Hugging Face's logo

Bleking
/
llava-v1.6-vicuna-13b_anyres

PEFT
Safetensors
llava_llama
Model card Files Community
Bleking commited on
Commit
55ad74c
·
1 Parent(s): 7090375

llava-v1.6-vicuna-13b_anyres

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-256/README.md +202 -0
  5. checkpoint-256/adapter_config.json +34 -0
  6. checkpoint-256/adapter_model.safetensors +3 -0
  7. checkpoint-256/global_step256/zero_pp_rank_0_mp_rank_00_model_states.pt +3 -0
  8. checkpoint-256/global_step256/zero_pp_rank_0_mp_rank_00_optim_states.pt +3 -0
  9. checkpoint-256/global_step256/zero_pp_rank_1_mp_rank_00_model_states.pt +3 -0
  10. checkpoint-256/global_step256/zero_pp_rank_1_mp_rank_00_optim_states.pt +3 -0
  11. checkpoint-256/global_step256/zero_pp_rank_2_mp_rank_00_model_states.pt +3 -0
  12. checkpoint-256/global_step256/zero_pp_rank_2_mp_rank_00_optim_states.pt +3 -0
  13. checkpoint-256/global_step256/zero_pp_rank_3_mp_rank_00_model_states.pt +3 -0
  14. checkpoint-256/global_step256/zero_pp_rank_3_mp_rank_00_optim_states.pt +3 -0
  15. checkpoint-256/latest +1 -0
  16. checkpoint-256/rng_state_0.pth +3 -0
  17. checkpoint-256/rng_state_1.pth +3 -0
  18. checkpoint-256/rng_state_2.pth +3 -0
  19. checkpoint-256/rng_state_3.pth +3 -0
  20. checkpoint-256/special_tokens_map.json +24 -0
  21. checkpoint-256/tokenizer.model +3 -0
  22. checkpoint-256/tokenizer_config.json +43 -0
  23. checkpoint-256/trainer_state.json +3873 -0
  24. checkpoint-256/training_args.bin +3 -0
  25. checkpoint-256/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 +77 -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-v1.6-vicuna-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-v1.6-vicuna-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
+ "o_proj",
24
+ "q_proj",
25
+ "k_proj",
26
+ "gate_proj",
27
+ "up_proj",
28
+ "down_proj",
29
+ "v_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 sha256:582f98ea28b87a902c530b0ebdd7bdc472bdb58830ca975afdb7caab0b5d85fb
3
+ size 65046168
checkpoint-256/README.md ADDED
@@ -0,0 +1,202 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ base_model: liuhaotian/llava-v1.6-vicuna-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-256/adapter_config.json ADDED
@@ -0,0 +1,34 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "alpha_pattern": {},
3
+ "auto_mapping": null,
4
+ "base_model_name_or_path": "liuhaotian/llava-v1.6-vicuna-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
+ "gate_proj",
24
+ "up_proj",
25
+ "down_proj",
26
+ "k_proj",
27
+ "q_proj",
28
+ "v_proj",
29
+ "o_proj"
30
+ ],
31
+ "task_type": "CAUSAL_LM",
32
+ "use_dora": false,
33
+ "use_rslora": false
34
+ }
checkpoint-256/adapter_model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:582f98ea28b87a902c530b0ebdd7bdc472bdb58830ca975afdb7caab0b5d85fb
3
+ size 65046168
checkpoint-256/global_step256/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 sha256:465bf9bcd385ef29a3aded4895139c4327721f5e279f0c5928a4350b26901d89
3
+ size 775138
checkpoint-256/global_step256/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 sha256:10c68ac2bffff7c6cd1da2bd544a0b27610ccc8747cf54ddfb246cdf24ae5f3b
3
+ size 191825901
checkpoint-256/global_step256/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:a3144132b88f2cfa0c1803ff5ffa79ab3b81f0b89322c27f987f899c8ff3914e
3
+ size 775138
checkpoint-256/global_step256/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 sha256:3c6d3e5ddc3618ab6c11874ff9c3aaa56b9014dabc1ff40a2dae3695964202c5
3
+ size 191825901
checkpoint-256/global_step256/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 sha256:20274aa80392ac7f568845aa97ec6813a7365b6d537f0f55ac2a56b0cc84bbe6
3
+ size 775138
checkpoint-256/global_step256/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 sha256:6b7c5723c3fa9c44b9378020f471d5b5ab1102856a7c288397a1809227947d98
3
+ size 191825901
checkpoint-256/global_step256/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 sha256:2d07a7fb704a5010ca2b049a5b38e040049ce1bb62dd2ad63debef296372b740
3
+ size 775138
checkpoint-256/global_step256/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 sha256:16b92dff9100643c9a0e96243df4bc82354374bb04eaa1541637da45bc3a7bf4
3
+ size 191825901
checkpoint-256/latest ADDED
@@ -0,0 +1 @@
 
 
1
+ global_step256
checkpoint-256/rng_state_0.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a82dfa4cdf25f3321dc609ee6cef399ec5eb5a776c270e602eed6aa9dbcbae97
3
+ size 14960
checkpoint-256/rng_state_1.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9a29c2706ec243f7c4af15e34459ed1c098adc1768fbab00b1ca5adede23f126
3
+ size 14960
checkpoint-256/rng_state_2.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2a9cdb67e15cdb7445043e4cafda66c8a090655a1f350cb975bf3231ca39c36c
3
+ size 14960
checkpoint-256/rng_state_3.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:962c98fdd1ccee0e253d9a2f1d5f6ee58bdbf1d8f2b333298226b9554d259ac0
3
+ size 14960
checkpoint-256/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-256/tokenizer.model ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347
3
+ size 499723
checkpoint-256/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-256/trainer_state.json ADDED
@@ -0,0 +1,3873 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "best_metric": 0.6575854420661926,
3
+ "best_model_checkpoint": "./checkpoints/llava-v1.6-vicuna-13b_anyres/checkpoint-256",
4
+ "epoch": 8.0,
5
+ "eval_steps": 1.0,
6
+ "global_step": 256,
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": 0.5230235555406132,
14
+ "learning_rate": 0.0,
15
+ "loss": 1.5809,
16
+ "step": 1
17
+ },
18
+ {
19
+ "epoch": 0.03125,
20
+ "eval_loss": 1.6275018453598022,
21
+ "eval_runtime": 82.059,
22
+ "eval_samples_per_second": 2.437,
23
+ "eval_steps_per_second": 0.305,
24
+ "step": 1
25
+ },
26
+ {
27
+ "epoch": 0.0625,
28
+ "grad_norm": 0.5095402010892089,
29
+ "learning_rate": 2e-05,
30
+ "loss": 1.4958,
31
+ "step": 2
32
+ },
33
+ {
34
+ "epoch": 0.0625,
35
+ "eval_loss": 1.6275018453598022,
36
+ "eval_runtime": 76.5747,
37
+ "eval_samples_per_second": 2.612,
38
+ "eval_steps_per_second": 0.326,
39
+ "step": 2
40
+ },
41
+ {
42
+ "epoch": 0.09375,
43
+ "grad_norm": 0.4998514282504938,
44
+ "learning_rate": 2e-05,
45
+ "loss": 1.5552,
46
+ "step": 3
47
+ },
48
+ {
49
+ "epoch": 0.09375,
50
+ "eval_loss": 1.5956931114196777,
51
+ "eval_runtime": 76.1563,
52
+ "eval_samples_per_second": 2.626,
53
+ "eval_steps_per_second": 0.328,
54
+ "step": 3
55
+ },
56
+ {
57
+ "epoch": 0.125,
58
+ "grad_norm": 0.4280580315108126,
59
+ "learning_rate": 2e-05,
60
+ "loss": 1.4846,
61
+ "step": 4
62
+ },
63
+ {
64
+ "epoch": 0.125,
65
+ "eval_loss": 1.5584176778793335,
66
+ "eval_runtime": 76.1235,
67
+ "eval_samples_per_second": 2.627,
68
+ "eval_steps_per_second": 0.328,
69
+ "step": 4
70
+ },
71
+ {
72
+ "epoch": 0.15625,
73
+ "grad_norm": 0.5678499435986384,
74
+ "learning_rate": 2e-05,
75
+ "loss": 1.5036,
76
+ "step": 5
77
+ },
78
+ {
79
+ "epoch": 0.15625,
80
+ "eval_loss": 1.5207562446594238,
81
+ "eval_runtime": 76.1514,
82
+ "eval_samples_per_second": 2.626,
83
+ "eval_steps_per_second": 0.328,
84
+ "step": 5
85
+ },
86
+ {
87
+ "epoch": 0.1875,
88
+ "grad_norm": 0.5368461657542534,
89
+ "learning_rate": 2e-05,
90
+ "loss": 1.476,
91
+ "step": 6
92
+ },
93
+ {
94
+ "epoch": 0.1875,
95
+ "eval_loss": 1.4807783365249634,
96
+ "eval_runtime": 77.3444,
97
+ "eval_samples_per_second": 2.586,
98
+ "eval_steps_per_second": 0.323,
99
+ "step": 6
100
+ },
101
+ {
102
+ "epoch": 0.21875,
103
+ "grad_norm": 0.5549950083087136,
104
+ "learning_rate": 2e-05,
105
+ "loss": 1.4358,
106
+ "step": 7
107
+ },
108
+ {
109
+ "epoch": 0.21875,
110
+ "eval_loss": 1.4411544799804688,
111
+ "eval_runtime": 77.066,
112
+ "eval_samples_per_second": 2.595,
113
+ "eval_steps_per_second": 0.324,
114
+ "step": 7
115
+ },
116
+ {
117
+ "epoch": 0.25,
118
+ "grad_norm": 0.5549950083087136,
119
+ "learning_rate": 2e-05,
120
+ "loss": 1.4369,
121
+ "step": 8
122
+ },
123
+ {
124
+ "epoch": 0.25,
125
+ "eval_loss": 1.4411544799804688,
126
+ "eval_runtime": 77.2807,
127
+ "eval_samples_per_second": 2.588,
128
+ "eval_steps_per_second": 0.323,
129
+ "step": 8
130
+ },
131
+ {
132
+ "epoch": 0.28125,
133
+ "grad_norm": 0.5292240951443854,
134
+ "learning_rate": 2e-05,
135
+ "loss": 1.4471,
136
+ "step": 9
137
+ },
138
+ {
139
+ "epoch": 0.28125,
140
+ "eval_loss": 1.4036556482315063,
141
+ "eval_runtime": 78.1562,
142
+ "eval_samples_per_second": 2.559,
143
+ "eval_steps_per_second": 0.32,
144
+ "step": 9
145
+ },
146
+ {
147
+ "epoch": 0.3125,
148
+ "grad_norm": 0.5292240951443854,
149
+ "learning_rate": 2e-05,
150
+ "loss": 1.3666,
151
+ "step": 10
152
+ },
153
+ {
154
+ "epoch": 0.3125,
155
+ "eval_loss": 1.4036556482315063,
156
+ "eval_runtime": 77.1645,
157
+ "eval_samples_per_second": 2.592,
158
+ "eval_steps_per_second": 0.324,
159
+ "step": 10
160
+ },
161
+ {
162
+ "epoch": 0.34375,
163
+ "grad_norm": 0.5292240951443854,
164
+ "learning_rate": 2e-05,
165
+ "loss": 1.4149,
166
+ "step": 11
167
+ },
168
+ {
169
+ "epoch": 0.34375,
170
+ "eval_loss": 1.4036556482315063,
171
+ "eval_runtime": 78.7627,
172
+ "eval_samples_per_second": 2.539,
173
+ "eval_steps_per_second": 0.317,
174
+ "step": 11
175
+ },
176
+ {
177
+ "epoch": 0.375,
178
+ "grad_norm": 0.684588966714067,
179
+ "learning_rate": 2e-05,
180
+ "loss": 1.3883,
181
+ "step": 12
182
+ },
183
+ {
184
+ "epoch": 0.375,
185
+ "eval_loss": 1.3679308891296387,
186
+ "eval_runtime": 78.4315,
187
+ "eval_samples_per_second": 2.55,
188
+ "eval_steps_per_second": 0.319,
189
+ "step": 12
190
+ },
191
+ {
192
+ "epoch": 0.40625,
193
+ "grad_norm": 0.6261826769491422,
194
+ "learning_rate": 2e-05,
195
+ "loss": 1.4271,
196
+ "step": 13
197
+ },
198
+ {
199
+ "epoch": 0.40625,
200
+ "eval_loss": 1.3369851112365723,
201
+ "eval_runtime": 78.685,
202
+ "eval_samples_per_second": 2.542,
203
+ "eval_steps_per_second": 0.318,
204
+ "step": 13
205
+ },
206
+ {
207
+ "epoch": 0.4375,
208
+ "grad_norm": 0.6261826769491422,
209
+ "learning_rate": 2e-05,
210
+ "loss": 1.2495,
211
+ "step": 14
212
+ },
213
+ {
214
+ "epoch": 0.4375,
215
+ "eval_loss": 1.3369851112365723,
216
+ "eval_runtime": 78.0511,
217
+ "eval_samples_per_second": 2.562,
218
+ "eval_steps_per_second": 0.32,
219
+ "step": 14
220
+ },
221
+ {
222
+ "epoch": 0.46875,
223
+ "grad_norm": 0.6028103951693778,
224
+ "learning_rate": 2e-05,
225
+ "loss": 1.3513,
226
+ "step": 15
227
+ },
228
+ {
229
+ "epoch": 0.46875,
230
+ "eval_loss": 1.3032653331756592,
231
+ "eval_runtime": 78.0271,
232
+ "eval_samples_per_second": 2.563,
233
+ "eval_steps_per_second": 0.32,
234
+ "step": 15
235
+ },
236
+ {
237
+ "epoch": 0.5,
238
+ "grad_norm": 0.769290402283396,
239
+ "learning_rate": 2e-05,
240
+ "loss": 1.3117,
241
+ "step": 16
242
+ },
243
+ {
244
+ "epoch": 0.5,
245
+ "eval_loss": 1.2661188840866089,
246
+ "eval_runtime": 78.1857,
247
+ "eval_samples_per_second": 2.558,
248
+ "eval_steps_per_second": 0.32,
249
+ "step": 16
250
+ },
251
+ {
252
+ "epoch": 0.53125,
253
+ "grad_norm": 1.3279338025863765,
254
+ "learning_rate": 2e-05,
255
+ "loss": 1.2768,
256
+ "step": 17
257
+ },
258
+ {
259
+ "epoch": 0.53125,
260
+ "eval_loss": 1.2299447059631348,
261
+ "eval_runtime": 78.2064,
262
+ "eval_samples_per_second": 2.557,
263
+ "eval_steps_per_second": 0.32,
264
+ "step": 17
265
+ },
266
+ {
267
+ "epoch": 0.5625,
268
+ "grad_norm": 0.7410327159336384,
269
+ "learning_rate": 2e-05,
270
+ "loss": 1.256,
271
+ "step": 18
272
+ },
273
+ {
274
+ "epoch": 0.5625,
275
+ "eval_loss": 1.2044258117675781,
276
+ "eval_runtime": 78.072,
277
+ "eval_samples_per_second": 2.562,
278
+ "eval_steps_per_second": 0.32,
279
+ "step": 18
280
+ },
281
+ {
282
+ "epoch": 0.59375,
283
+ "grad_norm": 0.44078820770408506,
284
+ "learning_rate": 2e-05,
285
+ "loss": 1.1252,
286
+ "step": 19
287
+ },
288
+ {
289
+ "epoch": 0.59375,
290
+ "eval_loss": 1.1826122999191284,
291
+ "eval_runtime": 78.7312,
292
+ "eval_samples_per_second": 2.54,
293
+ "eval_steps_per_second": 0.318,
294
+ "step": 19
295
+ },
296
+ {
297
+ "epoch": 0.625,
298
+ "grad_norm": 0.49020841613371097,
299
+ "learning_rate": 2e-05,
300
+ "loss": 1.2249,
301
+ "step": 20
302
+ },
303
+ {
304
+ "epoch": 0.625,
305
+ "eval_loss": 1.1616511344909668,
306
+ "eval_runtime": 78.2736,
307
+ "eval_samples_per_second": 2.555,
308
+ "eval_steps_per_second": 0.319,
309
+ "step": 20
310
+ },
311
+ {
312
+ "epoch": 0.65625,
313
+ "grad_norm": 0.43031322695269714,
314
+ "learning_rate": 2e-05,
315
+ "loss": 1.1466,
316
+ "step": 21
317
+ },
318
+ {
319
+ "epoch": 0.65625,
320
+ "eval_loss": 1.1410629749298096,
321
+ "eval_runtime": 79.6432,
322
+ "eval_samples_per_second": 2.511,
323
+ "eval_steps_per_second": 0.314,
324
+ "step": 21
325
+ },
326
+ {
327
+ "epoch": 0.6875,
328
+ "grad_norm": 0.45632085445955545,
329
+ "learning_rate": 2e-05,
330
+ "loss": 1.1951,
331
+ "step": 22
332
+ },
333
+ {
334
+ "epoch": 0.6875,
335
+ "eval_loss": 1.1204684972763062,
336
+ "eval_runtime": 79.0609,
337
+ "eval_samples_per_second": 2.53,
338
+ "eval_steps_per_second": 0.316,
339
+ "step": 22
340
+ },
341
+ {
342
+ "epoch": 0.71875,
343
+ "grad_norm": 0.40048586945364495,
344
+ "learning_rate": 2e-05,
345
+ "loss": 1.1826,
346
+ "step": 23
347
+ },
348
+ {
349
+ "epoch": 0.71875,
350
+ "eval_loss": 1.1002545356750488,
351
+ "eval_runtime": 82.8578,
352
+ "eval_samples_per_second": 2.414,
353
+ "eval_steps_per_second": 0.302,
354
+ "step": 23
355
+ },
356
+ {
357
+ "epoch": 0.75,
358
+ "grad_norm": 0.3703033261027938,
359
+ "learning_rate": 2e-05,
360
+ "loss": 1.1543,
361
+ "step": 24
362
+ },
363
+ {
364
+ "epoch": 0.75,
365
+ "eval_loss": 1.0805977582931519,
366
+ "eval_runtime": 76.1407,
367
+ "eval_samples_per_second": 2.627,
368
+ "eval_steps_per_second": 0.328,
369
+ "step": 24
370
+ },
371
+ {
372
+ "epoch": 0.78125,
373
+ "grad_norm": 0.3986313105418924,
374
+ "learning_rate": 2e-05,
375
+ "loss": 1.1046,
376
+ "step": 25
377
+ },
378
+ {
379
+ "epoch": 0.78125,
380
+ "eval_loss": 1.0610157251358032,
381
+ "eval_runtime": 76.3083,
382
+ "eval_samples_per_second": 2.621,
383
+ "eval_steps_per_second": 0.328,
384
+ "step": 25
385
+ },
386
+ {
387
+ "epoch": 0.8125,
388
+ "grad_norm": 0.36265027203577943,
389
+ "learning_rate": 2e-05,
390
+ "loss": 1.1048,
391
+ "step": 26
392
+ },
393
+ {
394
+ "epoch": 0.8125,
395
+ "eval_loss": 1.0421289205551147,
396
+ "eval_runtime": 77.2186,
397
+ "eval_samples_per_second": 2.59,
398
+ "eval_steps_per_second": 0.324,
399
+ "step": 26
400
+ },
401
+ {
402
+ "epoch": 0.84375,
403
+ "grad_norm": 0.3881748990218768,
404
+ "learning_rate": 2e-05,
405
+ "loss": 1.0425,
406
+ "step": 27
407
+ },
408
+ {
409
+ "epoch": 0.84375,
410
+ "eval_loss": 1.0240073204040527,
411
+ "eval_runtime": 77.8662,
412
+ "eval_samples_per_second": 2.569,
413
+ "eval_steps_per_second": 0.321,
414
+ "step": 27
415
+ },
416
+ {
417
+ "epoch": 0.875,
418
+ "grad_norm": 0.3734031294324286,
419
+ "learning_rate": 2e-05,
420
+ "loss": 1.0484,
421
+ "step": 28
422
+ },
423
+ {
424
+ "epoch": 0.875,
425
+ "eval_loss": 1.0066957473754883,
426
+ "eval_runtime": 77.269,
427
+ "eval_samples_per_second": 2.588,
428
+ "eval_steps_per_second": 0.324,
429
+ "step": 28
430
+ },
431
+ {
432
+ "epoch": 0.90625,
433
+ "grad_norm": 0.29695383079342563,
434
+ "learning_rate": 2e-05,
435
+ "loss": 1.0387,
436
+ "step": 29
437
+ },
438
+ {
439
+ "epoch": 0.90625,
440
+ "eval_loss": 0.9906074404716492,
441
+ "eval_runtime": 77.2245,
442
+ "eval_samples_per_second": 2.59,
443
+ "eval_steps_per_second": 0.324,
444
+ "step": 29
445
+ },
446
+ {
447
+ "epoch": 0.9375,
448
+ "grad_norm": 0.29273146875026623,
449
+ "learning_rate": 2e-05,
450
+ "loss": 1.0568,
451
+ "step": 30
452
+ },
453
+ {
454
+ "epoch": 0.9375,
455
+ "eval_loss": 0.975755512714386,
456
+ "eval_runtime": 78.0056,
457
+ "eval_samples_per_second": 2.564,
458
+ "eval_steps_per_second": 0.32,
459
+ "step": 30
460
+ },
461
+ {
462
+ "epoch": 0.96875,
463
+ "grad_norm": 0.35070440686850546,
464
+ "learning_rate": 2e-05,
465
+ "loss": 0.9114,
466
+ "step": 31
467
+ },
468
+ {
469
+ "epoch": 0.96875,
470
+ "eval_loss": 0.9615123271942139,
471
+ "eval_runtime": 77.9051,
472
+ "eval_samples_per_second": 2.567,
473
+ "eval_steps_per_second": 0.321,
474
+ "step": 31
475
+ },
476
+ {
477
+ "epoch": 1.0,
478
+ "grad_norm": 0.30846157140439384,
479
+ "learning_rate": 2e-05,
480
+ "loss": 0.9941,
481
+ "step": 32
482
+ },
483
+ {
484
+ "epoch": 1.0,
485
+ "eval_loss": 0.9480571150779724,
486
+ "eval_runtime": 77.2322,
487
+ "eval_samples_per_second": 2.59,
488
+ "eval_steps_per_second": 0.324,
489
+ "step": 32
490
+ },
491
+ {
492
+ "epoch": 1.03125,
493
+ "grad_norm": 0.2950381371932973,
494
+ "learning_rate": 2e-05,
495
+ "loss": 1.0297,
496
+ "step": 33
497
+ },
498
+ {
499
+ "epoch": 1.03125,
500
+ "eval_loss": 0.9356330037117004,
501
+ "eval_runtime": 81.8443,
502
+ "eval_samples_per_second": 2.444,
503
+ "eval_steps_per_second": 0.305,
504
+ "step": 33
505
+ },
506
+ {
507
+ "epoch": 1.0625,
508
+ "grad_norm": 0.27080038065834283,
509
+ "learning_rate": 2e-05,
510
+ "loss": 1.021,
511
+ "step": 34
512
+ },
513
+ {
514
+ "epoch": 1.0625,
515
+ "eval_loss": 0.9245791435241699,
516
+ "eval_runtime": 76.2071,
517
+ "eval_samples_per_second": 2.624,
518
+ "eval_steps_per_second": 0.328,
519
+ "step": 34
520
+ },
521
+ {
522
+ "epoch": 1.09375,
523
+ "grad_norm": 0.23165081252649894,
524
+ "learning_rate": 2e-05,
525
+ "loss": 1.0366,
526
+ "step": 35
527
+ },
528
+ {
529
+ "epoch": 1.09375,
530
+ "eval_loss": 0.9151126146316528,
531
+ "eval_runtime": 77.0412,
532
+ "eval_samples_per_second": 2.596,
533
+ "eval_steps_per_second": 0.325,
534
+ "step": 35
535
+ },
536
+ {
537
+ "epoch": 1.125,
538
+ "grad_norm": 0.4033780922500775,
539
+ "learning_rate": 2e-05,
540
+ "loss": 1.0127,
541
+ "step": 36
542
+ },
543
+ {
544
+ "epoch": 1.125,
545
+ "eval_loss": 0.9063960313796997,
546
+ "eval_runtime": 76.9327,
547
+ "eval_samples_per_second": 2.6,
548
+ "eval_steps_per_second": 0.325,
549
+ "step": 36
550
+ },
551
+ {
552
+ "epoch": 1.15625,
553
+ "grad_norm": 0.2398039831439168,
554
+ "learning_rate": 2e-05,
555
+ "loss": 0.9418,
556
+ "step": 37
557
+ },
558
+ {
559
+ "epoch": 1.15625,
560
+ "eval_loss": 0.8982363939285278,
561
+ "eval_runtime": 76.1234,
562
+ "eval_samples_per_second": 2.627,
563
+ "eval_steps_per_second": 0.328,
564
+ "step": 37
565
+ },
566
+ {
567
+ "epoch": 1.1875,
568
+ "grad_norm": 0.28793451241246804,
569
+ "learning_rate": 2e-05,
570
+ "loss": 0.9643,
571
+ "step": 38
572
+ },
573
+ {
574
+ "epoch": 1.1875,
575
+ "eval_loss": 0.8908895254135132,
576
+ "eval_runtime": 76.2877,
577
+ "eval_samples_per_second": 2.622,
578
+ "eval_steps_per_second": 0.328,
579
+ "step": 38
580
+ },
581
+ {
582
+ "epoch": 1.21875,
583
+ "grad_norm": 0.2927691606307197,
584
+ "learning_rate": 2e-05,
585
+ "loss": 1.0087,
586
+ "step": 39
587
+ },
588
+ {
589
+ "epoch": 1.21875,
590
+ "eval_loss": 0.8845618367195129,
591
+ "eval_runtime": 76.2282,
592
+ "eval_samples_per_second": 2.624,
593
+ "eval_steps_per_second": 0.328,
594
+ "step": 39
595
+ },
596
+ {
597
+ "epoch": 1.25,
598
+ "grad_norm": 0.26410982001408806,
599
+ "learning_rate": 2e-05,
600
+ "loss": 0.986,
601
+ "step": 40
602
+ },
603
+ {
604
+ "epoch": 1.25,
605
+ "eval_loss": 0.8784474730491638,
606
+ "eval_runtime": 76.2512,
607
+ "eval_samples_per_second": 2.623,
608
+ "eval_steps_per_second": 0.328,
609
+ "step": 40
610
+ },
611
+ {
612
+ "epoch": 1.28125,
613
+ "grad_norm": 0.29182630949665306,
614
+ "learning_rate": 2e-05,
615
+ "loss": 0.9711,
616
+ "step": 41
617
+ },
618
+ {
619
+ "epoch": 1.28125,
620
+ "eval_loss": 0.8725223541259766,
621
+ "eval_runtime": 77.1229,
622
+ "eval_samples_per_second": 2.593,
623
+ "eval_steps_per_second": 0.324,
624
+ "step": 41
625
+ },
626
+ {
627
+ "epoch": 1.3125,
628
+ "grad_norm": 0.36402838796832665,
629
+ "learning_rate": 2e-05,
630
+ "loss": 0.9263,
631
+ "step": 42
632
+ },
633
+ {
634
+ "epoch": 1.3125,
635
+ "eval_loss": 0.8662790060043335,
636
+ "eval_runtime": 77.2362,
637
+ "eval_samples_per_second": 2.589,
638
+ "eval_steps_per_second": 0.324,
639
+ "step": 42
640
+ },
641
+ {
642
+ "epoch": 1.34375,
643
+ "grad_norm": 0.29338184478895163,
644
+ "learning_rate": 2e-05,
645
+ "loss": 0.8947,
646
+ "step": 43
647
+ },
648
+ {
649
+ "epoch": 1.34375,
650
+ "eval_loss": 0.8600431680679321,
651
+ "eval_runtime": 77.1213,
652
+ "eval_samples_per_second": 2.593,
653
+ "eval_steps_per_second": 0.324,
654
+ "step": 43
655
+ },
656
+ {
657
+ "epoch": 1.375,
658
+ "grad_norm": 0.2201714229702277,
659
+ "learning_rate": 2e-05,
660
+ "loss": 0.9059,
661
+ "step": 44
662
+ },
663
+ {
664
+ "epoch": 1.375,
665
+ "eval_loss": 0.8545799255371094,
666
+ "eval_runtime": 77.991,
667
+ "eval_samples_per_second": 2.564,
668
+ "eval_steps_per_second": 0.321,
669
+ "step": 44
670
+ },
671
+ {
672
+ "epoch": 1.40625,
673
+ "grad_norm": 0.2254966625243654,
674
+ "learning_rate": 2e-05,
675
+ "loss": 0.8942,
676
+ "step": 45
677
+ },
678
+ {
679
+ "epoch": 1.40625,
680
+ "eval_loss": 0.8497399687767029,
681
+ "eval_runtime": 77.2698,
682
+ "eval_samples_per_second": 2.588,
683
+ "eval_steps_per_second": 0.324,
684
+ "step": 45
685
+ },
686
+ {
687
+ "epoch": 1.4375,
688
+ "grad_norm": 0.21753318432075458,
689
+ "learning_rate": 2e-05,
690
+ "loss": 0.9376,
691
+ "step": 46
692
+ },
693
+ {
694
+ "epoch": 1.4375,
695
+ "eval_loss": 0.8452473282814026,
696
+ "eval_runtime": 77.0568,
697
+ "eval_samples_per_second": 2.595,
698
+ "eval_steps_per_second": 0.324,
699
+ "step": 46
700
+ },
701
+ {
702
+ "epoch": 1.46875,
703
+ "grad_norm": 0.21449718265972945,
704
+ "learning_rate": 2e-05,
705
+ "loss": 0.9369,
706
+ "step": 47
707
+ },
708
+ {
709
+ "epoch": 1.46875,
710
+ "eval_loss": 0.841134786605835,
711
+ "eval_runtime": 77.225,
712
+ "eval_samples_per_second": 2.59,
713
+ "eval_steps_per_second": 0.324,
714
+ "step": 47
715
+ },
716
+ {
717
+ "epoch": 1.5,
718
+ "grad_norm": 0.2109063266748924,
719
+ "learning_rate": 2e-05,
720
+ "loss": 0.8511,
721
+ "step": 48
722
+ },
723
+ {
724
+ "epoch": 1.5,
725
+ "eval_loss": 0.8373770117759705,
726
+ "eval_runtime": 76.2309,
727
+ "eval_samples_per_second": 2.624,
728
+ "eval_steps_per_second": 0.328,
729
+ "step": 48
730
+ },
731
+ {
732
+ "epoch": 1.53125,
733
+ "grad_norm": 0.232838633689838,
734
+ "learning_rate": 2e-05,
735
+ "loss": 0.8694,
736
+ "step": 49
737
+ },
738
+ {
739
+ "epoch": 1.53125,
740
+ "eval_loss": 0.8338289856910706,
741
+ "eval_runtime": 76.277,
742
+ "eval_samples_per_second": 2.622,
743
+ "eval_steps_per_second": 0.328,
744
+ "step": 49
745
+ },
746
+ {
747
+ "epoch": 1.5625,
748
+ "grad_norm": 0.4189704940803984,
749
+ "learning_rate": 2e-05,
750
+ "loss": 0.8464,
751
+ "step": 50
752
+ },
753
+ {
754
+ "epoch": 1.5625,
755
+ "eval_loss": 0.8297132849693298,
756
+ "eval_runtime": 76.2872,
757
+ "eval_samples_per_second": 2.622,
758
+ "eval_steps_per_second": 0.328,
759
+ "step": 50
760
+ },
761
+ {
762
+ "epoch": 1.59375,
763
+ "grad_norm": 0.2171618165123276,
764
+ "learning_rate": 2e-05,
765
+ "loss": 0.8785,
766
+ "step": 51
767
+ },
768
+ {
769
+ "epoch": 1.59375,
770
+ "eval_loss": 0.8257431983947754,
771
+ "eval_runtime": 76.2639,
772
+ "eval_samples_per_second": 2.622,
773
+ "eval_steps_per_second": 0.328,
774
+ "step": 51
775
+ },
776
+ {
777
+ "epoch": 1.625,
778
+ "grad_norm": 0.21934651037670305,
779
+ "learning_rate": 2e-05,
780
+ "loss": 0.7645,
781
+ "step": 52
782
+ },
783
+ {
784
+ "epoch": 1.625,
785
+ "eval_loss": 0.8223557472229004,
786
+ "eval_runtime": 76.2383,
787
+ "eval_samples_per_second": 2.623,
788
+ "eval_steps_per_second": 0.328,
789
+ "step": 52
790
+ },
791
+ {
792
+ "epoch": 1.65625,
793
+ "grad_norm": 0.24183530733164746,
794
+ "learning_rate": 2e-05,
795
+ "loss": 0.9218,
796
+ "step": 53
797
+ },
798
+ {
799
+ "epoch": 1.65625,
800
+ "eval_loss": 0.8189653158187866,
801
+ "eval_runtime": 76.9819,
802
+ "eval_samples_per_second": 2.598,
803
+ "eval_steps_per_second": 0.325,
804
+ "step": 53
805
+ },
806
+ {
807
+ "epoch": 1.6875,
808
+ "grad_norm": 0.23450930244279267,
809
+ "learning_rate": 2e-05,
810
+ "loss": 0.8896,
811
+ "step": 54
812
+ },
813
+ {
814
+ "epoch": 1.6875,
815
+ "eval_loss": 0.8152530193328857,
816
+ "eval_runtime": 76.2378,
817
+ "eval_samples_per_second": 2.623,
818
+ "eval_steps_per_second": 0.328,
819
+ "step": 54
820
+ },
821
+ {
822
+ "epoch": 1.71875,
823
+ "grad_norm": 0.22081665899796085,
824
+ "learning_rate": 2e-05,
825
+ "loss": 0.8798,
826
+ "step": 55
827
+ },
828
+ {
829
+ "epoch": 1.71875,
830
+ "eval_loss": 0.8122122287750244,
831
+ "eval_runtime": 76.289,
832
+ "eval_samples_per_second": 2.622,
833
+ "eval_steps_per_second": 0.328,
834
+ "step": 55
835
+ },
836
+ {
837
+ "epoch": 1.75,
838
+ "grad_norm": 0.21311746114111046,
839
+ "learning_rate": 2e-05,
840
+ "loss": 0.9482,
841
+ "step": 56
842
+ },
843
+ {
844
+ "epoch": 1.75,
845
+ "eval_loss": 0.8092318773269653,
846
+ "eval_runtime": 77.8321,
847
+ "eval_samples_per_second": 2.57,
848
+ "eval_steps_per_second": 0.321,
849
+ "step": 56
850
+ },
851
+ {
852
+ "epoch": 1.78125,
853
+ "grad_norm": 0.2496565307107556,
854
+ "learning_rate": 2e-05,
855
+ "loss": 0.8917,
856
+ "step": 57
857
+ },
858
+ {
859
+ "epoch": 1.78125,
860
+ "eval_loss": 0.8070546984672546,
861
+ "eval_runtime": 77.2651,
862
+ "eval_samples_per_second": 2.588,
863
+ "eval_steps_per_second": 0.324,
864
+ "step": 57
865
+ },
866
+ {
867
+ "epoch": 1.8125,
868
+ "grad_norm": 0.2137866456424736,
869
+ "learning_rate": 2e-05,
870
+ "loss": 0.909,
871
+ "step": 58
872
+ },
873
+ {
874
+ "epoch": 1.8125,
875
+ "eval_loss": 0.8049566745758057,
876
+ "eval_runtime": 78.0925,
877
+ "eval_samples_per_second": 2.561,
878
+ "eval_steps_per_second": 0.32,
879
+ "step": 58
880
+ },
881
+ {
882
+ "epoch": 1.84375,
883
+ "grad_norm": 0.22567502859345095,
884
+ "learning_rate": 2e-05,
885
+ "loss": 0.8611,
886
+ "step": 59
887
+ },
888
+ {
889
+ "epoch": 1.84375,
890
+ "eval_loss": 0.8028810024261475,
891
+ "eval_runtime": 78.0553,
892
+ "eval_samples_per_second": 2.562,
893
+ "eval_steps_per_second": 0.32,
894
+ "step": 59
895
+ },
896
+ {
897
+ "epoch": 1.875,
898
+ "grad_norm": 0.23303796552302508,
899
+ "learning_rate": 2e-05,
900
+ "loss": 0.9209,
901
+ "step": 60
902
+ },
903
+ {
904
+ "epoch": 1.875,
905
+ "eval_loss": 0.800568699836731,
906
+ "eval_runtime": 78.052,
907
+ "eval_samples_per_second": 2.562,
908
+ "eval_steps_per_second": 0.32,
909
+ "step": 60
910
+ },
911
+ {
912
+ "epoch": 1.90625,
913
+ "grad_norm": 0.24566727726974544,
914
+ "learning_rate": 2e-05,
915
+ "loss": 0.8239,
916
+ "step": 61
917
+ },
918
+ {
919
+ "epoch": 1.90625,
920
+ "eval_loss": 0.7976545691490173,
921
+ "eval_runtime": 77.3056,
922
+ "eval_samples_per_second": 2.587,
923
+ "eval_steps_per_second": 0.323,
924
+ "step": 61
925
+ },
926
+ {
927
+ "epoch": 1.9375,
928
+ "grad_norm": 0.23014192522354907,
929
+ "learning_rate": 2e-05,
930
+ "loss": 0.8814,
931
+ "step": 62
932
+ },
933
+ {
934
+ "epoch": 1.9375,
935
+ "eval_loss": 0.7945474982261658,
936
+ "eval_runtime": 77.3398,
937
+ "eval_samples_per_second": 2.586,
938
+ "eval_steps_per_second": 0.323,
939
+ "step": 62
940
+ },
941
+ {
942
+ "epoch": 1.96875,
943
+ "grad_norm": 0.23042819102671622,
944
+ "learning_rate": 2e-05,
945
+ "loss": 0.9064,
946
+ "step": 63
947
+ },
948
+ {
949
+ "epoch": 1.96875,
950
+ "eval_loss": 0.7918359637260437,
951
+ "eval_runtime": 77.4272,
952
+ "eval_samples_per_second": 2.583,
953
+ "eval_steps_per_second": 0.323,
954
+ "step": 63
955
+ },
956
+ {
957
+ "epoch": 2.0,
958
+ "grad_norm": 0.23940667173206315,
959
+ "learning_rate": 2e-05,
960
+ "loss": 0.8658,
961
+ "step": 64
962
+ },
963
+ {
964
+ "epoch": 2.0,
965
+ "eval_loss": 0.7891160845756531,
966
+ "eval_runtime": 77.3236,
967
+ "eval_samples_per_second": 2.587,
968
+ "eval_steps_per_second": 0.323,
969
+ "step": 64
970
+ },
971
+ {
972
+ "epoch": 2.03125,
973
+ "grad_norm": 0.22630342930143643,
974
+ "learning_rate": 2e-05,
975
+ "loss": 0.8403,
976
+ "step": 65
977
+ },
978
+ {
979
+ "epoch": 2.03125,
980
+ "eval_loss": 0.7859742641448975,
981
+ "eval_runtime": 77.2001,
982
+ "eval_samples_per_second": 2.591,
983
+ "eval_steps_per_second": 0.324,
984
+ "step": 65
985
+ },
986
+ {
987
+ "epoch": 2.0625,
988
+ "grad_norm": 0.20949240460260976,
989
+ "learning_rate": 2e-05,
990
+ "loss": 0.8472,
991
+ "step": 66
992
+ },
993
+ {
994
+ "epoch": 2.0625,
995
+ "eval_loss": 0.7834083437919617,
996
+ "eval_runtime": 78.9646,
997
+ "eval_samples_per_second": 2.533,
998
+ "eval_steps_per_second": 0.317,
999
+ "step": 66
1000
+ },
1001
+ {
1002
+ "epoch": 2.09375,
1003
+ "grad_norm": 0.22714400479820654,
1004
+ "learning_rate": 2e-05,
1005
+ "loss": 0.841,
1006
+ "step": 67
1007
+ },
1008
+ {
1009
+ "epoch": 2.09375,
1010
+ "eval_loss": 0.7805308699607849,
1011
+ "eval_runtime": 78.7552,
1012
+ "eval_samples_per_second": 2.54,
1013
+ "eval_steps_per_second": 0.317,
1014
+ "step": 67
1015
+ },
1016
+ {
1017
+ "epoch": 2.125,
1018
+ "grad_norm": 0.23345123077006047,
1019
+ "learning_rate": 2e-05,
1020
+ "loss": 0.9028,
1021
+ "step": 68
1022
+ },
1023
+ {
1024
+ "epoch": 2.125,
1025
+ "eval_loss": 0.7779514789581299,
1026
+ "eval_runtime": 78.3387,
1027
+ "eval_samples_per_second": 2.553,
1028
+ "eval_steps_per_second": 0.319,
1029
+ "step": 68
1030
+ },
1031
+ {
1032
+ "epoch": 2.15625,
1033
+ "grad_norm": 0.251841542575211,
1034
+ "learning_rate": 2e-05,
1035
+ "loss": 0.8381,
1036
+ "step": 69
1037
+ },
1038
+ {
1039
+ "epoch": 2.15625,
1040
+ "eval_loss": 0.7756664752960205,
1041
+ "eval_runtime": 78.3109,
1042
+ "eval_samples_per_second": 2.554,
1043
+ "eval_steps_per_second": 0.319,
1044
+ "step": 69
1045
+ },
1046
+ {
1047
+ "epoch": 2.1875,
1048
+ "grad_norm": 0.23548386839773608,
1049
+ "learning_rate": 2e-05,
1050
+ "loss": 0.7914,
1051
+ "step": 70
1052
+ },
1053
+ {
1054
+ "epoch": 2.1875,
1055
+ "eval_loss": 0.7733604907989502,
1056
+ "eval_runtime": 78.9712,
1057
+ "eval_samples_per_second": 2.533,
1058
+ "eval_steps_per_second": 0.317,
1059
+ "step": 70
1060
+ },
1061
+ {
1062
+ "epoch": 2.21875,
1063
+ "grad_norm": 0.23262740912668387,
1064
+ "learning_rate": 2e-05,
1065
+ "loss": 0.8778,
1066
+ "step": 71
1067
+ },
1068
+ {
1069
+ "epoch": 2.21875,
1070
+ "eval_loss": 0.771755576133728,
1071
+ "eval_runtime": 78.2633,
1072
+ "eval_samples_per_second": 2.555,
1073
+ "eval_steps_per_second": 0.319,
1074
+ "step": 71
1075
+ },
1076
+ {
1077
+ "epoch": 2.25,
1078
+ "grad_norm": 0.22075289612357513,
1079
+ "learning_rate": 2e-05,
1080
+ "loss": 0.7945,
1081
+ "step": 72
1082
+ },
1083
+ {
1084
+ "epoch": 2.25,
1085
+ "eval_loss": 0.7705450654029846,
1086
+ "eval_runtime": 78.3151,
1087
+ "eval_samples_per_second": 2.554,
1088
+ "eval_steps_per_second": 0.319,
1089
+ "step": 72
1090
+ },
1091
+ {
1092
+ "epoch": 2.28125,
1093
+ "grad_norm": 0.25520381955936466,
1094
+ "learning_rate": 2e-05,
1095
+ "loss": 0.8387,
1096
+ "step": 73
1097
+ },
1098
+ {
1099
+ "epoch": 2.28125,
1100
+ "eval_loss": 0.7695029973983765,
1101
+ "eval_runtime": 78.2901,
1102
+ "eval_samples_per_second": 2.555,
1103
+ "eval_steps_per_second": 0.319,
1104
+ "step": 73
1105
+ },
1106
+ {
1107
+ "epoch": 2.3125,
1108
+ "grad_norm": 0.2047305385827267,
1109
+ "learning_rate": 2e-05,
1110
+ "loss": 0.8404,
1111
+ "step": 74
1112
+ },
1113
+ {
1114
+ "epoch": 2.3125,
1115
+ "eval_loss": 0.7684457302093506,
1116
+ "eval_runtime": 78.3875,
1117
+ "eval_samples_per_second": 2.551,
1118
+ "eval_steps_per_second": 0.319,
1119
+ "step": 74
1120
+ },
1121
+ {
1122
+ "epoch": 2.34375,
1123
+ "grad_norm": 0.2262323045133288,
1124
+ "learning_rate": 2e-05,
1125
+ "loss": 0.8811,
1126
+ "step": 75
1127
+ },
1128
+ {
1129
+ "epoch": 2.34375,
1130
+ "eval_loss": 0.7671162486076355,
1131
+ "eval_runtime": 78.202,
1132
+ "eval_samples_per_second": 2.557,
1133
+ "eval_steps_per_second": 0.32,
1134
+ "step": 75
1135
+ },
1136
+ {
1137
+ "epoch": 2.375,
1138
+ "grad_norm": 0.21885464923925876,
1139
+ "learning_rate": 2e-05,
1140
+ "loss": 0.7942,
1141
+ "step": 76
1142
+ },
1143
+ {
1144
+ "epoch": 2.375,
1145
+ "eval_loss": 0.7658494710922241,
1146
+ "eval_runtime": 78.1746,
1147
+ "eval_samples_per_second": 2.558,
1148
+ "eval_steps_per_second": 0.32,
1149
+ "step": 76
1150
+ },
1151
+ {
1152
+ "epoch": 2.40625,
1153
+ "grad_norm": 0.21717306953626966,
1154
+ "learning_rate": 2e-05,
1155
+ "loss": 0.8497,
1156
+ "step": 77
1157
+ },
1158
+ {
1159
+ "epoch": 2.40625,
1160
+ "eval_loss": 0.7642120122909546,
1161
+ "eval_runtime": 78.2026,
1162
+ "eval_samples_per_second": 2.557,
1163
+ "eval_steps_per_second": 0.32,
1164
+ "step": 77
1165
+ },
1166
+ {
1167
+ "epoch": 2.4375,
1168
+ "grad_norm": 0.2530725583748258,
1169
+ "learning_rate": 2e-05,
1170
+ "loss": 0.8584,
1171
+ "step": 78
1172
+ },
1173
+ {
1174
+ "epoch": 2.4375,
1175
+ "eval_loss": 0.7625510692596436,
1176
+ "eval_runtime": 78.1991,
1177
+ "eval_samples_per_second": 2.558,
1178
+ "eval_steps_per_second": 0.32,
1179
+ "step": 78
1180
+ },
1181
+ {
1182
+ "epoch": 2.46875,
1183
+ "grad_norm": 0.25354787036627263,
1184
+ "learning_rate": 2e-05,
1185
+ "loss": 0.8569,
1186
+ "step": 79
1187
+ },
1188
+ {
1189
+ "epoch": 2.46875,
1190
+ "eval_loss": 0.7616268396377563,
1191
+ "eval_runtime": 78.2915,
1192
+ "eval_samples_per_second": 2.555,
1193
+ "eval_steps_per_second": 0.319,
1194
+ "step": 79
1195
+ },
1196
+ {
1197
+ "epoch": 2.5,
1198
+ "grad_norm": 0.2800865746664007,
1199
+ "learning_rate": 2e-05,
1200
+ "loss": 0.9116,
1201
+ "step": 80
1202
+ },
1203
+ {
1204
+ "epoch": 2.5,
1205
+ "eval_loss": 0.7603214979171753,
1206
+ "eval_runtime": 78.2749,
1207
+ "eval_samples_per_second": 2.555,
1208
+ "eval_steps_per_second": 0.319,
1209
+ "step": 80
1210
+ },
1211
+ {
1212
+ "epoch": 2.53125,
1213
+ "grad_norm": 0.268139688449618,
1214
+ "learning_rate": 2e-05,
1215
+ "loss": 0.8397,
1216
+ "step": 81
1217
+ },
1218
+ {
1219
+ "epoch": 2.53125,
1220
+ "eval_loss": 0.7584869265556335,
1221
+ "eval_runtime": 79.1445,
1222
+ "eval_samples_per_second": 2.527,
1223
+ "eval_steps_per_second": 0.316,
1224
+ "step": 81
1225
+ },
1226
+ {
1227
+ "epoch": 2.5625,
1228
+ "grad_norm": 0.3128648654463789,
1229
+ "learning_rate": 2e-05,
1230
+ "loss": 0.8888,
1231
+ "step": 82
1232
+ },
1233
+ {
1234
+ "epoch": 2.5625,
1235
+ "eval_loss": 0.7566561102867126,
1236
+ "eval_runtime": 79.2089,
1237
+ "eval_samples_per_second": 2.525,
1238
+ "eval_steps_per_second": 0.316,
1239
+ "step": 82
1240
+ },
1241
+ {
1242
+ "epoch": 2.59375,
1243
+ "grad_norm": 0.2502355211215609,
1244
+ "learning_rate": 2e-05,
1245
+ "loss": 0.8346,
1246
+ "step": 83
1247
+ },
1248
+ {
1249
+ "epoch": 2.59375,
1250
+ "eval_loss": 0.7547345161437988,
1251
+ "eval_runtime": 79.2691,
1252
+ "eval_samples_per_second": 2.523,
1253
+ "eval_steps_per_second": 0.315,
1254
+ "step": 83
1255
+ },
1256
+ {
1257
+ "epoch": 2.625,
1258
+ "grad_norm": 0.25281184629018644,
1259
+ "learning_rate": 2e-05,
1260
+ "loss": 0.795,
1261
+ "step": 84
1262
+ },
1263
+ {
1264
+ "epoch": 2.625,
1265
+ "eval_loss": 0.7527951598167419,
1266
+ "eval_runtime": 79.4068,
1267
+ "eval_samples_per_second": 2.519,
1268
+ "eval_steps_per_second": 0.315,
1269
+ "step": 84
1270
+ },
1271
+ {
1272
+ "epoch": 2.65625,
1273
+ "grad_norm": 0.24246729562645003,
1274
+ "learning_rate": 2e-05,
1275
+ "loss": 0.7649,
1276
+ "step": 85
1277
+ },
1278
+ {
1279
+ "epoch": 2.65625,
1280
+ "eval_loss": 0.7509815096855164,
1281
+ "eval_runtime": 79.1612,
1282
+ "eval_samples_per_second": 2.526,
1283
+ "eval_steps_per_second": 0.316,
1284
+ "step": 85
1285
+ },
1286
+ {
1287
+ "epoch": 2.6875,
1288
+ "grad_norm": 0.27005475109453947,
1289
+ "learning_rate": 2e-05,
1290
+ "loss": 0.7964,
1291
+ "step": 86
1292
+ },
1293
+ {
1294
+ "epoch": 2.6875,
1295
+ "eval_loss": 0.7485950589179993,
1296
+ "eval_runtime": 80.0714,
1297
+ "eval_samples_per_second": 2.498,
1298
+ "eval_steps_per_second": 0.312,
1299
+ "step": 86
1300
+ },
1301
+ {
1302
+ "epoch": 2.71875,
1303
+ "grad_norm": 0.2723492355800971,
1304
+ "learning_rate": 2e-05,
1305
+ "loss": 0.8117,
1306
+ "step": 87
1307
+ },
1308
+ {
1309
+ "epoch": 2.71875,
1310
+ "eval_loss": 0.7459420561790466,
1311
+ "eval_runtime": 79.4075,
1312
+ "eval_samples_per_second": 2.519,
1313
+ "eval_steps_per_second": 0.315,
1314
+ "step": 87
1315
+ },
1316
+ {
1317
+ "epoch": 2.75,
1318
+ "grad_norm": 0.2946493898427159,
1319
+ "learning_rate": 2e-05,
1320
+ "loss": 0.8986,
1321
+ "step": 88
1322
+ },
1323
+ {
1324
+ "epoch": 2.75,
1325
+ "eval_loss": 0.7436455488204956,
1326
+ "eval_runtime": 79.3721,
1327
+ "eval_samples_per_second": 2.52,
1328
+ "eval_steps_per_second": 0.315,
1329
+ "step": 88
1330
+ },
1331
+ {
1332
+ "epoch": 2.78125,
1333
+ "grad_norm": 0.26411214734213284,
1334
+ "learning_rate": 2e-05,
1335
+ "loss": 0.8145,
1336
+ "step": 89
1337
+ },
1338
+ {
1339
+ "epoch": 2.78125,
1340
+ "eval_loss": 0.7424752712249756,
1341
+ "eval_runtime": 79.2988,
1342
+ "eval_samples_per_second": 2.522,
1343
+ "eval_steps_per_second": 0.315,
1344
+ "step": 89
1345
+ },
1346
+ {
1347
+ "epoch": 2.8125,
1348
+ "grad_norm": 0.27115747269014817,
1349
+ "learning_rate": 2e-05,
1350
+ "loss": 0.8457,
1351
+ "step": 90
1352
+ },
1353
+ {
1354
+ "epoch": 2.8125,
1355
+ "eval_loss": 0.7416408658027649,
1356
+ "eval_runtime": 79.4004,
1357
+ "eval_samples_per_second": 2.519,
1358
+ "eval_steps_per_second": 0.315,
1359
+ "step": 90
1360
+ },
1361
+ {
1362
+ "epoch": 2.84375,
1363
+ "grad_norm": 0.25831877964821937,
1364
+ "learning_rate": 2e-05,
1365
+ "loss": 0.7568,
1366
+ "step": 91
1367
+ },
1368
+ {
1369
+ "epoch": 2.84375,
1370
+ "eval_loss": 0.7404463291168213,
1371
+ "eval_runtime": 81.7767,
1372
+ "eval_samples_per_second": 2.446,
1373
+ "eval_steps_per_second": 0.306,
1374
+ "step": 91
1375
+ },
1376
+ {
1377
+ "epoch": 2.875,
1378
+ "grad_norm": 0.31273388454942935,
1379
+ "learning_rate": 2e-05,
1380
+ "loss": 0.8562,
1381
+ "step": 92
1382
+ },
1383
+ {
1384
+ "epoch": 2.875,
1385
+ "eval_loss": 0.7384185791015625,
1386
+ "eval_runtime": 82.3443,
1387
+ "eval_samples_per_second": 2.429,
1388
+ "eval_steps_per_second": 0.304,
1389
+ "step": 92
1390
+ },
1391
+ {
1392
+ "epoch": 2.90625,
1393
+ "grad_norm": 0.2838267071008901,
1394
+ "learning_rate": 2e-05,
1395
+ "loss": 0.7869,
1396
+ "step": 93
1397
+ },
1398
+ {
1399
+ "epoch": 2.90625,
1400
+ "eval_loss": 0.7366807460784912,
1401
+ "eval_runtime": 82.2622,
1402
+ "eval_samples_per_second": 2.431,
1403
+ "eval_steps_per_second": 0.304,
1404
+ "step": 93
1405
+ },
1406
+ {
1407
+ "epoch": 2.9375,
1408
+ "grad_norm": 0.28625827941831467,
1409
+ "learning_rate": 2e-05,
1410
+ "loss": 0.8618,
1411
+ "step": 94
1412
+ },
1413
+ {
1414
+ "epoch": 2.9375,
1415
+ "eval_loss": 0.7357398867607117,
1416
+ "eval_runtime": 81.9471,
1417
+ "eval_samples_per_second": 2.441,
1418
+ "eval_steps_per_second": 0.305,
1419
+ "step": 94
1420
+ },
1421
+ {
1422
+ "epoch": 2.96875,
1423
+ "grad_norm": 0.25548002643954326,
1424
+ "learning_rate": 2e-05,
1425
+ "loss": 0.8085,
1426
+ "step": 95
1427
+ },
1428
+ {
1429
+ "epoch": 2.96875,
1430
+ "eval_loss": 0.7356534004211426,
1431
+ "eval_runtime": 82.1186,
1432
+ "eval_samples_per_second": 2.436,
1433
+ "eval_steps_per_second": 0.304,
1434
+ "step": 95
1435
+ },
1436
+ {
1437
+ "epoch": 3.0,
1438
+ "grad_norm": 0.27081450830961107,
1439
+ "learning_rate": 2e-05,
1440
+ "loss": 0.7684,
1441
+ "step": 96
1442
+ },
1443
+ {
1444
+ "epoch": 3.0,
1445
+ "eval_loss": 0.7346957921981812,
1446
+ "eval_runtime": 81.5463,
1447
+ "eval_samples_per_second": 2.453,
1448
+ "eval_steps_per_second": 0.307,
1449
+ "step": 96
1450
+ },
1451
+ {
1452
+ "epoch": 3.03125,
1453
+ "grad_norm": 0.2985486737236676,
1454
+ "learning_rate": 2e-05,
1455
+ "loss": 0.7274,
1456
+ "step": 97
1457
+ },
1458
+ {
1459
+ "epoch": 3.03125,
1460
+ "eval_loss": 0.7325752377510071,
1461
+ "eval_runtime": 81.7804,
1462
+ "eval_samples_per_second": 2.446,
1463
+ "eval_steps_per_second": 0.306,
1464
+ "step": 97
1465
+ },
1466
+ {
1467
+ "epoch": 3.0625,
1468
+ "grad_norm": 0.29149719690624026,
1469
+ "learning_rate": 2e-05,
1470
+ "loss": 0.8119,
1471
+ "step": 98
1472
+ },
1473
+ {
1474
+ "epoch": 3.0625,
1475
+ "eval_loss": 0.7298976182937622,
1476
+ "eval_runtime": 76.2764,
1477
+ "eval_samples_per_second": 2.622,
1478
+ "eval_steps_per_second": 0.328,
1479
+ "step": 98
1480
+ },
1481
+ {
1482
+ "epoch": 3.09375,
1483
+ "grad_norm": 0.25227859825215865,
1484
+ "learning_rate": 2e-05,
1485
+ "loss": 0.7888,
1486
+ "step": 99
1487
+ },
1488
+ {
1489
+ "epoch": 3.09375,
1490
+ "eval_loss": 0.727373480796814,
1491
+ "eval_runtime": 76.2418,
1492
+ "eval_samples_per_second": 2.623,
1493
+ "eval_steps_per_second": 0.328,
1494
+ "step": 99
1495
+ },
1496
+ {
1497
+ "epoch": 3.125,
1498
+ "grad_norm": 0.27316954971752555,
1499
+ "learning_rate": 2e-05,
1500
+ "loss": 0.8224,
1501
+ "step": 100
1502
+ },
1503
+ {
1504
+ "epoch": 3.125,
1505
+ "eval_loss": 0.7254325747489929,
1506
+ "eval_runtime": 76.1474,
1507
+ "eval_samples_per_second": 2.626,
1508
+ "eval_steps_per_second": 0.328,
1509
+ "step": 100
1510
+ },
1511
+ {
1512
+ "epoch": 3.15625,
1513
+ "grad_norm": 0.24239788607957785,
1514
+ "learning_rate": 2e-05,
1515
+ "loss": 0.7535,
1516
+ "step": 101
1517
+ },
1518
+ {
1519
+ "epoch": 3.15625,
1520
+ "eval_loss": 0.724058985710144,
1521
+ "eval_runtime": 76.2391,
1522
+ "eval_samples_per_second": 2.623,
1523
+ "eval_steps_per_second": 0.328,
1524
+ "step": 101
1525
+ },
1526
+ {
1527
+ "epoch": 3.1875,
1528
+ "grad_norm": 0.25648385925427025,
1529
+ "learning_rate": 2e-05,
1530
+ "loss": 0.8195,
1531
+ "step": 102
1532
+ },
1533
+ {
1534
+ "epoch": 3.1875,
1535
+ "eval_loss": 0.7235870957374573,
1536
+ "eval_runtime": 76.9134,
1537
+ "eval_samples_per_second": 2.6,
1538
+ "eval_steps_per_second": 0.325,
1539
+ "step": 102
1540
+ },
1541
+ {
1542
+ "epoch": 3.21875,
1543
+ "grad_norm": 0.29620170789161204,
1544
+ "learning_rate": 2e-05,
1545
+ "loss": 0.8224,
1546
+ "step": 103
1547
+ },
1548
+ {
1549
+ "epoch": 3.21875,
1550
+ "eval_loss": 0.7228152751922607,
1551
+ "eval_runtime": 76.095,
1552
+ "eval_samples_per_second": 2.628,
1553
+ "eval_steps_per_second": 0.329,
1554
+ "step": 103
1555
+ },
1556
+ {
1557
+ "epoch": 3.25,
1558
+ "grad_norm": 0.3484116181139593,
1559
+ "learning_rate": 2e-05,
1560
+ "loss": 0.7478,
1561
+ "step": 104
1562
+ },
1563
+ {
1564
+ "epoch": 3.25,
1565
+ "eval_loss": 0.7209363579750061,
1566
+ "eval_runtime": 76.9377,
1567
+ "eval_samples_per_second": 2.6,
1568
+ "eval_steps_per_second": 0.325,
1569
+ "step": 104
1570
+ },
1571
+ {
1572
+ "epoch": 3.28125,
1573
+ "grad_norm": 0.25212350156184643,
1574
+ "learning_rate": 2e-05,
1575
+ "loss": 0.7885,
1576
+ "step": 105
1577
+ },
1578
+ {
1579
+ "epoch": 3.28125,
1580
+ "eval_loss": 0.7197096347808838,
1581
+ "eval_runtime": 76.2008,
1582
+ "eval_samples_per_second": 2.625,
1583
+ "eval_steps_per_second": 0.328,
1584
+ "step": 105
1585
+ },
1586
+ {
1587
+ "epoch": 3.3125,
1588
+ "grad_norm": 0.264200147608962,
1589
+ "learning_rate": 2e-05,
1590
+ "loss": 0.8371,
1591
+ "step": 106
1592
+ },
1593
+ {
1594
+ "epoch": 3.3125,
1595
+ "eval_loss": 0.7197055220603943,
1596
+ "eval_runtime": 78.1542,
1597
+ "eval_samples_per_second": 2.559,
1598
+ "eval_steps_per_second": 0.32,
1599
+ "step": 106
1600
+ },
1601
+ {
1602
+ "epoch": 3.34375,
1603
+ "grad_norm": 0.3309431084940201,
1604
+ "learning_rate": 2e-05,
1605
+ "loss": 0.6999,
1606
+ "step": 107
1607
+ },
1608
+ {
1609
+ "epoch": 3.34375,
1610
+ "eval_loss": 0.7187016010284424,
1611
+ "eval_runtime": 78.4259,
1612
+ "eval_samples_per_second": 2.55,
1613
+ "eval_steps_per_second": 0.319,
1614
+ "step": 107
1615
+ },
1616
+ {
1617
+ "epoch": 3.375,
1618
+ "grad_norm": 0.3131644456919823,
1619
+ "learning_rate": 2e-05,
1620
+ "loss": 0.7587,
1621
+ "step": 108
1622
+ },
1623
+ {
1624
+ "epoch": 3.375,
1625
+ "eval_loss": 0.717018187046051,
1626
+ "eval_runtime": 78.4558,
1627
+ "eval_samples_per_second": 2.549,
1628
+ "eval_steps_per_second": 0.319,
1629
+ "step": 108
1630
+ },
1631
+ {
1632
+ "epoch": 3.40625,
1633
+ "grad_norm": 0.33527684120780293,
1634
+ "learning_rate": 2e-05,
1635
+ "loss": 0.7468,
1636
+ "step": 109
1637
+ },
1638
+ {
1639
+ "epoch": 3.40625,
1640
+ "eval_loss": 0.7147062420845032,
1641
+ "eval_runtime": 78.2334,
1642
+ "eval_samples_per_second": 2.556,
1643
+ "eval_steps_per_second": 0.32,
1644
+ "step": 109
1645
+ },
1646
+ {
1647
+ "epoch": 3.4375,
1648
+ "grad_norm": 0.29542683956231724,
1649
+ "learning_rate": 2e-05,
1650
+ "loss": 0.7477,
1651
+ "step": 110
1652
+ },
1653
+ {
1654
+ "epoch": 3.4375,
1655
+ "eval_loss": 0.7130224704742432,
1656
+ "eval_runtime": 79.1179,
1657
+ "eval_samples_per_second": 2.528,
1658
+ "eval_steps_per_second": 0.316,
1659
+ "step": 110
1660
+ },
1661
+ {
1662
+ "epoch": 3.46875,
1663
+ "grad_norm": 0.31128698002926114,
1664
+ "learning_rate": 2e-05,
1665
+ "loss": 0.8153,
1666
+ "step": 111
1667
+ },
1668
+ {
1669
+ "epoch": 3.46875,
1670
+ "eval_loss": 0.7120551466941833,
1671
+ "eval_runtime": 80.292,
1672
+ "eval_samples_per_second": 2.491,
1673
+ "eval_steps_per_second": 0.311,
1674
+ "step": 111
1675
+ },
1676
+ {
1677
+ "epoch": 3.5,
1678
+ "grad_norm": 0.32502558864214215,
1679
+ "learning_rate": 2e-05,
1680
+ "loss": 0.8043,
1681
+ "step": 112
1682
+ },
1683
+ {
1684
+ "epoch": 3.5,
1685
+ "eval_loss": 0.7117202877998352,
1686
+ "eval_runtime": 79.7539,
1687
+ "eval_samples_per_second": 2.508,
1688
+ "eval_steps_per_second": 0.313,
1689
+ "step": 112
1690
+ },
1691
+ {
1692
+ "epoch": 3.53125,
1693
+ "grad_norm": 0.34335720855758517,
1694
+ "learning_rate": 2e-05,
1695
+ "loss": 0.871,
1696
+ "step": 113
1697
+ },
1698
+ {
1699
+ "epoch": 3.53125,
1700
+ "eval_loss": 0.7117029428482056,
1701
+ "eval_runtime": 80.0281,
1702
+ "eval_samples_per_second": 2.499,
1703
+ "eval_steps_per_second": 0.312,
1704
+ "step": 113
1705
+ },
1706
+ {
1707
+ "epoch": 3.5625,
1708
+ "grad_norm": 0.31951931695644,
1709
+ "learning_rate": 2e-05,
1710
+ "loss": 0.7453,
1711
+ "step": 114
1712
+ },
1713
+ {
1714
+ "epoch": 3.5625,
1715
+ "eval_loss": 0.7116554379463196,
1716
+ "eval_runtime": 79.7209,
1717
+ "eval_samples_per_second": 2.509,
1718
+ "eval_steps_per_second": 0.314,
1719
+ "step": 114
1720
+ },
1721
+ {
1722
+ "epoch": 3.59375,
1723
+ "grad_norm": 0.28067192963874266,
1724
+ "learning_rate": 2e-05,
1725
+ "loss": 0.8045,
1726
+ "step": 115
1727
+ },
1728
+ {
1729
+ "epoch": 3.59375,
1730
+ "eval_loss": 0.7118353843688965,
1731
+ "eval_runtime": 80.0195,
1732
+ "eval_samples_per_second": 2.499,
1733
+ "eval_steps_per_second": 0.312,
1734
+ "step": 115
1735
+ },
1736
+ {
1737
+ "epoch": 3.625,
1738
+ "grad_norm": 0.2739718257400276,
1739
+ "learning_rate": 2e-05,
1740
+ "loss": 0.775,
1741
+ "step": 116
1742
+ },
1743
+ {
1744
+ "epoch": 3.625,
1745
+ "eval_loss": 0.7122579216957092,
1746
+ "eval_runtime": 76.2052,
1747
+ "eval_samples_per_second": 2.624,
1748
+ "eval_steps_per_second": 0.328,
1749
+ "step": 116
1750
+ },
1751
+ {
1752
+ "epoch": 3.65625,
1753
+ "grad_norm": 0.31401723658881836,
1754
+ "learning_rate": 2e-05,
1755
+ "loss": 0.7826,
1756
+ "step": 117
1757
+ },
1758
+ {
1759
+ "epoch": 3.65625,
1760
+ "eval_loss": 0.7118574380874634,
1761
+ "eval_runtime": 76.1509,
1762
+ "eval_samples_per_second": 2.626,
1763
+ "eval_steps_per_second": 0.328,
1764
+ "step": 117
1765
+ },
1766
+ {
1767
+ "epoch": 3.6875,
1768
+ "grad_norm": 0.36925964858634625,
1769
+ "learning_rate": 2e-05,
1770
+ "loss": 0.7884,
1771
+ "step": 118
1772
+ },
1773
+ {
1774
+ "epoch": 3.6875,
1775
+ "eval_loss": 0.710691511631012,
1776
+ "eval_runtime": 76.2305,
1777
+ "eval_samples_per_second": 2.624,
1778
+ "eval_steps_per_second": 0.328,
1779
+ "step": 118
1780
+ },
1781
+ {
1782
+ "epoch": 3.71875,
1783
+ "grad_norm": 0.3050583880654791,
1784
+ "learning_rate": 2e-05,
1785
+ "loss": 0.8402,
1786
+ "step": 119
1787
+ },
1788
+ {
1789
+ "epoch": 3.71875,
1790
+ "eval_loss": 0.7096763849258423,
1791
+ "eval_runtime": 77.0581,
1792
+ "eval_samples_per_second": 2.595,
1793
+ "eval_steps_per_second": 0.324,
1794
+ "step": 119
1795
+ },
1796
+ {
1797
+ "epoch": 3.75,
1798
+ "grad_norm": 0.2648625651290031,
1799
+ "learning_rate": 2e-05,
1800
+ "loss": 0.7889,
1801
+ "step": 120
1802
+ },
1803
+ {
1804
+ "epoch": 3.75,
1805
+ "eval_loss": 0.7094223499298096,
1806
+ "eval_runtime": 76.1379,
1807
+ "eval_samples_per_second": 2.627,
1808
+ "eval_steps_per_second": 0.328,
1809
+ "step": 120
1810
+ },
1811
+ {
1812
+ "epoch": 3.78125,
1813
+ "grad_norm": 0.3107221696449271,
1814
+ "learning_rate": 2e-05,
1815
+ "loss": 0.7615,
1816
+ "step": 121
1817
+ },
1818
+ {
1819
+ "epoch": 3.78125,
1820
+ "eval_loss": 0.7081363201141357,
1821
+ "eval_runtime": 76.626,
1822
+ "eval_samples_per_second": 2.61,
1823
+ "eval_steps_per_second": 0.326,
1824
+ "step": 121
1825
+ },
1826
+ {
1827
+ "epoch": 3.8125,
1828
+ "grad_norm": 0.3455151299995048,
1829
+ "learning_rate": 2e-05,
1830
+ "loss": 0.8342,
1831
+ "step": 122
1832
+ },
1833
+ {
1834
+ "epoch": 3.8125,
1835
+ "eval_loss": 0.7063001990318298,
1836
+ "eval_runtime": 77.0293,
1837
+ "eval_samples_per_second": 2.596,
1838
+ "eval_steps_per_second": 0.325,
1839
+ "step": 122
1840
+ },
1841
+ {
1842
+ "epoch": 3.84375,
1843
+ "grad_norm": 0.28847071926472523,
1844
+ "learning_rate": 2e-05,
1845
+ "loss": 0.7477,
1846
+ "step": 123
1847
+ },
1848
+ {
1849
+ "epoch": 3.84375,
1850
+ "eval_loss": 0.7044610381126404,
1851
+ "eval_runtime": 76.2385,
1852
+ "eval_samples_per_second": 2.623,
1853
+ "eval_steps_per_second": 0.328,
1854
+ "step": 123
1855
+ },
1856
+ {
1857
+ "epoch": 3.875,
1858
+ "grad_norm": 0.26753816515069856,
1859
+ "learning_rate": 2e-05,
1860
+ "loss": 0.7653,
1861
+ "step": 124
1862
+ },
1863
+ {
1864
+ "epoch": 3.875,
1865
+ "eval_loss": 0.7033799886703491,
1866
+ "eval_runtime": 76.1985,
1867
+ "eval_samples_per_second": 2.625,
1868
+ "eval_steps_per_second": 0.328,
1869
+ "step": 124
1870
+ },
1871
+ {
1872
+ "epoch": 3.90625,
1873
+ "grad_norm": 0.3465046292893005,
1874
+ "learning_rate": 2e-05,
1875
+ "loss": 0.8144,
1876
+ "step": 125
1877
+ },
1878
+ {
1879
+ "epoch": 3.90625,
1880
+ "eval_loss": 0.7021930813789368,
1881
+ "eval_runtime": 76.2234,
1882
+ "eval_samples_per_second": 2.624,
1883
+ "eval_steps_per_second": 0.328,
1884
+ "step": 125
1885
+ },
1886
+ {
1887
+ "epoch": 3.9375,
1888
+ "grad_norm": 0.3451690427620698,
1889
+ "learning_rate": 2e-05,
1890
+ "loss": 0.7871,
1891
+ "step": 126
1892
+ },
1893
+ {
1894
+ "epoch": 3.9375,
1895
+ "eval_loss": 0.7013542652130127,
1896
+ "eval_runtime": 78.0752,
1897
+ "eval_samples_per_second": 2.562,
1898
+ "eval_steps_per_second": 0.32,
1899
+ "step": 126
1900
+ },
1901
+ {
1902
+ "epoch": 3.96875,
1903
+ "grad_norm": 0.31571858642673567,
1904
+ "learning_rate": 2e-05,
1905
+ "loss": 0.7568,
1906
+ "step": 127
1907
+ },
1908
+ {
1909
+ "epoch": 3.96875,
1910
+ "eval_loss": 0.7007560729980469,
1911
+ "eval_runtime": 78.3558,
1912
+ "eval_samples_per_second": 2.552,
1913
+ "eval_steps_per_second": 0.319,
1914
+ "step": 127
1915
+ },
1916
+ {
1917
+ "epoch": 4.0,
1918
+ "grad_norm": 0.3247003540270338,
1919
+ "learning_rate": 2e-05,
1920
+ "loss": 0.6714,
1921
+ "step": 128
1922
+ },
1923
+ {
1924
+ "epoch": 4.0,
1925
+ "eval_loss": 0.6999780535697937,
1926
+ "eval_runtime": 78.9788,
1927
+ "eval_samples_per_second": 2.532,
1928
+ "eval_steps_per_second": 0.317,
1929
+ "step": 128
1930
+ },
1931
+ {
1932
+ "epoch": 4.03125,
1933
+ "grad_norm": 0.2814983490019739,
1934
+ "learning_rate": 2e-05,
1935
+ "loss": 0.7797,
1936
+ "step": 129
1937
+ },
1938
+ {
1939
+ "epoch": 4.03125,
1940
+ "eval_loss": 0.6998200416564941,
1941
+ "eval_runtime": 78.3093,
1942
+ "eval_samples_per_second": 2.554,
1943
+ "eval_steps_per_second": 0.319,
1944
+ "step": 129
1945
+ },
1946
+ {
1947
+ "epoch": 4.0625,
1948
+ "grad_norm": 0.31961631715145106,
1949
+ "learning_rate": 2e-05,
1950
+ "loss": 0.7993,
1951
+ "step": 130
1952
+ },
1953
+ {
1954
+ "epoch": 4.0625,
1955
+ "eval_loss": 0.6995271444320679,
1956
+ "eval_runtime": 78.2172,
1957
+ "eval_samples_per_second": 2.557,
1958
+ "eval_steps_per_second": 0.32,
1959
+ "step": 130
1960
+ },
1961
+ {
1962
+ "epoch": 4.09375,
1963
+ "grad_norm": 0.32333364662215863,
1964
+ "learning_rate": 2e-05,
1965
+ "loss": 0.7896,
1966
+ "step": 131
1967
+ },
1968
+ {
1969
+ "epoch": 4.09375,
1970
+ "eval_loss": 0.6992727518081665,
1971
+ "eval_runtime": 79.0125,
1972
+ "eval_samples_per_second": 2.531,
1973
+ "eval_steps_per_second": 0.316,
1974
+ "step": 131
1975
+ },
1976
+ {
1977
+ "epoch": 4.125,
1978
+ "grad_norm": 0.3255859640449829,
1979
+ "learning_rate": 2e-05,
1980
+ "loss": 0.7542,
1981
+ "step": 132
1982
+ },
1983
+ {
1984
+ "epoch": 4.125,
1985
+ "eval_loss": 0.6988572478294373,
1986
+ "eval_runtime": 79.0,
1987
+ "eval_samples_per_second": 2.532,
1988
+ "eval_steps_per_second": 0.316,
1989
+ "step": 132
1990
+ },
1991
+ {
1992
+ "epoch": 4.15625,
1993
+ "grad_norm": 0.3307068947429175,
1994
+ "learning_rate": 2e-05,
1995
+ "loss": 0.8416,
1996
+ "step": 133
1997
+ },
1998
+ {
1999
+ "epoch": 4.15625,
2000
+ "eval_loss": 0.6981343030929565,
2001
+ "eval_runtime": 78.3309,
2002
+ "eval_samples_per_second": 2.553,
2003
+ "eval_steps_per_second": 0.319,
2004
+ "step": 133
2005
+ },
2006
+ {
2007
+ "epoch": 4.1875,
2008
+ "grad_norm": 0.3842303818116732,
2009
+ "learning_rate": 2e-05,
2010
+ "loss": 0.7605,
2011
+ "step": 134
2012
+ },
2013
+ {
2014
+ "epoch": 4.1875,
2015
+ "eval_loss": 0.6968980431556702,
2016
+ "eval_runtime": 78.5608,
2017
+ "eval_samples_per_second": 2.546,
2018
+ "eval_steps_per_second": 0.318,
2019
+ "step": 134
2020
+ },
2021
+ {
2022
+ "epoch": 4.21875,
2023
+ "grad_norm": 0.331839472419003,
2024
+ "learning_rate": 2e-05,
2025
+ "loss": 0.7643,
2026
+ "step": 135
2027
+ },
2028
+ {
2029
+ "epoch": 4.21875,
2030
+ "eval_loss": 0.6955949664115906,
2031
+ "eval_runtime": 78.3566,
2032
+ "eval_samples_per_second": 2.552,
2033
+ "eval_steps_per_second": 0.319,
2034
+ "step": 135
2035
+ },
2036
+ {
2037
+ "epoch": 4.25,
2038
+ "grad_norm": 0.31864813130499836,
2039
+ "learning_rate": 2e-05,
2040
+ "loss": 0.7369,
2041
+ "step": 136
2042
+ },
2043
+ {
2044
+ "epoch": 4.25,
2045
+ "eval_loss": 0.6951528787612915,
2046
+ "eval_runtime": 79.7802,
2047
+ "eval_samples_per_second": 2.507,
2048
+ "eval_steps_per_second": 0.313,
2049
+ "step": 136
2050
+ },
2051
+ {
2052
+ "epoch": 4.28125,
2053
+ "grad_norm": 0.352549164434451,
2054
+ "learning_rate": 2e-05,
2055
+ "loss": 0.7332,
2056
+ "step": 137
2057
+ },
2058
+ {
2059
+ "epoch": 4.28125,
2060
+ "eval_loss": 0.6947290897369385,
2061
+ "eval_runtime": 79.8171,
2062
+ "eval_samples_per_second": 2.506,
2063
+ "eval_steps_per_second": 0.313,
2064
+ "step": 137
2065
+ },
2066
+ {
2067
+ "epoch": 4.3125,
2068
+ "grad_norm": 0.37128812818896284,
2069
+ "learning_rate": 2e-05,
2070
+ "loss": 0.7542,
2071
+ "step": 138
2072
+ },
2073
+ {
2074
+ "epoch": 4.3125,
2075
+ "eval_loss": 0.6937370300292969,
2076
+ "eval_runtime": 79.7782,
2077
+ "eval_samples_per_second": 2.507,
2078
+ "eval_steps_per_second": 0.313,
2079
+ "step": 138
2080
+ },
2081
+ {
2082
+ "epoch": 4.34375,
2083
+ "grad_norm": 0.3348014941412048,
2084
+ "learning_rate": 2e-05,
2085
+ "loss": 0.7079,
2086
+ "step": 139
2087
+ },
2088
+ {
2089
+ "epoch": 4.34375,
2090
+ "eval_loss": 0.692456066608429,
2091
+ "eval_runtime": 79.9308,
2092
+ "eval_samples_per_second": 2.502,
2093
+ "eval_steps_per_second": 0.313,
2094
+ "step": 139
2095
+ },
2096
+ {
2097
+ "epoch": 4.375,
2098
+ "grad_norm": 0.34411051658527964,
2099
+ "learning_rate": 2e-05,
2100
+ "loss": 0.7465,
2101
+ "step": 140
2102
+ },
2103
+ {
2104
+ "epoch": 4.375,
2105
+ "eval_loss": 0.6915809512138367,
2106
+ "eval_runtime": 79.943,
2107
+ "eval_samples_per_second": 2.502,
2108
+ "eval_steps_per_second": 0.313,
2109
+ "step": 140
2110
+ },
2111
+ {
2112
+ "epoch": 4.40625,
2113
+ "grad_norm": 0.3373909601921749,
2114
+ "learning_rate": 2e-05,
2115
+ "loss": 0.7648,
2116
+ "step": 141
2117
+ },
2118
+ {
2119
+ "epoch": 4.40625,
2120
+ "eval_loss": 0.6912103295326233,
2121
+ "eval_runtime": 79.8515,
2122
+ "eval_samples_per_second": 2.505,
2123
+ "eval_steps_per_second": 0.313,
2124
+ "step": 141
2125
+ },
2126
+ {
2127
+ "epoch": 4.4375,
2128
+ "grad_norm": 0.33253827371305456,
2129
+ "learning_rate": 2e-05,
2130
+ "loss": 0.7224,
2131
+ "step": 142
2132
+ },
2133
+ {
2134
+ "epoch": 4.4375,
2135
+ "eval_loss": 0.6912806630134583,
2136
+ "eval_runtime": 80.6475,
2137
+ "eval_samples_per_second": 2.48,
2138
+ "eval_steps_per_second": 0.31,
2139
+ "step": 142
2140
+ },
2141
+ {
2142
+ "epoch": 4.46875,
2143
+ "grad_norm": 0.38458075172588313,
2144
+ "learning_rate": 2e-05,
2145
+ "loss": 0.7261,
2146
+ "step": 143
2147
+ },
2148
+ {
2149
+ "epoch": 4.46875,
2150
+ "eval_loss": 0.6905419230461121,
2151
+ "eval_runtime": 80.2606,
2152
+ "eval_samples_per_second": 2.492,
2153
+ "eval_steps_per_second": 0.311,
2154
+ "step": 143
2155
+ },
2156
+ {
2157
+ "epoch": 4.5,
2158
+ "grad_norm": 0.31351962640463144,
2159
+ "learning_rate": 2e-05,
2160
+ "loss": 0.6909,
2161
+ "step": 144
2162
+ },
2163
+ {
2164
+ "epoch": 4.5,
2165
+ "eval_loss": 0.6898491382598877,
2166
+ "eval_runtime": 79.9965,
2167
+ "eval_samples_per_second": 2.5,
2168
+ "eval_steps_per_second": 0.313,
2169
+ "step": 144
2170
+ },
2171
+ {
2172
+ "epoch": 4.53125,
2173
+ "grad_norm": 0.35474372115704583,
2174
+ "learning_rate": 2e-05,
2175
+ "loss": 0.7605,
2176
+ "step": 145
2177
+ },
2178
+ {
2179
+ "epoch": 4.53125,
2180
+ "eval_loss": 0.6893147230148315,
2181
+ "eval_runtime": 1475.5758,
2182
+ "eval_samples_per_second": 0.136,
2183
+ "eval_steps_per_second": 0.017,
2184
+ "step": 145
2185
+ },
2186
+ {
2187
+ "epoch": 4.5625,
2188
+ "grad_norm": 0.3479568917421202,
2189
+ "learning_rate": 2e-05,
2190
+ "loss": 0.6638,
2191
+ "step": 146
2192
+ },
2193
+ {
2194
+ "epoch": 4.5625,
2195
+ "eval_loss": 0.6884538531303406,
2196
+ "eval_runtime": 84.6835,
2197
+ "eval_samples_per_second": 2.362,
2198
+ "eval_steps_per_second": 0.295,
2199
+ "step": 146
2200
+ },
2201
+ {
2202
+ "epoch": 4.59375,
2203
+ "grad_norm": 0.3421823344428645,
2204
+ "learning_rate": 2e-05,
2205
+ "loss": 0.7339,
2206
+ "step": 147
2207
+ },
2208
+ {
2209
+ "epoch": 4.59375,
2210
+ "eval_loss": 0.6873475909233093,
2211
+ "eval_runtime": 83.3138,
2212
+ "eval_samples_per_second": 2.401,
2213
+ "eval_steps_per_second": 0.3,
2214
+ "step": 147
2215
+ },
2216
+ {
2217
+ "epoch": 4.625,
2218
+ "grad_norm": 0.3642187020830788,
2219
+ "learning_rate": 2e-05,
2220
+ "loss": 0.6825,
2221
+ "step": 148
2222
+ },
2223
+ {
2224
+ "epoch": 4.625,
2225
+ "eval_loss": 0.6858401298522949,
2226
+ "eval_runtime": 82.1066,
2227
+ "eval_samples_per_second": 2.436,
2228
+ "eval_steps_per_second": 0.304,
2229
+ "step": 148
2230
+ },
2231
+ {
2232
+ "epoch": 4.65625,
2233
+ "grad_norm": 0.35097547901391785,
2234
+ "learning_rate": 2e-05,
2235
+ "loss": 0.7986,
2236
+ "step": 149
2237
+ },
2238
+ {
2239
+ "epoch": 4.65625,
2240
+ "eval_loss": 0.6848779320716858,
2241
+ "eval_runtime": 84.4076,
2242
+ "eval_samples_per_second": 2.369,
2243
+ "eval_steps_per_second": 0.296,
2244
+ "step": 149
2245
+ },
2246
+ {
2247
+ "epoch": 4.6875,
2248
+ "grad_norm": 0.3568694843794629,
2249
+ "learning_rate": 2e-05,
2250
+ "loss": 0.7176,
2251
+ "step": 150
2252
+ },
2253
+ {
2254
+ "epoch": 4.6875,
2255
+ "eval_loss": 0.6842290759086609,
2256
+ "eval_runtime": 82.5945,
2257
+ "eval_samples_per_second": 2.421,
2258
+ "eval_steps_per_second": 0.303,
2259
+ "step": 150
2260
+ },
2261
+ {
2262
+ "epoch": 4.71875,
2263
+ "grad_norm": 0.34258633585260334,
2264
+ "learning_rate": 2e-05,
2265
+ "loss": 0.7363,
2266
+ "step": 151
2267
+ },
2268
+ {
2269
+ "epoch": 4.71875,
2270
+ "eval_loss": 0.6838659048080444,
2271
+ "eval_runtime": 85.9626,
2272
+ "eval_samples_per_second": 2.327,
2273
+ "eval_steps_per_second": 0.291,
2274
+ "step": 151
2275
+ },
2276
+ {
2277
+ "epoch": 4.75,
2278
+ "grad_norm": 0.42319523894659655,
2279
+ "learning_rate": 2e-05,
2280
+ "loss": 0.7675,
2281
+ "step": 152
2282
+ },
2283
+ {
2284
+ "epoch": 4.75,
2285
+ "eval_loss": 0.6830299496650696,
2286
+ "eval_runtime": 85.7189,
2287
+ "eval_samples_per_second": 2.333,
2288
+ "eval_steps_per_second": 0.292,
2289
+ "step": 152
2290
+ },
2291
+ {
2292
+ "epoch": 4.78125,
2293
+ "grad_norm": 0.3632195533127194,
2294
+ "learning_rate": 2e-05,
2295
+ "loss": 0.715,
2296
+ "step": 153
2297
+ },
2298
+ {
2299
+ "epoch": 4.78125,
2300
+ "eval_loss": 0.6826379895210266,
2301
+ "eval_runtime": 87.8244,
2302
+ "eval_samples_per_second": 2.277,
2303
+ "eval_steps_per_second": 0.285,
2304
+ "step": 153
2305
+ },
2306
+ {
2307
+ "epoch": 4.8125,
2308
+ "grad_norm": 0.3738308004604413,
2309
+ "learning_rate": 2e-05,
2310
+ "loss": 0.7344,
2311
+ "step": 154
2312
+ },
2313
+ {
2314
+ "epoch": 4.8125,
2315
+ "eval_loss": 0.6826817393302917,
2316
+ "eval_runtime": 86.5822,
2317
+ "eval_samples_per_second": 2.31,
2318
+ "eval_steps_per_second": 0.289,
2319
+ "step": 154
2320
+ },
2321
+ {
2322
+ "epoch": 4.84375,
2323
+ "grad_norm": 0.3618696330632776,
2324
+ "learning_rate": 2e-05,
2325
+ "loss": 0.6632,
2326
+ "step": 155
2327
+ },
2328
+ {
2329
+ "epoch": 4.84375,
2330
+ "eval_loss": 0.6827967166900635,
2331
+ "eval_runtime": 82.1829,
2332
+ "eval_samples_per_second": 2.434,
2333
+ "eval_steps_per_second": 0.304,
2334
+ "step": 155
2335
+ },
2336
+ {
2337
+ "epoch": 4.875,
2338
+ "grad_norm": 0.38901912569992203,
2339
+ "learning_rate": 2e-05,
2340
+ "loss": 0.7788,
2341
+ "step": 156
2342
+ },
2343
+ {
2344
+ "epoch": 4.875,
2345
+ "eval_loss": 0.6821711659431458,
2346
+ "eval_runtime": 84.4511,
2347
+ "eval_samples_per_second": 2.368,
2348
+ "eval_steps_per_second": 0.296,
2349
+ "step": 156
2350
+ },
2351
+ {
2352
+ "epoch": 4.90625,
2353
+ "grad_norm": 0.3516096507348829,
2354
+ "learning_rate": 2e-05,
2355
+ "loss": 0.7794,
2356
+ "step": 157
2357
+ },
2358
+ {
2359
+ "epoch": 4.90625,
2360
+ "eval_loss": 0.6819837689399719,
2361
+ "eval_runtime": 84.1594,
2362
+ "eval_samples_per_second": 2.376,
2363
+ "eval_steps_per_second": 0.297,
2364
+ "step": 157
2365
+ },
2366
+ {
2367
+ "epoch": 4.9375,
2368
+ "grad_norm": 0.36066902463794986,
2369
+ "learning_rate": 2e-05,
2370
+ "loss": 0.7674,
2371
+ "step": 158
2372
+ },
2373
+ {
2374
+ "epoch": 4.9375,
2375
+ "eval_loss": 0.6817716956138611,
2376
+ "eval_runtime": 83.8929,
2377
+ "eval_samples_per_second": 2.384,
2378
+ "eval_steps_per_second": 0.298,
2379
+ "step": 158
2380
+ },
2381
+ {
2382
+ "epoch": 4.96875,
2383
+ "grad_norm": 0.36641784926154175,
2384
+ "learning_rate": 2e-05,
2385
+ "loss": 0.7116,
2386
+ "step": 159
2387
+ },
2388
+ {
2389
+ "epoch": 4.96875,
2390
+ "eval_loss": 0.6816902160644531,
2391
+ "eval_runtime": 84.4431,
2392
+ "eval_samples_per_second": 2.368,
2393
+ "eval_steps_per_second": 0.296,
2394
+ "step": 159
2395
+ },
2396
+ {
2397
+ "epoch": 5.0,
2398
+ "grad_norm": 0.4020716293225933,
2399
+ "learning_rate": 2e-05,
2400
+ "loss": 0.7142,
2401
+ "step": 160
2402
+ },
2403
+ {
2404
+ "epoch": 5.0,
2405
+ "eval_loss": 0.6811469793319702,
2406
+ "eval_runtime": 86.0681,
2407
+ "eval_samples_per_second": 2.324,
2408
+ "eval_steps_per_second": 0.29,
2409
+ "step": 160
2410
+ },
2411
+ {
2412
+ "epoch": 5.03125,
2413
+ "grad_norm": 0.38360882669254054,
2414
+ "learning_rate": 2e-05,
2415
+ "loss": 0.6756,
2416
+ "step": 161
2417
+ },
2418
+ {
2419
+ "epoch": 5.03125,
2420
+ "eval_loss": 0.6798409223556519,
2421
+ "eval_runtime": 81.9903,
2422
+ "eval_samples_per_second": 2.439,
2423
+ "eval_steps_per_second": 0.305,
2424
+ "step": 161
2425
+ },
2426
+ {
2427
+ "epoch": 5.0625,
2428
+ "grad_norm": 0.34966156213066135,
2429
+ "learning_rate": 2e-05,
2430
+ "loss": 0.827,
2431
+ "step": 162
2432
+ },
2433
+ {
2434
+ "epoch": 5.0625,
2435
+ "eval_loss": 0.6788859367370605,
2436
+ "eval_runtime": 76.1753,
2437
+ "eval_samples_per_second": 2.626,
2438
+ "eval_steps_per_second": 0.328,
2439
+ "step": 162
2440
+ },
2441
+ {
2442
+ "epoch": 5.09375,
2443
+ "grad_norm": 0.41140842939901384,
2444
+ "learning_rate": 2e-05,
2445
+ "loss": 0.6409,
2446
+ "step": 163
2447
+ },
2448
+ {
2449
+ "epoch": 5.09375,
2450
+ "eval_loss": 0.6787077188491821,
2451
+ "eval_runtime": 76.2239,
2452
+ "eval_samples_per_second": 2.624,
2453
+ "eval_steps_per_second": 0.328,
2454
+ "step": 163
2455
+ },
2456
+ {
2457
+ "epoch": 5.125,
2458
+ "grad_norm": 0.4222084070163774,
2459
+ "learning_rate": 2e-05,
2460
+ "loss": 0.7774,
2461
+ "step": 164
2462
+ },
2463
+ {
2464
+ "epoch": 5.125,
2465
+ "eval_loss": 0.6796822547912598,
2466
+ "eval_runtime": 76.2141,
2467
+ "eval_samples_per_second": 2.624,
2468
+ "eval_steps_per_second": 0.328,
2469
+ "step": 164
2470
+ },
2471
+ {
2472
+ "epoch": 5.15625,
2473
+ "grad_norm": 0.4644454724424921,
2474
+ "learning_rate": 2e-05,
2475
+ "loss": 0.6057,
2476
+ "step": 165
2477
+ },
2478
+ {
2479
+ "epoch": 5.15625,
2480
+ "eval_loss": 0.6794346570968628,
2481
+ "eval_runtime": 76.3216,
2482
+ "eval_samples_per_second": 2.62,
2483
+ "eval_steps_per_second": 0.328,
2484
+ "step": 165
2485
+ },
2486
+ {
2487
+ "epoch": 5.1875,
2488
+ "grad_norm": 0.46128725263272996,
2489
+ "learning_rate": 2e-05,
2490
+ "loss": 0.7158,
2491
+ "step": 166
2492
+ },
2493
+ {
2494
+ "epoch": 5.1875,
2495
+ "eval_loss": 0.6791612505912781,
2496
+ "eval_runtime": 78.4909,
2497
+ "eval_samples_per_second": 2.548,
2498
+ "eval_steps_per_second": 0.319,
2499
+ "step": 166
2500
+ },
2501
+ {
2502
+ "epoch": 5.21875,
2503
+ "grad_norm": 0.37300666872025545,
2504
+ "learning_rate": 2e-05,
2505
+ "loss": 0.7363,
2506
+ "step": 167
2507
+ },
2508
+ {
2509
+ "epoch": 5.21875,
2510
+ "eval_loss": 0.6788016557693481,
2511
+ "eval_runtime": 78.5697,
2512
+ "eval_samples_per_second": 2.546,
2513
+ "eval_steps_per_second": 0.318,
2514
+ "step": 167
2515
+ },
2516
+ {
2517
+ "epoch": 5.25,
2518
+ "grad_norm": 0.41454648576180214,
2519
+ "learning_rate": 2e-05,
2520
+ "loss": 0.7759,
2521
+ "step": 168
2522
+ },
2523
+ {
2524
+ "epoch": 5.25,
2525
+ "eval_loss": 0.6787048578262329,
2526
+ "eval_runtime": 78.5317,
2527
+ "eval_samples_per_second": 2.547,
2528
+ "eval_steps_per_second": 0.318,
2529
+ "step": 168
2530
+ },
2531
+ {
2532
+ "epoch": 5.28125,
2533
+ "grad_norm": 0.40724665091386236,
2534
+ "learning_rate": 2e-05,
2535
+ "loss": 0.6944,
2536
+ "step": 169
2537
+ },
2538
+ {
2539
+ "epoch": 5.28125,
2540
+ "eval_loss": 0.679679811000824,
2541
+ "eval_runtime": 78.6899,
2542
+ "eval_samples_per_second": 2.542,
2543
+ "eval_steps_per_second": 0.318,
2544
+ "step": 169
2545
+ },
2546
+ {
2547
+ "epoch": 5.3125,
2548
+ "grad_norm": 0.3875110486208986,
2549
+ "learning_rate": 2e-05,
2550
+ "loss": 0.6634,
2551
+ "step": 170
2552
+ },
2553
+ {
2554
+ "epoch": 5.3125,
2555
+ "eval_loss": 0.6819935441017151,
2556
+ "eval_runtime": 78.3617,
2557
+ "eval_samples_per_second": 2.552,
2558
+ "eval_steps_per_second": 0.319,
2559
+ "step": 170
2560
+ },
2561
+ {
2562
+ "epoch": 5.34375,
2563
+ "grad_norm": 0.47956532155617193,
2564
+ "learning_rate": 2e-05,
2565
+ "loss": 0.687,
2566
+ "step": 171
2567
+ },
2568
+ {
2569
+ "epoch": 5.34375,
2570
+ "eval_loss": 0.6825206875801086,
2571
+ "eval_runtime": 78.4435,
2572
+ "eval_samples_per_second": 2.55,
2573
+ "eval_steps_per_second": 0.319,
2574
+ "step": 171
2575
+ },
2576
+ {
2577
+ "epoch": 5.375,
2578
+ "grad_norm": 0.4599359590587781,
2579
+ "learning_rate": 2e-05,
2580
+ "loss": 0.7718,
2581
+ "step": 172
2582
+ },
2583
+ {
2584
+ "epoch": 5.375,
2585
+ "eval_loss": 0.6816768050193787,
2586
+ "eval_runtime": 78.3005,
2587
+ "eval_samples_per_second": 2.554,
2588
+ "eval_steps_per_second": 0.319,
2589
+ "step": 172
2590
+ },
2591
+ {
2592
+ "epoch": 5.40625,
2593
+ "grad_norm": 0.4057490487995386,
2594
+ "learning_rate": 2e-05,
2595
+ "loss": 0.7292,
2596
+ "step": 173
2597
+ },
2598
+ {
2599
+ "epoch": 5.40625,
2600
+ "eval_loss": 0.6806090474128723,
2601
+ "eval_runtime": 78.3313,
2602
+ "eval_samples_per_second": 2.553,
2603
+ "eval_steps_per_second": 0.319,
2604
+ "step": 173
2605
+ },
2606
+ {
2607
+ "epoch": 5.4375,
2608
+ "grad_norm": 0.4143979315360467,
2609
+ "learning_rate": 2e-05,
2610
+ "loss": 0.7697,
2611
+ "step": 174
2612
+ },
2613
+ {
2614
+ "epoch": 5.4375,
2615
+ "eval_loss": 0.6795693039894104,
2616
+ "eval_runtime": 78.4526,
2617
+ "eval_samples_per_second": 2.549,
2618
+ "eval_steps_per_second": 0.319,
2619
+ "step": 174
2620
+ },
2621
+ {
2622
+ "epoch": 5.46875,
2623
+ "grad_norm": 0.4219663662343445,
2624
+ "learning_rate": 2e-05,
2625
+ "loss": 0.7534,
2626
+ "step": 175
2627
+ },
2628
+ {
2629
+ "epoch": 5.46875,
2630
+ "eval_loss": 0.6793847680091858,
2631
+ "eval_runtime": 78.8009,
2632
+ "eval_samples_per_second": 2.538,
2633
+ "eval_steps_per_second": 0.317,
2634
+ "step": 175
2635
+ },
2636
+ {
2637
+ "epoch": 5.5,
2638
+ "grad_norm": 0.4491811321927657,
2639
+ "learning_rate": 2e-05,
2640
+ "loss": 0.7004,
2641
+ "step": 176
2642
+ },
2643
+ {
2644
+ "epoch": 5.5,
2645
+ "eval_loss": 0.6775352358818054,
2646
+ "eval_runtime": 80.0685,
2647
+ "eval_samples_per_second": 2.498,
2648
+ "eval_steps_per_second": 0.312,
2649
+ "step": 176
2650
+ },
2651
+ {
2652
+ "epoch": 5.53125,
2653
+ "grad_norm": 0.46366516532638885,
2654
+ "learning_rate": 2e-05,
2655
+ "loss": 0.7357,
2656
+ "step": 177
2657
+ },
2658
+ {
2659
+ "epoch": 5.53125,
2660
+ "eval_loss": 0.6748698949813843,
2661
+ "eval_runtime": 80.0487,
2662
+ "eval_samples_per_second": 2.498,
2663
+ "eval_steps_per_second": 0.312,
2664
+ "step": 177
2665
+ },
2666
+ {
2667
+ "epoch": 5.5625,
2668
+ "grad_norm": 0.3815188640227797,
2669
+ "learning_rate": 2e-05,
2670
+ "loss": 0.7592,
2671
+ "step": 178
2672
+ },
2673
+ {
2674
+ "epoch": 5.5625,
2675
+ "eval_loss": 0.6728273034095764,
2676
+ "eval_runtime": 80.0318,
2677
+ "eval_samples_per_second": 2.499,
2678
+ "eval_steps_per_second": 0.312,
2679
+ "step": 178
2680
+ },
2681
+ {
2682
+ "epoch": 5.59375,
2683
+ "grad_norm": 0.41025429416666304,
2684
+ "learning_rate": 2e-05,
2685
+ "loss": 0.6585,
2686
+ "step": 179
2687
+ },
2688
+ {
2689
+ "epoch": 5.59375,
2690
+ "eval_loss": 0.6718859672546387,
2691
+ "eval_runtime": 79.8801,
2692
+ "eval_samples_per_second": 2.504,
2693
+ "eval_steps_per_second": 0.313,
2694
+ "step": 179
2695
+ },
2696
+ {
2697
+ "epoch": 5.625,
2698
+ "grad_norm": 0.40652817592240054,
2699
+ "learning_rate": 2e-05,
2700
+ "loss": 0.6611,
2701
+ "step": 180
2702
+ },
2703
+ {
2704
+ "epoch": 5.625,
2705
+ "eval_loss": 0.6715708374977112,
2706
+ "eval_runtime": 76.7261,
2707
+ "eval_samples_per_second": 2.607,
2708
+ "eval_steps_per_second": 0.326,
2709
+ "step": 180
2710
+ },
2711
+ {
2712
+ "epoch": 5.65625,
2713
+ "grad_norm": 0.40753961326688415,
2714
+ "learning_rate": 2e-05,
2715
+ "loss": 0.6779,
2716
+ "step": 181
2717
+ },
2718
+ {
2719
+ "epoch": 5.65625,
2720
+ "eval_loss": 0.6719761490821838,
2721
+ "eval_runtime": 77.0136,
2722
+ "eval_samples_per_second": 2.597,
2723
+ "eval_steps_per_second": 0.325,
2724
+ "step": 181
2725
+ },
2726
+ {
2727
+ "epoch": 5.6875,
2728
+ "grad_norm": 0.4232811980671673,
2729
+ "learning_rate": 2e-05,
2730
+ "loss": 0.6475,
2731
+ "step": 182
2732
+ },
2733
+ {
2734
+ "epoch": 5.6875,
2735
+ "eval_loss": 0.6724664568901062,
2736
+ "eval_runtime": 76.9731,
2737
+ "eval_samples_per_second": 2.598,
2738
+ "eval_steps_per_second": 0.325,
2739
+ "step": 182
2740
+ },
2741
+ {
2742
+ "epoch": 5.71875,
2743
+ "grad_norm": 0.5132756318549849,
2744
+ "learning_rate": 2e-05,
2745
+ "loss": 0.6801,
2746
+ "step": 183
2747
+ },
2748
+ {
2749
+ "epoch": 5.71875,
2750
+ "eval_loss": 0.6723365783691406,
2751
+ "eval_runtime": 76.4132,
2752
+ "eval_samples_per_second": 2.617,
2753
+ "eval_steps_per_second": 0.327,
2754
+ "step": 183
2755
+ },
2756
+ {
2757
+ "epoch": 5.75,
2758
+ "grad_norm": 0.43526879230161264,
2759
+ "learning_rate": 2e-05,
2760
+ "loss": 0.6673,
2761
+ "step": 184
2762
+ },
2763
+ {
2764
+ "epoch": 5.75,
2765
+ "eval_loss": 0.672926664352417,
2766
+ "eval_runtime": 76.1936,
2767
+ "eval_samples_per_second": 2.625,
2768
+ "eval_steps_per_second": 0.328,
2769
+ "step": 184
2770
+ },
2771
+ {
2772
+ "epoch": 5.78125,
2773
+ "grad_norm": 0.46965560853038507,
2774
+ "learning_rate": 2e-05,
2775
+ "loss": 0.7074,
2776
+ "step": 185
2777
+ },
2778
+ {
2779
+ "epoch": 5.78125,
2780
+ "eval_loss": 0.6731134057044983,
2781
+ "eval_runtime": 76.2345,
2782
+ "eval_samples_per_second": 2.623,
2783
+ "eval_steps_per_second": 0.328,
2784
+ "step": 185
2785
+ },
2786
+ {
2787
+ "epoch": 5.8125,
2788
+ "grad_norm": 0.4733296318676217,
2789
+ "learning_rate": 2e-05,
2790
+ "loss": 0.6791,
2791
+ "step": 186
2792
+ },
2793
+ {
2794
+ "epoch": 5.8125,
2795
+ "eval_loss": 0.6726363301277161,
2796
+ "eval_runtime": 78.3939,
2797
+ "eval_samples_per_second": 2.551,
2798
+ "eval_steps_per_second": 0.319,
2799
+ "step": 186
2800
+ },
2801
+ {
2802
+ "epoch": 5.84375,
2803
+ "grad_norm": 0.4662943253655961,
2804
+ "learning_rate": 2e-05,
2805
+ "loss": 0.7371,
2806
+ "step": 187
2807
+ },
2808
+ {
2809
+ "epoch": 5.84375,
2810
+ "eval_loss": 0.6726526021957397,
2811
+ "eval_runtime": 79.1834,
2812
+ "eval_samples_per_second": 2.526,
2813
+ "eval_steps_per_second": 0.316,
2814
+ "step": 187
2815
+ },
2816
+ {
2817
+ "epoch": 5.875,
2818
+ "grad_norm": 0.4420962889993382,
2819
+ "learning_rate": 2e-05,
2820
+ "loss": 0.675,
2821
+ "step": 188
2822
+ },
2823
+ {
2824
+ "epoch": 5.875,
2825
+ "eval_loss": 0.6727125644683838,
2826
+ "eval_runtime": 78.252,
2827
+ "eval_samples_per_second": 2.556,
2828
+ "eval_steps_per_second": 0.319,
2829
+ "step": 188
2830
+ },
2831
+ {
2832
+ "epoch": 5.90625,
2833
+ "grad_norm": 0.4345166976944551,
2834
+ "learning_rate": 2e-05,
2835
+ "loss": 0.6748,
2836
+ "step": 189
2837
+ },
2838
+ {
2839
+ "epoch": 5.90625,
2840
+ "eval_loss": 0.6725904941558838,
2841
+ "eval_runtime": 78.3914,
2842
+ "eval_samples_per_second": 2.551,
2843
+ "eval_steps_per_second": 0.319,
2844
+ "step": 189
2845
+ },
2846
+ {
2847
+ "epoch": 5.9375,
2848
+ "grad_norm": 0.45109463315374526,
2849
+ "learning_rate": 2e-05,
2850
+ "loss": 0.7024,
2851
+ "step": 190
2852
+ },
2853
+ {
2854
+ "epoch": 5.9375,
2855
+ "eval_loss": 0.6718384027481079,
2856
+ "eval_runtime": 78.4361,
2857
+ "eval_samples_per_second": 2.55,
2858
+ "eval_steps_per_second": 0.319,
2859
+ "step": 190
2860
+ },
2861
+ {
2862
+ "epoch": 5.96875,
2863
+ "grad_norm": 0.42953871838795626,
2864
+ "learning_rate": 2e-05,
2865
+ "loss": 0.6904,
2866
+ "step": 191
2867
+ },
2868
+ {
2869
+ "epoch": 5.96875,
2870
+ "eval_loss": 0.6703083515167236,
2871
+ "eval_runtime": 78.3863,
2872
+ "eval_samples_per_second": 2.551,
2873
+ "eval_steps_per_second": 0.319,
2874
+ "step": 191
2875
+ },
2876
+ {
2877
+ "epoch": 6.0,
2878
+ "grad_norm": 0.4248607379284984,
2879
+ "learning_rate": 2e-05,
2880
+ "loss": 0.6659,
2881
+ "step": 192
2882
+ },
2883
+ {
2884
+ "epoch": 6.0,
2885
+ "eval_loss": 0.6693080067634583,
2886
+ "eval_runtime": 78.4373,
2887
+ "eval_samples_per_second": 2.55,
2888
+ "eval_steps_per_second": 0.319,
2889
+ "step": 192
2890
+ },
2891
+ {
2892
+ "epoch": 6.03125,
2893
+ "grad_norm": 0.42839417453459494,
2894
+ "learning_rate": 2e-05,
2895
+ "loss": 0.7457,
2896
+ "step": 193
2897
+ },
2898
+ {
2899
+ "epoch": 6.03125,
2900
+ "eval_loss": 0.6689594984054565,
2901
+ "eval_runtime": 78.4169,
2902
+ "eval_samples_per_second": 2.55,
2903
+ "eval_steps_per_second": 0.319,
2904
+ "step": 193
2905
+ },
2906
+ {
2907
+ "epoch": 6.0625,
2908
+ "grad_norm": 0.4216922788166874,
2909
+ "learning_rate": 2e-05,
2910
+ "loss": 0.7189,
2911
+ "step": 194
2912
+ },
2913
+ {
2914
+ "epoch": 6.0625,
2915
+ "eval_loss": 0.6689300537109375,
2916
+ "eval_runtime": 78.9793,
2917
+ "eval_samples_per_second": 2.532,
2918
+ "eval_steps_per_second": 0.317,
2919
+ "step": 194
2920
+ },
2921
+ {
2922
+ "epoch": 6.09375,
2923
+ "grad_norm": 0.45199575791858004,
2924
+ "learning_rate": 2e-05,
2925
+ "loss": 0.6438,
2926
+ "step": 195
2927
+ },
2928
+ {
2929
+ "epoch": 6.09375,
2930
+ "eval_loss": 0.6690151691436768,
2931
+ "eval_runtime": 78.5002,
2932
+ "eval_samples_per_second": 2.548,
2933
+ "eval_steps_per_second": 0.318,
2934
+ "step": 195
2935
+ },
2936
+ {
2937
+ "epoch": 6.125,
2938
+ "grad_norm": 0.4166923177293841,
2939
+ "learning_rate": 2e-05,
2940
+ "loss": 0.6885,
2941
+ "step": 196
2942
+ },
2943
+ {
2944
+ "epoch": 6.125,
2945
+ "eval_loss": 0.6688613891601562,
2946
+ "eval_runtime": 80.5497,
2947
+ "eval_samples_per_second": 2.483,
2948
+ "eval_steps_per_second": 0.31,
2949
+ "step": 196
2950
+ },
2951
+ {
2952
+ "epoch": 6.15625,
2953
+ "grad_norm": 0.45164281863366285,
2954
+ "learning_rate": 2e-05,
2955
+ "loss": 0.7197,
2956
+ "step": 197
2957
+ },
2958
+ {
2959
+ "epoch": 6.15625,
2960
+ "eval_loss": 0.6687932014465332,
2961
+ "eval_runtime": 80.1482,
2962
+ "eval_samples_per_second": 2.495,
2963
+ "eval_steps_per_second": 0.312,
2964
+ "step": 197
2965
+ },
2966
+ {
2967
+ "epoch": 6.1875,
2968
+ "grad_norm": 0.45653924787504446,
2969
+ "learning_rate": 2e-05,
2970
+ "loss": 0.776,
2971
+ "step": 198
2972
+ },
2973
+ {
2974
+ "epoch": 6.1875,
2975
+ "eval_loss": 0.6690963506698608,
2976
+ "eval_runtime": 80.4464,
2977
+ "eval_samples_per_second": 2.486,
2978
+ "eval_steps_per_second": 0.311,
2979
+ "step": 198
2980
+ },
2981
+ {
2982
+ "epoch": 6.21875,
2983
+ "grad_norm": 0.4966562341334706,
2984
+ "learning_rate": 2e-05,
2985
+ "loss": 0.6532,
2986
+ "step": 199
2987
+ },
2988
+ {
2989
+ "epoch": 6.21875,
2990
+ "eval_loss": 0.669116735458374,
2991
+ "eval_runtime": 79.8294,
2992
+ "eval_samples_per_second": 2.505,
2993
+ "eval_steps_per_second": 0.313,
2994
+ "step": 199
2995
+ },
2996
+ {
2997
+ "epoch": 6.25,
2998
+ "grad_norm": 0.4838469303220975,
2999
+ "learning_rate": 2e-05,
3000
+ "loss": 0.6883,
3001
+ "step": 200
3002
+ },
3003
+ {
3004
+ "epoch": 6.25,
3005
+ "eval_loss": 0.6693156957626343,
3006
+ "eval_runtime": 80.25,
3007
+ "eval_samples_per_second": 2.492,
3008
+ "eval_steps_per_second": 0.312,
3009
+ "step": 200
3010
+ },
3011
+ {
3012
+ "epoch": 6.28125,
3013
+ "grad_norm": 0.4836820906895964,
3014
+ "learning_rate": 2e-05,
3015
+ "loss": 0.7106,
3016
+ "step": 201
3017
+ },
3018
+ {
3019
+ "epoch": 6.28125,
3020
+ "eval_loss": 0.6704170107841492,
3021
+ "eval_runtime": 79.9636,
3022
+ "eval_samples_per_second": 2.501,
3023
+ "eval_steps_per_second": 0.313,
3024
+ "step": 201
3025
+ },
3026
+ {
3027
+ "epoch": 6.3125,
3028
+ "grad_norm": 0.4945855983140219,
3029
+ "learning_rate": 2e-05,
3030
+ "loss": 0.6336,
3031
+ "step": 202
3032
+ },
3033
+ {
3034
+ "epoch": 6.3125,
3035
+ "eval_loss": 0.6708824038505554,
3036
+ "eval_runtime": 80.8044,
3037
+ "eval_samples_per_second": 2.475,
3038
+ "eval_steps_per_second": 0.309,
3039
+ "step": 202
3040
+ },
3041
+ {
3042
+ "epoch": 6.34375,
3043
+ "grad_norm": 0.44587847230103017,
3044
+ "learning_rate": 2e-05,
3045
+ "loss": 0.7811,
3046
+ "step": 203
3047
+ },
3048
+ {
3049
+ "epoch": 6.34375,
3050
+ "eval_loss": 0.6723968982696533,
3051
+ "eval_runtime": 80.1715,
3052
+ "eval_samples_per_second": 2.495,
3053
+ "eval_steps_per_second": 0.312,
3054
+ "step": 203
3055
+ },
3056
+ {
3057
+ "epoch": 6.375,
3058
+ "grad_norm": 0.5351063503195825,
3059
+ "learning_rate": 2e-05,
3060
+ "loss": 0.6222,
3061
+ "step": 204
3062
+ },
3063
+ {
3064
+ "epoch": 6.375,
3065
+ "eval_loss": 0.672196626663208,
3066
+ "eval_runtime": 79.927,
3067
+ "eval_samples_per_second": 2.502,
3068
+ "eval_steps_per_second": 0.313,
3069
+ "step": 204
3070
+ },
3071
+ {
3072
+ "epoch": 6.40625,
3073
+ "grad_norm": 0.4742985088010474,
3074
+ "learning_rate": 2e-05,
3075
+ "loss": 0.6157,
3076
+ "step": 205
3077
+ },
3078
+ {
3079
+ "epoch": 6.40625,
3080
+ "eval_loss": 0.671062171459198,
3081
+ "eval_runtime": 80.1997,
3082
+ "eval_samples_per_second": 2.494,
3083
+ "eval_steps_per_second": 0.312,
3084
+ "step": 205
3085
+ },
3086
+ {
3087
+ "epoch": 6.4375,
3088
+ "grad_norm": 0.5188882333349506,
3089
+ "learning_rate": 2e-05,
3090
+ "loss": 0.6462,
3091
+ "step": 206
3092
+ },
3093
+ {
3094
+ "epoch": 6.4375,
3095
+ "eval_loss": 0.6701972484588623,
3096
+ "eval_runtime": 81.6643,
3097
+ "eval_samples_per_second": 2.449,
3098
+ "eval_steps_per_second": 0.306,
3099
+ "step": 206
3100
+ },
3101
+ {
3102
+ "epoch": 6.46875,
3103
+ "grad_norm": 0.45328063593983603,
3104
+ "learning_rate": 2e-05,
3105
+ "loss": 0.7058,
3106
+ "step": 207
3107
+ },
3108
+ {
3109
+ "epoch": 6.46875,
3110
+ "eval_loss": 0.6699164509773254,
3111
+ "eval_runtime": 81.2228,
3112
+ "eval_samples_per_second": 2.462,
3113
+ "eval_steps_per_second": 0.308,
3114
+ "step": 207
3115
+ },
3116
+ {
3117
+ "epoch": 6.5,
3118
+ "grad_norm": 0.5197645538332801,
3119
+ "learning_rate": 2e-05,
3120
+ "loss": 0.6462,
3121
+ "step": 208
3122
+ },
3123
+ {
3124
+ "epoch": 6.5,
3125
+ "eval_loss": 0.6702597141265869,
3126
+ "eval_runtime": 81.1451,
3127
+ "eval_samples_per_second": 2.465,
3128
+ "eval_steps_per_second": 0.308,
3129
+ "step": 208
3130
+ },
3131
+ {
3132
+ "epoch": 6.53125,
3133
+ "grad_norm": 0.5762528184834232,
3134
+ "learning_rate": 2e-05,
3135
+ "loss": 0.6259,
3136
+ "step": 209
3137
+ },
3138
+ {
3139
+ "epoch": 6.53125,
3140
+ "eval_loss": 0.6696366667747498,
3141
+ "eval_runtime": 81.1643,
3142
+ "eval_samples_per_second": 2.464,
3143
+ "eval_steps_per_second": 0.308,
3144
+ "step": 209
3145
+ },
3146
+ {
3147
+ "epoch": 6.5625,
3148
+ "grad_norm": 0.5249503180293145,
3149
+ "learning_rate": 2e-05,
3150
+ "loss": 0.6045,
3151
+ "step": 210
3152
+ },
3153
+ {
3154
+ "epoch": 6.5625,
3155
+ "eval_loss": 0.6688054800033569,
3156
+ "eval_runtime": 80.9492,
3157
+ "eval_samples_per_second": 2.471,
3158
+ "eval_steps_per_second": 0.309,
3159
+ "step": 210
3160
+ },
3161
+ {
3162
+ "epoch": 6.59375,
3163
+ "grad_norm": 0.543503888655844,
3164
+ "learning_rate": 2e-05,
3165
+ "loss": 0.6496,
3166
+ "step": 211
3167
+ },
3168
+ {
3169
+ "epoch": 6.59375,
3170
+ "eval_loss": 0.6689916849136353,
3171
+ "eval_runtime": 81.6473,
3172
+ "eval_samples_per_second": 2.45,
3173
+ "eval_steps_per_second": 0.306,
3174
+ "step": 211
3175
+ },
3176
+ {
3177
+ "epoch": 6.625,
3178
+ "grad_norm": 0.48119553592193554,
3179
+ "learning_rate": 2e-05,
3180
+ "loss": 0.6211,
3181
+ "step": 212
3182
+ },
3183
+ {
3184
+ "epoch": 6.625,
3185
+ "eval_loss": 0.6703050136566162,
3186
+ "eval_runtime": 81.9207,
3187
+ "eval_samples_per_second": 2.441,
3188
+ "eval_steps_per_second": 0.305,
3189
+ "step": 212
3190
+ },
3191
+ {
3192
+ "epoch": 6.65625,
3193
+ "grad_norm": 0.5153356086819314,
3194
+ "learning_rate": 2e-05,
3195
+ "loss": 0.7135,
3196
+ "step": 213
3197
+ },
3198
+ {
3199
+ "epoch": 6.65625,
3200
+ "eval_loss": 0.6702842116355896,
3201
+ "eval_runtime": 81.1503,
3202
+ "eval_samples_per_second": 2.465,
3203
+ "eval_steps_per_second": 0.308,
3204
+ "step": 213
3205
+ },
3206
+ {
3207
+ "epoch": 6.6875,
3208
+ "grad_norm": 0.5249915042825578,
3209
+ "learning_rate": 2e-05,
3210
+ "loss": 0.6635,
3211
+ "step": 214
3212
+ },
3213
+ {
3214
+ "epoch": 6.6875,
3215
+ "eval_loss": 0.6687333583831787,
3216
+ "eval_runtime": 81.6743,
3217
+ "eval_samples_per_second": 2.449,
3218
+ "eval_steps_per_second": 0.306,
3219
+ "step": 214
3220
+ },
3221
+ {
3222
+ "epoch": 6.71875,
3223
+ "grad_norm": 0.5204840219868723,
3224
+ "learning_rate": 2e-05,
3225
+ "loss": 0.6701,
3226
+ "step": 215
3227
+ },
3228
+ {
3229
+ "epoch": 6.71875,
3230
+ "eval_loss": 0.6657728552818298,
3231
+ "eval_runtime": 81.106,
3232
+ "eval_samples_per_second": 2.466,
3233
+ "eval_steps_per_second": 0.308,
3234
+ "step": 215
3235
+ },
3236
+ {
3237
+ "epoch": 6.75,
3238
+ "grad_norm": 0.5266935225120133,
3239
+ "learning_rate": 2e-05,
3240
+ "loss": 0.6637,
3241
+ "step": 216
3242
+ },
3243
+ {
3244
+ "epoch": 6.75,
3245
+ "eval_loss": 0.6641908884048462,
3246
+ "eval_runtime": 82.2613,
3247
+ "eval_samples_per_second": 2.431,
3248
+ "eval_steps_per_second": 0.304,
3249
+ "step": 216
3250
+ },
3251
+ {
3252
+ "epoch": 6.78125,
3253
+ "grad_norm": 0.5438859451742696,
3254
+ "learning_rate": 2e-05,
3255
+ "loss": 0.6168,
3256
+ "step": 217
3257
+ },
3258
+ {
3259
+ "epoch": 6.78125,
3260
+ "eval_loss": 0.6652233600616455,
3261
+ "eval_runtime": 82.042,
3262
+ "eval_samples_per_second": 2.438,
3263
+ "eval_steps_per_second": 0.305,
3264
+ "step": 217
3265
+ },
3266
+ {
3267
+ "epoch": 6.8125,
3268
+ "grad_norm": 0.5716385253433929,
3269
+ "learning_rate": 2e-05,
3270
+ "loss": 0.6062,
3271
+ "step": 218
3272
+ },
3273
+ {
3274
+ "epoch": 6.8125,
3275
+ "eval_loss": 0.6656240820884705,
3276
+ "eval_runtime": 81.233,
3277
+ "eval_samples_per_second": 2.462,
3278
+ "eval_steps_per_second": 0.308,
3279
+ "step": 218
3280
+ },
3281
+ {
3282
+ "epoch": 6.84375,
3283
+ "grad_norm": 1.0572787630142522,
3284
+ "learning_rate": 2e-05,
3285
+ "loss": 0.7037,
3286
+ "step": 219
3287
+ },
3288
+ {
3289
+ "epoch": 6.84375,
3290
+ "eval_loss": 0.6645559072494507,
3291
+ "eval_runtime": 81.2099,
3292
+ "eval_samples_per_second": 2.463,
3293
+ "eval_steps_per_second": 0.308,
3294
+ "step": 219
3295
+ },
3296
+ {
3297
+ "epoch": 6.875,
3298
+ "grad_norm": 0.5924889323251107,
3299
+ "learning_rate": 2e-05,
3300
+ "loss": 0.712,
3301
+ "step": 220
3302
+ },
3303
+ {
3304
+ "epoch": 6.875,
3305
+ "eval_loss": 0.6619111895561218,
3306
+ "eval_runtime": 81.7826,
3307
+ "eval_samples_per_second": 2.446,
3308
+ "eval_steps_per_second": 0.306,
3309
+ "step": 220
3310
+ },
3311
+ {
3312
+ "epoch": 6.90625,
3313
+ "grad_norm": 0.5290576915218269,
3314
+ "learning_rate": 2e-05,
3315
+ "loss": 0.6659,
3316
+ "step": 221
3317
+ },
3318
+ {
3319
+ "epoch": 6.90625,
3320
+ "eval_loss": 0.6609540581703186,
3321
+ "eval_runtime": 82.9922,
3322
+ "eval_samples_per_second": 2.41,
3323
+ "eval_steps_per_second": 0.301,
3324
+ "step": 221
3325
+ },
3326
+ {
3327
+ "epoch": 6.9375,
3328
+ "grad_norm": 0.5831209517049147,
3329
+ "learning_rate": 2e-05,
3330
+ "loss": 0.6547,
3331
+ "step": 222
3332
+ },
3333
+ {
3334
+ "epoch": 6.9375,
3335
+ "eval_loss": 0.660676896572113,
3336
+ "eval_runtime": 83.6541,
3337
+ "eval_samples_per_second": 2.391,
3338
+ "eval_steps_per_second": 0.299,
3339
+ "step": 222
3340
+ },
3341
+ {
3342
+ "epoch": 6.96875,
3343
+ "grad_norm": 0.5320966369511158,
3344
+ "learning_rate": 2e-05,
3345
+ "loss": 0.6968,
3346
+ "step": 223
3347
+ },
3348
+ {
3349
+ "epoch": 6.96875,
3350
+ "eval_loss": 0.6618594527244568,
3351
+ "eval_runtime": 83.1148,
3352
+ "eval_samples_per_second": 2.406,
3353
+ "eval_steps_per_second": 0.301,
3354
+ "step": 223
3355
+ },
3356
+ {
3357
+ "epoch": 7.0,
3358
+ "grad_norm": 0.5829636446837394,
3359
+ "learning_rate": 2e-05,
3360
+ "loss": 0.7407,
3361
+ "step": 224
3362
+ },
3363
+ {
3364
+ "epoch": 7.0,
3365
+ "eval_loss": 0.6635661125183105,
3366
+ "eval_runtime": 82.8183,
3367
+ "eval_samples_per_second": 2.415,
3368
+ "eval_steps_per_second": 0.302,
3369
+ "step": 224
3370
+ },
3371
+ {
3372
+ "epoch": 7.03125,
3373
+ "grad_norm": 0.4975095056459566,
3374
+ "learning_rate": 2e-05,
3375
+ "loss": 0.6535,
3376
+ "step": 225
3377
+ },
3378
+ {
3379
+ "epoch": 7.03125,
3380
+ "eval_loss": 0.6641671657562256,
3381
+ "eval_runtime": 83.0267,
3382
+ "eval_samples_per_second": 2.409,
3383
+ "eval_steps_per_second": 0.301,
3384
+ "step": 225
3385
+ },
3386
+ {
3387
+ "epoch": 7.0625,
3388
+ "grad_norm": 0.5625698523064815,
3389
+ "learning_rate": 2e-05,
3390
+ "loss": 0.6012,
3391
+ "step": 226
3392
+ },
3393
+ {
3394
+ "epoch": 7.0625,
3395
+ "eval_loss": 0.6639044880867004,
3396
+ "eval_runtime": 83.3881,
3397
+ "eval_samples_per_second": 2.398,
3398
+ "eval_steps_per_second": 0.3,
3399
+ "step": 226
3400
+ },
3401
+ {
3402
+ "epoch": 7.09375,
3403
+ "grad_norm": 0.5436196850683295,
3404
+ "learning_rate": 2e-05,
3405
+ "loss": 0.6485,
3406
+ "step": 227
3407
+ },
3408
+ {
3409
+ "epoch": 7.09375,
3410
+ "eval_loss": 0.6651788353919983,
3411
+ "eval_runtime": 82.7096,
3412
+ "eval_samples_per_second": 2.418,
3413
+ "eval_steps_per_second": 0.302,
3414
+ "step": 227
3415
+ },
3416
+ {
3417
+ "epoch": 7.125,
3418
+ "grad_norm": 0.5598906287609361,
3419
+ "learning_rate": 2e-05,
3420
+ "loss": 0.6142,
3421
+ "step": 228
3422
+ },
3423
+ {
3424
+ "epoch": 7.125,
3425
+ "eval_loss": 0.6688636541366577,
3426
+ "eval_runtime": 82.601,
3427
+ "eval_samples_per_second": 2.421,
3428
+ "eval_steps_per_second": 0.303,
3429
+ "step": 228
3430
+ },
3431
+ {
3432
+ "epoch": 7.15625,
3433
+ "grad_norm": 0.7572979310697923,
3434
+ "learning_rate": 2e-05,
3435
+ "loss": 0.6221,
3436
+ "step": 229
3437
+ },
3438
+ {
3439
+ "epoch": 7.15625,
3440
+ "eval_loss": 0.6699694991111755,
3441
+ "eval_runtime": 82.6032,
3442
+ "eval_samples_per_second": 2.421,
3443
+ "eval_steps_per_second": 0.303,
3444
+ "step": 229
3445
+ },
3446
+ {
3447
+ "epoch": 7.1875,
3448
+ "grad_norm": 0.6173309690580897,
3449
+ "learning_rate": 2e-05,
3450
+ "loss": 0.5919,
3451
+ "step": 230
3452
+ },
3453
+ {
3454
+ "epoch": 7.1875,
3455
+ "eval_loss": 0.6706527471542358,
3456
+ "eval_runtime": 82.9732,
3457
+ "eval_samples_per_second": 2.41,
3458
+ "eval_steps_per_second": 0.301,
3459
+ "step": 230
3460
+ },
3461
+ {
3462
+ "epoch": 7.21875,
3463
+ "grad_norm": 0.643241771517866,
3464
+ "learning_rate": 2e-05,
3465
+ "loss": 0.7081,
3466
+ "step": 231
3467
+ },
3468
+ {
3469
+ "epoch": 7.21875,
3470
+ "eval_loss": 0.6700320243835449,
3471
+ "eval_runtime": 84.5621,
3472
+ "eval_samples_per_second": 2.365,
3473
+ "eval_steps_per_second": 0.296,
3474
+ "step": 231
3475
+ },
3476
+ {
3477
+ "epoch": 7.25,
3478
+ "grad_norm": 0.577638137570571,
3479
+ "learning_rate": 2e-05,
3480
+ "loss": 0.6873,
3481
+ "step": 232
3482
+ },
3483
+ {
3484
+ "epoch": 7.25,
3485
+ "eval_loss": 0.669111430644989,
3486
+ "eval_runtime": 84.5124,
3487
+ "eval_samples_per_second": 2.367,
3488
+ "eval_steps_per_second": 0.296,
3489
+ "step": 232
3490
+ },
3491
+ {
3492
+ "epoch": 7.28125,
3493
+ "grad_norm": 0.7229488296023369,
3494
+ "learning_rate": 2e-05,
3495
+ "loss": 0.6301,
3496
+ "step": 233
3497
+ },
3498
+ {
3499
+ "epoch": 7.28125,
3500
+ "eval_loss": 0.6664154529571533,
3501
+ "eval_runtime": 84.6437,
3502
+ "eval_samples_per_second": 2.363,
3503
+ "eval_steps_per_second": 0.295,
3504
+ "step": 233
3505
+ },
3506
+ {
3507
+ "epoch": 7.3125,
3508
+ "grad_norm": 0.5827815449039045,
3509
+ "learning_rate": 2e-05,
3510
+ "loss": 0.669,
3511
+ "step": 234
3512
+ },
3513
+ {
3514
+ "epoch": 7.3125,
3515
+ "eval_loss": 0.6641202569007874,
3516
+ "eval_runtime": 84.489,
3517
+ "eval_samples_per_second": 2.367,
3518
+ "eval_steps_per_second": 0.296,
3519
+ "step": 234
3520
+ },
3521
+ {
3522
+ "epoch": 7.34375,
3523
+ "grad_norm": 0.57507354017269,
3524
+ "learning_rate": 2e-05,
3525
+ "loss": 0.6474,
3526
+ "step": 235
3527
+ },
3528
+ {
3529
+ "epoch": 7.34375,
3530
+ "eval_loss": 0.6623325347900391,
3531
+ "eval_runtime": 84.5536,
3532
+ "eval_samples_per_second": 2.365,
3533
+ "eval_steps_per_second": 0.296,
3534
+ "step": 235
3535
+ },
3536
+ {
3537
+ "epoch": 7.375,
3538
+ "grad_norm": 0.5810844862533651,
3539
+ "learning_rate": 2e-05,
3540
+ "loss": 0.6048,
3541
+ "step": 236
3542
+ },
3543
+ {
3544
+ "epoch": 7.375,
3545
+ "eval_loss": 0.6619194746017456,
3546
+ "eval_runtime": 84.2296,
3547
+ "eval_samples_per_second": 2.374,
3548
+ "eval_steps_per_second": 0.297,
3549
+ "step": 236
3550
+ },
3551
+ {
3552
+ "epoch": 7.40625,
3553
+ "grad_norm": 0.6075032415813726,
3554
+ "learning_rate": 2e-05,
3555
+ "loss": 0.6529,
3556
+ "step": 237
3557
+ },
3558
+ {
3559
+ "epoch": 7.40625,
3560
+ "eval_loss": 0.6626202464103699,
3561
+ "eval_runtime": 84.9703,
3562
+ "eval_samples_per_second": 2.354,
3563
+ "eval_steps_per_second": 0.294,
3564
+ "step": 237
3565
+ },
3566
+ {
3567
+ "epoch": 7.4375,
3568
+ "grad_norm": 0.6402642234375245,
3569
+ "learning_rate": 2e-05,
3570
+ "loss": 0.6433,
3571
+ "step": 238
3572
+ },
3573
+ {
3574
+ "epoch": 7.4375,
3575
+ "eval_loss": 0.663289487361908,
3576
+ "eval_runtime": 84.8924,
3577
+ "eval_samples_per_second": 2.356,
3578
+ "eval_steps_per_second": 0.294,
3579
+ "step": 238
3580
+ },
3581
+ {
3582
+ "epoch": 7.46875,
3583
+ "grad_norm": 0.6335996982657431,
3584
+ "learning_rate": 2e-05,
3585
+ "loss": 0.6815,
3586
+ "step": 239
3587
+ },
3588
+ {
3589
+ "epoch": 7.46875,
3590
+ "eval_loss": 0.6636109948158264,
3591
+ "eval_runtime": 85.0551,
3592
+ "eval_samples_per_second": 2.351,
3593
+ "eval_steps_per_second": 0.294,
3594
+ "step": 239
3595
+ },
3596
+ {
3597
+ "epoch": 7.5,
3598
+ "grad_norm": 0.5796846795848909,
3599
+ "learning_rate": 2e-05,
3600
+ "loss": 0.6236,
3601
+ "step": 240
3602
+ },
3603
+ {
3604
+ "epoch": 7.5,
3605
+ "eval_loss": 0.6652829051017761,
3606
+ "eval_runtime": 84.7574,
3607
+ "eval_samples_per_second": 2.36,
3608
+ "eval_steps_per_second": 0.295,
3609
+ "step": 240
3610
+ },
3611
+ {
3612
+ "epoch": 7.53125,
3613
+ "grad_norm": 0.5380402145760035,
3614
+ "learning_rate": 2e-05,
3615
+ "loss": 0.6564,
3616
+ "step": 241
3617
+ },
3618
+ {
3619
+ "epoch": 7.53125,
3620
+ "eval_loss": 0.6676375865936279,
3621
+ "eval_runtime": 86.2058,
3622
+ "eval_samples_per_second": 2.32,
3623
+ "eval_steps_per_second": 0.29,
3624
+ "step": 241
3625
+ },
3626
+ {
3627
+ "epoch": 7.5625,
3628
+ "grad_norm": 0.5964298255824012,
3629
+ "learning_rate": 2e-05,
3630
+ "loss": 0.6475,
3631
+ "step": 242
3632
+ },
3633
+ {
3634
+ "epoch": 7.5625,
3635
+ "eval_loss": 0.6698520183563232,
3636
+ "eval_runtime": 85.8955,
3637
+ "eval_samples_per_second": 2.328,
3638
+ "eval_steps_per_second": 0.291,
3639
+ "step": 242
3640
+ },
3641
+ {
3642
+ "epoch": 7.59375,
3643
+ "grad_norm": 0.561279296875,
3644
+ "learning_rate": 2e-05,
3645
+ "loss": 0.6395,
3646
+ "step": 243
3647
+ },
3648
+ {
3649
+ "epoch": 7.59375,
3650
+ "eval_loss": 0.6705803871154785,
3651
+ "eval_runtime": 86.0036,
3652
+ "eval_samples_per_second": 2.325,
3653
+ "eval_steps_per_second": 0.291,
3654
+ "step": 243
3655
+ },
3656
+ {
3657
+ "epoch": 7.625,
3658
+ "grad_norm": 0.6757292755073548,
3659
+ "learning_rate": 2e-05,
3660
+ "loss": 0.7074,
3661
+ "step": 244
3662
+ },
3663
+ {
3664
+ "epoch": 7.625,
3665
+ "eval_loss": 0.6679538488388062,
3666
+ "eval_runtime": 85.5379,
3667
+ "eval_samples_per_second": 2.338,
3668
+ "eval_steps_per_second": 0.292,
3669
+ "step": 244
3670
+ },
3671
+ {
3672
+ "epoch": 7.65625,
3673
+ "grad_norm": 0.659077163070129,
3674
+ "learning_rate": 2e-05,
3675
+ "loss": 0.6078,
3676
+ "step": 245
3677
+ },
3678
+ {
3679
+ "epoch": 7.65625,
3680
+ "eval_loss": 0.6667564511299133,
3681
+ "eval_runtime": 85.752,
3682
+ "eval_samples_per_second": 2.332,
3683
+ "eval_steps_per_second": 0.292,
3684
+ "step": 245
3685
+ },
3686
+ {
3687
+ "epoch": 7.6875,
3688
+ "grad_norm": 0.6215405566454576,
3689
+ "learning_rate": 2e-05,
3690
+ "loss": 0.6603,
3691
+ "step": 246
3692
+ },
3693
+ {
3694
+ "epoch": 7.6875,
3695
+ "eval_loss": 0.665945291519165,
3696
+ "eval_runtime": 92.3086,
3697
+ "eval_samples_per_second": 2.167,
3698
+ "eval_steps_per_second": 0.271,
3699
+ "step": 246
3700
+ },
3701
+ {
3702
+ "epoch": 7.71875,
3703
+ "grad_norm": 0.6130534921490498,
3704
+ "learning_rate": 2e-05,
3705
+ "loss": 0.6435,
3706
+ "step": 247
3707
+ },
3708
+ {
3709
+ "epoch": 7.71875,
3710
+ "eval_loss": 0.6661685109138489,
3711
+ "eval_runtime": 87.1917,
3712
+ "eval_samples_per_second": 2.294,
3713
+ "eval_steps_per_second": 0.287,
3714
+ "step": 247
3715
+ },
3716
+ {
3717
+ "epoch": 7.75,
3718
+ "grad_norm": 0.6025415602868736,
3719
+ "learning_rate": 2e-05,
3720
+ "loss": 0.6308,
3721
+ "step": 248
3722
+ },
3723
+ {
3724
+ "epoch": 7.75,
3725
+ "eval_loss": 0.6658704280853271,
3726
+ "eval_runtime": 86.8233,
3727
+ "eval_samples_per_second": 2.304,
3728
+ "eval_steps_per_second": 0.288,
3729
+ "step": 248
3730
+ },
3731
+ {
3732
+ "epoch": 7.78125,
3733
+ "grad_norm": 0.6901593792019413,
3734
+ "learning_rate": 2e-05,
3735
+ "loss": 0.6777,
3736
+ "step": 249
3737
+ },
3738
+ {
3739
+ "epoch": 7.78125,
3740
+ "eval_loss": 0.6652414202690125,
3741
+ "eval_runtime": 86.7625,
3742
+ "eval_samples_per_second": 2.305,
3743
+ "eval_steps_per_second": 0.288,
3744
+ "step": 249
3745
+ },
3746
+ {
3747
+ "epoch": 7.8125,
3748
+ "grad_norm": 0.6436454697341579,
3749
+ "learning_rate": 2e-05,
3750
+ "loss": 0.6912,
3751
+ "step": 250
3752
+ },
3753
+ {
3754
+ "epoch": 7.8125,
3755
+ "eval_loss": 0.6654212474822998,
3756
+ "eval_runtime": 86.871,
3757
+ "eval_samples_per_second": 2.302,
3758
+ "eval_steps_per_second": 0.288,
3759
+ "step": 250
3760
+ },
3761
+ {
3762
+ "epoch": 7.84375,
3763
+ "grad_norm": 0.649040103024529,
3764
+ "learning_rate": 2e-05,
3765
+ "loss": 0.6025,
3766
+ "step": 251
3767
+ },
3768
+ {
3769
+ "epoch": 7.84375,
3770
+ "eval_loss": 0.6654068231582642,
3771
+ "eval_runtime": 86.7458,
3772
+ "eval_samples_per_second": 2.306,
3773
+ "eval_steps_per_second": 0.288,
3774
+ "step": 251
3775
+ },
3776
+ {
3777
+ "epoch": 7.875,
3778
+ "grad_norm": 0.6595522131680224,
3779
+ "learning_rate": 2e-05,
3780
+ "loss": 0.5973,
3781
+ "step": 252
3782
+ },
3783
+ {
3784
+ "epoch": 7.875,
3785
+ "eval_loss": 0.6644830107688904,
3786
+ "eval_runtime": 86.8739,
3787
+ "eval_samples_per_second": 2.302,
3788
+ "eval_steps_per_second": 0.288,
3789
+ "step": 252
3790
+ },
3791
+ {
3792
+ "epoch": 7.90625,
3793
+ "grad_norm": 0.6689891717273936,
3794
+ "learning_rate": 2e-05,
3795
+ "loss": 0.687,
3796
+ "step": 253
3797
+ },
3798
+ {
3799
+ "epoch": 7.90625,
3800
+ "eval_loss": 0.6616199612617493,
3801
+ "eval_runtime": 86.8222,
3802
+ "eval_samples_per_second": 2.304,
3803
+ "eval_steps_per_second": 0.288,
3804
+ "step": 253
3805
+ },
3806
+ {
3807
+ "epoch": 7.9375,
3808
+ "grad_norm": 0.6306846778314292,
3809
+ "learning_rate": 2e-05,
3810
+ "loss": 0.6599,
3811
+ "step": 254
3812
+ },
3813
+ {
3814
+ "epoch": 7.9375,
3815
+ "eval_loss": 0.6592965126037598,
3816
+ "eval_runtime": 86.8577,
3817
+ "eval_samples_per_second": 2.303,
3818
+ "eval_steps_per_second": 0.288,
3819
+ "step": 254
3820
+ },
3821
+ {
3822
+ "epoch": 7.96875,
3823
+ "grad_norm": 0.6021327993890785,
3824
+ "learning_rate": 2e-05,
3825
+ "loss": 0.575,
3826
+ "step": 255
3827
+ },
3828
+ {
3829
+ "epoch": 7.96875,
3830
+ "eval_loss": 0.6580593585968018,
3831
+ "eval_runtime": 86.7582,
3832
+ "eval_samples_per_second": 2.305,
3833
+ "eval_steps_per_second": 0.288,
3834
+ "step": 255
3835
+ },
3836
+ {
3837
+ "epoch": 8.0,
3838
+ "grad_norm": 0.6174712675568311,
3839
+ "learning_rate": 2e-05,
3840
+ "loss": 0.6341,
3841
+ "step": 256
3842
+ },
3843
+ {
3844
+ "epoch": 8.0,
3845
+ "eval_loss": 0.6575854420661926,
3846
+ "eval_runtime": 76.7634,
3847
+ "eval_samples_per_second": 2.605,
3848
+ "eval_steps_per_second": 0.326,
3849
+ "step": 256
3850
+ }
3851
+ ],
3852
+ "logging_steps": 1.0,
3853
+ "max_steps": 256,
3854
+ "num_input_tokens_seen": 0,
3855
+ "num_train_epochs": 8,
3856
+ "save_steps": 5,
3857
+ "stateful_callbacks": {
3858
+ "TrainerControl": {
3859
+ "args": {
3860
+ "should_epoch_stop": false,
3861
+ "should_evaluate": false,
3862
+ "should_log": false,
3863
+ "should_save": true,
3864
+ "should_training_stop": true
3865
+ },
3866
+ "attributes": {}
3867
+ }
3868
+ },
3869
+ "total_flos": 489287119011840.0,
3870
+ "train_batch_size": 8,
3871
+ "trial_name": null,
3872
+ "trial_params": null
3873
+ }
checkpoint-256/training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f188470aed89e37f0d7f17497d5475eb84bc639c6ba047e7db9629674c365735
3
+ size 8312
checkpoint-256/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-v1.6-vicuna-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-v1.6-vicuna-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
+ "o_proj",
24
+ "q_proj",
25
+ "k_proj",
26
+ "gate_proj",
27
+ "up_proj",
28
+ "down_proj",
29
+ "v_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 sha256:6431e6a5485bf9971a8a505d2ce8ac8f1350546005403146f8f3bab2c6c30f02
3
+ size 65046168
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 sha256:800218503888cfc6c6dacef7580b008ce5c28bc77e6d725edc53eb163eb213f4
3
+ size 775138
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 sha256:02fd758d84cad0221cd366926fce830c888c4a25795ffa592310e55569f8be57
3
+ size 191825901
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 sha256:5b02e6ef69e3a0ba9ce109c85ba6b611ee8a5dcaa7acfb914f2610b48ae72d5b
3
+ size 775138
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:c6b3cb3c0b9269d0f01562ac97cf7ed7cfd9c803885af1dde56b9a6eda5fd47d
3
+ size 191825901
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 sha256:86079448784de87bb5a32ee61f4e61ec37c91808e8572ee637f70dab7710618a
3
+ size 775138
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 sha256:6cde72030703e6207c5148a37bfc458cf75f4464644ba5922dd636d1ed703699
3
+ size 191825901
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 sha256:975b4b22694bd9e2244d3d9338e3df16410609bec48870a280c085c9d0ea6385
3
+ size 775138
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 sha256:2e7e632b974b0cd0cffd82ac220632d1d22d44d1da09616b219d5f96b07e535c
3
+ size 191825901
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 sha256:03af8976c83af29b26ac3bdd42a804bb9e4d7c51eff643b3ad188c88c846c088
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 sha256:65a0552a908836fab6a8e6f840ac7d8de6dafa58227414f46353830c2cac6eae
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 sha256:05eeeac05a2df77ec2d80d022a5d4c0d3a738fa5f3f0c7f6560893b766f6a722
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 sha256:70a680c028586c6979517de8d634e39a4908c3611cac7b62d70f1bcb83f6c594
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:df3df8f53bf051656f1ae89d4681a26c113103914ee95e8a97646c6c5c824188
3
+ size 8312
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 +1,77 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_attn_implementation_autoset": true,
3
+ "_name_or_path": "liuhaotian/llava-v1.6-vicuna-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": "anyres",
17
+ "image_crop_resolution": 224,
18
+ "image_grid_pinpoints": [
19
+ [
20
+ 336,
21
+ 672
22
+ ],
23
+ [
24
+ 672,
25
+ 336
26
+ ],
27
+ [
28
+ 672,
29
+ 672
30
+ ],
31
+ [
32
+ 1008,
33
+ 336
34
+ ],
35
+ [
36
+ 336,
37
+ 1008
38
+ ]
39
+ ],
40
+ "image_split_resolution": 224,
41
+ "initializer_range": 0.02,
42
+ "intermediate_size": 13824,
43
+ "max_length": 4096,
44
+ "max_position_embeddings": 4096,
45
+ "mlp_bias": false,
46
+ "mm_hidden_size": 1024,
47
+ "mm_patch_merge_type": "flat",
48
+ "mm_projector_lr": 2e-05,
49
+ "mm_projector_type": "mlp2x_gelu",
50
+ "mm_resampler_type": null,
51
+ "mm_use_im_patch_token": false,
52
+ "mm_use_im_start_end": false,
53
+ "mm_vision_select_feature": "patch",
54
+ "mm_vision_select_layer": -2,
55
+ "mm_vision_tower": "openai/clip-vit-large-patch14-336",
56
+ "mm_vision_tower_lr": 2e-06,
57
+ "model_type": "llava_llama",
58
+ "num_attention_heads": 40,
59
+ "num_hidden_layers": 40,
60
+ "num_key_value_heads": 40,
61
+ "pad_token_id": 0,
62
+ "pretraining_tp": 1,
63
+ "rms_norm_eps": 1e-05,
64
+ "rope_scaling": null,
65
+ "rope_theta": 10000.0,
66
+ "tie_word_embeddings": false,
67
+ "tokenizer_model_max_length": 2048,
68
+ "tokenizer_padding_side": "right",
69
+ "torch_dtype": "bfloat16",
70
+ "transformers_version": "4.46.3",
71
+ "tune_mm_mlp_adapter": false,
72
+ "tune_mm_vision_resampler": false,
73
+ "unfreeze_mm_vision_tower": true,
74
+ "use_cache": true,
75
+ "use_mm_proj": true,
76
+ "vocab_size": 32000
77
+ }
non_lora_trainables.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:05b685cfc78d68bd5ca3e549eb69e061a0e258fd61a17b196e3ef3876ec7cda3
3
+ size 62937264
optimizer.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7758883238817def488bea1dc14bb3c1a10225fa8d5b92dc0ada5d92c0cdf52b
3
+ size 191824418