JH-debug commited on
Commit
d6567a6
1 Parent(s): ce733d6

Upload folder using huggingface_hub

Browse files
config.json ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "Taekyoon/llama2-ko-7b-test",
3
+ "architectures": [
4
+ "LlamaForCausalLM"
5
+ ],
6
+ "bos_token_id": 1,
7
+ "eos_token_id": 2,
8
+ "hidden_act": "silu",
9
+ "hidden_size": 4096,
10
+ "initializer_range": 0.02,
11
+ "intermediate_size": 11008,
12
+ "max_position_embeddings": 4096,
13
+ "model_type": "llama",
14
+ "num_attention_heads": 32,
15
+ "num_hidden_layers": 32,
16
+ "num_key_value_heads": 32,
17
+ "pad_token_id": 0,
18
+ "pretraining_tp": 1,
19
+ "rms_norm_eps": 1e-06,
20
+ "rope_scaling": null,
21
+ "rope_theta": 10000.0,
22
+ "tie_word_embeddings": false,
23
+ "torch_dtype": "float32",
24
+ "transformers_version": "4.33.3",
25
+ "use_cache": false,
26
+ "vocab_size": 46337
27
+ }
generation_config.json ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "bos_token_id": 1,
4
+ "eos_token_id": 2,
5
+ "pad_token_id": 0,
6
+ "transformers_version": "4.33.3"
7
+ }
latest ADDED
@@ -0,0 +1 @@
 
 
1
+ global_step630
pytorch_model-00001-of-00003.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0dac4ee3ff6bb7e117d9769b75c47f464179ebdb36e63dbb412a7f250c965e3f
3
+ size 9932524870
pytorch_model-00002-of-00003.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:23ba26d3ddfd9839c78ad3f64ae6e2492ff0a7d18d183abd3305bf363c7405ac
3
+ size 9894794265
pytorch_model-00003-of-00003.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0c8b4b7a705b891c7b2830e241a1a1a88a83780ef2339139d30fa3b581095af3
3
+ size 7596239190
pytorch_model.bin.index.json ADDED
@@ -0,0 +1,298 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "metadata": {
3
+ "total_size": 27423457280
4
+ },
5
+ "weight_map": {
6
+ "lm_head.weight": "pytorch_model-00003-of-00003.bin",
7
+ "model.embed_tokens.weight": "pytorch_model-00001-of-00003.bin",
8
+ "model.layers.0.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
9
+ "model.layers.0.mlp.down_proj.weight": "pytorch_model-00001-of-00003.bin",
10
+ "model.layers.0.mlp.gate_proj.weight": "pytorch_model-00001-of-00003.bin",
11
+ "model.layers.0.mlp.up_proj.weight": "pytorch_model-00001-of-00003.bin",
12
+ "model.layers.0.post_attention_layernorm.weight": "pytorch_model-00001-of-00003.bin",
13
+ "model.layers.0.self_attn.k_proj.weight": "pytorch_model-00001-of-00003.bin",
14
+ "model.layers.0.self_attn.o_proj.weight": "pytorch_model-00001-of-00003.bin",
15
+ "model.layers.0.self_attn.q_proj.weight": "pytorch_model-00001-of-00003.bin",
16
+ "model.layers.0.self_attn.v_proj.weight": "pytorch_model-00001-of-00003.bin",
17
+ "model.layers.1.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
18
+ "model.layers.1.mlp.down_proj.weight": "pytorch_model-00001-of-00003.bin",
19
+ "model.layers.1.mlp.gate_proj.weight": "pytorch_model-00001-of-00003.bin",
20
+ "model.layers.1.mlp.up_proj.weight": "pytorch_model-00001-of-00003.bin",
21
+ "model.layers.1.post_attention_layernorm.weight": "pytorch_model-00001-of-00003.bin",
22
+ "model.layers.1.self_attn.k_proj.weight": "pytorch_model-00001-of-00003.bin",
23
+ "model.layers.1.self_attn.o_proj.weight": "pytorch_model-00001-of-00003.bin",
24
+ "model.layers.1.self_attn.q_proj.weight": "pytorch_model-00001-of-00003.bin",
25
+ "model.layers.1.self_attn.v_proj.weight": "pytorch_model-00001-of-00003.bin",
26
+ "model.layers.10.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
27
+ "model.layers.10.mlp.down_proj.weight": "pytorch_model-00001-of-00003.bin",
28
+ "model.layers.10.mlp.gate_proj.weight": "pytorch_model-00001-of-00003.bin",
29
+ "model.layers.10.mlp.up_proj.weight": "pytorch_model-00001-of-00003.bin",
30
+ "model.layers.10.post_attention_layernorm.weight": "pytorch_model-00001-of-00003.bin",
31
+ "model.layers.10.self_attn.k_proj.weight": "pytorch_model-00001-of-00003.bin",
32
+ "model.layers.10.self_attn.o_proj.weight": "pytorch_model-00001-of-00003.bin",
33
+ "model.layers.10.self_attn.q_proj.weight": "pytorch_model-00001-of-00003.bin",
34
+ "model.layers.10.self_attn.v_proj.weight": "pytorch_model-00001-of-00003.bin",
35
+ "model.layers.11.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
36
+ "model.layers.11.mlp.down_proj.weight": "pytorch_model-00002-of-00003.bin",
37
+ "model.layers.11.mlp.gate_proj.weight": "pytorch_model-00002-of-00003.bin",
38
+ "model.layers.11.mlp.up_proj.weight": "pytorch_model-00002-of-00003.bin",
39
+ "model.layers.11.post_attention_layernorm.weight": "pytorch_model-00002-of-00003.bin",
40
+ "model.layers.11.self_attn.k_proj.weight": "pytorch_model-00001-of-00003.bin",
41
+ "model.layers.11.self_attn.o_proj.weight": "pytorch_model-00001-of-00003.bin",
42
+ "model.layers.11.self_attn.q_proj.weight": "pytorch_model-00001-of-00003.bin",
43
+ "model.layers.11.self_attn.v_proj.weight": "pytorch_model-00001-of-00003.bin",
44
+ "model.layers.12.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
45
+ "model.layers.12.mlp.down_proj.weight": "pytorch_model-00002-of-00003.bin",
46
+ "model.layers.12.mlp.gate_proj.weight": "pytorch_model-00002-of-00003.bin",
47
+ "model.layers.12.mlp.up_proj.weight": "pytorch_model-00002-of-00003.bin",
48
+ "model.layers.12.post_attention_layernorm.weight": "pytorch_model-00002-of-00003.bin",
49
+ "model.layers.12.self_attn.k_proj.weight": "pytorch_model-00002-of-00003.bin",
50
+ "model.layers.12.self_attn.o_proj.weight": "pytorch_model-00002-of-00003.bin",
51
+ "model.layers.12.self_attn.q_proj.weight": "pytorch_model-00002-of-00003.bin",
52
+ "model.layers.12.self_attn.v_proj.weight": "pytorch_model-00002-of-00003.bin",
53
+ "model.layers.13.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
54
+ "model.layers.13.mlp.down_proj.weight": "pytorch_model-00002-of-00003.bin",
55
+ "model.layers.13.mlp.gate_proj.weight": "pytorch_model-00002-of-00003.bin",
56
+ "model.layers.13.mlp.up_proj.weight": "pytorch_model-00002-of-00003.bin",
57
+ "model.layers.13.post_attention_layernorm.weight": "pytorch_model-00002-of-00003.bin",
58
+ "model.layers.13.self_attn.k_proj.weight": "pytorch_model-00002-of-00003.bin",
59
+ "model.layers.13.self_attn.o_proj.weight": "pytorch_model-00002-of-00003.bin",
60
+ "model.layers.13.self_attn.q_proj.weight": "pytorch_model-00002-of-00003.bin",
61
+ "model.layers.13.self_attn.v_proj.weight": "pytorch_model-00002-of-00003.bin",
62
+ "model.layers.14.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
63
+ "model.layers.14.mlp.down_proj.weight": "pytorch_model-00002-of-00003.bin",
64
+ "model.layers.14.mlp.gate_proj.weight": "pytorch_model-00002-of-00003.bin",
65
+ "model.layers.14.mlp.up_proj.weight": "pytorch_model-00002-of-00003.bin",
66
+ "model.layers.14.post_attention_layernorm.weight": "pytorch_model-00002-of-00003.bin",
67
+ "model.layers.14.self_attn.k_proj.weight": "pytorch_model-00002-of-00003.bin",
68
+ "model.layers.14.self_attn.o_proj.weight": "pytorch_model-00002-of-00003.bin",
69
+ "model.layers.14.self_attn.q_proj.weight": "pytorch_model-00002-of-00003.bin",
70
+ "model.layers.14.self_attn.v_proj.weight": "pytorch_model-00002-of-00003.bin",
71
+ "model.layers.15.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
72
+ "model.layers.15.mlp.down_proj.weight": "pytorch_model-00002-of-00003.bin",
73
+ "model.layers.15.mlp.gate_proj.weight": "pytorch_model-00002-of-00003.bin",
74
+ "model.layers.15.mlp.up_proj.weight": "pytorch_model-00002-of-00003.bin",
75
+ "model.layers.15.post_attention_layernorm.weight": "pytorch_model-00002-of-00003.bin",
76
+ "model.layers.15.self_attn.k_proj.weight": "pytorch_model-00002-of-00003.bin",
77
+ "model.layers.15.self_attn.o_proj.weight": "pytorch_model-00002-of-00003.bin",
78
+ "model.layers.15.self_attn.q_proj.weight": "pytorch_model-00002-of-00003.bin",
79
+ "model.layers.15.self_attn.v_proj.weight": "pytorch_model-00002-of-00003.bin",
80
+ "model.layers.16.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
81
+ "model.layers.16.mlp.down_proj.weight": "pytorch_model-00002-of-00003.bin",
82
+ "model.layers.16.mlp.gate_proj.weight": "pytorch_model-00002-of-00003.bin",
83
+ "model.layers.16.mlp.up_proj.weight": "pytorch_model-00002-of-00003.bin",
84
+ "model.layers.16.post_attention_layernorm.weight": "pytorch_model-00002-of-00003.bin",
85
+ "model.layers.16.self_attn.k_proj.weight": "pytorch_model-00002-of-00003.bin",
86
+ "model.layers.16.self_attn.o_proj.weight": "pytorch_model-00002-of-00003.bin",
87
+ "model.layers.16.self_attn.q_proj.weight": "pytorch_model-00002-of-00003.bin",
88
+ "model.layers.16.self_attn.v_proj.weight": "pytorch_model-00002-of-00003.bin",
89
+ "model.layers.17.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
90
+ "model.layers.17.mlp.down_proj.weight": "pytorch_model-00002-of-00003.bin",
91
+ "model.layers.17.mlp.gate_proj.weight": "pytorch_model-00002-of-00003.bin",
92
+ "model.layers.17.mlp.up_proj.weight": "pytorch_model-00002-of-00003.bin",
93
+ "model.layers.17.post_attention_layernorm.weight": "pytorch_model-00002-of-00003.bin",
94
+ "model.layers.17.self_attn.k_proj.weight": "pytorch_model-00002-of-00003.bin",
95
+ "model.layers.17.self_attn.o_proj.weight": "pytorch_model-00002-of-00003.bin",
96
+ "model.layers.17.self_attn.q_proj.weight": "pytorch_model-00002-of-00003.bin",
97
+ "model.layers.17.self_attn.v_proj.weight": "pytorch_model-00002-of-00003.bin",
98
+ "model.layers.18.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
99
+ "model.layers.18.mlp.down_proj.weight": "pytorch_model-00002-of-00003.bin",
100
+ "model.layers.18.mlp.gate_proj.weight": "pytorch_model-00002-of-00003.bin",
101
+ "model.layers.18.mlp.up_proj.weight": "pytorch_model-00002-of-00003.bin",
102
+ "model.layers.18.post_attention_layernorm.weight": "pytorch_model-00002-of-00003.bin",
103
+ "model.layers.18.self_attn.k_proj.weight": "pytorch_model-00002-of-00003.bin",
104
+ "model.layers.18.self_attn.o_proj.weight": "pytorch_model-00002-of-00003.bin",
105
+ "model.layers.18.self_attn.q_proj.weight": "pytorch_model-00002-of-00003.bin",
106
+ "model.layers.18.self_attn.v_proj.weight": "pytorch_model-00002-of-00003.bin",
107
+ "model.layers.19.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
108
+ "model.layers.19.mlp.down_proj.weight": "pytorch_model-00002-of-00003.bin",
109
+ "model.layers.19.mlp.gate_proj.weight": "pytorch_model-00002-of-00003.bin",
110
+ "model.layers.19.mlp.up_proj.weight": "pytorch_model-00002-of-00003.bin",
111
+ "model.layers.19.post_attention_layernorm.weight": "pytorch_model-00002-of-00003.bin",
112
+ "model.layers.19.self_attn.k_proj.weight": "pytorch_model-00002-of-00003.bin",
113
+ "model.layers.19.self_attn.o_proj.weight": "pytorch_model-00002-of-00003.bin",
114
+ "model.layers.19.self_attn.q_proj.weight": "pytorch_model-00002-of-00003.bin",
115
+ "model.layers.19.self_attn.v_proj.weight": "pytorch_model-00002-of-00003.bin",
116
+ "model.layers.2.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
117
+ "model.layers.2.mlp.down_proj.weight": "pytorch_model-00001-of-00003.bin",
118
+ "model.layers.2.mlp.gate_proj.weight": "pytorch_model-00001-of-00003.bin",
119
+ "model.layers.2.mlp.up_proj.weight": "pytorch_model-00001-of-00003.bin",
120
+ "model.layers.2.post_attention_layernorm.weight": "pytorch_model-00001-of-00003.bin",
121
+ "model.layers.2.self_attn.k_proj.weight": "pytorch_model-00001-of-00003.bin",
122
+ "model.layers.2.self_attn.o_proj.weight": "pytorch_model-00001-of-00003.bin",
123
+ "model.layers.2.self_attn.q_proj.weight": "pytorch_model-00001-of-00003.bin",
124
+ "model.layers.2.self_attn.v_proj.weight": "pytorch_model-00001-of-00003.bin",
125
+ "model.layers.20.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
126
+ "model.layers.20.mlp.down_proj.weight": "pytorch_model-00002-of-00003.bin",
127
+ "model.layers.20.mlp.gate_proj.weight": "pytorch_model-00002-of-00003.bin",
128
+ "model.layers.20.mlp.up_proj.weight": "pytorch_model-00002-of-00003.bin",
129
+ "model.layers.20.post_attention_layernorm.weight": "pytorch_model-00002-of-00003.bin",
130
+ "model.layers.20.self_attn.k_proj.weight": "pytorch_model-00002-of-00003.bin",
131
+ "model.layers.20.self_attn.o_proj.weight": "pytorch_model-00002-of-00003.bin",
132
+ "model.layers.20.self_attn.q_proj.weight": "pytorch_model-00002-of-00003.bin",
133
+ "model.layers.20.self_attn.v_proj.weight": "pytorch_model-00002-of-00003.bin",
134
+ "model.layers.21.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
135
+ "model.layers.21.mlp.down_proj.weight": "pytorch_model-00002-of-00003.bin",
136
+ "model.layers.21.mlp.gate_proj.weight": "pytorch_model-00002-of-00003.bin",
137
+ "model.layers.21.mlp.up_proj.weight": "pytorch_model-00002-of-00003.bin",
138
+ "model.layers.21.post_attention_layernorm.weight": "pytorch_model-00002-of-00003.bin",
139
+ "model.layers.21.self_attn.k_proj.weight": "pytorch_model-00002-of-00003.bin",
140
+ "model.layers.21.self_attn.o_proj.weight": "pytorch_model-00002-of-00003.bin",
141
+ "model.layers.21.self_attn.q_proj.weight": "pytorch_model-00002-of-00003.bin",
142
+ "model.layers.21.self_attn.v_proj.weight": "pytorch_model-00002-of-00003.bin",
143
+ "model.layers.22.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
144
+ "model.layers.22.mlp.down_proj.weight": "pytorch_model-00002-of-00003.bin",
145
+ "model.layers.22.mlp.gate_proj.weight": "pytorch_model-00002-of-00003.bin",
146
+ "model.layers.22.mlp.up_proj.weight": "pytorch_model-00002-of-00003.bin",
147
+ "model.layers.22.post_attention_layernorm.weight": "pytorch_model-00002-of-00003.bin",
148
+ "model.layers.22.self_attn.k_proj.weight": "pytorch_model-00002-of-00003.bin",
149
+ "model.layers.22.self_attn.o_proj.weight": "pytorch_model-00002-of-00003.bin",
150
+ "model.layers.22.self_attn.q_proj.weight": "pytorch_model-00002-of-00003.bin",
151
+ "model.layers.22.self_attn.v_proj.weight": "pytorch_model-00002-of-00003.bin",
152
+ "model.layers.23.input_layernorm.weight": "pytorch_model-00003-of-00003.bin",
153
+ "model.layers.23.mlp.down_proj.weight": "pytorch_model-00003-of-00003.bin",
154
+ "model.layers.23.mlp.gate_proj.weight": "pytorch_model-00002-of-00003.bin",
155
+ "model.layers.23.mlp.up_proj.weight": "pytorch_model-00003-of-00003.bin",
156
+ "model.layers.23.post_attention_layernorm.weight": "pytorch_model-00003-of-00003.bin",
157
+ "model.layers.23.self_attn.k_proj.weight": "pytorch_model-00002-of-00003.bin",
158
+ "model.layers.23.self_attn.o_proj.weight": "pytorch_model-00002-of-00003.bin",
159
+ "model.layers.23.self_attn.q_proj.weight": "pytorch_model-00002-of-00003.bin",
160
+ "model.layers.23.self_attn.v_proj.weight": "pytorch_model-00002-of-00003.bin",
161
+ "model.layers.24.input_layernorm.weight": "pytorch_model-00003-of-00003.bin",
162
+ "model.layers.24.mlp.down_proj.weight": "pytorch_model-00003-of-00003.bin",
163
+ "model.layers.24.mlp.gate_proj.weight": "pytorch_model-00003-of-00003.bin",
164
+ "model.layers.24.mlp.up_proj.weight": "pytorch_model-00003-of-00003.bin",
165
+ "model.layers.24.post_attention_layernorm.weight": "pytorch_model-00003-of-00003.bin",
166
+ "model.layers.24.self_attn.k_proj.weight": "pytorch_model-00003-of-00003.bin",
167
+ "model.layers.24.self_attn.o_proj.weight": "pytorch_model-00003-of-00003.bin",
168
+ "model.layers.24.self_attn.q_proj.weight": "pytorch_model-00003-of-00003.bin",
169
+ "model.layers.24.self_attn.v_proj.weight": "pytorch_model-00003-of-00003.bin",
170
+ "model.layers.25.input_layernorm.weight": "pytorch_model-00003-of-00003.bin",
171
+ "model.layers.25.mlp.down_proj.weight": "pytorch_model-00003-of-00003.bin",
172
+ "model.layers.25.mlp.gate_proj.weight": "pytorch_model-00003-of-00003.bin",
173
+ "model.layers.25.mlp.up_proj.weight": "pytorch_model-00003-of-00003.bin",
174
+ "model.layers.25.post_attention_layernorm.weight": "pytorch_model-00003-of-00003.bin",
175
+ "model.layers.25.self_attn.k_proj.weight": "pytorch_model-00003-of-00003.bin",
176
+ "model.layers.25.self_attn.o_proj.weight": "pytorch_model-00003-of-00003.bin",
177
+ "model.layers.25.self_attn.q_proj.weight": "pytorch_model-00003-of-00003.bin",
178
+ "model.layers.25.self_attn.v_proj.weight": "pytorch_model-00003-of-00003.bin",
179
+ "model.layers.26.input_layernorm.weight": "pytorch_model-00003-of-00003.bin",
180
+ "model.layers.26.mlp.down_proj.weight": "pytorch_model-00003-of-00003.bin",
181
+ "model.layers.26.mlp.gate_proj.weight": "pytorch_model-00003-of-00003.bin",
182
+ "model.layers.26.mlp.up_proj.weight": "pytorch_model-00003-of-00003.bin",
183
+ "model.layers.26.post_attention_layernorm.weight": "pytorch_model-00003-of-00003.bin",
184
+ "model.layers.26.self_attn.k_proj.weight": "pytorch_model-00003-of-00003.bin",
185
+ "model.layers.26.self_attn.o_proj.weight": "pytorch_model-00003-of-00003.bin",
186
+ "model.layers.26.self_attn.q_proj.weight": "pytorch_model-00003-of-00003.bin",
187
+ "model.layers.26.self_attn.v_proj.weight": "pytorch_model-00003-of-00003.bin",
188
+ "model.layers.27.input_layernorm.weight": "pytorch_model-00003-of-00003.bin",
189
+ "model.layers.27.mlp.down_proj.weight": "pytorch_model-00003-of-00003.bin",
190
+ "model.layers.27.mlp.gate_proj.weight": "pytorch_model-00003-of-00003.bin",
191
+ "model.layers.27.mlp.up_proj.weight": "pytorch_model-00003-of-00003.bin",
192
+ "model.layers.27.post_attention_layernorm.weight": "pytorch_model-00003-of-00003.bin",
193
+ "model.layers.27.self_attn.k_proj.weight": "pytorch_model-00003-of-00003.bin",
194
+ "model.layers.27.self_attn.o_proj.weight": "pytorch_model-00003-of-00003.bin",
195
+ "model.layers.27.self_attn.q_proj.weight": "pytorch_model-00003-of-00003.bin",
196
+ "model.layers.27.self_attn.v_proj.weight": "pytorch_model-00003-of-00003.bin",
197
+ "model.layers.28.input_layernorm.weight": "pytorch_model-00003-of-00003.bin",
198
+ "model.layers.28.mlp.down_proj.weight": "pytorch_model-00003-of-00003.bin",
199
+ "model.layers.28.mlp.gate_proj.weight": "pytorch_model-00003-of-00003.bin",
200
+ "model.layers.28.mlp.up_proj.weight": "pytorch_model-00003-of-00003.bin",
201
+ "model.layers.28.post_attention_layernorm.weight": "pytorch_model-00003-of-00003.bin",
202
+ "model.layers.28.self_attn.k_proj.weight": "pytorch_model-00003-of-00003.bin",
203
+ "model.layers.28.self_attn.o_proj.weight": "pytorch_model-00003-of-00003.bin",
204
+ "model.layers.28.self_attn.q_proj.weight": "pytorch_model-00003-of-00003.bin",
205
+ "model.layers.28.self_attn.v_proj.weight": "pytorch_model-00003-of-00003.bin",
206
+ "model.layers.29.input_layernorm.weight": "pytorch_model-00003-of-00003.bin",
207
+ "model.layers.29.mlp.down_proj.weight": "pytorch_model-00003-of-00003.bin",
208
+ "model.layers.29.mlp.gate_proj.weight": "pytorch_model-00003-of-00003.bin",
209
+ "model.layers.29.mlp.up_proj.weight": "pytorch_model-00003-of-00003.bin",
210
+ "model.layers.29.post_attention_layernorm.weight": "pytorch_model-00003-of-00003.bin",
211
+ "model.layers.29.self_attn.k_proj.weight": "pytorch_model-00003-of-00003.bin",
212
+ "model.layers.29.self_attn.o_proj.weight": "pytorch_model-00003-of-00003.bin",
213
+ "model.layers.29.self_attn.q_proj.weight": "pytorch_model-00003-of-00003.bin",
214
+ "model.layers.29.self_attn.v_proj.weight": "pytorch_model-00003-of-00003.bin",
215
+ "model.layers.3.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
216
+ "model.layers.3.mlp.down_proj.weight": "pytorch_model-00001-of-00003.bin",
217
+ "model.layers.3.mlp.gate_proj.weight": "pytorch_model-00001-of-00003.bin",
218
+ "model.layers.3.mlp.up_proj.weight": "pytorch_model-00001-of-00003.bin",
219
+ "model.layers.3.post_attention_layernorm.weight": "pytorch_model-00001-of-00003.bin",
220
+ "model.layers.3.self_attn.k_proj.weight": "pytorch_model-00001-of-00003.bin",
221
+ "model.layers.3.self_attn.o_proj.weight": "pytorch_model-00001-of-00003.bin",
222
+ "model.layers.3.self_attn.q_proj.weight": "pytorch_model-00001-of-00003.bin",
223
+ "model.layers.3.self_attn.v_proj.weight": "pytorch_model-00001-of-00003.bin",
224
+ "model.layers.30.input_layernorm.weight": "pytorch_model-00003-of-00003.bin",
225
+ "model.layers.30.mlp.down_proj.weight": "pytorch_model-00003-of-00003.bin",
226
+ "model.layers.30.mlp.gate_proj.weight": "pytorch_model-00003-of-00003.bin",
227
+ "model.layers.30.mlp.up_proj.weight": "pytorch_model-00003-of-00003.bin",
228
+ "model.layers.30.post_attention_layernorm.weight": "pytorch_model-00003-of-00003.bin",
229
+ "model.layers.30.self_attn.k_proj.weight": "pytorch_model-00003-of-00003.bin",
230
+ "model.layers.30.self_attn.o_proj.weight": "pytorch_model-00003-of-00003.bin",
231
+ "model.layers.30.self_attn.q_proj.weight": "pytorch_model-00003-of-00003.bin",
232
+ "model.layers.30.self_attn.v_proj.weight": "pytorch_model-00003-of-00003.bin",
233
+ "model.layers.31.input_layernorm.weight": "pytorch_model-00003-of-00003.bin",
234
+ "model.layers.31.mlp.down_proj.weight": "pytorch_model-00003-of-00003.bin",
235
+ "model.layers.31.mlp.gate_proj.weight": "pytorch_model-00003-of-00003.bin",
236
+ "model.layers.31.mlp.up_proj.weight": "pytorch_model-00003-of-00003.bin",
237
+ "model.layers.31.post_attention_layernorm.weight": "pytorch_model-00003-of-00003.bin",
238
+ "model.layers.31.self_attn.k_proj.weight": "pytorch_model-00003-of-00003.bin",
239
+ "model.layers.31.self_attn.o_proj.weight": "pytorch_model-00003-of-00003.bin",
240
+ "model.layers.31.self_attn.q_proj.weight": "pytorch_model-00003-of-00003.bin",
241
+ "model.layers.31.self_attn.v_proj.weight": "pytorch_model-00003-of-00003.bin",
242
+ "model.layers.4.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
243
+ "model.layers.4.mlp.down_proj.weight": "pytorch_model-00001-of-00003.bin",
244
+ "model.layers.4.mlp.gate_proj.weight": "pytorch_model-00001-of-00003.bin",
245
+ "model.layers.4.mlp.up_proj.weight": "pytorch_model-00001-of-00003.bin",
246
+ "model.layers.4.post_attention_layernorm.weight": "pytorch_model-00001-of-00003.bin",
247
+ "model.layers.4.self_attn.k_proj.weight": "pytorch_model-00001-of-00003.bin",
248
+ "model.layers.4.self_attn.o_proj.weight": "pytorch_model-00001-of-00003.bin",
249
+ "model.layers.4.self_attn.q_proj.weight": "pytorch_model-00001-of-00003.bin",
250
+ "model.layers.4.self_attn.v_proj.weight": "pytorch_model-00001-of-00003.bin",
251
+ "model.layers.5.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
252
+ "model.layers.5.mlp.down_proj.weight": "pytorch_model-00001-of-00003.bin",
253
+ "model.layers.5.mlp.gate_proj.weight": "pytorch_model-00001-of-00003.bin",
254
+ "model.layers.5.mlp.up_proj.weight": "pytorch_model-00001-of-00003.bin",
255
+ "model.layers.5.post_attention_layernorm.weight": "pytorch_model-00001-of-00003.bin",
256
+ "model.layers.5.self_attn.k_proj.weight": "pytorch_model-00001-of-00003.bin",
257
+ "model.layers.5.self_attn.o_proj.weight": "pytorch_model-00001-of-00003.bin",
258
+ "model.layers.5.self_attn.q_proj.weight": "pytorch_model-00001-of-00003.bin",
259
+ "model.layers.5.self_attn.v_proj.weight": "pytorch_model-00001-of-00003.bin",
260
+ "model.layers.6.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
261
+ "model.layers.6.mlp.down_proj.weight": "pytorch_model-00001-of-00003.bin",
262
+ "model.layers.6.mlp.gate_proj.weight": "pytorch_model-00001-of-00003.bin",
263
+ "model.layers.6.mlp.up_proj.weight": "pytorch_model-00001-of-00003.bin",
264
+ "model.layers.6.post_attention_layernorm.weight": "pytorch_model-00001-of-00003.bin",
265
+ "model.layers.6.self_attn.k_proj.weight": "pytorch_model-00001-of-00003.bin",
266
+ "model.layers.6.self_attn.o_proj.weight": "pytorch_model-00001-of-00003.bin",
267
+ "model.layers.6.self_attn.q_proj.weight": "pytorch_model-00001-of-00003.bin",
268
+ "model.layers.6.self_attn.v_proj.weight": "pytorch_model-00001-of-00003.bin",
269
+ "model.layers.7.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
270
+ "model.layers.7.mlp.down_proj.weight": "pytorch_model-00001-of-00003.bin",
271
+ "model.layers.7.mlp.gate_proj.weight": "pytorch_model-00001-of-00003.bin",
272
+ "model.layers.7.mlp.up_proj.weight": "pytorch_model-00001-of-00003.bin",
273
+ "model.layers.7.post_attention_layernorm.weight": "pytorch_model-00001-of-00003.bin",
274
+ "model.layers.7.self_attn.k_proj.weight": "pytorch_model-00001-of-00003.bin",
275
+ "model.layers.7.self_attn.o_proj.weight": "pytorch_model-00001-of-00003.bin",
276
+ "model.layers.7.self_attn.q_proj.weight": "pytorch_model-00001-of-00003.bin",
277
+ "model.layers.7.self_attn.v_proj.weight": "pytorch_model-00001-of-00003.bin",
278
+ "model.layers.8.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
279
+ "model.layers.8.mlp.down_proj.weight": "pytorch_model-00001-of-00003.bin",
280
+ "model.layers.8.mlp.gate_proj.weight": "pytorch_model-00001-of-00003.bin",
281
+ "model.layers.8.mlp.up_proj.weight": "pytorch_model-00001-of-00003.bin",
282
+ "model.layers.8.post_attention_layernorm.weight": "pytorch_model-00001-of-00003.bin",
283
+ "model.layers.8.self_attn.k_proj.weight": "pytorch_model-00001-of-00003.bin",
284
+ "model.layers.8.self_attn.o_proj.weight": "pytorch_model-00001-of-00003.bin",
285
+ "model.layers.8.self_attn.q_proj.weight": "pytorch_model-00001-of-00003.bin",
286
+ "model.layers.8.self_attn.v_proj.weight": "pytorch_model-00001-of-00003.bin",
287
+ "model.layers.9.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
288
+ "model.layers.9.mlp.down_proj.weight": "pytorch_model-00001-of-00003.bin",
289
+ "model.layers.9.mlp.gate_proj.weight": "pytorch_model-00001-of-00003.bin",
290
+ "model.layers.9.mlp.up_proj.weight": "pytorch_model-00001-of-00003.bin",
291
+ "model.layers.9.post_attention_layernorm.weight": "pytorch_model-00001-of-00003.bin",
292
+ "model.layers.9.self_attn.k_proj.weight": "pytorch_model-00001-of-00003.bin",
293
+ "model.layers.9.self_attn.o_proj.weight": "pytorch_model-00001-of-00003.bin",
294
+ "model.layers.9.self_attn.q_proj.weight": "pytorch_model-00001-of-00003.bin",
295
+ "model.layers.9.self_attn.v_proj.weight": "pytorch_model-00001-of-00003.bin",
296
+ "model.norm.weight": "pytorch_model-00003-of-00003.bin"
297
+ }
298
+ }
rng_state_0.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:982bd0c16d66c3eece7beee8bddaa14b1c90bf8f69dc67202bcca3b4b9a47099
3
+ size 15920
rng_state_1.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b9ab3d337a4af62c2d871c438e02f107e0b3599e7ff42bdbae70fee57532bf66
3
+ size 15920
rng_state_2.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:374fb9bca844476287dfb22c7aed2b0726676bbfe4ef929aed1f904c45c1b993
3
+ size 15920
rng_state_3.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1b0074aee8d6261bdc69c1bb9347591668abc62c8dbadf7b6c6e44e1e45e725e
3
+ size 15920
special_tokens_map.json ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token": "<s>",
3
+ "eos_token": "</s>",
4
+ "pad_token": "<|end_of_text|>",
5
+ "unk_token": "</s>"
6
+ }
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer_config.json ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token": {
3
+ "__type": "AddedToken",
4
+ "content": "<s>",
5
+ "lstrip": false,
6
+ "normalized": false,
7
+ "rstrip": false,
8
+ "single_word": false
9
+ },
10
+ "clean_up_tokenization_spaces": false,
11
+ "eos_token": {
12
+ "__type": "AddedToken",
13
+ "content": "</s>",
14
+ "lstrip": false,
15
+ "normalized": false,
16
+ "rstrip": false,
17
+ "single_word": false
18
+ },
19
+ "legacy": false,
20
+ "model_max_length": 4096,
21
+ "pad_token": null,
22
+ "padding_side": "left",
23
+ "sp_model_kwargs": {},
24
+ "tokenizer_class": "LlamaTokenizer",
25
+ "truncation_side": "left",
26
+ "unk_token": {
27
+ "__type": "AddedToken",
28
+ "content": "<unk>",
29
+ "lstrip": false,
30
+ "normalized": false,
31
+ "rstrip": false,
32
+ "single_word": false
33
+ },
34
+ "use_default_system_prompt": true
35
+ }
trainer_state.json ADDED
@@ -0,0 +1,775 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "best_metric": null,
3
+ "best_model_checkpoint": null,
4
+ "epoch": 4.977777777777778,
5
+ "eval_steps": 500,
6
+ "global_step": 630,
7
+ "is_hyper_param_search": false,
8
+ "is_local_process_zero": true,
9
+ "is_world_process_zero": true,
10
+ "log_history": [
11
+ {
12
+ "epoch": 0.04,
13
+ "learning_rate": 7.769179709180454e-06,
14
+ "loss": 1.3862,
15
+ "step": 5
16
+ },
17
+ {
18
+ "epoch": 0.08,
19
+ "learning_rate": 1.111518328538394e-05,
20
+ "loss": 1.201,
21
+ "step": 10
22
+ },
23
+ {
24
+ "epoch": 0.12,
25
+ "learning_rate": 1.307246990474186e-05,
26
+ "loss": 1.213,
27
+ "step": 15
28
+ },
29
+ {
30
+ "epoch": 0.16,
31
+ "learning_rate": 1.4461186861587421e-05,
32
+ "loss": 1.2115,
33
+ "step": 20
34
+ },
35
+ {
36
+ "epoch": 0.2,
37
+ "learning_rate": 1.5538359418360907e-05,
38
+ "loss": 1.1947,
39
+ "step": 25
40
+ },
41
+ {
42
+ "epoch": 0.24,
43
+ "learning_rate": 1.6418473480945343e-05,
44
+ "loss": 1.1575,
45
+ "step": 30
46
+ },
47
+ {
48
+ "epoch": 0.28,
49
+ "learning_rate": 1.7162599318057647e-05,
50
+ "loss": 1.1609,
51
+ "step": 35
52
+ },
53
+ {
54
+ "epoch": 0.32,
55
+ "learning_rate": 1.7807190437790904e-05,
56
+ "loss": 1.1704,
57
+ "step": 40
58
+ },
59
+ {
60
+ "epoch": 0.36,
61
+ "learning_rate": 1.8375760100303265e-05,
62
+ "loss": 1.1499,
63
+ "step": 45
64
+ },
65
+ {
66
+ "epoch": 0.4,
67
+ "learning_rate": 1.8884362994564394e-05,
68
+ "loss": 1.1641,
69
+ "step": 50
70
+ },
71
+ {
72
+ "epoch": 0.43,
73
+ "learning_rate": 1.9344450276772258e-05,
74
+ "loss": 1.1783,
75
+ "step": 55
76
+ },
77
+ {
78
+ "epoch": 0.47,
79
+ "learning_rate": 1.9764477057148826e-05,
80
+ "loss": 1.1496,
81
+ "step": 60
82
+ },
83
+ {
84
+ "epoch": 0.51,
85
+ "learning_rate": 2e-05,
86
+ "loss": 1.1579,
87
+ "step": 65
88
+ },
89
+ {
90
+ "epoch": 0.55,
91
+ "learning_rate": 2e-05,
92
+ "loss": 1.1834,
93
+ "step": 70
94
+ },
95
+ {
96
+ "epoch": 0.59,
97
+ "learning_rate": 2e-05,
98
+ "loss": 1.1981,
99
+ "step": 75
100
+ },
101
+ {
102
+ "epoch": 0.63,
103
+ "learning_rate": 2e-05,
104
+ "loss": 1.1672,
105
+ "step": 80
106
+ },
107
+ {
108
+ "epoch": 0.67,
109
+ "learning_rate": 2e-05,
110
+ "loss": 1.1731,
111
+ "step": 85
112
+ },
113
+ {
114
+ "epoch": 0.71,
115
+ "learning_rate": 2e-05,
116
+ "loss": 1.1754,
117
+ "step": 90
118
+ },
119
+ {
120
+ "epoch": 0.75,
121
+ "learning_rate": 2e-05,
122
+ "loss": 1.1599,
123
+ "step": 95
124
+ },
125
+ {
126
+ "epoch": 0.79,
127
+ "learning_rate": 2e-05,
128
+ "loss": 1.1552,
129
+ "step": 100
130
+ },
131
+ {
132
+ "epoch": 0.83,
133
+ "learning_rate": 2e-05,
134
+ "loss": 1.1653,
135
+ "step": 105
136
+ },
137
+ {
138
+ "epoch": 0.87,
139
+ "learning_rate": 2e-05,
140
+ "loss": 1.1196,
141
+ "step": 110
142
+ },
143
+ {
144
+ "epoch": 0.91,
145
+ "learning_rate": 2e-05,
146
+ "loss": 1.1786,
147
+ "step": 115
148
+ },
149
+ {
150
+ "epoch": 0.95,
151
+ "learning_rate": 2e-05,
152
+ "loss": 1.1622,
153
+ "step": 120
154
+ },
155
+ {
156
+ "epoch": 0.99,
157
+ "learning_rate": 2e-05,
158
+ "loss": 1.1342,
159
+ "step": 125
160
+ },
161
+ {
162
+ "epoch": 1.03,
163
+ "learning_rate": 2e-05,
164
+ "loss": 0.8944,
165
+ "step": 130
166
+ },
167
+ {
168
+ "epoch": 1.07,
169
+ "learning_rate": 2e-05,
170
+ "loss": 0.7323,
171
+ "step": 135
172
+ },
173
+ {
174
+ "epoch": 1.11,
175
+ "learning_rate": 2e-05,
176
+ "loss": 0.7555,
177
+ "step": 140
178
+ },
179
+ {
180
+ "epoch": 1.15,
181
+ "learning_rate": 2e-05,
182
+ "loss": 0.725,
183
+ "step": 145
184
+ },
185
+ {
186
+ "epoch": 1.19,
187
+ "learning_rate": 2e-05,
188
+ "loss": 0.6937,
189
+ "step": 150
190
+ },
191
+ {
192
+ "epoch": 1.22,
193
+ "learning_rate": 2e-05,
194
+ "loss": 0.6984,
195
+ "step": 155
196
+ },
197
+ {
198
+ "epoch": 1.26,
199
+ "learning_rate": 2e-05,
200
+ "loss": 0.7137,
201
+ "step": 160
202
+ },
203
+ {
204
+ "epoch": 1.3,
205
+ "learning_rate": 2e-05,
206
+ "loss": 0.7295,
207
+ "step": 165
208
+ },
209
+ {
210
+ "epoch": 1.34,
211
+ "learning_rate": 2e-05,
212
+ "loss": 0.7126,
213
+ "step": 170
214
+ },
215
+ {
216
+ "epoch": 1.38,
217
+ "learning_rate": 2e-05,
218
+ "loss": 0.7208,
219
+ "step": 175
220
+ },
221
+ {
222
+ "epoch": 1.42,
223
+ "learning_rate": 2e-05,
224
+ "loss": 0.7109,
225
+ "step": 180
226
+ },
227
+ {
228
+ "epoch": 1.46,
229
+ "learning_rate": 2e-05,
230
+ "loss": 0.7367,
231
+ "step": 185
232
+ },
233
+ {
234
+ "epoch": 1.5,
235
+ "learning_rate": 2e-05,
236
+ "loss": 0.7144,
237
+ "step": 190
238
+ },
239
+ {
240
+ "epoch": 1.54,
241
+ "learning_rate": 2e-05,
242
+ "loss": 0.7257,
243
+ "step": 195
244
+ },
245
+ {
246
+ "epoch": 1.58,
247
+ "learning_rate": 2e-05,
248
+ "loss": 0.7346,
249
+ "step": 200
250
+ },
251
+ {
252
+ "epoch": 1.62,
253
+ "learning_rate": 2e-05,
254
+ "loss": 0.7454,
255
+ "step": 205
256
+ },
257
+ {
258
+ "epoch": 1.66,
259
+ "learning_rate": 2e-05,
260
+ "loss": 0.7115,
261
+ "step": 210
262
+ },
263
+ {
264
+ "epoch": 1.7,
265
+ "learning_rate": 2e-05,
266
+ "loss": 0.7137,
267
+ "step": 215
268
+ },
269
+ {
270
+ "epoch": 1.74,
271
+ "learning_rate": 2e-05,
272
+ "loss": 0.713,
273
+ "step": 220
274
+ },
275
+ {
276
+ "epoch": 1.78,
277
+ "learning_rate": 2e-05,
278
+ "loss": 0.7364,
279
+ "step": 225
280
+ },
281
+ {
282
+ "epoch": 1.82,
283
+ "learning_rate": 2e-05,
284
+ "loss": 0.7385,
285
+ "step": 230
286
+ },
287
+ {
288
+ "epoch": 1.86,
289
+ "learning_rate": 2e-05,
290
+ "loss": 0.7398,
291
+ "step": 235
292
+ },
293
+ {
294
+ "epoch": 1.9,
295
+ "learning_rate": 2e-05,
296
+ "loss": 0.7134,
297
+ "step": 240
298
+ },
299
+ {
300
+ "epoch": 1.94,
301
+ "learning_rate": 2e-05,
302
+ "loss": 0.7282,
303
+ "step": 245
304
+ },
305
+ {
306
+ "epoch": 1.98,
307
+ "learning_rate": 2e-05,
308
+ "loss": 0.7697,
309
+ "step": 250
310
+ },
311
+ {
312
+ "epoch": 2.01,
313
+ "learning_rate": 2e-05,
314
+ "loss": 0.5946,
315
+ "step": 255
316
+ },
317
+ {
318
+ "epoch": 2.05,
319
+ "learning_rate": 2e-05,
320
+ "loss": 0.3129,
321
+ "step": 260
322
+ },
323
+ {
324
+ "epoch": 2.09,
325
+ "learning_rate": 2e-05,
326
+ "loss": 0.2843,
327
+ "step": 265
328
+ },
329
+ {
330
+ "epoch": 2.13,
331
+ "learning_rate": 2e-05,
332
+ "loss": 0.276,
333
+ "step": 270
334
+ },
335
+ {
336
+ "epoch": 2.17,
337
+ "learning_rate": 2e-05,
338
+ "loss": 0.2661,
339
+ "step": 275
340
+ },
341
+ {
342
+ "epoch": 2.21,
343
+ "learning_rate": 2e-05,
344
+ "loss": 0.2728,
345
+ "step": 280
346
+ },
347
+ {
348
+ "epoch": 2.25,
349
+ "learning_rate": 2e-05,
350
+ "loss": 0.2701,
351
+ "step": 285
352
+ },
353
+ {
354
+ "epoch": 2.29,
355
+ "learning_rate": 2e-05,
356
+ "loss": 0.2752,
357
+ "step": 290
358
+ },
359
+ {
360
+ "epoch": 2.33,
361
+ "learning_rate": 2e-05,
362
+ "loss": 0.2667,
363
+ "step": 295
364
+ },
365
+ {
366
+ "epoch": 2.37,
367
+ "learning_rate": 2e-05,
368
+ "loss": 0.272,
369
+ "step": 300
370
+ },
371
+ {
372
+ "epoch": 2.41,
373
+ "learning_rate": 2e-05,
374
+ "loss": 0.2684,
375
+ "step": 305
376
+ },
377
+ {
378
+ "epoch": 2.45,
379
+ "learning_rate": 2e-05,
380
+ "loss": 0.2735,
381
+ "step": 310
382
+ },
383
+ {
384
+ "epoch": 2.49,
385
+ "learning_rate": 2e-05,
386
+ "loss": 0.2728,
387
+ "step": 315
388
+ },
389
+ {
390
+ "epoch": 2.53,
391
+ "learning_rate": 2e-05,
392
+ "loss": 0.28,
393
+ "step": 320
394
+ },
395
+ {
396
+ "epoch": 2.57,
397
+ "learning_rate": 2e-05,
398
+ "loss": 0.2989,
399
+ "step": 325
400
+ },
401
+ {
402
+ "epoch": 2.61,
403
+ "learning_rate": 2e-05,
404
+ "loss": 0.2837,
405
+ "step": 330
406
+ },
407
+ {
408
+ "epoch": 2.65,
409
+ "learning_rate": 2e-05,
410
+ "loss": 0.2823,
411
+ "step": 335
412
+ },
413
+ {
414
+ "epoch": 2.69,
415
+ "learning_rate": 2e-05,
416
+ "loss": 0.2947,
417
+ "step": 340
418
+ },
419
+ {
420
+ "epoch": 2.73,
421
+ "learning_rate": 2e-05,
422
+ "loss": 0.2859,
423
+ "step": 345
424
+ },
425
+ {
426
+ "epoch": 2.77,
427
+ "learning_rate": 2e-05,
428
+ "loss": 0.2936,
429
+ "step": 350
430
+ },
431
+ {
432
+ "epoch": 2.8,
433
+ "learning_rate": 2e-05,
434
+ "loss": 0.2924,
435
+ "step": 355
436
+ },
437
+ {
438
+ "epoch": 2.84,
439
+ "learning_rate": 2e-05,
440
+ "loss": 0.2818,
441
+ "step": 360
442
+ },
443
+ {
444
+ "epoch": 2.88,
445
+ "learning_rate": 2e-05,
446
+ "loss": 0.2869,
447
+ "step": 365
448
+ },
449
+ {
450
+ "epoch": 2.92,
451
+ "learning_rate": 2e-05,
452
+ "loss": 0.2992,
453
+ "step": 370
454
+ },
455
+ {
456
+ "epoch": 2.96,
457
+ "learning_rate": 2e-05,
458
+ "loss": 0.2946,
459
+ "step": 375
460
+ },
461
+ {
462
+ "epoch": 3.0,
463
+ "learning_rate": 2e-05,
464
+ "loss": 0.2801,
465
+ "step": 380
466
+ },
467
+ {
468
+ "epoch": 3.04,
469
+ "learning_rate": 2e-05,
470
+ "loss": 0.081,
471
+ "step": 385
472
+ },
473
+ {
474
+ "epoch": 3.08,
475
+ "learning_rate": 2e-05,
476
+ "loss": 0.0829,
477
+ "step": 390
478
+ },
479
+ {
480
+ "epoch": 3.12,
481
+ "learning_rate": 2e-05,
482
+ "loss": 0.077,
483
+ "step": 395
484
+ },
485
+ {
486
+ "epoch": 3.16,
487
+ "learning_rate": 2e-05,
488
+ "loss": 0.0764,
489
+ "step": 400
490
+ },
491
+ {
492
+ "epoch": 3.2,
493
+ "learning_rate": 2e-05,
494
+ "loss": 0.076,
495
+ "step": 405
496
+ },
497
+ {
498
+ "epoch": 3.24,
499
+ "learning_rate": 2e-05,
500
+ "loss": 0.0781,
501
+ "step": 410
502
+ },
503
+ {
504
+ "epoch": 3.28,
505
+ "learning_rate": 2e-05,
506
+ "loss": 0.074,
507
+ "step": 415
508
+ },
509
+ {
510
+ "epoch": 3.32,
511
+ "learning_rate": 2e-05,
512
+ "loss": 0.0779,
513
+ "step": 420
514
+ },
515
+ {
516
+ "epoch": 3.36,
517
+ "learning_rate": 2e-05,
518
+ "loss": 0.0754,
519
+ "step": 425
520
+ },
521
+ {
522
+ "epoch": 3.4,
523
+ "learning_rate": 2e-05,
524
+ "loss": 0.0795,
525
+ "step": 430
526
+ },
527
+ {
528
+ "epoch": 3.44,
529
+ "learning_rate": 2e-05,
530
+ "loss": 0.0747,
531
+ "step": 435
532
+ },
533
+ {
534
+ "epoch": 3.48,
535
+ "learning_rate": 2e-05,
536
+ "loss": 0.0793,
537
+ "step": 440
538
+ },
539
+ {
540
+ "epoch": 3.52,
541
+ "learning_rate": 2e-05,
542
+ "loss": 0.0838,
543
+ "step": 445
544
+ },
545
+ {
546
+ "epoch": 3.56,
547
+ "learning_rate": 2e-05,
548
+ "loss": 0.0773,
549
+ "step": 450
550
+ },
551
+ {
552
+ "epoch": 3.6,
553
+ "learning_rate": 2e-05,
554
+ "loss": 0.0735,
555
+ "step": 455
556
+ },
557
+ {
558
+ "epoch": 3.63,
559
+ "learning_rate": 2e-05,
560
+ "loss": 0.0787,
561
+ "step": 460
562
+ },
563
+ {
564
+ "epoch": 3.67,
565
+ "learning_rate": 2e-05,
566
+ "loss": 0.0863,
567
+ "step": 465
568
+ },
569
+ {
570
+ "epoch": 3.71,
571
+ "learning_rate": 2e-05,
572
+ "loss": 0.085,
573
+ "step": 470
574
+ },
575
+ {
576
+ "epoch": 3.75,
577
+ "learning_rate": 2e-05,
578
+ "loss": 0.0791,
579
+ "step": 475
580
+ },
581
+ {
582
+ "epoch": 3.79,
583
+ "learning_rate": 2e-05,
584
+ "loss": 0.0795,
585
+ "step": 480
586
+ },
587
+ {
588
+ "epoch": 3.83,
589
+ "learning_rate": 2e-05,
590
+ "loss": 0.0844,
591
+ "step": 485
592
+ },
593
+ {
594
+ "epoch": 3.87,
595
+ "learning_rate": 2e-05,
596
+ "loss": 0.0839,
597
+ "step": 490
598
+ },
599
+ {
600
+ "epoch": 3.91,
601
+ "learning_rate": 2e-05,
602
+ "loss": 0.0845,
603
+ "step": 495
604
+ },
605
+ {
606
+ "epoch": 3.95,
607
+ "learning_rate": 2e-05,
608
+ "loss": 0.0816,
609
+ "step": 500
610
+ },
611
+ {
612
+ "epoch": 3.99,
613
+ "learning_rate": 2e-05,
614
+ "loss": 0.09,
615
+ "step": 505
616
+ },
617
+ {
618
+ "epoch": 4.03,
619
+ "learning_rate": 2e-05,
620
+ "loss": 0.0484,
621
+ "step": 510
622
+ },
623
+ {
624
+ "epoch": 4.07,
625
+ "learning_rate": 2e-05,
626
+ "loss": 0.033,
627
+ "step": 515
628
+ },
629
+ {
630
+ "epoch": 4.11,
631
+ "learning_rate": 2e-05,
632
+ "loss": 0.0322,
633
+ "step": 520
634
+ },
635
+ {
636
+ "epoch": 4.15,
637
+ "learning_rate": 2e-05,
638
+ "loss": 0.0298,
639
+ "step": 525
640
+ },
641
+ {
642
+ "epoch": 4.19,
643
+ "learning_rate": 2e-05,
644
+ "loss": 0.0342,
645
+ "step": 530
646
+ },
647
+ {
648
+ "epoch": 4.23,
649
+ "learning_rate": 2e-05,
650
+ "loss": 0.0327,
651
+ "step": 535
652
+ },
653
+ {
654
+ "epoch": 4.27,
655
+ "learning_rate": 2e-05,
656
+ "loss": 0.0345,
657
+ "step": 540
658
+ },
659
+ {
660
+ "epoch": 4.31,
661
+ "learning_rate": 2e-05,
662
+ "loss": 0.0301,
663
+ "step": 545
664
+ },
665
+ {
666
+ "epoch": 4.35,
667
+ "learning_rate": 2e-05,
668
+ "loss": 0.0312,
669
+ "step": 550
670
+ },
671
+ {
672
+ "epoch": 4.39,
673
+ "learning_rate": 2e-05,
674
+ "loss": 0.0321,
675
+ "step": 555
676
+ },
677
+ {
678
+ "epoch": 4.42,
679
+ "learning_rate": 2e-05,
680
+ "loss": 0.0298,
681
+ "step": 560
682
+ },
683
+ {
684
+ "epoch": 4.46,
685
+ "learning_rate": 2e-05,
686
+ "loss": 0.0315,
687
+ "step": 565
688
+ },
689
+ {
690
+ "epoch": 4.5,
691
+ "learning_rate": 2e-05,
692
+ "loss": 0.0319,
693
+ "step": 570
694
+ },
695
+ {
696
+ "epoch": 4.54,
697
+ "learning_rate": 2e-05,
698
+ "loss": 0.0331,
699
+ "step": 575
700
+ },
701
+ {
702
+ "epoch": 4.58,
703
+ "learning_rate": 2e-05,
704
+ "loss": 0.0322,
705
+ "step": 580
706
+ },
707
+ {
708
+ "epoch": 4.62,
709
+ "learning_rate": 2e-05,
710
+ "loss": 0.0326,
711
+ "step": 585
712
+ },
713
+ {
714
+ "epoch": 4.66,
715
+ "learning_rate": 2e-05,
716
+ "loss": 0.0333,
717
+ "step": 590
718
+ },
719
+ {
720
+ "epoch": 4.7,
721
+ "learning_rate": 2e-05,
722
+ "loss": 0.0322,
723
+ "step": 595
724
+ },
725
+ {
726
+ "epoch": 4.74,
727
+ "learning_rate": 2e-05,
728
+ "loss": 0.0309,
729
+ "step": 600
730
+ },
731
+ {
732
+ "epoch": 4.78,
733
+ "learning_rate": 2e-05,
734
+ "loss": 0.0312,
735
+ "step": 605
736
+ },
737
+ {
738
+ "epoch": 4.82,
739
+ "learning_rate": 2e-05,
740
+ "loss": 0.0296,
741
+ "step": 610
742
+ },
743
+ {
744
+ "epoch": 4.86,
745
+ "learning_rate": 2e-05,
746
+ "loss": 0.0316,
747
+ "step": 615
748
+ },
749
+ {
750
+ "epoch": 4.9,
751
+ "learning_rate": 2e-05,
752
+ "loss": 0.0325,
753
+ "step": 620
754
+ },
755
+ {
756
+ "epoch": 4.94,
757
+ "learning_rate": 2e-05,
758
+ "loss": 0.0299,
759
+ "step": 625
760
+ },
761
+ {
762
+ "epoch": 4.98,
763
+ "learning_rate": 2e-05,
764
+ "loss": 0.0305,
765
+ "step": 630
766
+ }
767
+ ],
768
+ "logging_steps": 5,
769
+ "max_steps": 630,
770
+ "num_train_epochs": 5,
771
+ "save_steps": 100,
772
+ "total_flos": 519292402286592.0,
773
+ "trial_name": null,
774
+ "trial_params": null
775
+ }
training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:579531ef6fb97d3b55179f34627da65865b3a9bcf5be850eb44c0a809edda493
3
+ size 6648
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)