neox.layers.0.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
23
+ "gpt_neox.layers.1.attention.dense.bias": "model-00001-of-00002.safetensors",
24
+ "gpt_neox.layers.1.attention.dense.weight": "model-00001-of-00002.safetensors",
25
+ "gpt_neox.layers.1.attention.query_key_value.bias": "model-00001-of-00002.safetensors",
26
+ "gpt_neox.layers.1.attention.query_key_value.weight": "model-00001-of-00002.safetensors",
27
+ "gpt_neox.layers.1.input_layernorm.bias": "model-00001-of-00002.safetensors",
28
+ "gpt_neox.layers.1.input_layernorm.weight": "model-00001-of-00002.safetensors",
29
+ "gpt_neox.layers.1.mlp.dense_4h_to_h.bias": "model-00001-of-00002.safetensors",
30
+ "gpt_neox.layers.1.mlp.dense_4h_to_h.weight": "model-00001-of-00002.safetensors",
31
+ "gpt_neox.layers.1.mlp.dense_h_to_4h.bias": "model-00001-of-00002.safetensors",
32
+ "gpt_neox.layers.1.mlp.dense_h_to_4h.weight": "model-00001-of-00002.safetensors",
33
+ "gpt_neox.layers.1.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
34
+ "gpt_neox.layers.1.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
35
+ "gpt_neox.layers.10.attention.dense.bias": "model-00001-of-00002.safetensors",
36
+ "gpt_neox.layers.10.attention.dense.weight": "model-00001-of-00002.safetensors",
37
+ "gpt_neox.layers.10.attention.query_key_value.bias": "model-00001-of-00002.safetensors",
38
+ "gpt_neox.layers.10.attention.query_key_value.weight": "model-00001-of-00002.safetensors",
39
+ "gpt_neox.layers.10.input_layernorm.bias": "model-00001-of-00002.safetensors",
40
+ "gpt_neox.layers.10.input_layernorm.weight": "model-00001-of-00002.safetensors",
41
+ "gpt_neox.layers.10.mlp.dense_4h_to_h.bias": "model-00001-of-00002.safetensors",
42
+ "gpt_neox.layers.10.mlp.dense_4h_to_h.weight": "model-00001-of-00002.safetensors",
43
+ "gpt_neox.layers.10.mlp.dense_h_to_4h.bias": "model-00001-of-00002.safetensors",
44
+ "gpt_neox.layers.10.mlp.dense_h_to_4h.weight": "model-00001-of-00002.safetensors",
45
+ "gpt_neox.layers.10.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
46
+ "gpt_neox.layers.10.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
47
+ "gpt_neox.layers.11.attention.dense.bias": "model-00001-of-00002.safetensors",
48
+ "gpt_neox.layers.11.attention.dense.weight": "model-00001-of-00002.safetensors",
49
+ "gpt_neox.layers.11.attention.query_key_value.bias": "model-00001-of-00002.safetensors",
50
+ "gpt_neox.layers.11.attention.query_key_value.weight": "model-00001-of-00002.safetensors",
51
+ "gpt_neox.layers.11.input_layernorm.bias": "model-00001-of-00002.safetensors",
52
+ "gpt_neox.layers.11.input_layernorm.weight": "model-00001-of-00002.safetensors",
53
+ "gpt_neox.layers.11.mlp.dense_4h_to_h.bias": "model-00001-of-00002.safetensors",
54
+ "gpt_neox.layers.11.mlp.dense_4h_to_h.weight": "model-00001-of-00002.safetensors",
55
+ "gpt_neox.layers.11.mlp.dense_h_to_4h.bias": "model-00001-of-00002.safetensors",
56
+ "gpt_neox.layers.11.mlp.dense_h_to_4h.weight": "model-00001-of-00002.safetensors",
57
+ "gpt_neox.layers.11.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
58
+ "gpt_neox.layers.11.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
59
+ "gpt_neox.layers.12.attention.dense.bias": "model-00001-of-00002.safetensors",
60
+ "gpt_neox.layers.12.attention.dense.weight": "model-00001-of-00002.safetensors",
61
+ "gpt_neox.layers.12.attention.query_key_value.bias": "model-00001-of-00002.safetensors",
62
+ "gpt_neox.layers.12.attention.query_key_value.weight": "model-00001-of-00002.safetensors",
63
+ "gpt_neox.layers.12.input_layernorm.bias": "model-00001-of-00002.safetensors",
64
+ "gpt_neox.layers.12.input_layernorm.weight": "model-00001-of-00002.safetensors",
65
+ "gpt_neox.layers.12.mlp.dense_4h_to_h.bias": "model-00001-of-00002.safetensors",
66
+ "gpt_neox.layers.12.mlp.dense_4h_to_h.weight": "model-00001-of-00002.safetensors",
67
+ "gpt_neox.layers.12.mlp.dense_h_to_4h.bias": "model-00001-of-00002.safetensors",
68
+ "gpt_neox.layers.12.mlp.dense_h_to_4h.weight": "model-00001-of-00002.safetensors",
69
+ "gpt_neox.layers.12.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
70
+ "gpt_neox.layers.12.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
71
+ "gpt_neox.layers.13.attention.dense.bias": "model-00001-of-00002.safetensors",
72
+ "gpt_neox.layers.13.attention.dense.weight": "model-00001-of-00002.safetensors",
73
+ "gpt_neox.layers.13.attention.query_key_value.bias": "model-00001-of-00002.safetensors",
74
+ "gpt_neox.layers.13.attention.query_key_value.weight": "model-00001-of-00002.safetensors",
75
+ "gpt_neox.layers.13.input_layernorm.bias": "model-00001-of-00002.safetensors",
76
+ "gpt_neox.layers.13.input_layernorm.weight": "model-00001-of-00002.safetensors",
77
+ "gpt_neox.layers.13.mlp.dense_4h_to_h.bias": "model-00001-of-00002.safetensors",
78
+ "gpt_neox.layers.13.mlp.dense_4h_to_h.weight": "model-00001-of-00002.safetensors",
79
+ "gpt_neox.layers.13.mlp.dense_h_to_4h.bias": "model-00001-of-00002.safetensors",
80
+ "gpt_neox.layers.13.mlp.dense_h_to_4h.weight": "model-00001-of-00002.safetensors",
81
+ "gpt_neox.layers.13.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
82
+ "gpt_neox.layers.13.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
83
+ "gpt_neox.layers.14.attention.dense.bias": "model-00001-of-00002.safetensors",
84
+ "gpt_neox.layers.14.attention.dense.weight": "model-00001-of-00002.safetensors",
85
+ "gpt_neox.layers.14.attention.query_key_value.bias": "model-00001-of-00002.safetensors",
86
+ "gpt_neox.layers.14.attention.query_key_value.weight": "model-00001-of-00002.safetensors",
87
+ "gpt_neox.layers.14.input_layernorm.bias": "model-00001-of-00002.safetensors",
88
+ "gpt_neox.layers.14.input_layernorm.weight": "model-00001-of-00002.safetensors",
89
+ "gpt_neox.layers.14.mlp.dense_4h_to_h.bias": "model-00001-of-00002.safetensors",
90
+ "gpt_neox.layers.14.mlp.dense_4h_to_h.weight": "model-00001-of-00002.safetensors",
91
+ "gpt_neox.layers.14.mlp.dense_h_to_4h.bias": "model-00001-of-00002.safetensors",
92
+ "gpt_neox.layers.14.mlp.dense_h_to_4h.weight": "model-00001-of-00002.safetensors",
93
+ "gpt_neox.layers.14.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
94
+ "gpt_neox.layers.14.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
95
+ "gpt_neox.layers.15.attention.dense.bias": "model-00001-of-00002.safetensors",
96
+ "gpt_neox.layers.15.attention.dense.weight": "model-00001-of-00002.safetensors",
97
+ "gpt_neox.layers.15.attention.query_key_value.bias": "model-00001-of-00002.safetensors",
98
+ "gpt_neox.layers.15.attention.query_key_value.weight": "model-00001-of-00002.safetensors",
99
+ "gpt_neox.layers.15.input_layernorm.bias": "model-00001-of-00002.safetensors",
100
+ "gpt_neox.layers.15.input_layernorm.weight": "model-00001-of-00002.safetensors",
101
+ "gpt_neox.layers.15.mlp.dense_4h_to_h.bias": "model-00001-of-00002.safetensors",
102
+ "gpt_neox.layers.15.mlp.dense_4h_to_h.weight": "model-00001-of-00002.safetensors",
103
+ "gpt_neox.layers.15.mlp.dense_h_to_4h.bias": "model-00001-of-00002.safetensors",
104
+ "gpt_neox.layers.15.mlp.dense_h_to_4h.weight": "model-00001-of-00002.safetensors",
105
+ "gpt_neox.layers.15.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
106
+ "gpt_neox.layers.15.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
107
+ "gpt_neox.layers.16.attention.dense.bias": "model-00001-of-00002.safetensors",
108
+ "gpt_neox.layers.16.attention.dense.weight": "model-00001-of-00002.safetensors",
109
+ "gpt_neox.layers.16.attention.query_key_value.bias": "model-00001-of-00002.safetensors",
110
+ "gpt_neox.layers.16.attention.query_key_value.weight": "model-00001-of-00002.safetensors",
111
+ "gpt_neox.layers.16.input_layernorm.bias": "model-00001-of-00002.safetensors",
112
+ "gpt_neox.layers.16.input_layernorm.weight": "model-00001-of-00002.safetensors",
113
+ "gpt_neox.layers.16.mlp.dense_4h_to_h.bias": "model-00001-of-00002.safetensors",
114
+ "gpt_neox.layers.16.mlp.dense_4h_to_h.weight": "model-00001-of-00002.safetensors",
115
+ "gpt_neox.layers.16.mlp.dense_h_to_4h.bias": "model-00001-of-00002.safetensors",
116
+ "gpt_neox.layers.16.mlp.dense_h_to_4h.weight": "model-00001-of-00002.safetensors",
117
+ "gpt_neox.layers.16.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
118
+ "gpt_neox.layers.16.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
119
+ "gpt_neox.layers.17.attention.dense.bias": "model-00001-of-00002.safetensors",
120
+ "gpt_neox.layers.17.attention.dense.weight": "model-00001-of-00002.safetensors",
121
+ "gpt_neox.layers.17.attention.query_key_value.bias": "model-00001-of-00002.safetensors",
122
+ "gpt_neox.layers.17.attention.query_key_value.weight": "model-00001-of-00002.safetensors",
123
+ "gpt_neox.layers.17.input_layernorm.bias": "model-00001-of-00002.safetensors",
124
+ "gpt_neox.layers.17.input_layernorm.weight": "model-00001-of-00002.safetensors",
125
+ "gpt_neox.layers.17.mlp.dense_4h_to_h.bias": "model-00001-of-00002.safetensors",
126
+ "gpt_neox.layers.17.mlp.dense_4h_to_h.weight": "model-00001-of-00002.safetensors",
127
+ "gpt_neox.layers.17.mlp.dense_h_to_4h.bias": "model-00001-of-00002.safetensors",
128
+ "gpt_neox.layers.17.mlp.dense_h_to_4h.weight": "model-00001-of-00002.safetensors",
129
+ "gpt_neox.layers.17.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
130
+ "gpt_neox.layers.17.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
131
+ "gpt_neox.layers.18.attention.dense.bias": "model-00001-of-00002.safetensors",
132
+ "gpt_neox.layers.18.attention.dense.weight": "model-00001-of-00002.safetensors",
133
+ "gpt_neox.layers.18.attention.query_key_value.bias": "model-00001-of-00002.safetensors",
134
+ "gpt_neox.layers.18.attention.query_key_value.weight": "model-00001-of-00002.safetensors",
135
+ "gpt_neox.layers.18.input_layernorm.bias": "model-00001-of-00002.safetensors",
136
+ "gpt_neox.layers.18.input_layernorm.weight": "model-00001-of-00002.safetensors",
137
+ "gpt_neox.layers.18.mlp.dense_4h_to_h.bias": "model-00001-of-00002.safetensors",
138
+ "gpt_neox.layers.18.mlp.dense_4h_to_h.weight": "model-00001-of-00002.safetensors",
139
+ "gpt_neox.layers.18.mlp.dense_h_to_4h.bias": "model-00001-of-00002.safetensors",
140
+ "gpt_neox.layers.18.mlp.dense_h_to_4h.weight": "model-00001-of-00002.safetensors",
141
+ "gpt_neox.layers.18.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
142
+ "gpt_neox.layers.18.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
143
+ "gpt_neox.layers.19.attention.dense.bias": "model-00001-of-00002.safetensors",
144
+ "gpt_neox.layers.19.attention.dense.weight": "model-00001-of-00002.safetensors",
145
+ "gpt_neox.layers.19.attention.query_key_value.bias": "model-00001-of-00002.safetensors",
146
+ "gpt_neox.layers.19.attention.query_key_value.weight": "model-00001-of-00002.safetensors",
147
+ "gpt_neox.layers.19.input_layernorm.bias": "model-00001-of-00002.safetensors",
148
+ "gpt_neox.layers.19.input_layernorm.weight": "model-00001-of-00002.safetensors",
149
+ "gpt_neox.layers.19.mlp.dense_4h_to_h.bias": "model-00001-of-00002.safetensors",
150
+ "gpt_neox.layers.19.mlp.dense_4h_to_h.weight": "model-00001-of-00002.safetensors",
151
+ "gpt_neox.layers.19.mlp.dense_h_to_4h.bias": "model-00001-of-00002.safetensors",
152
+ "gpt_neox.layers.19.mlp.dense_h_to_4h.weight": "model-00001-of-00002.safetensors",
153
+ "gpt_neox.layers.19.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
154
+ "gpt_neox.layers.19.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
155
+ "gpt_neox.layers.2.attention.dense.bias": "model-00001-of-00002.safetensors",
156
+ "gpt_neox.layers.2.attention.dense.weight": "model-00001-of-00002.safetensors",
157
+ "gpt_neox.layers.2.attention.query_key_value.bias": "model-00001-of-00002.safetensors",
158
+ "gpt_neox.layers.2.attention.query_key_value.weight": "model-00001-of-00002.safetensors",
159
+ "gpt_neox.layers.2.input_layernorm.bias": "model-00001-of-00002.safetensors",
160
+ "gpt_neox.layers.2.input_layernorm.weight": "model-00001-of-00002.safetensors",
161
+ "gpt_neox.layers.2.mlp.dense_4h_to_h.bias": "model-00001-of-00002.safetensors",
162
+ "gpt_neox.layers.2.mlp.dense_4h_to_h.weight": "model-00001-of-00002.safetensors",
163
+ "gpt_neox.layers.2.mlp.dense_h_to_4h.bias": "model-00001-of-00002.safetensors",
164
+ "gpt_neox.layers.2.mlp.dense_h_to_4h.weight": "model-00001-of-00002.safetensors",
165
+ "gpt_neox.layers.2.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
166
+ "gpt_neox.layers.2.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
167
+ "gpt_neox.layers.20.attention.dense.bias": "model-00001-of-00002.safetensors",
168
+ "gpt_neox.layers.20.attention.dense.weight": "model-00001-of-00002.safetensors",
169
+ "gpt_neox.layers.20.attention.query_key_value.bias": "model-00001-of-00002.safetensors",
170
+ "gpt_neox.layers.20.attention.query_key_value.weight": "model-00001-of-00002.safetensors",
171
+ "gpt_neox.layers.20.input_layernorm.bias": "model-00001-of-00002.safetensors",
172
+ "gpt_neox.layers.20.input_layernorm.weight": "model-00001-of-00002.safetensors",
173
+ "gpt_neox.layers.20.mlp.dense_4h_to_h.bias": "model-00001-of-00002.safetensors",
174
+ "gpt_neox.layers.20.mlp.dense_4h_to_h.weight": "model-00001-of-00002.safetensors",
175
+ "gpt_neox.layers.20.mlp.dense_h_to_4h.bias": "model-00001-of-00002.safetensors",
176
+ "gpt_neox.layers.20.mlp.dense_h_to_4h.weight": "model-00001-of-00002.safetensors",
177
+ "gpt_neox.layers.20.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
178
+ "gpt_neox.layers.20.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
179
+ "gpt_neox.layers.21.attention.dense.bias": "model-00001-of-00002.safetensors",
180
+ "gpt_neox.layers.21.attention.dense.weight": "model-00001-of-00002.safetensors",
181
+ "gpt_neox.layers.21.attention.query_key_value.bias": "model-00001-of-00002.safetensors",
182
+ "gpt_neox.layers.21.attention.query_key_value.weight": "model-00001-of-00002.safetensors",
183
+ "gpt_neox.layers.21.input_layernorm.bias": "model-00001-of-00002.safetensors",
184
+ "gpt_neox.layers.21.input_layernorm.weight": "model-00001-of-00002.safetensors",
185
+ "gpt_neox.layers.21.mlp.dense_4h_to_h.bias": "model-00001-of-00002.safetensors",
186
+ "gpt_neox.layers.21.mlp.dense_4h_to_h.weight": "model-00001-of-00002.safetensors",
187
+ "gpt_neox.layers.21.mlp.dense_h_to_4h.bias": "model-00001-of-00002.safetensors",
188
+ "gpt_neox.layers.21.mlp.dense_h_to_4h.weight": "model-00001-of-00002.safetensors",
189
+ "gpt_neox.layers.21.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
190
+ "gpt_neox.layers.21.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
191
+ "gpt_neox.layers.22.attention.dense.bias": "model-00001-of-00002.safetensors",
192
+ "gpt_neox.layers.22.attention.dense.weight": "model-00001-of-00002.safetensors",
193
+ "gpt_neox.layers.22.attention.query_key_value.bias": "model-00001-of-00002.safetensors",
194
+ "gpt_neox.layers.22.attention.query_key_value.weight": "model-00001-of-00002.safetensors",
195
+ "gpt_neox.layers.22.input_layernorm.bias": "model-00001-of-00002.safetensors",
196
+ "gpt_neox.layers.22.input_layernorm.weight": "model-00001-of-00002.safetensors",
197
+ "gpt_neox.layers.22.mlp.dense_4h_to_h.bias": "model-00001-of-00002.safetensors",
198
+ "gpt_neox.layers.22.mlp.dense_4h_to_h.weight": "model-00001-of-00002.safetensors",
199
+ "gpt_neox.layers.22.mlp.dense_h_to_4h.bias": "model-00001-of-00002.safetensors",
200
+ "gpt_neox.layers.22.mlp.dense_h_to_4h.weight": "model-00001-of-00002.safetensors",
201
+ "gpt_neox.layers.22.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
202
+ "gpt_neox.layers.22.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
203
+ "gpt_neox.layers.23.attention.dense.bias": "model-00001-of-00002.safetensors",
204
+ "gpt_neox.layers.23.attention.dense.weight": "model-00001-of-00002.safetensors",
205
+ "gpt_neox.layers.23.attention.query_key_value.bias": "model-00001-of-00002.safetensors",
206
+ "gpt_neox.layers.23.attention.query_key_value.weight": "model-00001-of-00002.safetensors",
207
+ "gpt_neox.layers.23.input_layernorm.bias": "model-00001-of-00002.safetensors",
208
+ "gpt_neox.layers.23.input_layernorm.weight": "model-00001-of-00002.safetensors",
209
+ "gpt_neox.layers.23.mlp.dense_4h_to_h.bias": "model-00001-of-00002.safetensors",
210
+ "gpt_neox.layers.23.mlp.dense_4h_to_h.weight": "model-00001-of-00002.safetensors",
211
+ "gpt_neox.layers.23.mlp.dense_h_to_4h.bias": "model-00001-of-00002.safetensors",
212
+ "gpt_neox.layers.23.mlp.dense_h_to_4h.weight": "model-00001-of-00002.safetensors",
213
+ "gpt_neox.layers.23.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
214
+ "gpt_neox.layers.23.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
215
+ "gpt_neox.layers.24.attention.dense.bias": "model-00001-of-00002.safetensors",
216
+ "gpt_neox.layers.24.attention.dense.weight": "model-00001-of-00002.safetensors",
217
+ "gpt_neox.layers.24.attention.query_key_value.bias": "model-00001-of-00002.safetensors",
218
+ "gpt_neox.layers.24.attention.query_key_value.weight": "model-00001-of-00002.safetensors",
219
+ "gpt_neox.layers.24.input_layernorm.bias": "model-00001-of-00002.safetensors",
220
+ "gpt_neox.layers.24.input_layernorm.weight": "model-00001-of-00002.safetensors",
221
+ "gpt_neox.layers.24.mlp.dense_4h_to_h.bias": "model-00001-of-00002.safetensors",
222
+ "gpt_neox.layers.24.mlp.dense_4h_to_h.weight": "model-00001-of-00002.safetensors",
223
+ "gpt_neox.layers.24.mlp.dense_h_to_4h.bias": "model-00001-of-00002.safetensors",
224
+ "gpt_neox.layers.24.mlp.dense_h_to_4h.weight": "model-00001-of-00002.safetensors",
225
+ "gpt_neox.layers.24.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
226
+ "gpt_neox.layers.24.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
227
+ "gpt_neox.layers.25.attention.dense.bias": "model-00001-of-00002.safetensors",
228
+ "gpt_neox.layers.25.attention.dense.weight": "model-00001-of-00002.safetensors",
229
+ "gpt_neox.layers.25.attention.query_key_value.bias": "model-00001-of-00002.safetensors",
230
+ "gpt_neox.layers.25.attention.query_key_value.weight": "model-00001-of-00002.safetensors",
231
+ "gpt_neox.layers.25.input_layernorm.bias": "model-00001-of-00002.safetensors",
232
+ "gpt_neox.layers.25.input_layernorm.weight": "model-00001-of-00002.safetensors",
233
+ "gpt_neox.layers.25.mlp.dense_4h_to_h.bias": "model-00001-of-00002.safetensors",
234
+ "gpt_neox.layers.25.mlp.dense_4h_to_h.weight": "model-00001-of-00002.safetensors",
235
+ "gpt_neox.layers.25.mlp.dense_h_to_4h.bias": "model-00001-of-00002.safetensors",
236
+ "gpt_neox.layers.25.mlp.dense_h_to_4h.weight": "model-00001-of-00002.safetensors",
237
+ "gpt_neox.layers.25.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
238
+ "gpt_neox.layers.25.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
239
+ "gpt_neox.layers.26.attention.dense.bias": "model-00001-of-00002.safetensors",
240
+ "gpt_neox.layers.26.attention.dense.weight": "model-00001-of-00002.safetensors",
241
+ "gpt_neox.layers.26.attention.query_key_value.bias": "model-00001-of-00002.safetensors",
242
+ "gpt_neox.layers.26.attention.query_key_value.weight": "model-00001-of-00002.safetensors",
243
+ "gpt_neox.layers.26.input_layernorm.bias": "model-00001-of-00002.safetensors",
244
+ "gpt_neox.layers.26.input_layernorm.weight": "model-00001-of-00002.safetensors",
245
+ "gpt_neox.layers.26.mlp.dense_4h_to_h.bias": "model-00001-of-00002.safetensors",
246
+ "gpt_neox.layers.26.mlp.dense_4h_to_h.weight": "model-00001-of-00002.safetensors",
247
+ "gpt_neox.layers.26.mlp.dense_h_to_4h.bias": "model-00001-of-00002.safetensors",
248
+ "gpt_neox.layers.26.mlp.dense_h_to_4h.weight": "model-00001-of-00002.safetensors",
249
+ "gpt_neox.layers.26.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
250
+ "gpt_neox.layers.26.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
251
+ "gpt_neox.layers.27.attention.dense.bias": "model-00001-of-00002.safetensors",
252
+ "gpt_neox.layers.27.attention.dense.weight": "model-00001-of-00002.safetensors",
253
+ "gpt_neox.layers.27.attention.query_key_value.bias": "model-00001-of-00002.safetensors",
254
+ "gpt_neox.layers.27.attention.query_key_value.weight": "model-00001-of-00002.safetensors",
255
+ "gpt_neox.layers.27.input_layernorm.bias": "model-00001-of-00002.safetensors",
256
+ "gpt_neox.layers.27.input_layernorm.weight": "model-00001-of-00002.safetensors",
257
+ "gpt_neox.layers.27.mlp.dense_4h_to_h.bias": "model-00001-of-00002.safetensors",
258
+ "gpt_neox.layers.27.mlp.dense_4h_to_h.weight": "model-00001-of-00002.safetensors",
259
+ "gpt_neox.layers.27.mlp.dense_h_to_4h.bias": "model-00001-of-00002.safetensors",
260
+ "gpt_neox.layers.27.mlp.dense_h_to_4h.weight": "model-00001-of-00002.safetensors",
261
+ "gpt_neox.layers.27.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
262
+ "gpt_neox.layers.27.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
263
+ "gpt_neox.layers.28.attention.dense.bias": "model-00001-of-00002.safetensors",
264
+ "gpt_neox.layers.28.attention.dense.weight": "model-00001-of-00002.safetensors",
265
+ "gpt_neox.layers.28.attention.query_key_value.bias": "model-00001-of-00002.safetensors",
266
+ "gpt_neox.layers.28.attention.query_key_value.weight": "model-00001-of-00002.safetensors",
267
+ "gpt_neox.layers.28.input_layernorm.bias": "model-00001-of-00002.safetensors",
268
+ "gpt_neox.layers.28.input_layernorm.weight": "model-00001-of-00002.safetensors",
269
+ "gpt_neox.layers.28.mlp.dense_4h_to_h.bias": "model-00001-of-00002.safetensors",
270
+ "gpt_neox.layers.28.mlp.dense_4h_to_h.weight": "model-00001-of-00002.safetensors",
271
+ "gpt_neox.layers.28.mlp.dense_h_to_4h.bias": "model-00001-of-00002.safetensors",
272
+ "gpt_neox.layers.28.mlp.dense_h_to_4h.weight": "model-00001-of-00002.safetensors",
273
+ "gpt_neox.layers.28.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
274
+ "gpt_neox.layers.28.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
275
+ "gpt_neox.layers.29.attention.dense.bias": "model-00001-of-00002.safetensors",
276
+ "gpt_neox.layers.29.attention.dense.weight": "model-00001-of-00002.safetensors",
277
+ "gpt_neox.layers.29.attention.query_key_value.bias": "model-00001-of-00002.safetensors",
278
+ "gpt_neox.layers.29.attention.query_key_value.weight": "model-00001-of-00002.safetensors",
279
+ "gpt_neox.layers.29.input_layernorm.bias": "model-00001-of-00002.safetensors",
280
+ "gpt_neox.layers.29.input_layernorm.weight": "model-00001-of-00002.safetensors",
281
+ "gpt_neox.layers.29.mlp.dense_4h_to_h.bias": "model-00001-of-00002.safetensors",
282
+ "gpt_neox.layers.29.mlp.dense_4h_to_h.weight": "model-00001-of-00002.safetensors",
283
+ "gpt_neox.layers.29.mlp.dense_h_to_4h.bias": "model-00001-of-00002.safetensors",
284
+ "gpt_neox.layers.29.mlp.dense_h_to_4h.weight": "model-00001-of-00002.safetensors",
285
+ "gpt_neox.layers.29.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
286
+ "gpt_neox.layers.29.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
287
+ "gpt_neox.layers.3.attention.dense.bias": "model-00001-of-00002.safetensors",
288
+ "gpt_neox.layers.3.attention.dense.weight": "model-00001-of-00002.safetensors",
289
+ "gpt_neox.layers.3.attention.query_key_value.bias": "model-00001-of-00002.safetensors",
290
+ "gpt_neox.layers.3.attention.query_key_value.weight": "model-00001-of-00002.safetensors",
291
+ "gpt_neox.layers.3.input_layernorm.bias": "model-00001-of-00002.safetensors",
292
+ "gpt_neox.layers.3.input_layernorm.weight": "model-00001-of-00002.safetensors",
293
+ "gpt_neox.layers.3.mlp.dense_4h_to_h.bias": "model-00001-of-00002.safetensors",
294
+ "gpt_neox.layers.3.mlp.dense_4h_to_h.weight": "model-00001-of-00002.safetensors",
295
+ "gpt_neox.layers.3.mlp.dense_h_to_4h.bias": "model-00001-of-00002.safetensors",
296
+ "gpt_neox.layers.3.mlp.dense_h_to_4h.weight": "model-00001-of-00002.safetensors",
297
+ "gpt_neox.layers.3.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
298
+ "gpt_neox.layers.3.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
299
+ "gpt_neox.layers.30.attention.dense.bias": "model-00002-of-00002.safetensors",
300
+ "gpt_neox.layers.30.attention.dense.weight": "model-00002-of-00002.safetensors",
301
+ "gpt_neox.layers.30.attention.query_key_value.bias": "model-00002-of-00002.safetensors",
302
+ "gpt_neox.layers.30.attention.query_key_value.weight": "model-00002-of-00002.safetensors",
303
+ "gpt_neox.layers.30.input_layernorm.bias": "model-00001-of-00002.safetensors",
304
+ "gpt_neox.layers.30.input_layernorm.weight": "model-00001-of-00002.safetensors",
305
+ "gpt_neox.layers.30.mlp.dense_4h_to_h.bias": "model-00002-of-00002.safetensors",
306
+ "gpt_neox.layers.30.mlp.dense_4h_to_h.weight": "model-00002-of-00002.safetensors",
307
+ "gpt_neox.layers.30.mlp.dense_h_to_4h.bias": "model-00002-of-00002.safetensors",
308
+ "gpt_neox.layers.30.mlp.dense_h_to_4h.weight": "model-00002-of-00002.safetensors",
309
+ "gpt_neox.layers.30.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
310
+ "gpt_neox.layers.30.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
311
+ "gpt_neox.layers.31.attention.dense.bias": "model-00002-of-00002.safetensors",
312
+ "gpt_neox.layers.31.attention.dense.weight": "model-00002-of-00002.safetensors",
313
+ "gpt_neox.layers.31.attention.query_key_value.bias": "model-00002-of-00002.safetensors",
314
+ "gpt_neox.layers.31.attention.query_key_value.weight": "model-00002-of-00002.safetensors",
315
+ "gpt_neox.layers.31.input_layernorm.bias": "model-00002-of-00002.safetensors",
316
+ "gpt_neox.layers.31.input_layernorm.weight": "model-00002-of-00002.safetensors",
317
+ "gpt_neox.layers.31.mlp.dense_4h_to_h.bias": "model-00002-of-00002.safetensors",
318
+ "gpt_neox.layers.31.mlp.dense_4h_to_h.weight": "model-00002-of-00002.safetensors",
319
+ "gpt_neox.layers.31.mlp.dense_h_to_4h.bias": "model-00002-of-00002.safetensors",
320
+ "gpt_neox.layers.31.mlp.dense_h_to_4h.weight": "model-00002-of-00002.safetensors",
321
+ "gpt_neox.layers.31.post_attention_layernorm.bias": "model-00002-of-00002.safetensors",
322
+ "gpt_neox.layers.31.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
323
+ "gpt_neox.layers.4.attention.dense.bias": "model-00001-of-00002.safetensors",
324
+ "gpt_neox.layers.4.attention.dense.weight": "model-00001-of-00002.safetensors",
325
+ "gpt_neox.layers.4.attention.query_key_value.bias": "model-00001-of-00002.safetensors",
326
+ "gpt_neox.layers.4.attention.query_key_value.weight": "model-00001-of-00002.safetensors",
327
+ "gpt_neox.layers.4.input_layernorm.bias": "model-00001-of-00002.safetensors",
328
+ "gpt_neox.layers.4.input_layernorm.weight": "model-00001-of-00002.safetensors",
329
+ "gpt_neox.layers.4.mlp.dense_4h_to_h.bias": "model-00001-of-00002.safetensors",
330
+ "gpt_neox.layers.4.mlp.dense_4h_to_h.weight": "model-00001-of-00002.safetensors",
331
+ "gpt_neox.layers.4.mlp.dense_h_to_4h.bias": "model-00001-of-00002.safetensors",
332
+ "gpt_neox.layers.4.mlp.dense_h_to_4h.weight": "model-00001-of-00002.safetensors",
333
+ "gpt_neox.layers.4.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
334
+ "gpt_neox.layers.4.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
335
+ "gpt_neox.layers.5.attention.dense.bias": "model-00001-of-00002.safetensors",
336
+ "gpt_neox.layers.5.attention.dense.weight": "model-00001-of-00002.safetensors",
337
+ "gpt_neox.layers.5.attention.query_key_value.bias": "model-00001-of-00002.safetensors",
338
+ "gpt_neox.layers.5.attention.query_key_value.weight": "model-00001-of-00002.safetensors",
339
+ "gpt_neox.layers.5.input_layernorm.bias": "model-00001-of-00002.safetensors",
340
+ "gpt_neox.layers.5.input_layernorm.weight": "model-00001-of-00002.safetensors",
341
+ "gpt_neox.layers.5.mlp.dense_4h_to_h.bias": "model-00001-of-00002.safetensors",
342
+ "gpt_neox.layers.5.mlp.dense_4h_to_h.weight": "model-00001-of-00002.safetensors",
343
+ "gpt_neox.layers.5.mlp.dense_h_to_4h.bias": "model-00001-of-00002.safetensors",
344
+ "gpt_neox.layers.5.mlp.dense_h_to_4h.weight": "model-00001-of-00002.safetensors",
345
+ "gpt_neox.layers.5.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
346
+ "gpt_neox.layers.5.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
347
+ "gpt_neox.layers.6.attention.dense.bias": "model-00001-of-00002.safetensors",
348
+ "gpt_neox.layers.6.attention.dense.weight": "model-00001-of-00002.safetensors",
349
+ "gpt_neox.layers.6.attention.query_key_value.bias": "model-00001-of-00002.safetensors",
350
+ "gpt_neox.layers.6.attention.query_key_value.weight": "model-00001-of-00002.safetensors",
351
+ "gpt_neox.layers.6.input_layernorm.bias": "model-00001-of-00002.safetensors",
352
+ "gpt_neox.layers.6.input_layernorm.weight": "model-00001-of-00002.safetensors",
353
+ "gpt_neox.layers.6.mlp.dense_4h_to_h.bias": "model-00001-of-00002.safetensors",
354
+ "gpt_neox.layers.6.mlp.dense_4h_to_h.weight": "model-00001-of-00002.safetensors",
355
+ "gpt_neox.layers.6.mlp.dense_h_to_4h.bias": "model-00001-of-00002.safetensors",
356
+ "gpt_neox.layers.6.mlp.dense_h_to_4h.weight": "model-00001-of-00002.safetensors",
357
+ "gpt_neox.layers.6.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
358
+ "gpt_neox.layers.6.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
359
+ "gpt_neox.layers.7.attention.dense.bias": "model-00001-of-00002.safetensors",
360
+ "gpt_neox.layers.7.attention.dense.weight": "model-00001-of-00002.safetensors",
361
+ "gpt_neox.layers.7.attention.query_key_value.bias": "model-00001-of-00002.safetensors",
362
+ "gpt_neox.layers.7.attention.query_key_value.weight": "model-00001-of-00002.safetensors",
363
+ "gpt_neox.layers.7.input_layernorm.bias": "model-00001-of-00002.safetensors",
364
+ "gpt_neox.layers.7.input_layernorm.weight": "model-00001-of-00002.safetensors",
365
+ "gpt_neox.layers.7.mlp.dense_4h_to_h.bias": "model-00001-of-00002.safetensors",
366
+ "gpt_neox.layers.7.mlp.dense_4h_to_h.weight": "model-00001-of-00002.safetensors",
367
+ "gpt_neox.layers.7.mlp.dense_h_to_4h.bias": "model-00001-of-00002.safetensors",
368
+ "gpt_neox.layers.7.mlp.dense_h_to_4h.weight": "model-00001-of-00002.safetensors",
369
+ "gpt_neox.layers.7.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
370
+ "gpt_neox.layers.7.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
371
+ "gpt_neox.layers.8.attention.dense.bias": "model-00001-of-00002.safetensors",
372
+ "gpt_neox.layers.8.attention.dense.weight": "model-00001-of-00002.safetensors",
373
+ "gpt_neox.layers.8.attention.query_key_value.bias": "model-00001-of-00002.safetensors",
374
+ "gpt_neox.layers.8.attention.query_key_value.weight": "model-00001-of-00002.safetensors",
375
+ "gpt_neox.layers.8.input_layernorm.bias": "model-00001-of-00002.safetensors",
376
+ "gpt_neox.layers.8.input_layernorm.weight": "model-00001-of-00002.safetensors",
377
+ "gpt_neox.layers.8.mlp.dense_4h_to_h.bias": "model-00001-of-00002.safetensors",
378
+ "gpt_neox.layers.8.mlp.dense_4h_to_h.weight": "model-00001-of-00002.safetensors",
379
+ "gpt_neox.layers.8.mlp.dense_h_to_4h.bias": "model-00001-of-00002.safetensors",
380
+ "gpt_neox.layers.8.mlp.dense_h_to_4h.weight": "model-00001-of-00002.safetensors",
381
+ "gpt_neox.layers.8.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
382
+ "gpt_neox.layers.8.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
383
+ "gpt_neox.layers.9.attention.dense.bias": "model-00001-of-00002.safetensors",
384
+ "gpt_neox.layers.9.attention.dense.weight": "model-00001-of-00002.safetensors",
385
+ "gpt_neox.layers.9.attention.query_key_value.bias": "model-00001-of-00002.safetensors",
386
+ "gpt_neox.layers.9.attention.query_key_value.weight": "model-00001-of-00002.safetensors",
387
+ "gpt_neox.layers.9.input_layernorm.bias": "model-00001-of-00002.safetensors",
388
+ "gpt_neox.layers.9.input_layernorm.weight": "model-00001-of-00002.safetensors",
389
+ "gpt_neox.layers.9.mlp.dense_4h_to_h.bias": "model-00001-of-00002.safetensors",
390
+ "gpt_neox.layers.9.mlp.dense_4h_to_h.weight": "model-00001-of-00002.safetensors",
391
+ "gpt_neox.layers.9.mlp.dense_h_to_4h.bias": "model-00001-of-00002.safetensors",
392
+ "gpt_neox.layers.9.mlp.dense_h_to_4h.weight": "model-00001-of-00002.safetensors",
393
+ "gpt_neox.layers.9.post_attention_layernorm.bias": "model-00001-of-00002.safetensors",
394
+ "gpt_neox.layers.9.post_attention_layernorm.weight": "model-00001-of-00002.safetensors"
395
+ }
396
+ }
checkpoint-400/rng_state_0.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ac8e33d1175d0decde28a8b9fbdfbd68d1ab0032d6e7000c9f9e2eebcc22e2a2
3
+ size 15152
checkpoint-400/scheduler.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:973239cd0eeb6221c8b6392b3d261581765ee3c5d4b47eea8506b1af3fec437a
3
+ size 1064
checkpoint-400/special_tokens_map.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token": {
3
+ "content": "<|endoftext|>",
4
+ "lstrip": false,
5
+ "normalized": false,
6
+ "rstrip": false,
7
+ "single_word": false
8
+ },
9
+ "eos_token": {
10
+ "content": "<|endoftext|>",
11
+ "lstrip": false,
12
+ "normalized": false,
13
+ "rstrip": false,
14
+ "single_word": false
15
+ },
16
+ "pad_token": "<|endoftext|>",
17
+ "unk_token": {
18
+ "content": "<|endoftext|>",
19
+ "lstrip": false,
20
+ "normalized": false,
21
+ "rstrip": false,
22
+ "single_word": false
23
+ }
24
+ }
checkpoint-400/tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
checkpoint-400/tokenizer_config.json ADDED
@@ -0,0 +1,214 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_bos_token": false,
3
+ "add_eos_token": false,
4
+ "add_prefix_space": false,
5
+ "added_tokens_decoder": {
6
+ "0": {
7
+ "content": "<|endoftext|>",
8
+ "lstrip": false,
9
+ "normalized": false,
10
+ "rstrip": false,
11
+ "single_word": false,
12
+ "special": true
13
+ },
14
+ "1": {
15
+ "content": "<|padding|>",
16
+ "lstrip": false,
17
+ "normalized": false,
18
+ "rstrip": false,
19
+ "single_word": false,
20
+ "special": true
21
+ },
22
+ "50254": {
23
+ "content": " ",
24
+ "lstrip": false,
25
+ "normalized": true,
26
+ "rstrip": false,
27
+ "single_word": false,
28
+ "special": false
29
+ },
30
+ "50255": {
31
+ "content": " ",
32
+ "lstrip": false,
33
+ "normalized": true,
34
+ "rstrip": false,
35
+ "single_word": false,
36
+ "special": false
37
+ },
38
+ "50256": {
39
+ "content": " ",
40
+ "lstrip": false,
41
+ "normalized": true,
42
+ "rstrip": false,
43
+ "single_word": false,
44
+ "special": false
45
+ },
46
+ "50257": {
47
+ "content": " ",
48
+ "lstrip": false,
49
+ "normalized": true,
50
+ "rstrip": false,
51
+ "single_word": false,
52
+ "special": false
53
+ },
54
+ "50258": {
55
+ "content": " ",
56
+ "lstrip": false,
57
+ "normalized": true,
58
+ "rstrip": false,
59
+ "single_word": false,
60
+ "special": false
61
+ },
62
+ "50259": {
63
+ "content": " ",
64
+ "lstrip": false,
65
+ "normalized": true,
66
+ "rstrip": false,
67
+ "single_word": false,
68
+ "special": false
69
+ },
70
+ "50260": {
71
+ "content": " ",
72
+ "lstrip": false,
73
+ "normalized": true,
74
+ "rstrip": false,
75
+ "single_word": false,
76
+ "special": false
77
+ },
78
+ "50261": {
79
+ "content": " ",
80
+ "lstrip": false,
81
+ "normalized": true,
82
+ "rstrip": false,
83
+ "single_word": false,
84
+ "special": false
85
+ },
86
+ "50262": {
87
+ "content": " ",
88
+ "lstrip": false,
89
+ "normalized": true,
90
+ "rstrip": false,
91
+ "single_word": false,
92
+ "special": false
93
+ },
94
+ "50263": {
95
+ "content": " ",
96
+ "lstrip": false,
97
+ "normalized": true,
98
+ "rstrip": false,
99
+ "single_word": false,
100
+ "special": false
101
+ },
102
+ "50264": {
103
+ "content": " ",
104
+ "lstrip": false,
105
+ "normalized": true,
106
+ "rstrip": false,
107
+ "single_word": false,
108
+ "special": false
109
+ },
110
+ "50265": {
111
+ "content": " ",
112
+ "lstrip": false,
113
+ "normalized": true,
114
+ "rstrip": false,
115
+ "single_word": false,
116
+ "special": false
117
+ },
118
+ "50266": {
119
+ "content": " ",
120
+ "lstrip": false,
121
+ "normalized": true,
122
+ "rstrip": false,
123
+ "single_word": false,
124
+ "special": false
125
+ },
126
+ "50267": {
127
+ "content": " ",
128
+ "lstrip": false,
129
+ "normalized": true,
130
+ "rstrip": false,
131
+ "single_word": false,
132
+ "special": false
133
+ },
134
+ "50268": {
135
+ "content": " ",
136
+ "lstrip": false,
137
+ "normalized": true,
138
+ "rstrip": false,
139
+ "single_word": false,
140
+ "special": false
141
+ },
142
+ "50269": {
143
+ "content": " ",
144
+ "lstrip": false,
145
+ "normalized": true,
146
+ "rstrip": false,
147
+ "single_word": false,
148
+ "special": false
149
+ },
150
+ "50270": {
151
+ "content": " ",
152
+ "lstrip": false,
153
+ "normalized": true,
154
+ "rstrip": false,
155
+ "single_word": false,
156
+ "special": false
157
+ },
158
+ "50271": {
159
+ "content": " ",
160
+ "lstrip": false,
161
+ "normalized": true,
162
+ "rstrip": false,
163
+ "single_word": false,
164
+ "special": false
165
+ },
166
+ "50272": {
167
+ "content": " ",
168
+ "lstrip": false,
169
+ "normalized": true,
170
+ "rstrip": false,
171
+ "single_word": false,
172
+ "special": false
173
+ },
174
+ "50273": {
175
+ "content": " ",
176
+ "lstrip": false,
177
+ "normalized": true,
178
+ "rstrip": false,
179
+ "single_word": false,
180
+ "special": false
181
+ },
182
+ "50274": {
183
+ "content": " ",
184
+ "lstrip": false,
185
+ "normalized": true,
186
+ "rstrip": false,
187
+ "single_word": false,
188
+ "special": false
189
+ },
190
+ "50275": {
191
+ "content": " ",
192
+ "lstrip": false,
193
+ "normalized": true,
194
+ "rstrip": false,
195
+ "single_word": false,
196
+ "special": false
197
+ },
198
+ "50276": {
199
+ "content": " ",
200
+ "lstrip": false,
201
+ "normalized": true,
202
+ "rstrip": false,
203
+ "single_word": false,
204
+ "special": false
205
+ }
206
+ },
207
+ "bos_token": "<|endoftext|>",
208
+ "clean_up_tokenization_spaces": true,
209
+ "eos_token": "<|endoftext|>",
210
+ "model_max_length": 1000000000000000019884624838656,
211
+ "pad_token": "<|endoftext|>",
212
+ "tokenizer_class": "GPTNeoXTokenizer",
213
+ "unk_token": "<|endoftext|>"
214
+ }
checkpoint-400/trainer_state.json ADDED
@@ -0,0 +1,329 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "best_metric": null,
3
+ "best_model_checkpoint": null,
4
+ "epoch": 8.556149732620321,
5
+ "eval_steps": 400,
6
+ "global_step": 400,
7
+ "is_hyper_param_search": false,
8
+ "is_local_process_zero": true,
9
+ "is_world_process_zero": true,
10
+ "log_history": [
11
+ {
12
+ "epoch": 0.0213903743315508,
13
+ "grad_norm": 0.0,
14
+ "learning_rate": 0.0,
15
+ "loss": 1.8558,
16
+ "step": 1
17
+ },
18
+ {
19
+ "epoch": 0.21390374331550802,
20
+ "grad_norm": 0.0,
21
+ "learning_rate": 0.0,
22
+ "loss": 2.0957,
23
+ "step": 10
24
+ },
25
+ {
26
+ "epoch": 0.42780748663101603,
27
+ "grad_norm": 63.96761703491211,
28
+ "learning_rate": 8e-08,
29
+ "loss": 2.1274,
30
+ "step": 20
31
+ },
32
+ {
33
+ "epoch": 0.6417112299465241,
34
+ "grad_norm": 56.0494270324707,
35
+ "learning_rate": 4.4e-07,
36
+ "loss": 2.0513,
37
+ "step": 30
38
+ },
39
+ {
40
+ "epoch": 0.8556149732620321,
41
+ "grad_norm": 31.82117462158203,
42
+ "learning_rate": 8.400000000000001e-07,
43
+ "loss": 1.6275,
44
+ "step": 40
45
+ },
46
+ {
47
+ "epoch": 1.0695187165775402,
48
+ "grad_norm": 27.546764373779297,
49
+ "learning_rate": 1.2400000000000002e-06,
50
+ "loss": 1.3922,
51
+ "step": 50
52
+ },
53
+ {
54
+ "epoch": 1.2834224598930482,
55
+ "grad_norm": 27.145185470581055,
56
+ "learning_rate": 1.6400000000000002e-06,
57
+ "loss": 1.0424,
58
+ "step": 60
59
+ },
60
+ {
61
+ "epoch": 1.4973262032085561,
62
+ "grad_norm": 33.02227020263672,
63
+ "learning_rate": 2.04e-06,
64
+ "loss": 1.0434,
65
+ "step": 70
66
+ },
67
+ {
68
+ "epoch": 1.7112299465240641,
69
+ "grad_norm": 23.329740524291992,
70
+ "learning_rate": 2.4400000000000004e-06,
71
+ "loss": 1.0585,
72
+ "step": 80
73
+ },
74
+ {
75
+ "epoch": 1.9251336898395723,
76
+ "grad_norm": 26.408071517944336,
77
+ "learning_rate": 2.84e-06,
78
+ "loss": 0.8849,
79
+ "step": 90
80
+ },
81
+ {
82
+ "epoch": 2.1390374331550803,
83
+ "grad_norm": 17.790225982666016,
84
+ "learning_rate": 3.2400000000000003e-06,
85
+ "loss": 0.5427,
86
+ "step": 100
87
+ },
88
+ {
89
+ "epoch": 2.3529411764705883,
90
+ "grad_norm": 29.846942901611328,
91
+ "learning_rate": 3.6400000000000003e-06,
92
+ "loss": 0.2784,
93
+ "step": 110
94
+ },
95
+ {
96
+ "epoch": 2.5668449197860963,
97
+ "grad_norm": 20.77630043029785,
98
+ "learning_rate": 4.04e-06,
99
+ "loss": 0.2569,
100
+ "step": 120
101
+ },
102
+ {
103
+ "epoch": 2.7807486631016043,
104
+ "grad_norm": 33.25271987915039,
105
+ "learning_rate": 4.440000000000001e-06,
106
+ "loss": 0.1836,
107
+ "step": 130
108
+ },
109
+ {
110
+ "epoch": 2.9946524064171123,
111
+ "grad_norm": 66.23503875732422,
112
+ "learning_rate": 4.84e-06,
113
+ "loss": 0.3042,
114
+ "step": 140
115
+ },
116
+ {
117
+ "epoch": 3.2085561497326203,
118
+ "grad_norm": 4.618898868560791,
119
+ "learning_rate": 5.240000000000001e-06,
120
+ "loss": 0.0455,
121
+ "step": 150
122
+ },
123
+ {
124
+ "epoch": 3.4224598930481283,
125
+ "grad_norm": 5.856906890869141,
126
+ "learning_rate": 5.64e-06,
127
+ "loss": 0.0558,
128
+ "step": 160
129
+ },
130
+ {
131
+ "epoch": 3.6363636363636362,
132
+ "grad_norm": 25.529407501220703,
133
+ "learning_rate": 6.040000000000001e-06,
134
+ "loss": 0.0835,
135
+ "step": 170
136
+ },
137
+ {
138
+ "epoch": 3.8502673796791442,
139
+ "grad_norm": 16.971477508544922,
140
+ "learning_rate": 6.440000000000001e-06,
141
+ "loss": 0.1126,
142
+ "step": 180
143
+ },
144
+ {
145
+ "epoch": 4.064171122994653,
146
+ "grad_norm": 9.28547191619873,
147
+ "learning_rate": 6.8400000000000014e-06,
148
+ "loss": 0.0838,
149
+ "step": 190
150
+ },
151
+ {
152
+ "epoch": 4.278074866310161,
153
+ "grad_norm": 17.296798706054688,
154
+ "learning_rate": 7.24e-06,
155
+ "loss": 0.0518,
156
+ "step": 200
157
+ },
158
+ {
159
+ "epoch": 4.491978609625669,
160
+ "grad_norm": 12.538177490234375,
161
+ "learning_rate": 7.640000000000001e-06,
162
+ "loss": 0.0532,
163
+ "step": 210
164
+ },
165
+ {
166
+ "epoch": 4.705882352941177,
167
+ "grad_norm": 24.446659088134766,
168
+ "learning_rate": 8.040000000000001e-06,
169
+ "loss": 0.1583,
170
+ "step": 220
171
+ },
172
+ {
173
+ "epoch": 4.919786096256685,
174
+ "grad_norm": 25.139781951904297,
175
+ "learning_rate": 8.44e-06,
176
+ "loss": 0.0894,
177
+ "step": 230
178
+ },
179
+ {
180
+ "epoch": 5.133689839572193,
181
+ "grad_norm": 2.361607074737549,
182
+ "learning_rate": 8.84e-06,
183
+ "loss": 0.0531,
184
+ "step": 240
185
+ },
186
+ {
187
+ "epoch": 5.347593582887701,
188
+ "grad_norm": 4.768627643585205,
189
+ "learning_rate": 9.240000000000001e-06,
190
+ "loss": 0.0664,
191
+ "step": 250
192
+ },
193
+ {
194
+ "epoch": 5.561497326203209,
195
+ "grad_norm": 37.02785873413086,
196
+ "learning_rate": 9.640000000000001e-06,
197
+ "loss": 0.1502,
198
+ "step": 260
199
+ },
200
+ {
201
+ "epoch": 5.775401069518717,
202
+ "grad_norm": 48.693763732910156,
203
+ "learning_rate": 1e-05,
204
+ "loss": 0.1953,
205
+ "step": 270
206
+ },
207
+ {
208
+ "epoch": 5.989304812834225,
209
+ "grad_norm": 37.84030532836914,
210
+ "learning_rate": 1e-05,
211
+ "loss": 0.1806,
212
+ "step": 280
213
+ },
214
+ {
215
+ "epoch": 6.2032085561497325,
216
+ "grad_norm": 39.74201965332031,
217
+ "learning_rate": 1e-05,
218
+ "loss": 0.0866,
219
+ "step": 290
220
+ },
221
+ {
222
+ "epoch": 6.4171122994652405,
223
+ "grad_norm": 7.8511834144592285,
224
+ "learning_rate": 1e-05,
225
+ "loss": 0.0994,
226
+ "step": 300
227
+ },
228
+ {
229
+ "epoch": 6.6310160427807485,
230
+ "grad_norm": 54.63271713256836,
231
+ "learning_rate": 1e-05,
232
+ "loss": 0.1712,
233
+ "step": 310
234
+ },
235
+ {
236
+ "epoch": 6.8449197860962565,
237
+ "grad_norm": 44.596153259277344,
238
+ "learning_rate": 1e-05,
239
+ "loss": 0.1442,
240
+ "step": 320
241
+ },
242
+ {
243
+ "epoch": 7.0588235294117645,
244
+ "grad_norm": 5.094964504241943,
245
+ "learning_rate": 1e-05,
246
+ "loss": 0.0526,
247
+ "step": 330
248
+ },
249
+ {
250
+ "epoch": 7.2727272727272725,
251
+ "grad_norm": 12.109665870666504,
252
+ "learning_rate": 1e-05,
253
+ "loss": 0.03,
254
+ "step": 340
255
+ },
256
+ {
257
+ "epoch": 7.4866310160427805,
258
+ "grad_norm": 34.3682975769043,
259
+ "learning_rate": 1e-05,
260
+ "loss": 0.0759,
261
+ "step": 350
262
+ },
263
+ {
264
+ "epoch": 7.7005347593582885,
265
+ "grad_norm": 2.3397929668426514,
266
+ "learning_rate": 1e-05,
267
+ "loss": 0.0685,
268
+ "step": 360
269
+ },
270
+ {
271
+ "epoch": 7.9144385026737964,
272
+ "grad_norm": 21.006542205810547,
273
+ "learning_rate": 1e-05,
274
+ "loss": 0.0637,
275
+ "step": 370
276
+ },
277
+ {
278
+ "epoch": 8.128342245989305,
279
+ "grad_norm": 2.5418567657470703,
280
+ "learning_rate": 1e-05,
281
+ "loss": 0.057,
282
+ "step": 380
283
+ },
284
+ {
285
+ "epoch": 8.342245989304812,
286
+ "grad_norm": 18.59604263305664,
287
+ "learning_rate": 1e-05,
288
+ "loss": 0.1096,
289
+ "step": 390
290
+ },
291
+ {
292
+ "epoch": 8.556149732620321,
293
+ "grad_norm": 52.85743713378906,
294
+ "learning_rate": 1e-05,
295
+ "loss": 0.1422,
296
+ "step": 400
297
+ },
298
+ {
299
+ "epoch": 8.556149732620321,
300
+ "eval_accuracy": 0.4,
301
+ "eval_loss": 4.6875,
302
+ "eval_runtime": 0.8724,
303
+ "eval_samples_per_second": 11.462,
304
+ "eval_steps_per_second": 1.146,
305
+ "step": 400
306
+ }
307
+ ],
308
+ "logging_steps": 10,
309
+ "max_steps": 2500,
310
+ "num_input_tokens_seen": 0,
311
+ "num_train_epochs": 55,
312
+ "save_steps": 400,
313
+ "stateful_callbacks": {
314
+ "TrainerControl": {
315
+ "args": {
316
+ "should_epoch_stop": false,
317
+ "should_evaluate": false,
318
+ "should_log": false,
319
+ "should_save": true,
320
+ "should_training_stop": false
321
+ },
322
+ "attributes": {}
323
+ }
324
+ },
325
+ "total_flos": 3.3017478812532736e+17,
326
+ "train_batch_size": 4,
327
+ "trial_name": null,
328
+ "trial_params": null
329
+ }
checkpoint-400/training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7ea5f51f632ffed5e26cc809a24c0a64055c95fd6a4bba189c05a309473a06bf
3
+ size 6136
checkpoint-400/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)