Hugging Face's logo

Bleking
/
llava-v1.5-13b

PEFT
Safetensors
llava_llama
Model card Files
xet
 Community
Bleking commited on
Commit
07ada15
·
1 Parent(s): e6dee11

push llava-v1.5-13b

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. README.md +202 -0
  2. adapter_config.json +34 -0
  3. adapter_model.safetensors +3 -0
  4. checkpoint-224/README.md +202 -0
  5. checkpoint-224/adapter_config.json +34 -0
  6. checkpoint-224/adapter_model.safetensors +3 -0
  7. checkpoint-224/global_step224/zero_pp_rank_0_mp_rank_00_model_states.pt +3 -0
  8. checkpoint-224/global_step224/zero_pp_rank_0_mp_rank_00_optim_states.pt +3 -0
  9. checkpoint-224/global_step224/zero_pp_rank_1_mp_rank_00_model_states.pt +3 -0
  10. checkpoint-224/global_step224/zero_pp_rank_1_mp_rank_00_optim_states.pt +3 -0
  11. checkpoint-224/global_step224/zero_pp_rank_2_mp_rank_00_model_states.pt +3 -0
  12. checkpoint-224/global_step224/zero_pp_rank_2_mp_rank_00_optim_states.pt +3 -0
  13. checkpoint-224/global_step224/zero_pp_rank_3_mp_rank_00_model_states.pt +3 -0
  14. checkpoint-224/global_step224/zero_pp_rank_3_mp_rank_00_optim_states.pt +3 -0
  15. checkpoint-224/latest +1 -0
  16. checkpoint-224/rng_state_0.pth +3 -0
  17. checkpoint-224/rng_state_1.pth +3 -0
  18. checkpoint-224/rng_state_2.pth +3 -0
  19. checkpoint-224/rng_state_3.pth +3 -0
  20. checkpoint-224/special_tokens_map.json +24 -0
  21. checkpoint-224/tokenizer.model +3 -0
  22. checkpoint-224/tokenizer_config.json +43 -0
  23. checkpoint-224/trainer_state.json +3393 -0
  24. checkpoint-224/training_args.bin +3 -0
  25. checkpoint-224/zero_to_fp32.py +604 -0
  26. checkpoint-320/README.md +202 -0
  27. checkpoint-320/adapter_config.json +34 -0
  28. checkpoint-320/adapter_model.safetensors +3 -0
  29. checkpoint-320/global_step320/zero_pp_rank_0_mp_rank_00_model_states.pt +3 -0
  30. checkpoint-320/global_step320/zero_pp_rank_0_mp_rank_00_optim_states.pt +3 -0
  31. checkpoint-320/global_step320/zero_pp_rank_1_mp_rank_00_model_states.pt +3 -0
  32. checkpoint-320/global_step320/zero_pp_rank_1_mp_rank_00_optim_states.pt +3 -0
  33. checkpoint-320/global_step320/zero_pp_rank_2_mp_rank_00_model_states.pt +3 -0
  34. checkpoint-320/global_step320/zero_pp_rank_2_mp_rank_00_optim_states.pt +3 -0
  35. checkpoint-320/global_step320/zero_pp_rank_3_mp_rank_00_model_states.pt +3 -0
  36. checkpoint-320/global_step320/zero_pp_rank_3_mp_rank_00_optim_states.pt +3 -0
  37. checkpoint-320/latest +1 -0
  38. checkpoint-320/rng_state_0.pth +3 -0
  39. checkpoint-320/rng_state_1.pth +3 -0
  40. checkpoint-320/rng_state_2.pth +3 -0
  41. checkpoint-320/rng_state_3.pth +3 -0
  42. checkpoint-320/special_tokens_map.json +24 -0
  43. checkpoint-320/tokenizer.model +3 -0
  44. checkpoint-320/tokenizer_config.json +43 -0
  45. checkpoint-320/trainer_state.json +0 -0
  46. checkpoint-320/training_args.bin +3 -0
  47. checkpoint-320/zero_to_fp32.py +604 -0
  48. config.json +52 -0
  49. non_lora_trainables.bin +3 -0
  50. optimizer.pt +3 -0
README.md ADDED
@@ -0,0 +1,202 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ base_model: liuhaotian/llava-v1.5-13b
3
+ library_name: peft
4
+ ---
5
+
6
+ # Model Card for Model ID
7
+
8
+ <!-- Provide a quick summary of what the model is/does. -->
9
+
10
+
11
+
12
+ ## Model Details
13
+
14
+ ### Model Description
15
+
16
+ <!-- Provide a longer summary of what this model is. -->
17
+
18
+
19
+
20
+ - **Developed by:** [More Information Needed]
21
+ - **Funded by [optional]:** [More Information Needed]
22
+ - **Shared by [optional]:** [More Information Needed]
23
+ - **Model type:** [More Information Needed]
24
+ - **Language(s) (NLP):** [More Information Needed]
25
+ - **License:** [More Information Needed]
26
+ - **Finetuned from model [optional]:** [More Information Needed]
27
+
28
+ ### Model Sources [optional]
29
+
30
+ <!-- Provide the basic links for the model. -->
31
+
32
+ - **Repository:** [More Information Needed]
33
+ - **Paper [optional]:** [More Information Needed]
34
+ - **Demo [optional]:** [More Information Needed]
35
+
36
+ ## Uses
37
+
38
+ <!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
39
+
40
+ ### Direct Use
41
+
42
+ <!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
43
+
44
+ [More Information Needed]
45
+
46
+ ### Downstream Use [optional]
47
+
48
+ <!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
49
+
50
+ [More Information Needed]
51
+
52
+ ### Out-of-Scope Use
53
+
54
+ <!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
55
+
56
+ [More Information Needed]
57
+
58
+ ## Bias, Risks, and Limitations
59
+
60
+ <!-- This section is meant to convey both technical and sociotechnical limitations. -->
61
+
62
+ [More Information Needed]
63
+
64
+ ### Recommendations
65
+
66
+ <!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
67
+
68
+ Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
69
+
70
+ ## How to Get Started with the Model
71
+
72
+ Use the code below to get started with the model.
73
+
74
+ [More Information Needed]
75
+
76
+ ## Training Details
77
+
78
+ ### Training Data
79
+
80
+ <!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
81
+
82
+ [More Information Needed]
83
+
84
+ ### Training Procedure
85
+
86
+ <!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
87
+
88
+ #### Preprocessing [optional]
89
+
90
+ [More Information Needed]
91
+
92
+
93
+ #### Training Hyperparameters
94
+
95
+ - **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
96
+
97
+ #### Speeds, Sizes, Times [optional]
98
+
99
+ <!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
100
+
101
+ [More Information Needed]
102
+
103
+ ## Evaluation
104
+
105
+ <!-- This section describes the evaluation protocols and provides the results. -->
106
+
107
+ ### Testing Data, Factors & Metrics
108
+
109
+ #### Testing Data
110
+
111
+ <!-- This should link to a Dataset Card if possible. -->
112
+
113
+ [More Information Needed]
114
+
115
+ #### Factors
116
+
117
+ <!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
118
+
119
+ [More Information Needed]
120
+
121
+ #### Metrics
122
+
123
+ <!-- These are the evaluation metrics being used, ideally with a description of why. -->
124
+
125
+ [More Information Needed]
126
+
127
+ ### Results
128
+
129
+ [More Information Needed]
130
+
131
+ #### Summary
132
+
133
+
134
+
135
+ ## Model Examination [optional]
136
+
137
+ <!-- Relevant interpretability work for the model goes here -->
138
+
139
+ [More Information Needed]
140
+
141
+ ## Environmental Impact
142
+
143
+ <!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
144
+
145
+ Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
146
+
147
+ - **Hardware Type:** [More Information Needed]
148
+ - **Hours used:** [More Information Needed]
149
+ - **Cloud Provider:** [More Information Needed]
150
+ - **Compute Region:** [More Information Needed]
151
+ - **Carbon Emitted:** [More Information Needed]
152
+
153
+ ## Technical Specifications [optional]
154
+
155
+ ### Model Architecture and Objective
156
+
157
+ [More Information Needed]
158
+
159
+ ### Compute Infrastructure
160
+
161
+ [More Information Needed]
162
+
163
+ #### Hardware
164
+
165
+ [More Information Needed]
166
+
167
+ #### Software
168
+
169
+ [More Information Needed]
170
+
171
+ ## Citation [optional]
172
+
173
+ <!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
174
+
175
+ **BibTeX:**
176
+
177
+ [More Information Needed]
178
+
179
+ **APA:**
180
+
181
+ [More Information Needed]
182
+
183
+ ## Glossary [optional]
184
+
185
+ <!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
186
+
187
+ [More Information Needed]
188
+
189
+ ## More Information [optional]
190
+
191
+ [More Information Needed]
192
+
193
+ ## Model Card Authors [optional]
194
+
195
+ [More Information Needed]
196
+
197
+ ## Model Card Contact
198
+
199
+ [More Information Needed]
200
+ ### Framework versions
201
+
202
+ - PEFT 0.13.2
adapter_config.json ADDED
@@ -0,0 +1,34 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "alpha_pattern": {},
3
+ "auto_mapping": null,
4
+ "base_model_name_or_path": "liuhaotian/llava-v1.5-13b",
5
+ "bias": "none",
6
+ "fan_in_fan_out": false,
7
+ "inference_mode": true,
8
+ "init_lora_weights": true,
9
+ "layer_replication": null,
10
+ "layers_pattern": null,
11
+ "layers_to_transform": null,
12
+ "loftq_config": {},
13
+ "lora_alpha": 16,
14
+ "lora_dropout": 0.05,
15
+ "megatron_config": null,
16
+ "megatron_core": "megatron.core",
17
+ "modules_to_save": null,
18
+ "peft_type": "LORA",
19
+ "r": 8,
20
+ "rank_pattern": {},
21
+ "revision": null,
22
+ "target_modules": [
23
+ "up_proj",
24
+ "k_proj",
25
+ "v_proj",
26
+ "gate_proj",
27
+ "o_proj",
28
+ "down_proj",
29
+ "q_proj"
30
+ ],
31
+ "task_type": "CAUSAL_LM",
32
+ "use_dora": false,
33
+ "use_rslora": false
34
+ }
adapter_model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c0ca6fa3684a839a6dd096790f5cb429d7bac913ab314c609dc0399ab43390ad
3
+ size 62660864
checkpoint-224/README.md ADDED
@@ -0,0 +1,202 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ base_model: liuhaotian/llava-v1.5-13b
3
+ library_name: peft
4
+ ---
5
+
6
+ # Model Card for Model ID
7
+
8
+ <!-- Provide a quick summary of what the model is/does. -->
9
+
10
+
11
+
12
+ ## Model Details
13
+
14
+ ### Model Description
15
+
16
+ <!-- Provide a longer summary of what this model is. -->
17
+
18
+
19
+
20
+ - **Developed by:** [More Information Needed]
21
+ - **Funded by [optional]:** [More Information Needed]
22
+ - **Shared by [optional]:** [More Information Needed]
23
+ - **Model type:** [More Information Needed]
24
+ - **Language(s) (NLP):** [More Information Needed]
25
+ - **License:** [More Information Needed]
26
+ - **Finetuned from model [optional]:** [More Information Needed]
27
+
28
+ ### Model Sources [optional]
29
+
30
+ <!-- Provide the basic links for the model. -->
31
+
32
+ - **Repository:** [More Information Needed]
33
+ - **Paper [optional]:** [More Information Needed]
34
+ - **Demo [optional]:** [More Information Needed]
35
+
36
+ ## Uses
37
+
38
+ <!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
39
+
40
+ ### Direct Use
41
+
42
+ <!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
43
+
44
+ [More Information Needed]
45
+
46
+ ### Downstream Use [optional]
47
+
48
+ <!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
49
+
50
+ [More Information Needed]
51
+
52
+ ### Out-of-Scope Use
53
+
54
+ <!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
55
+
56
+ [More Information Needed]
57
+
58
+ ## Bias, Risks, and Limitations
59
+
60
+ <!-- This section is meant to convey both technical and sociotechnical limitations. -->
61
+
62
+ [More Information Needed]
63
+
64
+ ### Recommendations
65
+
66
+ <!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
67
+
68
+ Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
69
+
70
+ ## How to Get Started with the Model
71
+
72
+ Use the code below to get started with the model.
73
+
74
+ [More Information Needed]
75
+
76
+ ## Training Details
77
+
78
+ ### Training Data
79
+
80
+ <!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
81
+
82
+ [More Information Needed]
83
+
84
+ ### Training Procedure
85
+
86
+ <!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
87
+
88
+ #### Preprocessing [optional]
89
+
90
+ [More Information Needed]
91
+
92
+
93
+ #### Training Hyperparameters
94
+
95
+ - **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
96
+
97
+ #### Speeds, Sizes, Times [optional]
98
+
99
+ <!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
100
+
101
+ [More Information Needed]
102
+
103
+ ## Evaluation
104
+
105
+ <!-- This section describes the evaluation protocols and provides the results. -->
106
+
107
+ ### Testing Data, Factors & Metrics
108
+
109
+ #### Testing Data
110
+
111
+ <!-- This should link to a Dataset Card if possible. -->
112
+
113
+ [More Information Needed]
114
+
115
+ #### Factors
116
+
117
+ <!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
118
+
119
+ [More Information Needed]
120
+
121
+ #### Metrics
122
+
123
+ <!-- These are the evaluation metrics being used, ideally with a description of why. -->
124
+
125
+ [More Information Needed]
126
+
127
+ ### Results
128
+
129
+ [More Information Needed]
130
+
131
+ #### Summary
132
+
133
+
134
+
135
+ ## Model Examination [optional]
136
+
137
+ <!-- Relevant interpretability work for the model goes here -->
138
+
139
+ [More Information Needed]
140
+
141
+ ## Environmental Impact
142
+
143
+ <!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
144
+
145
+ Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
146
+
147
+ - **Hardware Type:** [More Information Needed]
148
+ - **Hours used:** [More Information Needed]
149
+ - **Cloud Provider:** [More Information Needed]
150
+ - **Compute Region:** [More Information Needed]
151
+ - **Carbon Emitted:** [More Information Needed]
152
+
153
+ ## Technical Specifications [optional]
154
+
155
+ ### Model Architecture and Objective
156
+
157
+ [More Information Needed]
158
+
159
+ ### Compute Infrastructure
160
+
161
+ [More Information Needed]
162
+
163
+ #### Hardware
164
+
165
+ [More Information Needed]
166
+
167
+ #### Software
168
+
169
+ [More Information Needed]
170
+
171
+ ## Citation [optional]
172
+
173
+ <!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
174
+
175
+ **BibTeX:**
176
+
177
+ [More Information Needed]
178
+
179
+ **APA:**
180
+
181
+ [More Information Needed]
182
+
183
+ ## Glossary [optional]
184
+
185
+ <!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
186
+
187
+ [More Information Needed]
188
+
189
+ ## More Information [optional]
190
+
191
+ [More Information Needed]
192
+
193
+ ## Model Card Authors [optional]
194
+
195
+ [More Information Needed]
196
+
197
+ ## Model Card Contact
198
+
199
+ [More Information Needed]
200
+ ### Framework versions
201
+
202
+ - PEFT 0.13.2
checkpoint-224/adapter_config.json ADDED
@@ -0,0 +1,34 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "alpha_pattern": {},
3
+ "auto_mapping": null,
4
+ "base_model_name_or_path": "liuhaotian/llava-v1.5-13b",
5
+ "bias": "none",
6
+ "fan_in_fan_out": false,
7
+ "inference_mode": true,
8
+ "init_lora_weights": true,
9
+ "layer_replication": null,
10
+ "layers_pattern": null,
11
+ "layers_to_transform": null,
12
+ "loftq_config": {},
13
+ "lora_alpha": 16,
14
+ "lora_dropout": 0.05,
15
+ "megatron_config": null,
16
+ "megatron_core": "megatron.core",
17
+ "modules_to_save": null,
18
+ "peft_type": "LORA",
19
+ "r": 8,
20
+ "rank_pattern": {},
21
+ "revision": null,
22
+ "target_modules": [
23
+ "v_proj",
24
+ "down_proj",
25
+ "up_proj",
26
+ "q_proj",
27
+ "o_proj",
28
+ "k_proj",
29
+ "gate_proj"
30
+ ],
31
+ "task_type": "CAUSAL_LM",
32
+ "use_dora": false,
33
+ "use_rslora": false
34
+ }
checkpoint-224/adapter_model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c0ca6fa3684a839a6dd096790f5cb429d7bac913ab314c609dc0399ab43390ad
3
+ size 62660864
checkpoint-224/global_step224/zero_pp_rank_0_mp_rank_00_model_states.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:809c10f935f479188e529d8e510db416031e71c61c901246cbff27407af9f9a4
3
+ size 593618
checkpoint-224/global_step224/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 sha256:137299d61c0b7289153a620a6f2c7773a828e56c86835b354d73fcb65a3949d1
3
+ size 188286957
checkpoint-224/global_step224/zero_pp_rank_1_mp_rank_00_model_states.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e5cdba4711b4b0d505ef85e7f5337c47e7f5e2da1639a74c451238d88a5025df
3
+ size 593618
checkpoint-224/global_step224/zero_pp_rank_1_mp_rank_00_optim_states.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:344f6a277d819456eefe986b75f61b1caef404db8f10ce1a1aaf23e6ad464af7
3
+ size 188286957
checkpoint-224/global_step224/zero_pp_rank_2_mp_rank_00_model_states.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e32eaa7a0b327ca2f2cb690a2aed3e10e6c9155107947b0563bd7b1f17c58cee
3
+ size 593618
checkpoint-224/global_step224/zero_pp_rank_2_mp_rank_00_optim_states.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0b1f7a9eacac7373d9db9f2eb64e400b5f6c7767db0ec3bccb20f1292f75fd0d
3
+ size 188286957
checkpoint-224/global_step224/zero_pp_rank_3_mp_rank_00_model_states.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ad6eef1ecdff0f96311700f48b433a118a173b7c92e643450314d1e702f2503d
3
+ size 593618
checkpoint-224/global_step224/zero_pp_rank_3_mp_rank_00_optim_states.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:759ec840d614ccc1ef432f6d7ef14d5cf0c039a1c98e13bb80c6778f7d6ab6a8
3
+ size 188286957
checkpoint-224/latest ADDED
@@ -0,0 +1 @@
 
 
1
+ global_step224
checkpoint-224/rng_state_0.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b32a0d78d2c8a662384fae768e5baa216e31a9e0266798371b91daacdc50f66b
3
+ size 14960
checkpoint-224/rng_state_1.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b04c2612dcb2b3d3fd7c114e9a5fdd7ec6de1ea0b8cc1ac278cbf363b67e5eb5
3
+ size 14960
checkpoint-224/rng_state_2.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:652571be01e2aa6a8f0308d56ef329284a66de86ae1f13d2a777a17d7a5fbea8
3
+ size 14960
checkpoint-224/rng_state_3.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c7f2cb749c8a1d3504339650d317841f232205be1d955c4f0dab1072cd2ecfe7
3
+ size 14960
checkpoint-224/special_tokens_map.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token": {
3
+ "content": "<s>",
4
+ "lstrip": false,
5
+ "normalized": false,
6
+ "rstrip": false,
7
+ "single_word": false
8
+ },
9
+ "eos_token": {
10
+ "content": "</s>",
11
+ "lstrip": false,
12
+ "normalized": false,
13
+ "rstrip": false,
14
+ "single_word": false
15
+ },
16
+ "pad_token": "<unk>",
17
+ "unk_token": {
18
+ "content": "<unk>",
19
+ "lstrip": false,
20
+ "normalized": false,
21
+ "rstrip": false,
22
+ "single_word": false
23
+ }
24
+ }
checkpoint-224/tokenizer.model ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347
3
+ size 499723
checkpoint-224/tokenizer_config.json ADDED
@@ -0,0 +1,43 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_bos_token": true,
3
+ "add_eos_token": false,
4
+ "add_prefix_space": true,
5
+ "added_tokens_decoder": {
6
+ "0": {
7
+ "content": "<unk>",
8
+ "lstrip": false,
9
+ "normalized": false,
10
+ "rstrip": false,
11
+ "single_word": false,
12
+ "special": true
13
+ },
14
+ "1": {
15
+ "content": "<s>",
16
+ "lstrip": false,
17
+ "normalized": false,
18
+ "rstrip": false,
19
+ "single_word": false,
20
+ "special": true
21
+ },
22
+ "2": {
23
+ "content": "</s>",
24
+ "lstrip": false,
25
+ "normalized": false,
26
+ "rstrip": false,
27
+ "single_word": false,
28
+ "special": true
29
+ }
30
+ },
31
+ "bos_token": "<s>",
32
+ "clean_up_tokenization_spaces": false,
33
+ "eos_token": "</s>",
34
+ "legacy": false,
35
+ "model_max_length": 2048,
36
+ "pad_token": "<unk>",
37
+ "padding_side": "right",
38
+ "sp_model_kwargs": {},
39
+ "spaces_between_special_tokens": false,
40
+ "tokenizer_class": "LlamaTokenizer",
41
+ "unk_token": "<unk>",
42
+ "use_default_system_prompt": false
43
+ }
checkpoint-224/trainer_state.json ADDED
@@ -0,0 +1,3393 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "best_metric": 0.6895740032196045,
3
+ "best_model_checkpoint": "./checkpoints/llava-v1.5-13b/checkpoint-224",
4
+ "epoch": 7.0,
5
+ "eval_steps": 1.0,
6
+ "global_step": 224,
7
+ "is_hyper_param_search": false,
8
+ "is_local_process_zero": true,
9
+ "is_world_process_zero": true,
10
+ "log_history": [
11
+ {
12
+ "epoch": 0.03125,
13
+ "grad_norm": 0.2380081706918525,
14
+ "learning_rate": 0.0,
15
+ "loss": 1.2458,
16
+ "step": 1
17
+ },
18
+ {
19
+ "epoch": 0.03125,
20
+ "eval_loss": 1.3161638975143433,
21
+ "eval_runtime": 50.8995,
22
+ "eval_samples_per_second": 3.929,
23
+ "eval_steps_per_second": 0.255,
24
+ "step": 1
25
+ },
26
+ {
27
+ "epoch": 0.0625,
28
+ "grad_norm": 0.20429495268987705,
29
+ "learning_rate": 8.613531161467863e-06,
30
+ "loss": 1.2003,
31
+ "step": 2
32
+ },
33
+ {
34
+ "epoch": 0.0625,
35
+ "eval_loss": 1.3161638975143433,
36
+ "eval_runtime": 47.4818,
37
+ "eval_samples_per_second": 4.212,
38
+ "eval_steps_per_second": 0.274,
39
+ "step": 2
40
+ },
41
+ {
42
+ "epoch": 0.09375,
43
+ "grad_norm": 0.20616215800420787,
44
+ "learning_rate": 1.3652123889719709e-05,
45
+ "loss": 1.2622,
46
+ "step": 3
47
+ },
48
+ {
49
+ "epoch": 0.09375,
50
+ "eval_loss": 1.309991478919983,
51
+ "eval_runtime": 47.4152,
52
+ "eval_samples_per_second": 4.218,
53
+ "eval_steps_per_second": 0.274,
54
+ "step": 3
55
+ },
56
+ {
57
+ "epoch": 0.125,
58
+ "grad_norm": 0.20155595022101944,
59
+ "learning_rate": 1.7227062322935725e-05,
60
+ "loss": 1.2845,
61
+ "step": 4
62
+ },
63
+ {
64
+ "epoch": 0.125,
65
+ "eval_loss": 1.3013781309127808,
66
+ "eval_runtime": 47.4814,
67
+ "eval_samples_per_second": 4.212,
68
+ "eval_steps_per_second": 0.274,
69
+ "step": 4
70
+ },
71
+ {
72
+ "epoch": 0.15625,
73
+ "grad_norm": 0.21113117474989132,
74
+ "learning_rate": 2e-05,
75
+ "loss": 1.246,
76
+ "step": 5
77
+ },
78
+ {
79
+ "epoch": 0.15625,
80
+ "eval_loss": 1.2892160415649414,
81
+ "eval_runtime": 47.7209,
82
+ "eval_samples_per_second": 4.191,
83
+ "eval_steps_per_second": 0.272,
84
+ "step": 5
85
+ },
86
+ {
87
+ "epoch": 0.1875,
88
+ "grad_norm": 0.21377946631015488,
89
+ "learning_rate": 2e-05,
90
+ "loss": 1.2684,
91
+ "step": 6
92
+ },
93
+ {
94
+ "epoch": 0.1875,
95
+ "eval_loss": 1.2754532098770142,
96
+ "eval_runtime": 47.5781,
97
+ "eval_samples_per_second": 4.204,
98
+ "eval_steps_per_second": 0.273,
99
+ "step": 6
100
+ },
101
+ {
102
+ "epoch": 0.21875,
103
+ "grad_norm": 0.2284268997618767,
104
+ "learning_rate": 2e-05,
105
+ "loss": 1.2681,
106
+ "step": 7
107
+ },
108
+ {
109
+ "epoch": 0.21875,
110
+ "eval_loss": 1.2605774402618408,
111
+ "eval_runtime": 47.5326,
112
+ "eval_samples_per_second": 4.208,
113
+ "eval_steps_per_second": 0.273,
114
+ "step": 7
115
+ },
116
+ {
117
+ "epoch": 0.25,
118
+ "grad_norm": 0.23585343568544442,
119
+ "learning_rate": 2e-05,
120
+ "loss": 1.2407,
121
+ "step": 8
122
+ },
123
+ {
124
+ "epoch": 0.25,
125
+ "eval_loss": 1.244718313217163,
126
+ "eval_runtime": 47.5001,
127
+ "eval_samples_per_second": 4.211,
128
+ "eval_steps_per_second": 0.274,
129
+ "step": 8
130
+ },
131
+ {
132
+ "epoch": 0.28125,
133
+ "grad_norm": 0.23051191992462533,
134
+ "learning_rate": 2e-05,
135
+ "loss": 1.2766,
136
+ "step": 9
137
+ },
138
+ {
139
+ "epoch": 0.28125,
140
+ "eval_loss": 1.2285138368606567,
141
+ "eval_runtime": 47.4631,
142
+ "eval_samples_per_second": 4.214,
143
+ "eval_steps_per_second": 0.274,
144
+ "step": 9
145
+ },
146
+ {
147
+ "epoch": 0.3125,
148
+ "grad_norm": 0.22726394327484983,
149
+ "learning_rate": 2e-05,
150
+ "loss": 1.2024,
151
+ "step": 10
152
+ },
153
+ {
154
+ "epoch": 0.3125,
155
+ "eval_loss": 1.2118008136749268,
156
+ "eval_runtime": 47.4991,
157
+ "eval_samples_per_second": 4.211,
158
+ "eval_steps_per_second": 0.274,
159
+ "step": 10
160
+ },
161
+ {
162
+ "epoch": 0.34375,
163
+ "grad_norm": 0.25404890894461285,
164
+ "learning_rate": 2e-05,
165
+ "loss": 1.2742,
166
+ "step": 11
167
+ },
168
+ {
169
+ "epoch": 0.34375,
170
+ "eval_loss": 1.1942989826202393,
171
+ "eval_runtime": 49.2609,
172
+ "eval_samples_per_second": 4.06,
173
+ "eval_steps_per_second": 0.264,
174
+ "step": 11
175
+ },
176
+ {
177
+ "epoch": 0.375,
178
+ "grad_norm": 0.26336210916526287,
179
+ "learning_rate": 2e-05,
180
+ "loss": 1.2258,
181
+ "step": 12
182
+ },
183
+ {
184
+ "epoch": 0.375,
185
+ "eval_loss": 1.176426649093628,
186
+ "eval_runtime": 49.0639,
187
+ "eval_samples_per_second": 4.076,
188
+ "eval_steps_per_second": 0.265,
189
+ "step": 12
190
+ },
191
+ {
192
+ "epoch": 0.40625,
193
+ "grad_norm": 0.29637148470746666,
194
+ "learning_rate": 2e-05,
195
+ "loss": 1.2345,
196
+ "step": 13
197
+ },
198
+ {
199
+ "epoch": 0.40625,
200
+ "eval_loss": 1.1577811241149902,
201
+ "eval_runtime": 49.1352,
202
+ "eval_samples_per_second": 4.07,
203
+ "eval_steps_per_second": 0.265,
204
+ "step": 13
205
+ },
206
+ {
207
+ "epoch": 0.4375,
208
+ "grad_norm": 0.2841880377627424,
209
+ "learning_rate": 2e-05,
210
+ "loss": 1.0765,
211
+ "step": 14
212
+ },
213
+ {
214
+ "epoch": 0.4375,
215
+ "eval_loss": 1.1381279230117798,
216
+ "eval_runtime": 49.25,
217
+ "eval_samples_per_second": 4.061,
218
+ "eval_steps_per_second": 0.264,
219
+ "step": 14
220
+ },
221
+ {
222
+ "epoch": 0.46875,
223
+ "grad_norm": 0.2773140636191091,
224
+ "learning_rate": 2e-05,
225
+ "loss": 1.1812,
226
+ "step": 15
227
+ },
228
+ {
229
+ "epoch": 0.46875,
230
+ "eval_loss": 1.1178216934204102,
231
+ "eval_runtime": 49.0879,
232
+ "eval_samples_per_second": 4.074,
233
+ "eval_steps_per_second": 0.265,
234
+ "step": 15
235
+ },
236
+ {
237
+ "epoch": 0.5,
238
+ "grad_norm": 0.3568607365552051,
239
+ "learning_rate": 2e-05,
240
+ "loss": 1.1327,
241
+ "step": 16
242
+ },
243
+ {
244
+ "epoch": 0.5,
245
+ "eval_loss": 1.0954149961471558,
246
+ "eval_runtime": 48.6546,
247
+ "eval_samples_per_second": 4.111,
248
+ "eval_steps_per_second": 0.267,
249
+ "step": 16
250
+ },
251
+ {
252
+ "epoch": 0.53125,
253
+ "grad_norm": 0.32574391414112897,
254
+ "learning_rate": 2e-05,
255
+ "loss": 1.1162,
256
+ "step": 17
257
+ },
258
+ {
259
+ "epoch": 0.53125,
260
+ "eval_loss": 1.071275234222412,
261
+ "eval_runtime": 48.5618,
262
+ "eval_samples_per_second": 4.118,
263
+ "eval_steps_per_second": 0.268,
264
+ "step": 17
265
+ },
266
+ {
267
+ "epoch": 0.5625,
268
+ "grad_norm": 0.4256864144638081,
269
+ "learning_rate": 2e-05,
270
+ "loss": 1.1138,
271
+ "step": 18
272
+ },
273
+ {
274
+ "epoch": 0.5625,
275
+ "eval_loss": 1.0455905199050903,
276
+ "eval_runtime": 48.4981,
277
+ "eval_samples_per_second": 4.124,
278
+ "eval_steps_per_second": 0.268,
279
+ "step": 18
280
+ },
281
+ {
282
+ "epoch": 0.59375,
283
+ "grad_norm": 0.31230014132112643,
284
+ "learning_rate": 2e-05,
285
+ "loss": 1.0011,
286
+ "step": 19
287
+ },
288
+ {
289
+ "epoch": 0.59375,
290
+ "eval_loss": 1.0208789110183716,
291
+ "eval_runtime": 48.4675,
292
+ "eval_samples_per_second": 4.126,
293
+ "eval_steps_per_second": 0.268,
294
+ "step": 19
295
+ },
296
+ {
297
+ "epoch": 0.625,
298
+ "grad_norm": 0.3025724039243594,
299
+ "learning_rate": 2e-05,
300
+ "loss": 1.109,
301
+ "step": 20
302
+ },
303
+ {
304
+ "epoch": 0.625,
305
+ "eval_loss": 1.002480149269104,
306
+ "eval_runtime": 48.5265,
307
+ "eval_samples_per_second": 4.121,
308
+ "eval_steps_per_second": 0.268,
309
+ "step": 20
310
+ },
311
+ {
312
+ "epoch": 0.65625,
313
+ "grad_norm": 0.27787879590501874,
314
+ "learning_rate": 2e-05,
315
+ "loss": 1.0291,
316
+ "step": 21
317
+ },
318
+ {
319
+ "epoch": 0.65625,
320
+ "eval_loss": 0.9933492541313171,
321
+ "eval_runtime": 50.0369,
322
+ "eval_samples_per_second": 3.997,
323
+ "eval_steps_per_second": 0.26,
324
+ "step": 21
325
+ },
326
+ {
327
+ "epoch": 0.6875,
328
+ "grad_norm": 0.4231294067130801,
329
+ "learning_rate": 2e-05,
330
+ "loss": 1.0779,
331
+ "step": 22
332
+ },
333
+ {
334
+ "epoch": 0.6875,
335
+ "eval_loss": 0.9850385785102844,
336
+ "eval_runtime": 50.0062,
337
+ "eval_samples_per_second": 4.0,
338
+ "eval_steps_per_second": 0.26,
339
+ "step": 22
340
+ },
341
+ {
342
+ "epoch": 0.71875,
343
+ "grad_norm": 0.42130097437373987,
344
+ "learning_rate": 2e-05,
345
+ "loss": 1.0897,
346
+ "step": 23
347
+ },
348
+ {
349
+ "epoch": 0.71875,
350
+ "eval_loss": 0.9758670330047607,
351
+ "eval_runtime": 50.1031,
352
+ "eval_samples_per_second": 3.992,
353
+ "eval_steps_per_second": 0.259,
354
+ "step": 23
355
+ },
356
+ {
357
+ "epoch": 0.75,
358
+ "grad_norm": 0.27711808063263893,
359
+ "learning_rate": 2e-05,
360
+ "loss": 1.0739,
361
+ "step": 24
362
+ },
363
+ {
364
+ "epoch": 0.75,
365
+ "eval_loss": 0.9674506187438965,
366
+ "eval_runtime": 50.0337,
367
+ "eval_samples_per_second": 3.997,
368
+ "eval_steps_per_second": 0.26,
369
+ "step": 24
370
+ },
371
+ {
372
+ "epoch": 0.78125,
373
+ "grad_norm": 0.2879649409281791,
374
+ "learning_rate": 2e-05,
375
+ "loss": 1.0182,
376
+ "step": 25
377
+ },
378
+ {
379
+ "epoch": 0.78125,
380
+ "eval_loss": 0.9592065215110779,
381
+ "eval_runtime": 50.0709,
382
+ "eval_samples_per_second": 3.994,
383
+ "eval_steps_per_second": 0.26,
384
+ "step": 25
385
+ },
386
+ {
387
+ "epoch": 0.8125,
388
+ "grad_norm": 0.19327450826076825,
389
+ "learning_rate": 2e-05,
390
+ "loss": 1.0413,
391
+ "step": 26
392
+ },
393
+ {
394
+ "epoch": 0.8125,
395
+ "eval_loss": 0.9518552422523499,
396
+ "eval_runtime": 50.0572,
397
+ "eval_samples_per_second": 3.995,
398
+ "eval_steps_per_second": 0.26,
399
+ "step": 26
400
+ },
401
+ {
402
+ "epoch": 0.84375,
403
+ "grad_norm": 0.19707021382445633,
404
+ "learning_rate": 2e-05,
405
+ "loss": 0.9525,
406
+ "step": 27
407
+ },
408
+ {
409
+ "epoch": 0.84375,
410
+ "eval_loss": 0.9449941515922546,
411
+ "eval_runtime": 50.0515,
412
+ "eval_samples_per_second": 3.996,
413
+ "eval_steps_per_second": 0.26,
414
+ "step": 27
415
+ },
416
+ {
417
+ "epoch": 0.875,
418
+ "grad_norm": 0.2420270757641518,
419
+ "learning_rate": 2e-05,
420
+ "loss": 0.9658,
421
+ "step": 28
422
+ },
423
+ {
424
+ "epoch": 0.875,
425
+ "eval_loss": 0.9378474354743958,
426
+ "eval_runtime": 49.9299,
427
+ "eval_samples_per_second": 4.006,
428
+ "eval_steps_per_second": 0.26,
429
+ "step": 28
430
+ },
431
+ {
432
+ "epoch": 0.90625,
433
+ "grad_norm": 0.18074632782127534,
434
+ "learning_rate": 2e-05,
435
+ "loss": 0.9866,
436
+ "step": 29
437
+ },
438
+ {
439
+ "epoch": 0.90625,
440
+ "eval_loss": 0.93099045753479,
441
+ "eval_runtime": 50.0096,
442
+ "eval_samples_per_second": 3.999,
443
+ "eval_steps_per_second": 0.26,
444
+ "step": 29
445
+ },
446
+ {
447
+ "epoch": 0.9375,
448
+ "grad_norm": 0.1936051126921734,
449
+ "learning_rate": 2e-05,
450
+ "loss": 1.0128,
451
+ "step": 30
452
+ },
453
+ {
454
+ "epoch": 0.9375,
455
+ "eval_loss": 0.9244199991226196,
456
+ "eval_runtime": 50.2469,
457
+ "eval_samples_per_second": 3.98,
458
+ "eval_steps_per_second": 0.259,
459
+ "step": 30
460
+ },
461
+ {
462
+ "epoch": 0.96875,
463
+ "grad_norm": 0.26164254459782943,
464
+ "learning_rate": 2e-05,
465
+ "loss": 0.88,
466
+ "step": 31
467
+ },
468
+ {
469
+ "epoch": 0.96875,
470
+ "eval_loss": 0.9175177216529846,
471
+ "eval_runtime": 50.1695,
472
+ "eval_samples_per_second": 3.986,
473
+ "eval_steps_per_second": 0.259,
474
+ "step": 31
475
+ },
476
+ {
477
+ "epoch": 1.0,
478
+ "grad_norm": 0.18677152741688485,
479
+ "learning_rate": 2e-05,
480
+ "loss": 0.9569,
481
+ "step": 32
482
+ },
483
+ {
484
+ "epoch": 1.0,
485
+ "eval_loss": 0.9108598828315735,
486
+ "eval_runtime": 50.0387,
487
+ "eval_samples_per_second": 3.997,
488
+ "eval_steps_per_second": 0.26,
489
+ "step": 32
490
+ },
491
+ {
492
+ "epoch": 1.03125,
493
+ "grad_norm": 0.20486279036126417,
494
+ "learning_rate": 2e-05,
495
+ "loss": 1.0208,
496
+ "step": 33
497
+ },
498
+ {
499
+ "epoch": 1.03125,
500
+ "eval_loss": 0.9042049646377563,
501
+ "eval_runtime": 50.1472,
502
+ "eval_samples_per_second": 3.988,
503
+ "eval_steps_per_second": 0.259,
504
+ "step": 33
505
+ },
506
+ {
507
+ "epoch": 1.0625,
508
+ "grad_norm": 0.2004946169291112,
509
+ "learning_rate": 2e-05,
510
+ "loss": 0.9931,
511
+ "step": 34
512
+ },
513
+ {
514
+ "epoch": 1.0625,
515
+ "eval_loss": 0.8980298042297363,
516
+ "eval_runtime": 50.245,
517
+ "eval_samples_per_second": 3.98,
518
+ "eval_steps_per_second": 0.259,
519
+ "step": 34
520
+ },
521
+ {
522
+ "epoch": 1.09375,
523
+ "grad_norm": 0.1645872432258401,
524
+ "learning_rate": 2e-05,
525
+ "loss": 1.0184,
526
+ "step": 35
527
+ },
528
+ {
529
+ "epoch": 1.09375,
530
+ "eval_loss": 0.8924428820610046,
531
+ "eval_runtime": 50.3703,
532
+ "eval_samples_per_second": 3.971,
533
+ "eval_steps_per_second": 0.258,
534
+ "step": 35
535
+ },
536
+ {
537
+ "epoch": 1.125,
538
+ "grad_norm": 0.18293519304435016,
539
+ "learning_rate": 2e-05,
540
+ "loss": 1.0026,
541
+ "step": 36
542
+ },
543
+ {
544
+ "epoch": 1.125,
545
+ "eval_loss": 0.8870412707328796,
546
+ "eval_runtime": 50.0483,
547
+ "eval_samples_per_second": 3.996,
548
+ "eval_steps_per_second": 0.26,
549
+ "step": 36
550
+ },
551
+ {
552
+ "epoch": 1.15625,
553
+ "grad_norm": 0.17712548516246762,
554
+ "learning_rate": 2e-05,
555
+ "loss": 0.9387,
556
+ "step": 37
557
+ },
558
+ {
559
+ "epoch": 1.15625,
560
+ "eval_loss": 0.881915271282196,
561
+ "eval_runtime": 49.9751,
562
+ "eval_samples_per_second": 4.002,
563
+ "eval_steps_per_second": 0.26,
564
+ "step": 37
565
+ },
566
+ {
567
+ "epoch": 1.1875,
568
+ "grad_norm": 0.21472689311609464,
569
+ "learning_rate": 2e-05,
570
+ "loss": 0.958,
571
+ "step": 38
572
+ },
573
+ {
574
+ "epoch": 1.1875,
575
+ "eval_loss": 0.8768754601478577,
576
+ "eval_runtime": 50.1204,
577
+ "eval_samples_per_second": 3.99,
578
+ "eval_steps_per_second": 0.259,
579
+ "step": 38
580
+ },
581
+ {
582
+ "epoch": 1.21875,
583
+ "grad_norm": 0.21117297910005806,
584
+ "learning_rate": 2e-05,
585
+ "loss": 0.9922,
586
+ "step": 39
587
+ },
588
+ {
589
+ "epoch": 1.21875,
590
+ "eval_loss": 0.8718628883361816,
591
+ "eval_runtime": 50.1732,
592
+ "eval_samples_per_second": 3.986,
593
+ "eval_steps_per_second": 0.259,
594
+ "step": 39
595
+ },
596
+ {
597
+ "epoch": 1.25,
598
+ "grad_norm": 0.17835587003909165,
599
+ "learning_rate": 2e-05,
600
+ "loss": 0.9776,
601
+ "step": 40
602
+ },
603
+ {
604
+ "epoch": 1.25,
605
+ "eval_loss": 0.8669865131378174,
606
+ "eval_runtime": 50.1148,
607
+ "eval_samples_per_second": 3.991,
608
+ "eval_steps_per_second": 0.259,
609
+ "step": 40
610
+ },
611
+ {
612
+ "epoch": 1.28125,
613
+ "grad_norm": 0.2092736372483734,
614
+ "learning_rate": 2e-05,
615
+ "loss": 0.9731,
616
+ "step": 41
617
+ },
618
+ {
619
+ "epoch": 1.28125,
620
+ "eval_loss": 0.8619834780693054,
621
+ "eval_runtime": 50.052,
622
+ "eval_samples_per_second": 3.996,
623
+ "eval_steps_per_second": 0.26,
624
+ "step": 41
625
+ },
626
+ {
627
+ "epoch": 1.3125,
628
+ "grad_norm": 0.2338857391910308,
629
+ "learning_rate": 2e-05,
630
+ "loss": 0.9319,
631
+ "step": 42
632
+ },
633
+ {
634
+ "epoch": 1.3125,
635
+ "eval_loss": 0.8572126030921936,
636
+ "eval_runtime": 50.1212,
637
+ "eval_samples_per_second": 3.99,
638
+ "eval_steps_per_second": 0.259,
639
+ "step": 42
640
+ },
641
+ {
642
+ "epoch": 1.34375,
643
+ "grad_norm": 0.19168719284572813,
644
+ "learning_rate": 2e-05,
645
+ "loss": 0.9083,
646
+ "step": 43
647
+ },
648
+ {
649
+ "epoch": 1.34375,
650
+ "eval_loss": 0.8525611758232117,
651
+ "eval_runtime": 50.1733,
652
+ "eval_samples_per_second": 3.986,
653
+ "eval_steps_per_second": 0.259,
654
+ "step": 43
655
+ },
656
+ {
657
+ "epoch": 1.375,
658
+ "grad_norm": 0.20004868138433377,
659
+ "learning_rate": 2e-05,
660
+ "loss": 0.9118,
661
+ "step": 44
662
+ },
663
+ {
664
+ "epoch": 1.375,
665
+ "eval_loss": 0.8483461141586304,
666
+ "eval_runtime": 50.1083,
667
+ "eval_samples_per_second": 3.991,
668
+ "eval_steps_per_second": 0.259,
669
+ "step": 44
670
+ },
671
+ {
672
+ "epoch": 1.40625,
673
+ "grad_norm": 0.19012965506122342,
674
+ "learning_rate": 2e-05,
675
+ "loss": 0.8888,
676
+ "step": 45
677
+ },
678
+ {
679
+ "epoch": 1.40625,
680
+ "eval_loss": 0.8446614742279053,
681
+ "eval_runtime": 50.1171,
682
+ "eval_samples_per_second": 3.991,
683
+ "eval_steps_per_second": 0.259,
684
+ "step": 45
685
+ },
686
+ {
687
+ "epoch": 1.4375,
688
+ "grad_norm": 0.21187005706805245,
689
+ "learning_rate": 2e-05,
690
+ "loss": 0.9319,
691
+ "step": 46
692
+ },
693
+ {
694
+ "epoch": 1.4375,
695
+ "eval_loss": 0.8412036299705505,
696
+ "eval_runtime": 50.0918,
697
+ "eval_samples_per_second": 3.993,
698
+ "eval_steps_per_second": 0.26,
699
+ "step": 46
700
+ },
701
+ {
702
+ "epoch": 1.46875,
703
+ "grad_norm": 0.19673832205926584,
704
+ "learning_rate": 2e-05,
705
+ "loss": 0.9359,
706
+ "step": 47
707
+ },
708
+ {
709
+ "epoch": 1.46875,
710
+ "eval_loss": 0.8380417823791504,
711
+ "eval_runtime": 50.2214,
712
+ "eval_samples_per_second": 3.982,
713
+ "eval_steps_per_second": 0.259,
714
+ "step": 47
715
+ },
716
+ {
717
+ "epoch": 1.5,
718
+ "grad_norm": 0.21712294106174318,
719
+ "learning_rate": 2e-05,
720
+ "loss": 0.8511,
721
+ "step": 48
722
+ },
723
+ {
724
+ "epoch": 1.5,
725
+ "eval_loss": 0.8353021740913391,
726
+ "eval_runtime": 50.1617,
727
+ "eval_samples_per_second": 3.987,
728
+ "eval_steps_per_second": 0.259,
729
+ "step": 48
730
+ },
731
+ {
732
+ "epoch": 1.53125,
733
+ "grad_norm": 0.2138924779700934,
734
+ "learning_rate": 2e-05,
735
+ "loss": 0.8695,
736
+ "step": 49
737
+ },
738
+ {
739
+ "epoch": 1.53125,
740
+ "eval_loss": 0.8327407836914062,
741
+ "eval_runtime": 50.1442,
742
+ "eval_samples_per_second": 3.988,
743
+ "eval_steps_per_second": 0.259,
744
+ "step": 49
745
+ },
746
+ {
747
+ "epoch": 1.5625,
748
+ "grad_norm": 0.22387442384578618,
749
+ "learning_rate": 2e-05,
750
+ "loss": 0.8518,
751
+ "step": 50
752
+ },
753
+ {
754
+ "epoch": 1.5625,
755
+ "eval_loss": 0.8301742076873779,
756
+ "eval_runtime": 50.1867,
757
+ "eval_samples_per_second": 3.985,
758
+ "eval_steps_per_second": 0.259,
759
+ "step": 50
760
+ },
761
+ {
762
+ "epoch": 1.59375,
763
+ "grad_norm": 0.1975577146517192,
764
+ "learning_rate": 2e-05,
765
+ "loss": 0.8868,
766
+ "step": 51
767
+ },
768
+ {
769
+ "epoch": 1.59375,
770
+ "eval_loss": 0.8275265693664551,
771
+ "eval_runtime": 51.2257,
772
+ "eval_samples_per_second": 3.904,
773
+ "eval_steps_per_second": 0.254,
774
+ "step": 51
775
+ },
776
+ {
777
+ "epoch": 1.625,
778
+ "grad_norm": 0.21474817057286624,
779
+ "learning_rate": 2e-05,
780
+ "loss": 0.767,
781
+ "step": 52
782
+ },
783
+ {
784
+ "epoch": 1.625,
785
+ "eval_loss": 0.824796736240387,
786
+ "eval_runtime": 51.276,
787
+ "eval_samples_per_second": 3.9,
788
+ "eval_steps_per_second": 0.254,
789
+ "step": 52
790
+ },
791
+ {
792
+ "epoch": 1.65625,
793
+ "grad_norm": 0.21105651676755652,
794
+ "learning_rate": 2e-05,
795
+ "loss": 0.9219,
796
+ "step": 53
797
+ },
798
+ {
799
+ "epoch": 1.65625,
800
+ "eval_loss": 0.8221166729927063,
801
+ "eval_runtime": 51.141,
802
+ "eval_samples_per_second": 3.911,
803
+ "eval_steps_per_second": 0.254,
804
+ "step": 53
805
+ },
806
+ {
807
+ "epoch": 1.6875,
808
+ "grad_norm": 0.20706475184742085,
809
+ "learning_rate": 2e-05,
810
+ "loss": 0.8873,
811
+ "step": 54
812
+ },
813
+ {
814
+ "epoch": 1.6875,
815
+ "eval_loss": 0.819589376449585,
816
+ "eval_runtime": 51.0045,
817
+ "eval_samples_per_second": 3.921,
818
+ "eval_steps_per_second": 0.255,
819
+ "step": 54
820
+ },
821
+ {
822
+ "epoch": 1.71875,
823
+ "grad_norm": 0.21722220033855957,
824
+ "learning_rate": 2e-05,
825
+ "loss": 0.8956,
826
+ "step": 55
827
+ },
828
+ {
829
+ "epoch": 1.71875,
830
+ "eval_loss": 0.8176340460777283,
831
+ "eval_runtime": 51.1941,
832
+ "eval_samples_per_second": 3.907,
833
+ "eval_steps_per_second": 0.254,
834
+ "step": 55
835
+ },
836
+ {
837
+ "epoch": 1.75,
838
+ "grad_norm": 0.20669001221665667,
839
+ "learning_rate": 2e-05,
840
+ "loss": 0.9506,
841
+ "step": 56
842
+ },
843
+ {
844
+ "epoch": 1.75,
845
+ "eval_loss": 0.8158826231956482,
846
+ "eval_runtime": 52.1162,
847
+ "eval_samples_per_second": 3.838,
848
+ "eval_steps_per_second": 0.249,
849
+ "step": 56
850
+ },
851
+ {
852
+ "epoch": 1.78125,
853
+ "grad_norm": 0.22189732090066341,
854
+ "learning_rate": 2e-05,
855
+ "loss": 0.8955,
856
+ "step": 57
857
+ },
858
+ {
859
+ "epoch": 1.78125,
860
+ "eval_loss": 0.814656674861908,
861
+ "eval_runtime": 52.1361,
862
+ "eval_samples_per_second": 3.836,
863
+ "eval_steps_per_second": 0.249,
864
+ "step": 57
865
+ },
866
+ {
867
+ "epoch": 1.8125,
868
+ "grad_norm": 0.2030113892848459,
869
+ "learning_rate": 2e-05,
870
+ "loss": 0.9108,
871
+ "step": 58
872
+ },
873
+ {
874
+ "epoch": 1.8125,
875
+ "eval_loss": 0.813343346118927,
876
+ "eval_runtime": 52.2552,
877
+ "eval_samples_per_second": 3.827,
878
+ "eval_steps_per_second": 0.249,
879
+ "step": 58
880
+ },
881
+ {
882
+ "epoch": 1.84375,
883
+ "grad_norm": 0.2123201057569791,
884
+ "learning_rate": 2e-05,
885
+ "loss": 0.8779,
886
+ "step": 59
887
+ },
888
+ {
889
+ "epoch": 1.84375,
890
+ "eval_loss": 0.8116877675056458,
891
+ "eval_runtime": 52.1233,
892
+ "eval_samples_per_second": 3.837,
893
+ "eval_steps_per_second": 0.249,
894
+ "step": 59
895
+ },
896
+ {
897
+ "epoch": 1.875,
898
+ "grad_norm": 0.211551126937912,
899
+ "learning_rate": 2e-05,
900
+ "loss": 0.9294,
901
+ "step": 60
902
+ },
903
+ {
904
+ "epoch": 1.875,
905
+ "eval_loss": 0.8098442554473877,
906
+ "eval_runtime": 52.1091,
907
+ "eval_samples_per_second": 3.838,
908
+ "eval_steps_per_second": 0.249,
909
+ "step": 60
910
+ },
911
+ {
912
+ "epoch": 1.90625,
913
+ "grad_norm": 0.24981344981629752,
914
+ "learning_rate": 2e-05,
915
+ "loss": 0.8409,
916
+ "step": 61
917
+ },
918
+ {
919
+ "epoch": 1.90625,
920
+ "eval_loss": 0.8070770502090454,
921
+ "eval_runtime": 53.4187,
922
+ "eval_samples_per_second": 3.744,
923
+ "eval_steps_per_second": 0.243,
924
+ "step": 61
925
+ },
926
+ {
927
+ "epoch": 1.9375,
928
+ "grad_norm": 0.2341550589775159,
929
+ "learning_rate": 2e-05,
930
+ "loss": 0.888,
931
+ "step": 62
932
+ },
933
+ {
934
+ "epoch": 1.9375,
935
+ "eval_loss": 0.8040286898612976,
936
+ "eval_runtime": 53.2197,
937
+ "eval_samples_per_second": 3.758,
938
+ "eval_steps_per_second": 0.244,
939
+ "step": 62
940
+ },
941
+ {
942
+ "epoch": 1.96875,
943
+ "grad_norm": 0.2336241775649256,
944
+ "learning_rate": 2e-05,
945
+ "loss": 0.913,
946
+ "step": 63
947
+ },
948
+ {
949
+ "epoch": 1.96875,
950
+ "eval_loss": 0.8013430833816528,
951
+ "eval_runtime": 53.1784,
952
+ "eval_samples_per_second": 3.761,
953
+ "eval_steps_per_second": 0.244,
954
+ "step": 63
955
+ },
956
+ {
957
+ "epoch": 2.0,
958
+ "grad_norm": 0.2414390628081758,
959
+ "learning_rate": 2e-05,
960
+ "loss": 0.8754,
961
+ "step": 64
962
+ },
963
+ {
964
+ "epoch": 2.0,
965
+ "eval_loss": 0.7985894680023193,
966
+ "eval_runtime": 53.2454,
967
+ "eval_samples_per_second": 3.756,
968
+ "eval_steps_per_second": 0.244,
969
+ "step": 64
970
+ },
971
+ {
972
+ "epoch": 2.03125,
973
+ "grad_norm": 0.2484104465653703,
974
+ "learning_rate": 2e-05,
975
+ "loss": 0.8497,
976
+ "step": 65
977
+ },
978
+ {
979
+ "epoch": 2.03125,
980
+ "eval_loss": 0.7954932451248169,
981
+ "eval_runtime": 53.3794,
982
+ "eval_samples_per_second": 3.747,
983
+ "eval_steps_per_second": 0.244,
984
+ "step": 65
985
+ },
986
+ {
987
+ "epoch": 2.0625,
988
+ "grad_norm": 0.23859744120942086,
989
+ "learning_rate": 2e-05,
990
+ "loss": 0.8567,
991
+ "step": 66
992
+ },
993
+ {
994
+ "epoch": 2.0625,
995
+ "eval_loss": 0.7929843068122864,
996
+ "eval_runtime": 55.517,
997
+ "eval_samples_per_second": 3.602,
998
+ "eval_steps_per_second": 0.234,
999
+ "step": 66
1000
+ },
1001
+ {
1002
+ "epoch": 2.09375,
1003
+ "grad_norm": 0.24584758647855462,
1004
+ "learning_rate": 2e-05,
1005
+ "loss": 0.8489,
1006
+ "step": 67
1007
+ },
1008
+ {
1009
+ "epoch": 2.09375,
1010
+ "eval_loss": 0.7903321981430054,
1011
+ "eval_runtime": 55.4151,
1012
+ "eval_samples_per_second": 3.609,
1013
+ "eval_steps_per_second": 0.235,
1014
+ "step": 67
1015
+ },
1016
+ {
1017
+ "epoch": 2.125,
1018
+ "grad_norm": 0.2484917818304153,
1019
+ "learning_rate": 2e-05,
1020
+ "loss": 0.9122,
1021
+ "step": 68
1022
+ },
1023
+ {
1024
+ "epoch": 2.125,
1025
+ "eval_loss": 0.7877185344696045,
1026
+ "eval_runtime": 55.4069,
1027
+ "eval_samples_per_second": 3.61,
1028
+ "eval_steps_per_second": 0.235,
1029
+ "step": 68
1030
+ },
1031
+ {
1032
+ "epoch": 2.15625,
1033
+ "grad_norm": 0.2184614083026819,
1034
+ "learning_rate": 2e-05,
1035
+ "loss": 0.8355,
1036
+ "step": 69
1037
+ },
1038
+ {
1039
+ "epoch": 2.15625,
1040
+ "eval_loss": 0.7852210998535156,
1041
+ "eval_runtime": 55.3381,
1042
+ "eval_samples_per_second": 3.614,
1043
+ "eval_steps_per_second": 0.235,
1044
+ "step": 69
1045
+ },
1046
+ {
1047
+ "epoch": 2.1875,
1048
+ "grad_norm": 0.24978410070800153,
1049
+ "learning_rate": 2e-05,
1050
+ "loss": 0.7968,
1051
+ "step": 70
1052
+ },
1053
+ {
1054
+ "epoch": 2.1875,
1055
+ "eval_loss": 0.7827157378196716,
1056
+ "eval_runtime": 55.3708,
1057
+ "eval_samples_per_second": 3.612,
1058
+ "eval_steps_per_second": 0.235,
1059
+ "step": 70
1060
+ },
1061
+ {
1062
+ "epoch": 2.21875,
1063
+ "grad_norm": 0.23059883325890385,
1064
+ "learning_rate": 2e-05,
1065
+ "loss": 0.8783,
1066
+ "step": 71
1067
+ },
1068
+ {
1069
+ "epoch": 2.21875,
1070
+ "eval_loss": 0.7805906534194946,
1071
+ "eval_runtime": 55.6033,
1072
+ "eval_samples_per_second": 3.597,
1073
+ "eval_steps_per_second": 0.234,
1074
+ "step": 71
1075
+ },
1076
+ {
1077
+ "epoch": 2.25,
1078
+ "grad_norm": 0.23261007334915096,
1079
+ "learning_rate": 2e-05,
1080
+ "loss": 0.7956,
1081
+ "step": 72
1082
+ },
1083
+ {
1084
+ "epoch": 2.25,
1085
+ "eval_loss": 0.7786691784858704,
1086
+ "eval_runtime": 55.0913,
1087
+ "eval_samples_per_second": 3.63,
1088
+ "eval_steps_per_second": 0.236,
1089
+ "step": 72
1090
+ },
1091
+ {
1092
+ "epoch": 2.28125,
1093
+ "grad_norm": 0.25779598356574085,
1094
+ "learning_rate": 2e-05,
1095
+ "loss": 0.8426,
1096
+ "step": 73
1097
+ },
1098
+ {
1099
+ "epoch": 2.28125,
1100
+ "eval_loss": 0.7771151661872864,
1101
+ "eval_runtime": 55.0698,
1102
+ "eval_samples_per_second": 3.632,
1103
+ "eval_steps_per_second": 0.236,
1104
+ "step": 73
1105
+ },
1106
+ {
1107
+ "epoch": 2.3125,
1108
+ "grad_norm": 0.2288243335971112,
1109
+ "learning_rate": 2e-05,
1110
+ "loss": 0.8381,
1111
+ "step": 74
1112
+ },
1113
+ {
1114
+ "epoch": 2.3125,
1115
+ "eval_loss": 0.7756838202476501,
1116
+ "eval_runtime": 54.8412,
1117
+ "eval_samples_per_second": 3.647,
1118
+ "eval_steps_per_second": 0.237,
1119
+ "step": 74
1120
+ },
1121
+ {
1122
+ "epoch": 2.34375,
1123
+ "grad_norm": 0.24235644907977733,
1124
+ "learning_rate": 2e-05,
1125
+ "loss": 0.887,
1126
+ "step": 75
1127
+ },
1128
+ {
1129
+ "epoch": 2.34375,
1130
+ "eval_loss": 0.7739972472190857,
1131
+ "eval_runtime": 54.9718,
1132
+ "eval_samples_per_second": 3.638,
1133
+ "eval_steps_per_second": 0.236,
1134
+ "step": 75
1135
+ },
1136
+ {
1137
+ "epoch": 2.375,
1138
+ "grad_norm": 0.23666820017867402,
1139
+ "learning_rate": 2e-05,
1140
+ "loss": 0.8007,
1141
+ "step": 76
1142
+ },
1143
+ {
1144
+ "epoch": 2.375,
1145
+ "eval_loss": 0.7724328637123108,
1146
+ "eval_runtime": 55.0225,
1147
+ "eval_samples_per_second": 3.635,
1148
+ "eval_steps_per_second": 0.236,
1149
+ "step": 76
1150
+ },
1151
+ {
1152
+ "epoch": 2.40625,
1153
+ "grad_norm": 0.22815737396609181,
1154
+ "learning_rate": 2e-05,
1155
+ "loss": 0.8529,
1156
+ "step": 77
1157
+ },
1158
+ {
1159
+ "epoch": 2.40625,
1160
+ "eval_loss": 0.7710004448890686,
1161
+ "eval_runtime": 55.321,
1162
+ "eval_samples_per_second": 3.615,
1163
+ "eval_steps_per_second": 0.235,
1164
+ "step": 77
1165
+ },
1166
+ {
1167
+ "epoch": 2.4375,
1168
+ "grad_norm": 0.2701264871470739,
1169
+ "learning_rate": 2e-05,
1170
+ "loss": 0.8515,
1171
+ "step": 78
1172
+ },
1173
+ {
1174
+ "epoch": 2.4375,
1175
+ "eval_loss": 0.7695322632789612,
1176
+ "eval_runtime": 55.3045,
1177
+ "eval_samples_per_second": 3.616,
1178
+ "eval_steps_per_second": 0.235,
1179
+ "step": 78
1180
+ },
1181
+ {
1182
+ "epoch": 2.46875,
1183
+ "grad_norm": 0.24363813951328234,
1184
+ "learning_rate": 2e-05,
1185
+ "loss": 0.8587,
1186
+ "step": 79
1187
+ },
1188
+ {
1189
+ "epoch": 2.46875,
1190
+ "eval_loss": 0.7689024209976196,
1191
+ "eval_runtime": 55.3009,
1192
+ "eval_samples_per_second": 3.617,
1193
+ "eval_steps_per_second": 0.235,
1194
+ "step": 79
1195
+ },
1196
+ {
1197
+ "epoch": 2.5,
1198
+ "grad_norm": 0.30924701355253065,
1199
+ "learning_rate": 2e-05,
1200
+ "loss": 0.9076,
1201
+ "step": 80
1202
+ },
1203
+ {
1204
+ "epoch": 2.5,
1205
+ "eval_loss": 0.7676254510879517,
1206
+ "eval_runtime": 55.2365,
1207
+ "eval_samples_per_second": 3.621,
1208
+ "eval_steps_per_second": 0.235,
1209
+ "step": 80
1210
+ },
1211
+ {
1212
+ "epoch": 2.53125,
1213
+ "grad_norm": 0.2665188280221636,
1214
+ "learning_rate": 2e-05,
1215
+ "loss": 0.8445,
1216
+ "step": 81
1217
+ },
1218
+ {
1219
+ "epoch": 2.53125,
1220
+ "eval_loss": 0.7661146521568298,
1221
+ "eval_runtime": 55.2775,
1222
+ "eval_samples_per_second": 3.618,
1223
+ "eval_steps_per_second": 0.235,
1224
+ "step": 81
1225
+ },
1226
+ {
1227
+ "epoch": 2.5625,
1228
+ "grad_norm": 0.24674191720675534,
1229
+ "learning_rate": 2e-05,
1230
+ "loss": 0.8882,
1231
+ "step": 82
1232
+ },
1233
+ {
1234
+ "epoch": 2.5625,
1235
+ "eval_loss": 0.76513671875,
1236
+ "eval_runtime": 55.0857,
1237
+ "eval_samples_per_second": 3.631,
1238
+ "eval_steps_per_second": 0.236,
1239
+ "step": 82
1240
+ },
1241
+ {
1242
+ "epoch": 2.59375,
1243
+ "grad_norm": 0.2736689405531704,
1244
+ "learning_rate": 2e-05,
1245
+ "loss": 0.8336,
1246
+ "step": 83
1247
+ },
1248
+ {
1249
+ "epoch": 2.59375,
1250
+ "eval_loss": 0.764373779296875,
1251
+ "eval_runtime": 55.2069,
1252
+ "eval_samples_per_second": 3.623,
1253
+ "eval_steps_per_second": 0.235,
1254
+ "step": 83
1255
+ },
1256
+ {
1257
+ "epoch": 2.625,
1258
+ "grad_norm": 0.290841287198557,
1259
+ "learning_rate": 2e-05,
1260
+ "loss": 0.795,
1261
+ "step": 84
1262
+ },
1263
+ {
1264
+ "epoch": 2.625,
1265
+ "eval_loss": 0.7632084488868713,
1266
+ "eval_runtime": 55.1009,
1267
+ "eval_samples_per_second": 3.63,
1268
+ "eval_steps_per_second": 0.236,
1269
+ "step": 84
1270
+ },
1271
+ {
1272
+ "epoch": 2.65625,
1273
+ "grad_norm": 0.2912051076836381,
1274
+ "learning_rate": 2e-05,
1275
+ "loss": 0.772,
1276
+ "step": 85
1277
+ },
1278
+ {
1279
+ "epoch": 2.65625,
1280
+ "eval_loss": 0.7618446350097656,
1281
+ "eval_runtime": 55.3717,
1282
+ "eval_samples_per_second": 3.612,
1283
+ "eval_steps_per_second": 0.235,
1284
+ "step": 85
1285
+ },
1286
+ {
1287
+ "epoch": 2.6875,
1288
+ "grad_norm": 0.3169908538809109,
1289
+ "learning_rate": 2e-05,
1290
+ "loss": 0.8148,
1291
+ "step": 86
1292
+ },
1293
+ {
1294
+ "epoch": 2.6875,
1295
+ "eval_loss": 0.7599577307701111,
1296
+ "eval_runtime": 55.3931,
1297
+ "eval_samples_per_second": 3.611,
1298
+ "eval_steps_per_second": 0.235,
1299
+ "step": 86
1300
+ },
1301
+ {
1302
+ "epoch": 2.71875,
1303
+ "grad_norm": 0.28780549186847426,
1304
+ "learning_rate": 2e-05,
1305
+ "loss": 0.8154,
1306
+ "step": 87
1307
+ },
1308
+ {
1309
+ "epoch": 2.71875,
1310
+ "eval_loss": 0.7583369612693787,
1311
+ "eval_runtime": 55.1679,
1312
+ "eval_samples_per_second": 3.625,
1313
+ "eval_steps_per_second": 0.236,
1314
+ "step": 87
1315
+ },
1316
+ {
1317
+ "epoch": 2.75,
1318
+ "grad_norm": 0.30695250620091474,
1319
+ "learning_rate": 2e-05,
1320
+ "loss": 0.9032,
1321
+ "step": 88
1322
+ },
1323
+ {
1324
+ "epoch": 2.75,
1325
+ "eval_loss": 0.7571613192558289,
1326
+ "eval_runtime": 55.1779,
1327
+ "eval_samples_per_second": 3.625,
1328
+ "eval_steps_per_second": 0.236,
1329
+ "step": 88
1330
+ },
1331
+ {
1332
+ "epoch": 2.78125,
1333
+ "grad_norm": 0.2693887416759828,
1334
+ "learning_rate": 2e-05,
1335
+ "loss": 0.8106,
1336
+ "step": 89
1337
+ },
1338
+ {
1339
+ "epoch": 2.78125,
1340
+ "eval_loss": 0.7566004991531372,
1341
+ "eval_runtime": 55.1107,
1342
+ "eval_samples_per_second": 3.629,
1343
+ "eval_steps_per_second": 0.236,
1344
+ "step": 89
1345
+ },
1346
+ {
1347
+ "epoch": 2.8125,
1348
+ "grad_norm": 0.2887583627563198,
1349
+ "learning_rate": 2e-05,
1350
+ "loss": 0.8518,
1351
+ "step": 90
1352
+ },
1353
+ {
1354
+ "epoch": 2.8125,
1355
+ "eval_loss": 0.7558963298797607,
1356
+ "eval_runtime": 55.2153,
1357
+ "eval_samples_per_second": 3.622,
1358
+ "eval_steps_per_second": 0.235,
1359
+ "step": 90
1360
+ },
1361
+ {
1362
+ "epoch": 2.84375,
1363
+ "grad_norm": 0.3059402168979351,
1364
+ "learning_rate": 2e-05,
1365
+ "loss": 0.7727,
1366
+ "step": 91
1367
+ },
1368
+ {
1369
+ "epoch": 2.84375,
1370
+ "eval_loss": 0.7545350790023804,
1371
+ "eval_runtime": 55.3225,
1372
+ "eval_samples_per_second": 3.615,
1373
+ "eval_steps_per_second": 0.235,
1374
+ "step": 91
1375
+ },
1376
+ {
1377
+ "epoch": 2.875,
1378
+ "grad_norm": 0.3096260477909968,
1379
+ "learning_rate": 2e-05,
1380
+ "loss": 0.8477,
1381
+ "step": 92
1382
+ },
1383
+ {
1384
+ "epoch": 2.875,
1385
+ "eval_loss": 0.7526452541351318,
1386
+ "eval_runtime": 55.4311,
1387
+ "eval_samples_per_second": 3.608,
1388
+ "eval_steps_per_second": 0.235,
1389
+ "step": 92
1390
+ },
1391
+ {
1392
+ "epoch": 2.90625,
1393
+ "grad_norm": 0.31498884686525297,
1394
+ "learning_rate": 2e-05,
1395
+ "loss": 0.7982,
1396
+ "step": 93
1397
+ },
1398
+ {
1399
+ "epoch": 2.90625,
1400
+ "eval_loss": 0.7510760426521301,
1401
+ "eval_runtime": 55.4361,
1402
+ "eval_samples_per_second": 3.608,
1403
+ "eval_steps_per_second": 0.235,
1404
+ "step": 93
1405
+ },
1406
+ {
1407
+ "epoch": 2.9375,
1408
+ "grad_norm": 0.31302830623184313,
1409
+ "learning_rate": 2e-05,
1410
+ "loss": 0.871,
1411
+ "step": 94
1412
+ },
1413
+ {
1414
+ "epoch": 2.9375,
1415
+ "eval_loss": 0.7500898838043213,
1416
+ "eval_runtime": 55.3025,
1417
+ "eval_samples_per_second": 3.616,
1418
+ "eval_steps_per_second": 0.235,
1419
+ "step": 94
1420
+ },
1421
+ {
1422
+ "epoch": 2.96875,
1423
+ "grad_norm": 0.3132608568779145,
1424
+ "learning_rate": 2e-05,
1425
+ "loss": 0.8094,
1426
+ "step": 95
1427
+ },
1428
+ {
1429
+ "epoch": 2.96875,
1430
+ "eval_loss": 0.7498895525932312,
1431
+ "eval_runtime": 55.2402,
1432
+ "eval_samples_per_second": 3.621,
1433
+ "eval_steps_per_second": 0.235,
1434
+ "step": 95
1435
+ },
1436
+ {
1437
+ "epoch": 3.0,
1438
+ "grad_norm": 0.298645350091386,
1439
+ "learning_rate": 2e-05,
1440
+ "loss": 0.7673,
1441
+ "step": 96
1442
+ },
1443
+ {
1444
+ "epoch": 3.0,
1445
+ "eval_loss": 0.7493192553520203,
1446
+ "eval_runtime": 54.8718,
1447
+ "eval_samples_per_second": 3.645,
1448
+ "eval_steps_per_second": 0.237,
1449
+ "step": 96
1450
+ },
1451
+ {
1452
+ "epoch": 3.03125,
1453
+ "grad_norm": 0.34042584783125357,
1454
+ "learning_rate": 2e-05,
1455
+ "loss": 0.7336,
1456
+ "step": 97
1457
+ },
1458
+ {
1459
+ "epoch": 3.03125,
1460
+ "eval_loss": 0.7476670742034912,
1461
+ "eval_runtime": 54.9305,
1462
+ "eval_samples_per_second": 3.641,
1463
+ "eval_steps_per_second": 0.237,
1464
+ "step": 97
1465
+ },
1466
+ {
1467
+ "epoch": 3.0625,
1468
+ "grad_norm": 0.293099043801068,
1469
+ "learning_rate": 2e-05,
1470
+ "loss": 0.8088,
1471
+ "step": 98
1472
+ },
1473
+ {
1474
+ "epoch": 3.0625,
1475
+ "eval_loss": 0.745802640914917,
1476
+ "eval_runtime": 55.2051,
1477
+ "eval_samples_per_second": 3.623,
1478
+ "eval_steps_per_second": 0.235,
1479
+ "step": 98
1480
+ },
1481
+ {
1482
+ "epoch": 3.09375,
1483
+ "grad_norm": 0.3042839507858426,
1484
+ "learning_rate": 2e-05,
1485
+ "loss": 0.787,
1486
+ "step": 99
1487
+ },
1488
+ {
1489
+ "epoch": 3.09375,
1490
+ "eval_loss": 0.7439618110656738,
1491
+ "eval_runtime": 55.0065,
1492
+ "eval_samples_per_second": 3.636,
1493
+ "eval_steps_per_second": 0.236,
1494
+ "step": 99
1495
+ },
1496
+ {
1497
+ "epoch": 3.125,
1498
+ "grad_norm": 0.32992077073227005,
1499
+ "learning_rate": 2e-05,
1500
+ "loss": 0.8296,
1501
+ "step": 100
1502
+ },
1503
+ {
1504
+ "epoch": 3.125,
1505
+ "eval_loss": 0.7424842715263367,
1506
+ "eval_runtime": 55.1254,
1507
+ "eval_samples_per_second": 3.628,
1508
+ "eval_steps_per_second": 0.236,
1509
+ "step": 100
1510
+ },
1511
+ {
1512
+ "epoch": 3.15625,
1513
+ "grad_norm": 0.2798839747424062,
1514
+ "learning_rate": 2e-05,
1515
+ "loss": 0.7642,
1516
+ "step": 101
1517
+ },
1518
+ {
1519
+ "epoch": 3.15625,
1520
+ "eval_loss": 0.7414796948432922,
1521
+ "eval_runtime": 49.183,
1522
+ "eval_samples_per_second": 4.066,
1523
+ "eval_steps_per_second": 0.264,
1524
+ "step": 101
1525
+ },
1526
+ {
1527
+ "epoch": 3.1875,
1528
+ "grad_norm": 0.3046631191964983,
1529
+ "learning_rate": 2e-05,
1530
+ "loss": 0.8203,
1531
+ "step": 102
1532
+ },
1533
+ {
1534
+ "epoch": 3.1875,
1535
+ "eval_loss": 0.7410265207290649,
1536
+ "eval_runtime": 48.1541,
1537
+ "eval_samples_per_second": 4.153,
1538
+ "eval_steps_per_second": 0.27,
1539
+ "step": 102
1540
+ },
1541
+ {
1542
+ "epoch": 3.21875,
1543
+ "grad_norm": 0.3117517214859861,
1544
+ "learning_rate": 2e-05,
1545
+ "loss": 0.8222,
1546
+ "step": 103
1547
+ },
1548
+ {
1549
+ "epoch": 3.21875,
1550
+ "eval_loss": 0.7405675649642944,
1551
+ "eval_runtime": 47.7145,
1552
+ "eval_samples_per_second": 4.192,
1553
+ "eval_steps_per_second": 0.272,
1554
+ "step": 103
1555
+ },
1556
+ {
1557
+ "epoch": 3.25,
1558
+ "grad_norm": 0.3412709249466801,
1559
+ "learning_rate": 2e-05,
1560
+ "loss": 0.7459,
1561
+ "step": 104
1562
+ },
1563
+ {
1564
+ "epoch": 3.25,
1565
+ "eval_loss": 0.7395681738853455,
1566
+ "eval_runtime": 47.5855,
1567
+ "eval_samples_per_second": 4.203,
1568
+ "eval_steps_per_second": 0.273,
1569
+ "step": 104
1570
+ },
1571
+ {
1572
+ "epoch": 3.28125,
1573
+ "grad_norm": 0.2917443566507923,
1574
+ "learning_rate": 2e-05,
1575
+ "loss": 0.7849,
1576
+ "step": 105
1577
+ },
1578
+ {
1579
+ "epoch": 3.28125,
1580
+ "eval_loss": 0.7387100458145142,
1581
+ "eval_runtime": 47.6344,
1582
+ "eval_samples_per_second": 4.199,
1583
+ "eval_steps_per_second": 0.273,
1584
+ "step": 105
1585
+ },
1586
+ {
1587
+ "epoch": 3.3125,
1588
+ "grad_norm": 0.3054484743574741,
1589
+ "learning_rate": 2e-05,
1590
+ "loss": 0.8354,
1591
+ "step": 106
1592
+ },
1593
+ {
1594
+ "epoch": 3.3125,
1595
+ "eval_loss": 0.7384718060493469,
1596
+ "eval_runtime": 47.8373,
1597
+ "eval_samples_per_second": 4.181,
1598
+ "eval_steps_per_second": 0.272,
1599
+ "step": 106
1600
+ },
1601
+ {
1602
+ "epoch": 3.34375,
1603
+ "grad_norm": 0.34986630381114014,
1604
+ "learning_rate": 2e-05,
1605
+ "loss": 0.7069,
1606
+ "step": 107
1607
+ },
1608
+ {
1609
+ "epoch": 3.34375,
1610
+ "eval_loss": 0.737342357635498,
1611
+ "eval_runtime": 47.5763,
1612
+ "eval_samples_per_second": 4.204,
1613
+ "eval_steps_per_second": 0.273,
1614
+ "step": 107
1615
+ },
1616
+ {
1617
+ "epoch": 3.375,
1618
+ "grad_norm": 0.32324403145716496,
1619
+ "learning_rate": 2e-05,
1620
+ "loss": 0.767,
1621
+ "step": 108
1622
+ },
1623
+ {
1624
+ "epoch": 3.375,
1625
+ "eval_loss": 0.7360101938247681,
1626
+ "eval_runtime": 47.5774,
1627
+ "eval_samples_per_second": 4.204,
1628
+ "eval_steps_per_second": 0.273,
1629
+ "step": 108
1630
+ },
1631
+ {
1632
+ "epoch": 3.40625,
1633
+ "grad_norm": 0.3795969851258545,
1634
+ "learning_rate": 2e-05,
1635
+ "loss": 0.7556,
1636
+ "step": 109
1637
+ },
1638
+ {
1639
+ "epoch": 3.40625,
1640
+ "eval_loss": 0.7339167594909668,
1641
+ "eval_runtime": 47.5818,
1642
+ "eval_samples_per_second": 4.203,
1643
+ "eval_steps_per_second": 0.273,
1644
+ "step": 109
1645
+ },
1646
+ {
1647
+ "epoch": 3.4375,
1648
+ "grad_norm": 0.34401062275458993,
1649
+ "learning_rate": 2e-05,
1650
+ "loss": 0.7494,
1651
+ "step": 110
1652
+ },
1653
+ {
1654
+ "epoch": 3.4375,
1655
+ "eval_loss": 0.7321068644523621,
1656
+ "eval_runtime": 47.7643,
1657
+ "eval_samples_per_second": 4.187,
1658
+ "eval_steps_per_second": 0.272,
1659
+ "step": 110
1660
+ },
1661
+ {
1662
+ "epoch": 3.46875,
1663
+ "grad_norm": 0.3248480010385237,
1664
+ "learning_rate": 2e-05,
1665
+ "loss": 0.8103,
1666
+ "step": 111
1667
+ },
1668
+ {
1669
+ "epoch": 3.46875,
1670
+ "eval_loss": 0.7309197783470154,
1671
+ "eval_runtime": 49.5841,
1672
+ "eval_samples_per_second": 4.034,
1673
+ "eval_steps_per_second": 0.262,
1674
+ "step": 111
1675
+ },
1676
+ {
1677
+ "epoch": 3.5,
1678
+ "grad_norm": 0.3572409124813593,
1679
+ "learning_rate": 2e-05,
1680
+ "loss": 0.7972,
1681
+ "step": 112
1682
+ },
1683
+ {
1684
+ "epoch": 3.5,
1685
+ "eval_loss": 0.7301727533340454,
1686
+ "eval_runtime": 49.3728,
1687
+ "eval_samples_per_second": 4.051,
1688
+ "eval_steps_per_second": 0.263,
1689
+ "step": 112
1690
+ },
1691
+ {
1692
+ "epoch": 3.53125,
1693
+ "grad_norm": 0.37348522775103665,
1694
+ "learning_rate": 2e-05,
1695
+ "loss": 0.88,
1696
+ "step": 113
1697
+ },
1698
+ {
1699
+ "epoch": 3.53125,
1700
+ "eval_loss": 0.7292957305908203,
1701
+ "eval_runtime": 49.2192,
1702
+ "eval_samples_per_second": 4.063,
1703
+ "eval_steps_per_second": 0.264,
1704
+ "step": 113
1705
+ },
1706
+ {
1707
+ "epoch": 3.5625,
1708
+ "grad_norm": 0.37667450960329546,
1709
+ "learning_rate": 2e-05,
1710
+ "loss": 0.7518,
1711
+ "step": 114
1712
+ },
1713
+ {
1714
+ "epoch": 3.5625,
1715
+ "eval_loss": 0.728556215763092,
1716
+ "eval_runtime": 49.0971,
1717
+ "eval_samples_per_second": 4.074,
1718
+ "eval_steps_per_second": 0.265,
1719
+ "step": 114
1720
+ },
1721
+ {
1722
+ "epoch": 3.59375,
1723
+ "grad_norm": 0.3163628607304638,
1724
+ "learning_rate": 2e-05,
1725
+ "loss": 0.7948,
1726
+ "step": 115
1727
+ },
1728
+ {
1729
+ "epoch": 3.59375,
1730
+ "eval_loss": 0.7287828326225281,
1731
+ "eval_runtime": 49.0213,
1732
+ "eval_samples_per_second": 4.08,
1733
+ "eval_steps_per_second": 0.265,
1734
+ "step": 115
1735
+ },
1736
+ {
1737
+ "epoch": 3.625,
1738
+ "grad_norm": 0.3038899302084592,
1739
+ "learning_rate": 2e-05,
1740
+ "loss": 0.7791,
1741
+ "step": 116
1742
+ },
1743
+ {
1744
+ "epoch": 3.625,
1745
+ "eval_loss": 0.7294514179229736,
1746
+ "eval_runtime": 51.9137,
1747
+ "eval_samples_per_second": 3.853,
1748
+ "eval_steps_per_second": 0.25,
1749
+ "step": 116
1750
+ },
1751
+ {
1752
+ "epoch": 3.65625,
1753
+ "grad_norm": 0.3746448663122327,
1754
+ "learning_rate": 2e-05,
1755
+ "loss": 0.7863,
1756
+ "step": 117
1757
+ },
1758
+ {
1759
+ "epoch": 3.65625,
1760
+ "eval_loss": 0.7289304137229919,
1761
+ "eval_runtime": 51.3023,
1762
+ "eval_samples_per_second": 3.898,
1763
+ "eval_steps_per_second": 0.253,
1764
+ "step": 117
1765
+ },
1766
+ {
1767
+ "epoch": 3.6875,
1768
+ "grad_norm": 0.4058937381299434,
1769
+ "learning_rate": 2e-05,
1770
+ "loss": 0.7907,
1771
+ "step": 118
1772
+ },
1773
+ {
1774
+ "epoch": 3.6875,
1775
+ "eval_loss": 0.7281011343002319,
1776
+ "eval_runtime": 50.8635,
1777
+ "eval_samples_per_second": 3.932,
1778
+ "eval_steps_per_second": 0.256,
1779
+ "step": 118
1780
+ },
1781
+ {
1782
+ "epoch": 3.71875,
1783
+ "grad_norm": 0.31608065583227885,
1784
+ "learning_rate": 2e-05,
1785
+ "loss": 0.8348,
1786
+ "step": 119
1787
+ },
1788
+ {
1789
+ "epoch": 3.71875,
1790
+ "eval_loss": 0.7280247211456299,
1791
+ "eval_runtime": 50.4903,
1792
+ "eval_samples_per_second": 3.961,
1793
+ "eval_steps_per_second": 0.257,
1794
+ "step": 119
1795
+ },
1796
+ {
1797
+ "epoch": 3.75,
1798
+ "grad_norm": 0.3375768031046084,
1799
+ "learning_rate": 2e-05,
1800
+ "loss": 0.7783,
1801
+ "step": 120
1802
+ },
1803
+ {
1804
+ "epoch": 3.75,
1805
+ "eval_loss": 0.7281913757324219,
1806
+ "eval_runtime": 50.5906,
1807
+ "eval_samples_per_second": 3.953,
1808
+ "eval_steps_per_second": 0.257,
1809
+ "step": 120
1810
+ },
1811
+ {
1812
+ "epoch": 3.78125,
1813
+ "grad_norm": 0.36047493494859845,
1814
+ "learning_rate": 2e-05,
1815
+ "loss": 0.765,
1816
+ "step": 121
1817
+ },
1818
+ {
1819
+ "epoch": 3.78125,
1820
+ "eval_loss": 0.7269737124443054,
1821
+ "eval_runtime": 53.4722,
1822
+ "eval_samples_per_second": 3.74,
1823
+ "eval_steps_per_second": 0.243,
1824
+ "step": 121
1825
+ },
1826
+ {
1827
+ "epoch": 3.8125,
1828
+ "grad_norm": 0.389743860171921,
1829
+ "learning_rate": 2e-05,
1830
+ "loss": 0.8269,
1831
+ "step": 122
1832
+ },
1833
+ {
1834
+ "epoch": 3.8125,
1835
+ "eval_loss": 0.7251996397972107,
1836
+ "eval_runtime": 53.4986,
1837
+ "eval_samples_per_second": 3.738,
1838
+ "eval_steps_per_second": 0.243,
1839
+ "step": 122
1840
+ },
1841
+ {
1842
+ "epoch": 3.84375,
1843
+ "grad_norm": 0.33850935145960215,
1844
+ "learning_rate": 2e-05,
1845
+ "loss": 0.7497,
1846
+ "step": 123
1847
+ },
1848
+ {
1849
+ "epoch": 3.84375,
1850
+ "eval_loss": 0.723595142364502,
1851
+ "eval_runtime": 53.4196,
1852
+ "eval_samples_per_second": 3.744,
1853
+ "eval_steps_per_second": 0.243,
1854
+ "step": 123
1855
+ },
1856
+ {
1857
+ "epoch": 3.875,
1858
+ "grad_norm": 0.3166770012114478,
1859
+ "learning_rate": 2e-05,
1860
+ "loss": 0.7648,
1861
+ "step": 124
1862
+ },
1863
+ {
1864
+ "epoch": 3.875,
1865
+ "eval_loss": 0.7223578095436096,
1866
+ "eval_runtime": 52.6143,
1867
+ "eval_samples_per_second": 3.801,
1868
+ "eval_steps_per_second": 0.247,
1869
+ "step": 124
1870
+ },
1871
+ {
1872
+ "epoch": 3.90625,
1873
+ "grad_norm": 0.41948670305268276,
1874
+ "learning_rate": 2e-05,
1875
+ "loss": 0.8306,
1876
+ "step": 125
1877
+ },
1878
+ {
1879
+ "epoch": 3.90625,
1880
+ "eval_loss": 0.7206680774688721,
1881
+ "eval_runtime": 52.3885,
1882
+ "eval_samples_per_second": 3.818,
1883
+ "eval_steps_per_second": 0.248,
1884
+ "step": 125
1885
+ },
1886
+ {
1887
+ "epoch": 3.9375,
1888
+ "grad_norm": 0.35580041105853477,
1889
+ "learning_rate": 2e-05,
1890
+ "loss": 0.7945,
1891
+ "step": 126
1892
+ },
1893
+ {
1894
+ "epoch": 3.9375,
1895
+ "eval_loss": 0.7196171283721924,
1896
+ "eval_runtime": 55.1225,
1897
+ "eval_samples_per_second": 3.628,
1898
+ "eval_steps_per_second": 0.236,
1899
+ "step": 126
1900
+ },
1901
+ {
1902
+ "epoch": 3.96875,
1903
+ "grad_norm": 0.38411890663257114,
1904
+ "learning_rate": 2e-05,
1905
+ "loss": 0.7466,
1906
+ "step": 127
1907
+ },
1908
+ {
1909
+ "epoch": 3.96875,
1910
+ "eval_loss": 0.7188088297843933,
1911
+ "eval_runtime": 55.3068,
1912
+ "eval_samples_per_second": 3.616,
1913
+ "eval_steps_per_second": 0.235,
1914
+ "step": 127
1915
+ },
1916
+ {
1917
+ "epoch": 4.0,
1918
+ "grad_norm": 0.3682220575203032,
1919
+ "learning_rate": 2e-05,
1920
+ "loss": 0.6752,
1921
+ "step": 128
1922
+ },
1923
+ {
1924
+ "epoch": 4.0,
1925
+ "eval_loss": 0.7181470990180969,
1926
+ "eval_runtime": 53.9116,
1927
+ "eval_samples_per_second": 3.71,
1928
+ "eval_steps_per_second": 0.241,
1929
+ "step": 128
1930
+ },
1931
+ {
1932
+ "epoch": 4.03125,
1933
+ "grad_norm": 0.34160763542661665,
1934
+ "learning_rate": 2e-05,
1935
+ "loss": 0.7788,
1936
+ "step": 129
1937
+ },
1938
+ {
1939
+ "epoch": 4.03125,
1940
+ "eval_loss": 0.717949390411377,
1941
+ "eval_runtime": 53.8446,
1942
+ "eval_samples_per_second": 3.714,
1943
+ "eval_steps_per_second": 0.241,
1944
+ "step": 129
1945
+ },
1946
+ {
1947
+ "epoch": 4.0625,
1948
+ "grad_norm": 0.35709301353799944,
1949
+ "learning_rate": 2e-05,
1950
+ "loss": 0.8002,
1951
+ "step": 130
1952
+ },
1953
+ {
1954
+ "epoch": 4.0625,
1955
+ "eval_loss": 0.7179380655288696,
1956
+ "eval_runtime": 53.9299,
1957
+ "eval_samples_per_second": 3.709,
1958
+ "eval_steps_per_second": 0.241,
1959
+ "step": 130
1960
+ },
1961
+ {
1962
+ "epoch": 4.09375,
1963
+ "grad_norm": 0.3503147340749238,
1964
+ "learning_rate": 2e-05,
1965
+ "loss": 0.7789,
1966
+ "step": 131
1967
+ },
1968
+ {
1969
+ "epoch": 4.09375,
1970
+ "eval_loss": 0.7180312871932983,
1971
+ "eval_runtime": 53.4091,
1972
+ "eval_samples_per_second": 3.745,
1973
+ "eval_steps_per_second": 0.243,
1974
+ "step": 131
1975
+ },
1976
+ {
1977
+ "epoch": 4.125,
1978
+ "grad_norm": 0.3931715546229069,
1979
+ "learning_rate": 2e-05,
1980
+ "loss": 0.762,
1981
+ "step": 132
1982
+ },
1983
+ {
1984
+ "epoch": 4.125,
1985
+ "eval_loss": 0.717825710773468,
1986
+ "eval_runtime": 53.6366,
1987
+ "eval_samples_per_second": 3.729,
1988
+ "eval_steps_per_second": 0.242,
1989
+ "step": 132
1990
+ },
1991
+ {
1992
+ "epoch": 4.15625,
1993
+ "grad_norm": 0.36864033862644363,
1994
+ "learning_rate": 2e-05,
1995
+ "loss": 0.829,
1996
+ "step": 133
1997
+ },
1998
+ {
1999
+ "epoch": 4.15625,
2000
+ "eval_loss": 0.7178698182106018,
2001
+ "eval_runtime": 53.4891,
2002
+ "eval_samples_per_second": 3.739,
2003
+ "eval_steps_per_second": 0.243,
2004
+ "step": 133
2005
+ },
2006
+ {
2007
+ "epoch": 4.1875,
2008
+ "grad_norm": 0.41393587587462155,
2009
+ "learning_rate": 2e-05,
2010
+ "loss": 0.7624,
2011
+ "step": 134
2012
+ },
2013
+ {
2014
+ "epoch": 4.1875,
2015
+ "eval_loss": 0.7181968092918396,
2016
+ "eval_runtime": 53.5395,
2017
+ "eval_samples_per_second": 3.736,
2018
+ "eval_steps_per_second": 0.243,
2019
+ "step": 134
2020
+ },
2021
+ {
2022
+ "epoch": 4.21875,
2023
+ "grad_norm": 0.36727603900023204,
2024
+ "learning_rate": 2e-05,
2025
+ "loss": 0.7572,
2026
+ "step": 135
2027
+ },
2028
+ {
2029
+ "epoch": 4.21875,
2030
+ "eval_loss": 0.7187527418136597,
2031
+ "eval_runtime": 53.4818,
2032
+ "eval_samples_per_second": 3.74,
2033
+ "eval_steps_per_second": 0.243,
2034
+ "step": 135
2035
+ },
2036
+ {
2037
+ "epoch": 4.25,
2038
+ "grad_norm": 0.3684078795455007,
2039
+ "learning_rate": 2e-05,
2040
+ "loss": 0.7352,
2041
+ "step": 136
2042
+ },
2043
+ {
2044
+ "epoch": 4.25,
2045
+ "eval_loss": 0.7194793820381165,
2046
+ "eval_runtime": 53.4694,
2047
+ "eval_samples_per_second": 3.74,
2048
+ "eval_steps_per_second": 0.243,
2049
+ "step": 136
2050
+ },
2051
+ {
2052
+ "epoch": 4.28125,
2053
+ "grad_norm": 0.42414766562621153,
2054
+ "learning_rate": 2e-05,
2055
+ "loss": 0.7433,
2056
+ "step": 137
2057
+ },
2058
+ {
2059
+ "epoch": 4.28125,
2060
+ "eval_loss": 0.7189603447914124,
2061
+ "eval_runtime": 53.8049,
2062
+ "eval_samples_per_second": 3.717,
2063
+ "eval_steps_per_second": 0.242,
2064
+ "step": 137
2065
+ },
2066
+ {
2067
+ "epoch": 4.3125,
2068
+ "grad_norm": 0.40420796619211563,
2069
+ "learning_rate": 2e-05,
2070
+ "loss": 0.7466,
2071
+ "step": 138
2072
+ },
2073
+ {
2074
+ "epoch": 4.3125,
2075
+ "eval_loss": 0.7173956036567688,
2076
+ "eval_runtime": 53.4014,
2077
+ "eval_samples_per_second": 3.745,
2078
+ "eval_steps_per_second": 0.243,
2079
+ "step": 138
2080
+ },
2081
+ {
2082
+ "epoch": 4.34375,
2083
+ "grad_norm": 0.36419740641344456,
2084
+ "learning_rate": 2e-05,
2085
+ "loss": 0.7045,
2086
+ "step": 139
2087
+ },
2088
+ {
2089
+ "epoch": 4.34375,
2090
+ "eval_loss": 0.7153105139732361,
2091
+ "eval_runtime": 53.285,
2092
+ "eval_samples_per_second": 3.753,
2093
+ "eval_steps_per_second": 0.244,
2094
+ "step": 139
2095
+ },
2096
+ {
2097
+ "epoch": 4.375,
2098
+ "grad_norm": 0.384927357409491,
2099
+ "learning_rate": 2e-05,
2100
+ "loss": 0.7437,
2101
+ "step": 140
2102
+ },
2103
+ {
2104
+ "epoch": 4.375,
2105
+ "eval_loss": 0.7135314345359802,
2106
+ "eval_runtime": 53.4056,
2107
+ "eval_samples_per_second": 3.745,
2108
+ "eval_steps_per_second": 0.243,
2109
+ "step": 140
2110
+ },
2111
+ {
2112
+ "epoch": 4.40625,
2113
+ "grad_norm": 0.37218579680263697,
2114
+ "learning_rate": 2e-05,
2115
+ "loss": 0.7693,
2116
+ "step": 141
2117
+ },
2118
+ {
2119
+ "epoch": 4.40625,
2120
+ "eval_loss": 0.7120725512504578,
2121
+ "eval_runtime": 53.5467,
2122
+ "eval_samples_per_second": 3.735,
2123
+ "eval_steps_per_second": 0.243,
2124
+ "step": 141
2125
+ },
2126
+ {
2127
+ "epoch": 4.4375,
2128
+ "grad_norm": 0.38541382926033946,
2129
+ "learning_rate": 2e-05,
2130
+ "loss": 0.708,
2131
+ "step": 142
2132
+ },
2133
+ {
2134
+ "epoch": 4.4375,
2135
+ "eval_loss": 0.7110380530357361,
2136
+ "eval_runtime": 53.4119,
2137
+ "eval_samples_per_second": 3.744,
2138
+ "eval_steps_per_second": 0.243,
2139
+ "step": 142
2140
+ },
2141
+ {
2142
+ "epoch": 4.46875,
2143
+ "grad_norm": 0.4028726453247759,
2144
+ "learning_rate": 2e-05,
2145
+ "loss": 0.7263,
2146
+ "step": 143
2147
+ },
2148
+ {
2149
+ "epoch": 4.46875,
2150
+ "eval_loss": 0.7100683450698853,
2151
+ "eval_runtime": 53.4337,
2152
+ "eval_samples_per_second": 3.743,
2153
+ "eval_steps_per_second": 0.243,
2154
+ "step": 143
2155
+ },
2156
+ {
2157
+ "epoch": 4.5,
2158
+ "grad_norm": 0.3736204162232246,
2159
+ "learning_rate": 2e-05,
2160
+ "loss": 0.698,
2161
+ "step": 144
2162
+ },
2163
+ {
2164
+ "epoch": 4.5,
2165
+ "eval_loss": 0.7093971371650696,
2166
+ "eval_runtime": 53.4582,
2167
+ "eval_samples_per_second": 3.741,
2168
+ "eval_steps_per_second": 0.243,
2169
+ "step": 144
2170
+ },
2171
+ {
2172
+ "epoch": 4.53125,
2173
+ "grad_norm": 0.4179284798304916,
2174
+ "learning_rate": 2e-05,
2175
+ "loss": 0.7611,
2176
+ "step": 145
2177
+ },
2178
+ {
2179
+ "epoch": 4.53125,
2180
+ "eval_loss": 0.7089446783065796,
2181
+ "eval_runtime": 53.4752,
2182
+ "eval_samples_per_second": 3.74,
2183
+ "eval_steps_per_second": 0.243,
2184
+ "step": 145
2185
+ },
2186
+ {
2187
+ "epoch": 4.5625,
2188
+ "grad_norm": 0.4038858950888911,
2189
+ "learning_rate": 2e-05,
2190
+ "loss": 0.6652,
2191
+ "step": 146
2192
+ },
2193
+ {
2194
+ "epoch": 4.5625,
2195
+ "eval_loss": 0.7089542150497437,
2196
+ "eval_runtime": 53.4741,
2197
+ "eval_samples_per_second": 3.74,
2198
+ "eval_steps_per_second": 0.243,
2199
+ "step": 146
2200
+ },
2201
+ {
2202
+ "epoch": 4.59375,
2203
+ "grad_norm": 0.41740068710674544,
2204
+ "learning_rate": 2e-05,
2205
+ "loss": 0.7319,
2206
+ "step": 147
2207
+ },
2208
+ {
2209
+ "epoch": 4.59375,
2210
+ "eval_loss": 0.7090431451797485,
2211
+ "eval_runtime": 53.2419,
2212
+ "eval_samples_per_second": 3.756,
2213
+ "eval_steps_per_second": 0.244,
2214
+ "step": 147
2215
+ },
2216
+ {
2217
+ "epoch": 4.625,
2218
+ "grad_norm": 0.4288335811568808,
2219
+ "learning_rate": 2e-05,
2220
+ "loss": 0.6837,
2221
+ "step": 148
2222
+ },
2223
+ {
2224
+ "epoch": 4.625,
2225
+ "eval_loss": 0.7088204026222229,
2226
+ "eval_runtime": 53.3614,
2227
+ "eval_samples_per_second": 3.748,
2228
+ "eval_steps_per_second": 0.244,
2229
+ "step": 148
2230
+ },
2231
+ {
2232
+ "epoch": 4.65625,
2233
+ "grad_norm": 0.399955010119186,
2234
+ "learning_rate": 2e-05,
2235
+ "loss": 0.7989,
2236
+ "step": 149
2237
+ },
2238
+ {
2239
+ "epoch": 4.65625,
2240
+ "eval_loss": 0.7084855437278748,
2241
+ "eval_runtime": 53.4923,
2242
+ "eval_samples_per_second": 3.739,
2243
+ "eval_steps_per_second": 0.243,
2244
+ "step": 149
2245
+ },
2246
+ {
2247
+ "epoch": 4.6875,
2248
+ "grad_norm": 0.41794643164255846,
2249
+ "learning_rate": 2e-05,
2250
+ "loss": 0.7194,
2251
+ "step": 150
2252
+ },
2253
+ {
2254
+ "epoch": 4.6875,
2255
+ "eval_loss": 0.7080708146095276,
2256
+ "eval_runtime": 53.639,
2257
+ "eval_samples_per_second": 3.729,
2258
+ "eval_steps_per_second": 0.242,
2259
+ "step": 150
2260
+ },
2261
+ {
2262
+ "epoch": 4.71875,
2263
+ "grad_norm": 0.40953367303148197,
2264
+ "learning_rate": 2e-05,
2265
+ "loss": 0.7354,
2266
+ "step": 151
2267
+ },
2268
+ {
2269
+ "epoch": 4.71875,
2270
+ "eval_loss": 0.7077429890632629,
2271
+ "eval_runtime": 53.3837,
2272
+ "eval_samples_per_second": 3.746,
2273
+ "eval_steps_per_second": 0.244,
2274
+ "step": 151
2275
+ },
2276
+ {
2277
+ "epoch": 4.75,
2278
+ "grad_norm": 0.5012282841513718,
2279
+ "learning_rate": 2e-05,
2280
+ "loss": 0.7662,
2281
+ "step": 152
2282
+ },
2283
+ {
2284
+ "epoch": 4.75,
2285
+ "eval_loss": 0.7064151167869568,
2286
+ "eval_runtime": 53.3549,
2287
+ "eval_samples_per_second": 3.748,
2288
+ "eval_steps_per_second": 0.244,
2289
+ "step": 152
2290
+ },
2291
+ {
2292
+ "epoch": 4.78125,
2293
+ "grad_norm": 0.4210784420989087,
2294
+ "learning_rate": 2e-05,
2295
+ "loss": 0.7133,
2296
+ "step": 153
2297
+ },
2298
+ {
2299
+ "epoch": 4.78125,
2300
+ "eval_loss": 0.7052726745605469,
2301
+ "eval_runtime": 53.5059,
2302
+ "eval_samples_per_second": 3.738,
2303
+ "eval_steps_per_second": 0.243,
2304
+ "step": 153
2305
+ },
2306
+ {
2307
+ "epoch": 4.8125,
2308
+ "grad_norm": 0.43520348530514996,
2309
+ "learning_rate": 2e-05,
2310
+ "loss": 0.729,
2311
+ "step": 154
2312
+ },
2313
+ {
2314
+ "epoch": 4.8125,
2315
+ "eval_loss": 0.7045274972915649,
2316
+ "eval_runtime": 53.8352,
2317
+ "eval_samples_per_second": 3.715,
2318
+ "eval_steps_per_second": 0.241,
2319
+ "step": 154
2320
+ },
2321
+ {
2322
+ "epoch": 4.84375,
2323
+ "grad_norm": 0.4287647569802656,
2324
+ "learning_rate": 2e-05,
2325
+ "loss": 0.6727,
2326
+ "step": 155
2327
+ },
2328
+ {
2329
+ "epoch": 4.84375,
2330
+ "eval_loss": 0.7041358947753906,
2331
+ "eval_runtime": 53.7435,
2332
+ "eval_samples_per_second": 3.721,
2333
+ "eval_steps_per_second": 0.242,
2334
+ "step": 155
2335
+ },
2336
+ {
2337
+ "epoch": 4.875,
2338
+ "grad_norm": 0.41883715320456333,
2339
+ "learning_rate": 2e-05,
2340
+ "loss": 0.7755,
2341
+ "step": 156
2342
+ },
2343
+ {
2344
+ "epoch": 4.875,
2345
+ "eval_loss": 0.7037128210067749,
2346
+ "eval_runtime": 53.8035,
2347
+ "eval_samples_per_second": 3.717,
2348
+ "eval_steps_per_second": 0.242,
2349
+ "step": 156
2350
+ },
2351
+ {
2352
+ "epoch": 4.90625,
2353
+ "grad_norm": 0.40617584505395354,
2354
+ "learning_rate": 2e-05,
2355
+ "loss": 0.7776,
2356
+ "step": 157
2357
+ },
2358
+ {
2359
+ "epoch": 4.90625,
2360
+ "eval_loss": 0.703965425491333,
2361
+ "eval_runtime": 53.8731,
2362
+ "eval_samples_per_second": 3.712,
2363
+ "eval_steps_per_second": 0.241,
2364
+ "step": 157
2365
+ },
2366
+ {
2367
+ "epoch": 4.9375,
2368
+ "grad_norm": 0.4085802225532245,
2369
+ "learning_rate": 2e-05,
2370
+ "loss": 0.7628,
2371
+ "step": 158
2372
+ },
2373
+ {
2374
+ "epoch": 4.9375,
2375
+ "eval_loss": 0.7040860056877136,
2376
+ "eval_runtime": 53.9059,
2377
+ "eval_samples_per_second": 3.71,
2378
+ "eval_steps_per_second": 0.241,
2379
+ "step": 158
2380
+ },
2381
+ {
2382
+ "epoch": 4.96875,
2383
+ "grad_norm": 0.418039298119887,
2384
+ "learning_rate": 2e-05,
2385
+ "loss": 0.7221,
2386
+ "step": 159
2387
+ },
2388
+ {
2389
+ "epoch": 4.96875,
2390
+ "eval_loss": 0.7039948105812073,
2391
+ "eval_runtime": 53.7323,
2392
+ "eval_samples_per_second": 3.722,
2393
+ "eval_steps_per_second": 0.242,
2394
+ "step": 159
2395
+ },
2396
+ {
2397
+ "epoch": 5.0,
2398
+ "grad_norm": 0.46118870048713073,
2399
+ "learning_rate": 2e-05,
2400
+ "loss": 0.7029,
2401
+ "step": 160
2402
+ },
2403
+ {
2404
+ "epoch": 5.0,
2405
+ "eval_loss": 0.703814685344696,
2406
+ "eval_runtime": 53.8975,
2407
+ "eval_samples_per_second": 3.711,
2408
+ "eval_steps_per_second": 0.241,
2409
+ "step": 160
2410
+ },
2411
+ {
2412
+ "epoch": 5.03125,
2413
+ "grad_norm": 0.431474386110294,
2414
+ "learning_rate": 2e-05,
2415
+ "loss": 0.6772,
2416
+ "step": 161
2417
+ },
2418
+ {
2419
+ "epoch": 5.03125,
2420
+ "eval_loss": 0.7034456133842468,
2421
+ "eval_runtime": 51.1105,
2422
+ "eval_samples_per_second": 3.913,
2423
+ "eval_steps_per_second": 0.254,
2424
+ "step": 161
2425
+ },
2426
+ {
2427
+ "epoch": 5.0625,
2428
+ "grad_norm": 0.39618929325750435,
2429
+ "learning_rate": 2e-05,
2430
+ "loss": 0.8219,
2431
+ "step": 162
2432
+ },
2433
+ {
2434
+ "epoch": 5.0625,
2435
+ "eval_loss": 0.7042189240455627,
2436
+ "eval_runtime": 47.2927,
2437
+ "eval_samples_per_second": 4.229,
2438
+ "eval_steps_per_second": 0.275,
2439
+ "step": 162
2440
+ },
2441
+ {
2442
+ "epoch": 5.09375,
2443
+ "grad_norm": 0.4489132713249424,
2444
+ "learning_rate": 2e-05,
2445
+ "loss": 0.6387,
2446
+ "step": 163
2447
+ },
2448
+ {
2449
+ "epoch": 5.09375,
2450
+ "eval_loss": 0.7061256170272827,
2451
+ "eval_runtime": 47.387,
2452
+ "eval_samples_per_second": 4.221,
2453
+ "eval_steps_per_second": 0.274,
2454
+ "step": 163
2455
+ },
2456
+ {
2457
+ "epoch": 5.125,
2458
+ "grad_norm": 0.5100329637159183,
2459
+ "learning_rate": 2e-05,
2460
+ "loss": 0.7677,
2461
+ "step": 164
2462
+ },
2463
+ {
2464
+ "epoch": 5.125,
2465
+ "eval_loss": 0.708121657371521,
2466
+ "eval_runtime": 47.3311,
2467
+ "eval_samples_per_second": 4.226,
2468
+ "eval_steps_per_second": 0.275,
2469
+ "step": 164
2470
+ },
2471
+ {
2472
+ "epoch": 5.15625,
2473
+ "grad_norm": 0.525511631981176,
2474
+ "learning_rate": 2e-05,
2475
+ "loss": 0.5956,
2476
+ "step": 165
2477
+ },
2478
+ {
2479
+ "epoch": 5.15625,
2480
+ "eval_loss": 0.7091134786605835,
2481
+ "eval_runtime": 47.2978,
2482
+ "eval_samples_per_second": 4.229,
2483
+ "eval_steps_per_second": 0.275,
2484
+ "step": 165
2485
+ },
2486
+ {
2487
+ "epoch": 5.1875,
2488
+ "grad_norm": 0.534675354231597,
2489
+ "learning_rate": 2e-05,
2490
+ "loss": 0.7097,
2491
+ "step": 166
2492
+ },
2493
+ {
2494
+ "epoch": 5.1875,
2495
+ "eval_loss": 0.7097848653793335,
2496
+ "eval_runtime": 47.4095,
2497
+ "eval_samples_per_second": 4.219,
2498
+ "eval_steps_per_second": 0.274,
2499
+ "step": 166
2500
+ },
2501
+ {
2502
+ "epoch": 5.21875,
2503
+ "grad_norm": 0.47286903698857446,
2504
+ "learning_rate": 2e-05,
2505
+ "loss": 0.7371,
2506
+ "step": 167
2507
+ },
2508
+ {
2509
+ "epoch": 5.21875,
2510
+ "eval_loss": 0.7090296745300293,
2511
+ "eval_runtime": 47.4487,
2512
+ "eval_samples_per_second": 4.215,
2513
+ "eval_steps_per_second": 0.274,
2514
+ "step": 167
2515
+ },
2516
+ {
2517
+ "epoch": 5.25,
2518
+ "grad_norm": 0.4734705066820788,
2519
+ "learning_rate": 2e-05,
2520
+ "loss": 0.7652,
2521
+ "step": 168
2522
+ },
2523
+ {
2524
+ "epoch": 5.25,
2525
+ "eval_loss": 0.7079525589942932,
2526
+ "eval_runtime": 47.4101,
2527
+ "eval_samples_per_second": 4.219,
2528
+ "eval_steps_per_second": 0.274,
2529
+ "step": 168
2530
+ },
2531
+ {
2532
+ "epoch": 5.28125,
2533
+ "grad_norm": 0.46209764763985184,
2534
+ "learning_rate": 2e-05,
2535
+ "loss": 0.6852,
2536
+ "step": 169
2537
+ },
2538
+ {
2539
+ "epoch": 5.28125,
2540
+ "eval_loss": 0.7072803974151611,
2541
+ "eval_runtime": 47.3704,
2542
+ "eval_samples_per_second": 4.222,
2543
+ "eval_steps_per_second": 0.274,
2544
+ "step": 169
2545
+ },
2546
+ {
2547
+ "epoch": 5.3125,
2548
+ "grad_norm": 0.4828284708486433,
2549
+ "learning_rate": 2e-05,
2550
+ "loss": 0.6609,
2551
+ "step": 170
2552
+ },
2553
+ {
2554
+ "epoch": 5.3125,
2555
+ "eval_loss": 0.7068901062011719,
2556
+ "eval_runtime": 47.425,
2557
+ "eval_samples_per_second": 4.217,
2558
+ "eval_steps_per_second": 0.274,
2559
+ "step": 170
2560
+ },
2561
+ {
2562
+ "epoch": 5.34375,
2563
+ "grad_norm": 0.5230116179180577,
2564
+ "learning_rate": 2e-05,
2565
+ "loss": 0.6872,
2566
+ "step": 171
2567
+ },
2568
+ {
2569
+ "epoch": 5.34375,
2570
+ "eval_loss": 0.7058187127113342,
2571
+ "eval_runtime": 47.5711,
2572
+ "eval_samples_per_second": 4.204,
2573
+ "eval_steps_per_second": 0.273,
2574
+ "step": 171
2575
+ },
2576
+ {
2577
+ "epoch": 5.375,
2578
+ "grad_norm": 0.48081340678536255,
2579
+ "learning_rate": 2e-05,
2580
+ "loss": 0.7694,
2581
+ "step": 172
2582
+ },
2583
+ {
2584
+ "epoch": 5.375,
2585
+ "eval_loss": 0.7044984698295593,
2586
+ "eval_runtime": 47.4233,
2587
+ "eval_samples_per_second": 4.217,
2588
+ "eval_steps_per_second": 0.274,
2589
+ "step": 172
2590
+ },
2591
+ {
2592
+ "epoch": 5.40625,
2593
+ "grad_norm": 0.4787525602476421,
2594
+ "learning_rate": 2e-05,
2595
+ "loss": 0.7342,
2596
+ "step": 173
2597
+ },
2598
+ {
2599
+ "epoch": 5.40625,
2600
+ "eval_loss": 0.7032212018966675,
2601
+ "eval_runtime": 47.3534,
2602
+ "eval_samples_per_second": 4.224,
2603
+ "eval_steps_per_second": 0.275,
2604
+ "step": 173
2605
+ },
2606
+ {
2607
+ "epoch": 5.4375,
2608
+ "grad_norm": 0.4871847582306217,
2609
+ "learning_rate": 2e-05,
2610
+ "loss": 0.7562,
2611
+ "step": 174
2612
+ },
2613
+ {
2614
+ "epoch": 5.4375,
2615
+ "eval_loss": 0.7019696235656738,
2616
+ "eval_runtime": 47.382,
2617
+ "eval_samples_per_second": 4.221,
2618
+ "eval_steps_per_second": 0.274,
2619
+ "step": 174
2620
+ },
2621
+ {
2622
+ "epoch": 5.46875,
2623
+ "grad_norm": 0.47999745025553603,
2624
+ "learning_rate": 2e-05,
2625
+ "loss": 0.7534,
2626
+ "step": 175
2627
+ },
2628
+ {
2629
+ "epoch": 5.46875,
2630
+ "eval_loss": 0.7014529705047607,
2631
+ "eval_runtime": 47.4435,
2632
+ "eval_samples_per_second": 4.216,
2633
+ "eval_steps_per_second": 0.274,
2634
+ "step": 175
2635
+ },
2636
+ {
2637
+ "epoch": 5.5,
2638
+ "grad_norm": 0.5168030891996357,
2639
+ "learning_rate": 2e-05,
2640
+ "loss": 0.707,
2641
+ "step": 176
2642
+ },
2643
+ {
2644
+ "epoch": 5.5,
2645
+ "eval_loss": 0.6993884444236755,
2646
+ "eval_runtime": 47.4943,
2647
+ "eval_samples_per_second": 4.211,
2648
+ "eval_steps_per_second": 0.274,
2649
+ "step": 176
2650
+ },
2651
+ {
2652
+ "epoch": 5.53125,
2653
+ "grad_norm": 0.536450206978984,
2654
+ "learning_rate": 2e-05,
2655
+ "loss": 0.7318,
2656
+ "step": 177
2657
+ },
2658
+ {
2659
+ "epoch": 5.53125,
2660
+ "eval_loss": 0.6971662640571594,
2661
+ "eval_runtime": 47.4193,
2662
+ "eval_samples_per_second": 4.218,
2663
+ "eval_steps_per_second": 0.274,
2664
+ "step": 177
2665
+ },
2666
+ {
2667
+ "epoch": 5.5625,
2668
+ "grad_norm": 0.45352543205020696,
2669
+ "learning_rate": 2e-05,
2670
+ "loss": 0.7421,
2671
+ "step": 178
2672
+ },
2673
+ {
2674
+ "epoch": 5.5625,
2675
+ "eval_loss": 0.6962605118751526,
2676
+ "eval_runtime": 47.3798,
2677
+ "eval_samples_per_second": 4.221,
2678
+ "eval_steps_per_second": 0.274,
2679
+ "step": 178
2680
+ },
2681
+ {
2682
+ "epoch": 5.59375,
2683
+ "grad_norm": 0.5054883443109318,
2684
+ "learning_rate": 2e-05,
2685
+ "loss": 0.6668,
2686
+ "step": 179
2687
+ },
2688
+ {
2689
+ "epoch": 5.59375,
2690
+ "eval_loss": 0.6970357298851013,
2691
+ "eval_runtime": 47.3311,
2692
+ "eval_samples_per_second": 4.226,
2693
+ "eval_steps_per_second": 0.275,
2694
+ "step": 179
2695
+ },
2696
+ {
2697
+ "epoch": 5.625,
2698
+ "grad_norm": 0.49584660418833293,
2699
+ "learning_rate": 2e-05,
2700
+ "loss": 0.6548,
2701
+ "step": 180
2702
+ },
2703
+ {
2704
+ "epoch": 5.625,
2705
+ "eval_loss": 0.6980059146881104,
2706
+ "eval_runtime": 47.299,
2707
+ "eval_samples_per_second": 4.228,
2708
+ "eval_steps_per_second": 0.275,
2709
+ "step": 180
2710
+ },
2711
+ {
2712
+ "epoch": 5.65625,
2713
+ "grad_norm": 0.5114381326491793,
2714
+ "learning_rate": 2e-05,
2715
+ "loss": 0.6691,
2716
+ "step": 181
2717
+ },
2718
+ {
2719
+ "epoch": 5.65625,
2720
+ "eval_loss": 0.6995040774345398,
2721
+ "eval_runtime": 47.3887,
2722
+ "eval_samples_per_second": 4.22,
2723
+ "eval_steps_per_second": 0.274,
2724
+ "step": 181
2725
+ },
2726
+ {
2727
+ "epoch": 5.6875,
2728
+ "grad_norm": 0.48550125668870825,
2729
+ "learning_rate": 2e-05,
2730
+ "loss": 0.6525,
2731
+ "step": 182
2732
+ },
2733
+ {
2734
+ "epoch": 5.6875,
2735
+ "eval_loss": 0.7020326256752014,
2736
+ "eval_runtime": 47.3838,
2737
+ "eval_samples_per_second": 4.221,
2738
+ "eval_steps_per_second": 0.274,
2739
+ "step": 182
2740
+ },
2741
+ {
2742
+ "epoch": 5.71875,
2743
+ "grad_norm": 0.5860847796671736,
2744
+ "learning_rate": 2e-05,
2745
+ "loss": 0.674,
2746
+ "step": 183
2747
+ },
2748
+ {
2749
+ "epoch": 5.71875,
2750
+ "eval_loss": 0.7027825713157654,
2751
+ "eval_runtime": 47.3875,
2752
+ "eval_samples_per_second": 4.221,
2753
+ "eval_steps_per_second": 0.274,
2754
+ "step": 183
2755
+ },
2756
+ {
2757
+ "epoch": 5.75,
2758
+ "grad_norm": 0.5535582209035479,
2759
+ "learning_rate": 2e-05,
2760
+ "loss": 0.6643,
2761
+ "step": 184
2762
+ },
2763
+ {
2764
+ "epoch": 5.75,
2765
+ "eval_loss": 0.7025408148765564,
2766
+ "eval_runtime": 47.5534,
2767
+ "eval_samples_per_second": 4.206,
2768
+ "eval_steps_per_second": 0.273,
2769
+ "step": 184
2770
+ },
2771
+ {
2772
+ "epoch": 5.78125,
2773
+ "grad_norm": 0.5443574176405931,
2774
+ "learning_rate": 2e-05,
2775
+ "loss": 0.709,
2776
+ "step": 185
2777
+ },
2778
+ {
2779
+ "epoch": 5.78125,
2780
+ "eval_loss": 0.7007840871810913,
2781
+ "eval_runtime": 47.4469,
2782
+ "eval_samples_per_second": 4.215,
2783
+ "eval_steps_per_second": 0.274,
2784
+ "step": 185
2785
+ },
2786
+ {
2787
+ "epoch": 5.8125,
2788
+ "grad_norm": 0.563830259704143,
2789
+ "learning_rate": 2e-05,
2790
+ "loss": 0.6884,
2791
+ "step": 186
2792
+ },
2793
+ {
2794
+ "epoch": 5.8125,
2795
+ "eval_loss": 0.6979361176490784,
2796
+ "eval_runtime": 49.1203,
2797
+ "eval_samples_per_second": 4.072,
2798
+ "eval_steps_per_second": 0.265,
2799
+ "step": 186
2800
+ },
2801
+ {
2802
+ "epoch": 5.84375,
2803
+ "grad_norm": 0.5094956892765212,
2804
+ "learning_rate": 2e-05,
2805
+ "loss": 0.7318,
2806
+ "step": 187
2807
+ },
2808
+ {
2809
+ "epoch": 5.84375,
2810
+ "eval_loss": 0.6962587237358093,
2811
+ "eval_runtime": 49.1831,
2812
+ "eval_samples_per_second": 4.066,
2813
+ "eval_steps_per_second": 0.264,
2814
+ "step": 187
2815
+ },
2816
+ {
2817
+ "epoch": 5.875,
2818
+ "grad_norm": 0.5264819980742595,
2819
+ "learning_rate": 2e-05,
2820
+ "loss": 0.6746,
2821
+ "step": 188
2822
+ },
2823
+ {
2824
+ "epoch": 5.875,
2825
+ "eval_loss": 0.694776713848114,
2826
+ "eval_runtime": 49.1994,
2827
+ "eval_samples_per_second": 4.065,
2828
+ "eval_steps_per_second": 0.264,
2829
+ "step": 188
2830
+ },
2831
+ {
2832
+ "epoch": 5.90625,
2833
+ "grad_norm": 0.4737429304023209,
2834
+ "learning_rate": 2e-05,
2835
+ "loss": 0.664,
2836
+ "step": 189
2837
+ },
2838
+ {
2839
+ "epoch": 5.90625,
2840
+ "eval_loss": 0.6939517855644226,
2841
+ "eval_runtime": 49.2438,
2842
+ "eval_samples_per_second": 4.061,
2843
+ "eval_steps_per_second": 0.264,
2844
+ "step": 189
2845
+ },
2846
+ {
2847
+ "epoch": 5.9375,
2848
+ "grad_norm": 0.494163934813738,
2849
+ "learning_rate": 2e-05,
2850
+ "loss": 0.6978,
2851
+ "step": 190
2852
+ },
2853
+ {
2854
+ "epoch": 5.9375,
2855
+ "eval_loss": 0.6933834552764893,
2856
+ "eval_runtime": 49.3494,
2857
+ "eval_samples_per_second": 4.053,
2858
+ "eval_steps_per_second": 0.263,
2859
+ "step": 190
2860
+ },
2861
+ {
2862
+ "epoch": 5.96875,
2863
+ "grad_norm": 0.4945972278087299,
2864
+ "learning_rate": 2e-05,
2865
+ "loss": 0.6909,
2866
+ "step": 191
2867
+ },
2868
+ {
2869
+ "epoch": 5.96875,
2870
+ "eval_loss": 0.6924250721931458,
2871
+ "eval_runtime": 50.3255,
2872
+ "eval_samples_per_second": 3.974,
2873
+ "eval_steps_per_second": 0.258,
2874
+ "step": 191
2875
+ },
2876
+ {
2877
+ "epoch": 6.0,
2878
+ "grad_norm": 0.48872556688745233,
2879
+ "learning_rate": 2e-05,
2880
+ "loss": 0.6622,
2881
+ "step": 192
2882
+ },
2883
+ {
2884
+ "epoch": 6.0,
2885
+ "eval_loss": 0.6922193765640259,
2886
+ "eval_runtime": 50.4561,
2887
+ "eval_samples_per_second": 3.964,
2888
+ "eval_steps_per_second": 0.258,
2889
+ "step": 192
2890
+ },
2891
+ {
2892
+ "epoch": 6.03125,
2893
+ "grad_norm": 0.5013452255378538,
2894
+ "learning_rate": 2e-05,
2895
+ "loss": 0.7458,
2896
+ "step": 193
2897
+ },
2898
+ {
2899
+ "epoch": 6.03125,
2900
+ "eval_loss": 0.6931161284446716,
2901
+ "eval_runtime": 50.5049,
2902
+ "eval_samples_per_second": 3.96,
2903
+ "eval_steps_per_second": 0.257,
2904
+ "step": 193
2905
+ },
2906
+ {
2907
+ "epoch": 6.0625,
2908
+ "grad_norm": 0.48271161232093784,
2909
+ "learning_rate": 2e-05,
2910
+ "loss": 0.7171,
2911
+ "step": 194
2912
+ },
2913
+ {
2914
+ "epoch": 6.0625,
2915
+ "eval_loss": 0.6959040760993958,
2916
+ "eval_runtime": 50.2441,
2917
+ "eval_samples_per_second": 3.981,
2918
+ "eval_steps_per_second": 0.259,
2919
+ "step": 194
2920
+ },
2921
+ {
2922
+ "epoch": 6.09375,
2923
+ "grad_norm": 0.5414562703154852,
2924
+ "learning_rate": 2e-05,
2925
+ "loss": 0.6419,
2926
+ "step": 195
2927
+ },
2928
+ {
2929
+ "epoch": 6.09375,
2930
+ "eval_loss": 0.7000604271888733,
2931
+ "eval_runtime": 50.4261,
2932
+ "eval_samples_per_second": 3.966,
2933
+ "eval_steps_per_second": 0.258,
2934
+ "step": 195
2935
+ },
2936
+ {
2937
+ "epoch": 6.125,
2938
+ "grad_norm": 0.5074661247335385,
2939
+ "learning_rate": 2e-05,
2940
+ "loss": 0.6881,
2941
+ "step": 196
2942
+ },
2943
+ {
2944
+ "epoch": 6.125,
2945
+ "eval_loss": 0.7039622664451599,
2946
+ "eval_runtime": 51.5214,
2947
+ "eval_samples_per_second": 3.882,
2948
+ "eval_steps_per_second": 0.252,
2949
+ "step": 196
2950
+ },
2951
+ {
2952
+ "epoch": 6.15625,
2953
+ "grad_norm": 0.5603468534764365,
2954
+ "learning_rate": 2e-05,
2955
+ "loss": 0.7085,
2956
+ "step": 197
2957
+ },
2958
+ {
2959
+ "epoch": 6.15625,
2960
+ "eval_loss": 0.7055023312568665,
2961
+ "eval_runtime": 51.7102,
2962
+ "eval_samples_per_second": 3.868,
2963
+ "eval_steps_per_second": 0.251,
2964
+ "step": 197
2965
+ },
2966
+ {
2967
+ "epoch": 6.1875,
2968
+ "grad_norm": 0.5992190802422799,
2969
+ "learning_rate": 2e-05,
2970
+ "loss": 0.7614,
2971
+ "step": 198
2972
+ },
2973
+ {
2974
+ "epoch": 6.1875,
2975
+ "eval_loss": 0.7046856880187988,
2976
+ "eval_runtime": 51.5464,
2977
+ "eval_samples_per_second": 3.88,
2978
+ "eval_steps_per_second": 0.252,
2979
+ "step": 198
2980
+ },
2981
+ {
2982
+ "epoch": 6.21875,
2983
+ "grad_norm": 0.6293684167527106,
2984
+ "learning_rate": 2e-05,
2985
+ "loss": 0.6435,
2986
+ "step": 199
2987
+ },
2988
+ {
2989
+ "epoch": 6.21875,
2990
+ "eval_loss": 0.7021151781082153,
2991
+ "eval_runtime": 51.5328,
2992
+ "eval_samples_per_second": 3.881,
2993
+ "eval_steps_per_second": 0.252,
2994
+ "step": 199
2995
+ },
2996
+ {
2997
+ "epoch": 6.25,
2998
+ "grad_norm": 0.591265449241434,
2999
+ "learning_rate": 2e-05,
3000
+ "loss": 0.688,
3001
+ "step": 200
3002
+ },
3003
+ {
3004
+ "epoch": 6.25,
3005
+ "eval_loss": 0.7002359628677368,
3006
+ "eval_runtime": 51.5812,
3007
+ "eval_samples_per_second": 3.877,
3008
+ "eval_steps_per_second": 0.252,
3009
+ "step": 200
3010
+ },
3011
+ {
3012
+ "epoch": 6.28125,
3013
+ "grad_norm": 0.543141536526749,
3014
+ "learning_rate": 2e-05,
3015
+ "loss": 0.7027,
3016
+ "step": 201
3017
+ },
3018
+ {
3019
+ "epoch": 6.28125,
3020
+ "eval_loss": 0.6986366510391235,
3021
+ "eval_runtime": 52.6956,
3022
+ "eval_samples_per_second": 3.795,
3023
+ "eval_steps_per_second": 0.247,
3024
+ "step": 201
3025
+ },
3026
+ {
3027
+ "epoch": 6.3125,
3028
+ "grad_norm": 0.5679656300203245,
3029
+ "learning_rate": 2e-05,
3030
+ "loss": 0.625,
3031
+ "step": 202
3032
+ },
3033
+ {
3034
+ "epoch": 6.3125,
3035
+ "eval_loss": 0.698679506778717,
3036
+ "eval_runtime": 52.5102,
3037
+ "eval_samples_per_second": 3.809,
3038
+ "eval_steps_per_second": 0.248,
3039
+ "step": 202
3040
+ },
3041
+ {
3042
+ "epoch": 6.34375,
3043
+ "grad_norm": 0.5285839896523021,
3044
+ "learning_rate": 2e-05,
3045
+ "loss": 0.7687,
3046
+ "step": 203
3047
+ },
3048
+ {
3049
+ "epoch": 6.34375,
3050
+ "eval_loss": 0.7005956768989563,
3051
+ "eval_runtime": 52.6067,
3052
+ "eval_samples_per_second": 3.802,
3053
+ "eval_steps_per_second": 0.247,
3054
+ "step": 203
3055
+ },
3056
+ {
3057
+ "epoch": 6.375,
3058
+ "grad_norm": 0.6512964945211068,
3059
+ "learning_rate": 2e-05,
3060
+ "loss": 0.623,
3061
+ "step": 204
3062
+ },
3063
+ {
3064
+ "epoch": 6.375,
3065
+ "eval_loss": 0.7013595104217529,
3066
+ "eval_runtime": 52.5428,
3067
+ "eval_samples_per_second": 3.806,
3068
+ "eval_steps_per_second": 0.247,
3069
+ "step": 204
3070
+ },
3071
+ {
3072
+ "epoch": 6.40625,
3073
+ "grad_norm": 0.5295248631519638,
3074
+ "learning_rate": 2e-05,
3075
+ "loss": 0.5941,
3076
+ "step": 205
3077
+ },
3078
+ {
3079
+ "epoch": 6.40625,
3080
+ "eval_loss": 0.7016547322273254,
3081
+ "eval_runtime": 52.6142,
3082
+ "eval_samples_per_second": 3.801,
3083
+ "eval_steps_per_second": 0.247,
3084
+ "step": 205
3085
+ },
3086
+ {
3087
+ "epoch": 6.4375,
3088
+ "grad_norm": 0.6134157701434021,
3089
+ "learning_rate": 2e-05,
3090
+ "loss": 0.6506,
3091
+ "step": 206
3092
+ },
3093
+ {
3094
+ "epoch": 6.4375,
3095
+ "eval_loss": 0.7009623646736145,
3096
+ "eval_runtime": 52.1942,
3097
+ "eval_samples_per_second": 3.832,
3098
+ "eval_steps_per_second": 0.249,
3099
+ "step": 206
3100
+ },
3101
+ {
3102
+ "epoch": 6.46875,
3103
+ "grad_norm": 0.57886797614996,
3104
+ "learning_rate": 2e-05,
3105
+ "loss": 0.6983,
3106
+ "step": 207
3107
+ },
3108
+ {
3109
+ "epoch": 6.46875,
3110
+ "eval_loss": 0.6988092064857483,
3111
+ "eval_runtime": 52.2577,
3112
+ "eval_samples_per_second": 3.827,
3113
+ "eval_steps_per_second": 0.249,
3114
+ "step": 207
3115
+ },
3116
+ {
3117
+ "epoch": 6.5,
3118
+ "grad_norm": 0.5593482836944472,
3119
+ "learning_rate": 2e-05,
3120
+ "loss": 0.6348,
3121
+ "step": 208
3122
+ },
3123
+ {
3124
+ "epoch": 6.5,
3125
+ "eval_loss": 0.698823094367981,
3126
+ "eval_runtime": 52.2296,
3127
+ "eval_samples_per_second": 3.829,
3128
+ "eval_steps_per_second": 0.249,
3129
+ "step": 208
3130
+ },
3131
+ {
3132
+ "epoch": 6.53125,
3133
+ "grad_norm": 0.662802162179718,
3134
+ "learning_rate": 2e-05,
3135
+ "loss": 0.6206,
3136
+ "step": 209
3137
+ },
3138
+ {
3139
+ "epoch": 6.53125,
3140
+ "eval_loss": 0.6990167498588562,
3141
+ "eval_runtime": 52.4316,
3142
+ "eval_samples_per_second": 3.814,
3143
+ "eval_steps_per_second": 0.248,
3144
+ "step": 209
3145
+ },
3146
+ {
3147
+ "epoch": 6.5625,
3148
+ "grad_norm": 0.6874374231122908,
3149
+ "learning_rate": 2e-05,
3150
+ "loss": 0.6033,
3151
+ "step": 210
3152
+ },
3153
+ {
3154
+ "epoch": 6.5625,
3155
+ "eval_loss": 0.699796736240387,
3156
+ "eval_runtime": 52.3193,
3157
+ "eval_samples_per_second": 3.823,
3158
+ "eval_steps_per_second": 0.248,
3159
+ "step": 210
3160
+ },
3161
+ {
3162
+ "epoch": 6.59375,
3163
+ "grad_norm": 0.6625766736772473,
3164
+ "learning_rate": 2e-05,
3165
+ "loss": 0.6398,
3166
+ "step": 211
3167
+ },
3168
+ {
3169
+ "epoch": 6.59375,
3170
+ "eval_loss": 0.6989737153053284,
3171
+ "eval_runtime": 52.1885,
3172
+ "eval_samples_per_second": 3.832,
3173
+ "eval_steps_per_second": 0.249,
3174
+ "step": 211
3175
+ },
3176
+ {
3177
+ "epoch": 6.625,
3178
+ "grad_norm": 0.6563419096027812,
3179
+ "learning_rate": 2e-05,
3180
+ "loss": 0.6119,
3181
+ "step": 212
3182
+ },
3183
+ {
3184
+ "epoch": 6.625,
3185
+ "eval_loss": 0.6973609924316406,
3186
+ "eval_runtime": 52.1628,
3187
+ "eval_samples_per_second": 3.834,
3188
+ "eval_steps_per_second": 0.249,
3189
+ "step": 212
3190
+ },
3191
+ {
3192
+ "epoch": 6.65625,
3193
+ "grad_norm": 0.5796353226697397,
3194
+ "learning_rate": 2e-05,
3195
+ "loss": 0.7041,
3196
+ "step": 213
3197
+ },
3198
+ {
3199
+ "epoch": 6.65625,
3200
+ "eval_loss": 0.6957942247390747,
3201
+ "eval_runtime": 52.2028,
3202
+ "eval_samples_per_second": 3.831,
3203
+ "eval_steps_per_second": 0.249,
3204
+ "step": 213
3205
+ },
3206
+ {
3207
+ "epoch": 6.6875,
3208
+ "grad_norm": 0.5711947110504899,
3209
+ "learning_rate": 2e-05,
3210
+ "loss": 0.6465,
3211
+ "step": 214
3212
+ },
3213
+ {
3214
+ "epoch": 6.6875,
3215
+ "eval_loss": 0.696739673614502,
3216
+ "eval_runtime": 52.1849,
3217
+ "eval_samples_per_second": 3.833,
3218
+ "eval_steps_per_second": 0.249,
3219
+ "step": 214
3220
+ },
3221
+ {
3222
+ "epoch": 6.71875,
3223
+ "grad_norm": 0.6619502413653232,
3224
+ "learning_rate": 2e-05,
3225
+ "loss": 0.6563,
3226
+ "step": 215
3227
+ },
3228
+ {
3229
+ "epoch": 6.71875,
3230
+ "eval_loss": 0.6960940361022949,
3231
+ "eval_runtime": 52.0996,
3232
+ "eval_samples_per_second": 3.839,
3233
+ "eval_steps_per_second": 0.25,
3234
+ "step": 215
3235
+ },
3236
+ {
3237
+ "epoch": 6.75,
3238
+ "grad_norm": 0.6587126256919645,
3239
+ "learning_rate": 2e-05,
3240
+ "loss": 0.6505,
3241
+ "step": 216
3242
+ },
3243
+ {
3244
+ "epoch": 6.75,
3245
+ "eval_loss": 0.6959022283554077,
3246
+ "eval_runtime": 52.1062,
3247
+ "eval_samples_per_second": 3.838,
3248
+ "eval_steps_per_second": 0.249,
3249
+ "step": 216
3250
+ },
3251
+ {
3252
+ "epoch": 6.78125,
3253
+ "grad_norm": 0.648164277941964,
3254
+ "learning_rate": 2e-05,
3255
+ "loss": 0.5969,
3256
+ "step": 217
3257
+ },
3258
+ {
3259
+ "epoch": 6.78125,
3260
+ "eval_loss": 0.6999121308326721,
3261
+ "eval_runtime": 51.9356,
3262
+ "eval_samples_per_second": 3.851,
3263
+ "eval_steps_per_second": 0.25,
3264
+ "step": 217
3265
+ },
3266
+ {
3267
+ "epoch": 6.8125,
3268
+ "grad_norm": 0.6595860789738482,
3269
+ "learning_rate": 2e-05,
3270
+ "loss": 0.5945,
3271
+ "step": 218
3272
+ },
3273
+ {
3274
+ "epoch": 6.8125,
3275
+ "eval_loss": 0.7028067111968994,
3276
+ "eval_runtime": 52.2232,
3277
+ "eval_samples_per_second": 3.83,
3278
+ "eval_steps_per_second": 0.249,
3279
+ "step": 218
3280
+ },
3281
+ {
3282
+ "epoch": 6.84375,
3283
+ "grad_norm": 0.7116894779822719,
3284
+ "learning_rate": 2e-05,
3285
+ "loss": 0.7027,
3286
+ "step": 219
3287
+ },
3288
+ {
3289
+ "epoch": 6.84375,
3290
+ "eval_loss": 0.7035638689994812,
3291
+ "eval_runtime": 52.1471,
3292
+ "eval_samples_per_second": 3.835,
3293
+ "eval_steps_per_second": 0.249,
3294
+ "step": 219
3295
+ },
3296
+ {
3297
+ "epoch": 6.875,
3298
+ "grad_norm": 0.7581142336087988,
3299
+ "learning_rate": 2e-05,
3300
+ "loss": 0.7171,
3301
+ "step": 220
3302
+ },
3303
+ {
3304
+ "epoch": 6.875,
3305
+ "eval_loss": 0.6981176733970642,
3306
+ "eval_runtime": 52.1366,
3307
+ "eval_samples_per_second": 3.836,
3308
+ "eval_steps_per_second": 0.249,
3309
+ "step": 220
3310
+ },
3311
+ {
3312
+ "epoch": 6.90625,
3313
+ "grad_norm": 0.6261292745909233,
3314
+ "learning_rate": 2e-05,
3315
+ "loss": 0.658,
3316
+ "step": 221
3317
+ },
3318
+ {
3319
+ "epoch": 6.90625,
3320
+ "eval_loss": 0.6939045786857605,
3321
+ "eval_runtime": 52.2211,
3322
+ "eval_samples_per_second": 3.83,
3323
+ "eval_steps_per_second": 0.249,
3324
+ "step": 221
3325
+ },
3326
+ {
3327
+ "epoch": 6.9375,
3328
+ "grad_norm": 0.7256427809370966,
3329
+ "learning_rate": 2e-05,
3330
+ "loss": 0.6576,
3331
+ "step": 222
3332
+ },
3333
+ {
3334
+ "epoch": 6.9375,
3335
+ "eval_loss": 0.6904327273368835,
3336
+ "eval_runtime": 52.1829,
3337
+ "eval_samples_per_second": 3.833,
3338
+ "eval_steps_per_second": 0.249,
3339
+ "step": 222
3340
+ },
3341
+ {
3342
+ "epoch": 6.96875,
3343
+ "grad_norm": 0.6653711103404113,
3344
+ "learning_rate": 2e-05,
3345
+ "loss": 0.6938,
3346
+ "step": 223
3347
+ },
3348
+ {
3349
+ "epoch": 6.96875,
3350
+ "eval_loss": 0.6893274188041687,
3351
+ "eval_runtime": 51.899,
3352
+ "eval_samples_per_second": 3.854,
3353
+ "eval_steps_per_second": 0.25,
3354
+ "step": 223
3355
+ },
3356
+ {
3357
+ "epoch": 7.0,
3358
+ "grad_norm": 0.6730688267524797,
3359
+ "learning_rate": 2e-05,
3360
+ "loss": 0.7397,
3361
+ "step": 224
3362
+ },
3363
+ {
3364
+ "epoch": 7.0,
3365
+ "eval_loss": 0.6895740032196045,
3366
+ "eval_runtime": 52.1977,
3367
+ "eval_samples_per_second": 3.832,
3368
+ "eval_steps_per_second": 0.249,
3369
+ "step": 224
3370
+ }
3371
+ ],
3372
+ "logging_steps": 1.0,
3373
+ "max_steps": 224,
3374
+ "num_input_tokens_seen": 0,
3375
+ "num_train_epochs": 7,
3376
+ "save_steps": 5,
3377
+ "stateful_callbacks": {
3378
+ "TrainerControl": {
3379
+ "args": {
3380
+ "should_epoch_stop": false,
3381
+ "should_evaluate": false,
3382
+ "should_log": false,
3383
+ "should_save": true,
3384
+ "should_training_stop": true
3385
+ },
3386
+ "attributes": {}
3387
+ }
3388
+ },
3389
+ "total_flos": 322567586447360.0,
3390
+ "train_batch_size": 8,
3391
+ "trial_name": null,
3392
+ "trial_params": null
3393
+ }
checkpoint-224/training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:21fca50f49cefaafcd1ff13744949a11f0be41ae12da12aa7b74f1b7c0c2d5f2
3
+ size 8184
checkpoint-224/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)
checkpoint-320/README.md ADDED
@@ -0,0 +1,202 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ base_model: liuhaotian/llava-v1.5-13b
3
+ library_name: peft
4
+ ---
5
+
6
+ # Model Card for Model ID
7
+
8
+ <!-- Provide a quick summary of what the model is/does. -->
9
+
10
+
11
+
12
+ ## Model Details
13
+
14
+ ### Model Description
15
+
16
+ <!-- Provide a longer summary of what this model is. -->
17
+
18
+
19
+
20
+ - **Developed by:** [More Information Needed]
21
+ - **Funded by [optional]:** [More Information Needed]
22
+ - **Shared by [optional]:** [More Information Needed]
23
+ - **Model type:** [More Information Needed]
24
+ - **Language(s) (NLP):** [More Information Needed]
25
+ - **License:** [More Information Needed]
26
+ - **Finetuned from model [optional]:** [More Information Needed]
27
+
28
+ ### Model Sources [optional]
29
+
30
+ <!-- Provide the basic links for the model. -->
31
+
32
+ - **Repository:** [More Information Needed]
33
+ - **Paper [optional]:** [More Information Needed]
34
+ - **Demo [optional]:** [More Information Needed]
35
+
36
+ ## Uses
37
+
38
+ <!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
39
+
40
+ ### Direct Use
41
+
42
+ <!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
43
+
44
+ [More Information Needed]
45
+
46
+ ### Downstream Use [optional]
47
+
48
+ <!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
49
+
50
+ [More Information Needed]
51
+
52
+ ### Out-of-Scope Use
53
+
54
+ <!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
55
+
56
+ [More Information Needed]
57
+
58
+ ## Bias, Risks, and Limitations
59
+
60
+ <!-- This section is meant to convey both technical and sociotechnical limitations. -->
61
+
62
+ [More Information Needed]
63
+
64
+ ### Recommendations
65
+
66
+ <!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
67
+
68
+ Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
69
+
70
+ ## How to Get Started with the Model
71
+
72
+ Use the code below to get started with the model.
73
+
74
+ [More Information Needed]
75
+
76
+ ## Training Details
77
+
78
+ ### Training Data
79
+
80
+ <!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
81
+
82
+ [More Information Needed]
83
+
84
+ ### Training Procedure
85
+
86
+ <!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
87
+
88
+ #### Preprocessing [optional]
89
+
90
+ [More Information Needed]
91
+
92
+
93
+ #### Training Hyperparameters
94
+
95
+ - **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
96
+
97
+ #### Speeds, Sizes, Times [optional]
98
+
99
+ <!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
100
+
101
+ [More Information Needed]
102
+
103
+ ## Evaluation
104
+
105
+ <!-- This section describes the evaluation protocols and provides the results. -->
106
+
107
+ ### Testing Data, Factors & Metrics
108
+
109
+ #### Testing Data
110
+
111
+ <!-- This should link to a Dataset Card if possible. -->
112
+
113
+ [More Information Needed]
114
+
115
+ #### Factors
116
+
117
+ <!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
118
+
119
+ [More Information Needed]
120
+
121
+ #### Metrics
122
+
123
+ <!-- These are the evaluation metrics being used, ideally with a description of why. -->
124
+
125
+ [More Information Needed]
126
+
127
+ ### Results
128
+
129
+ [More Information Needed]
130
+
131
+ #### Summary
132
+
133
+
134
+
135
+ ## Model Examination [optional]
136
+
137
+ <!-- Relevant interpretability work for the model goes here -->
138
+
139
+ [More Information Needed]
140
+
141
+ ## Environmental Impact
142
+
143
+ <!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
144
+
145
+ Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
146
+
147
+ - **Hardware Type:** [More Information Needed]
148
+ - **Hours used:** [More Information Needed]
149
+ - **Cloud Provider:** [More Information Needed]
150
+ - **Compute Region:** [More Information Needed]
151
+ - **Carbon Emitted:** [More Information Needed]
152
+
153
+ ## Technical Specifications [optional]
154
+
155
+ ### Model Architecture and Objective
156
+
157
+ [More Information Needed]
158
+
159
+ ### Compute Infrastructure
160
+
161
+ [More Information Needed]
162
+
163
+ #### Hardware
164
+
165
+ [More Information Needed]
166
+
167
+ #### Software
168
+
169
+ [More Information Needed]
170
+
171
+ ## Citation [optional]
172
+
173
+ <!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
174
+
175
+ **BibTeX:**
176
+
177
+ [More Information Needed]
178
+
179
+ **APA:**
180
+
181
+ [More Information Needed]
182
+
183
+ ## Glossary [optional]
184
+
185
+ <!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
186
+
187
+ [More Information Needed]
188
+
189
+ ## More Information [optional]
190
+
191
+ [More Information Needed]
192
+
193
+ ## Model Card Authors [optional]
194
+
195
+ [More Information Needed]
196
+
197
+ ## Model Card Contact
198
+
199
+ [More Information Needed]
200
+ ### Framework versions
201
+
202
+ - PEFT 0.13.2
checkpoint-320/adapter_config.json ADDED
@@ -0,0 +1,34 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "alpha_pattern": {},
3
+ "auto_mapping": null,
4
+ "base_model_name_or_path": "liuhaotian/llava-v1.5-13b",
5
+ "bias": "none",
6
+ "fan_in_fan_out": false,
7
+ "inference_mode": true,
8
+ "init_lora_weights": true,
9
+ "layer_replication": null,
10
+ "layers_pattern": null,
11
+ "layers_to_transform": null,
12
+ "loftq_config": {},
13
+ "lora_alpha": 16,
14
+ "lora_dropout": 0.05,
15
+ "megatron_config": null,
16
+ "megatron_core": "megatron.core",
17
+ "modules_to_save": null,
18
+ "peft_type": "LORA",
19
+ "r": 8,
20
+ "rank_pattern": {},
21
+ "revision": null,
22
+ "target_modules": [
23
+ "up_proj",
24
+ "k_proj",
25
+ "v_proj",
26
+ "gate_proj",
27
+ "o_proj",
28
+ "down_proj",
29
+ "q_proj"
30
+ ],
31
+ "task_type": "CAUSAL_LM",
32
+ "use_dora": false,
33
+ "use_rslora": false
34
+ }
checkpoint-320/adapter_model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:623b69502fe51b1eed879b0e0e29485ea633f4af66c5d1cfbc7a6c33c093568a
3
+ size 62660864
checkpoint-320/global_step320/zero_pp_rank_0_mp_rank_00_model_states.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e8f38fd19d47d3a7f3f53c6d20cd95926e61f16c114ce7946b1d673f3041f837
3
+ size 593618
checkpoint-320/global_step320/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 sha256:3a7c8a298ac8838329dd66ff2e8fdac8ffabcefb449e22a3a1d11ee77376209a
3
+ size 188286957
checkpoint-320/global_step320/zero_pp_rank_1_mp_rank_00_model_states.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:91e813727135ac80358f97bf33e38bc7f766a71c143ccf5e7aae24f8dae837e0
3
+ size 593618
checkpoint-320/global_step320/zero_pp_rank_1_mp_rank_00_optim_states.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8244a7c9eca3b99273f88fdb9bd50bb1f1bb20c3143e5e6010bbab902181ed11
3
+ size 188286957
checkpoint-320/global_step320/zero_pp_rank_2_mp_rank_00_model_states.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:439ec584a84b9faf03f1b1df8b8fc29f633f526de043b549f21a36c5c753e21b
3
+ size 593618
checkpoint-320/global_step320/zero_pp_rank_2_mp_rank_00_optim_states.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:800813a53a6a794d85e69313e2602f05058273df63bfded7fc5a5cbf6f601774
3
+ size 188286957
checkpoint-320/global_step320/zero_pp_rank_3_mp_rank_00_model_states.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2203aa85ac0d0ab529317771b382ac283333ea0d10a636bdf695573d5c44cdff
3
+ size 593618
checkpoint-320/global_step320/zero_pp_rank_3_mp_rank_00_optim_states.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:92ba3aa4564dd89fac03a619fe9596391ee3bdbd8a1e7d499e87c5d460ee3ef2
3
+ size 188286957
checkpoint-320/latest ADDED
@@ -0,0 +1 @@
 
 
1
+ global_step320
checkpoint-320/rng_state_0.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:70e8e40bf1ad96295ca0d42d42672162cccbcbebb089a9837b3df939e99749c9
3
+ size 14960
checkpoint-320/rng_state_1.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:63f24b3a2fcd38868dad836e3fd055d6d1cd19f36948ca4f89760988afb3e439
3
+ size 14960
checkpoint-320/rng_state_2.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0a0998e05fda02393da198c76a9d294538ace12d82c157062e1ca1b94450cacc
3
+ size 14960
checkpoint-320/rng_state_3.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:86c573687454121dca6cfe001a0e146f7fee70895d84869ebe65004d1bfaa5c5
3
+ size 14960
checkpoint-320/special_tokens_map.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token": {
3
+ "content": "<s>",
4
+ "lstrip": false,
5
+ "normalized": false,
6
+ "rstrip": false,
7
+ "single_word": false
8
+ },
9
+ "eos_token": {
10
+ "content": "</s>",
11
+ "lstrip": false,
12
+ "normalized": false,
13
+ "rstrip": false,
14
+ "single_word": false
15
+ },
16
+ "pad_token": "<unk>",
17
+ "unk_token": {
18
+ "content": "<unk>",
19
+ "lstrip": false,
20
+ "normalized": false,
21
+ "rstrip": false,
22
+ "single_word": false
23
+ }
24
+ }
checkpoint-320/tokenizer.model ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347
3
+ size 499723
checkpoint-320/tokenizer_config.json ADDED
@@ -0,0 +1,43 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_bos_token": true,
3
+ "add_eos_token": false,
4
+ "add_prefix_space": true,
5
+ "added_tokens_decoder": {
6
+ "0": {
7
+ "content": "<unk>",
8
+ "lstrip": false,
9
+ "normalized": false,
10
+ "rstrip": false,
11
+ "single_word": false,
12
+ "special": true
13
+ },
14
+ "1": {
15
+ "content": "<s>",
16
+ "lstrip": false,
17
+ "normalized": false,
18
+ "rstrip": false,
19
+ "single_word": false,
20
+ "special": true
21
+ },
22
+ "2": {
23
+ "content": "</s>",
24
+ "lstrip": false,
25
+ "normalized": false,
26
+ "rstrip": false,
27
+ "single_word": false,
28
+ "special": true
29
+ }
30
+ },
31
+ "bos_token": "<s>",
32
+ "clean_up_tokenization_spaces": false,
33
+ "eos_token": "</s>",
34
+ "legacy": false,
35
+ "model_max_length": 2048,
36
+ "pad_token": "<unk>",
37
+ "padding_side": "right",
38
+ "sp_model_kwargs": {},
39
+ "spaces_between_special_tokens": false,
40
+ "tokenizer_class": "LlamaTokenizer",
41
+ "unk_token": "<unk>",
42
+ "use_default_system_prompt": false
43
+ }
checkpoint-320/trainer_state.json ADDED
The diff for this file is too large to render. See raw diff
 
checkpoint-320/training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:01c288b68aca2db1424771fefc72a3d0ade725c40fa055d7766bbc2e9001652d
3
+ size 8248
checkpoint-320/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)
config.json ADDED
@@ -0,0 +1,52 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_attn_implementation_autoset": true,
3
+ "_name_or_path": "liuhaotian/llava-v1.5-13b",
4
+ "architectures": [
5
+ "LlavaLlamaForCausalLM"
6
+ ],
7
+ "attention_bias": false,
8
+ "attention_dropout": 0.0,
9
+ "bos_token_id": 1,
10
+ "eos_token_id": 2,
11
+ "freeze_mm_mlp_adapter": false,
12
+ "freeze_mm_vision_resampler": false,
13
+ "head_dim": 128,
14
+ "hidden_act": "silu",
15
+ "hidden_size": 5120,
16
+ "image_aspect_ratio": "pad",
17
+ "initializer_range": 0.02,
18
+ "intermediate_size": 13824,
19
+ "max_length": 4096,
20
+ "max_position_embeddings": 4096,
21
+ "mlp_bias": false,
22
+ "mm_hidden_size": 1024,
23
+ "mm_patch_merge_type": "flat",
24
+ "mm_projector_lr": 2e-05,
25
+ "mm_projector_type": "mlp2x_gelu",
26
+ "mm_resampler_type": null,
27
+ "mm_use_im_patch_token": false,
28
+ "mm_use_im_start_end": false,
29
+ "mm_vision_select_feature": "patch",
30
+ "mm_vision_select_layer": -2,
31
+ "mm_vision_tower": "openai/clip-vit-large-patch14-336",
32
+ "model_type": "llava_llama",
33
+ "num_attention_heads": 40,
34
+ "num_hidden_layers": 40,
35
+ "num_key_value_heads": 40,
36
+ "pad_token_id": 0,
37
+ "pretraining_tp": 1,
38
+ "rms_norm_eps": 1e-05,
39
+ "rope_scaling": null,
40
+ "rope_theta": 10000.0,
41
+ "tie_word_embeddings": false,
42
+ "tokenizer_model_max_length": 2048,
43
+ "tokenizer_padding_side": "right",
44
+ "torch_dtype": "float16",
45
+ "transformers_version": "4.46.0",
46
+ "tune_mm_mlp_adapter": false,
47
+ "tune_mm_vision_resampler": false,
48
+ "unfreeze_mm_vision_tower": false,
49
+ "use_cache": true,
50
+ "use_mm_proj": true,
51
+ "vocab_size": 32000
52
+ }
non_lora_trainables.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c6df2742e15c72ec9864fc1273e3cf58fe5d383be2bd94bd9c4c3f9b0c636c0c
3
+ size 62937264
optimizer.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4f3e9883e556f25ad08adc0128c738f5af230da2703707d2fefab59aab93ff82
3
+ size 188285474