Upload entire folder 160 steps of 351 steps, half of 3 epochs.
Browse filesThis view is limited to 50 files because it contains too many changes.
See raw diff
- checkpoint-100/README.md +202 -0
- checkpoint-100/adapter_config.json +35 -0
- checkpoint-100/adapter_model.safetensors +3 -0
- checkpoint-100/global_step100/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt +3 -0
- checkpoint-100/global_step100/mp_rank_00_model_states.pt +3 -0
- checkpoint-100/latest +1 -0
- checkpoint-100/rng_state.pth +3 -0
- checkpoint-100/trainer_state.json +733 -0
- checkpoint-100/training_args.bin +3 -0
- checkpoint-100/zero_to_fp32.py +592 -0
- checkpoint-120/README.md +202 -0
- checkpoint-120/adapter_config.json +35 -0
- checkpoint-120/adapter_model.safetensors +3 -0
- checkpoint-120/global_step120/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt +3 -0
- checkpoint-120/global_step120/mp_rank_00_model_states.pt +3 -0
- checkpoint-120/latest +1 -0
- checkpoint-120/rng_state.pth +3 -0
- checkpoint-120/trainer_state.json +873 -0
- checkpoint-120/training_args.bin +3 -0
- checkpoint-120/zero_to_fp32.py +592 -0
- checkpoint-140/README.md +202 -0
- checkpoint-140/adapter_config.json +35 -0
- checkpoint-140/adapter_model.safetensors +3 -0
- checkpoint-140/global_step140/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt +3 -0
- checkpoint-140/global_step140/mp_rank_00_model_states.pt +3 -0
- checkpoint-140/latest +1 -0
- checkpoint-140/rng_state.pth +3 -0
- checkpoint-140/trainer_state.json +1013 -0
- checkpoint-140/training_args.bin +3 -0
- checkpoint-140/zero_to_fp32.py +592 -0
- checkpoint-160/README.md +202 -0
- checkpoint-160/adapter_config.json +35 -0
- checkpoint-160/adapter_model.safetensors +3 -0
- checkpoint-160/global_step160/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt +3 -0
- checkpoint-160/global_step160/mp_rank_00_model_states.pt +3 -0
- checkpoint-160/latest +1 -0
- checkpoint-160/rng_state.pth +3 -0
- checkpoint-160/trainer_state.json +1153 -0
- checkpoint-160/training_args.bin +3 -0
- checkpoint-160/zero_to_fp32.py +592 -0
- checkpoint-20/README.md +202 -0
- checkpoint-20/adapter_config.json +35 -0
- checkpoint-20/adapter_model.safetensors +3 -0
- checkpoint-20/global_step20/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt +3 -0
- checkpoint-20/global_step20/mp_rank_00_model_states.pt +3 -0
- checkpoint-20/latest +1 -0
- checkpoint-20/rng_state.pth +3 -0
- checkpoint-20/trainer_state.json +173 -0
- checkpoint-20/training_args.bin +3 -0
- checkpoint-20/zero_to_fp32.py +592 -0
checkpoint-100/README.md
ADDED
|
@@ -0,0 +1,202 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: peft
|
| 3 |
+
base_model: ../ckpts/Meta-Llama-3-8B-Instruct
|
| 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 0.11.1
|
checkpoint-100/adapter_config.json
ADDED
|
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"alpha_pattern": {},
|
| 3 |
+
"auto_mapping": null,
|
| 4 |
+
"base_model_name_or_path": "../ckpts/Meta-Llama-3-8B-Instruct",
|
| 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": 32,
|
| 14 |
+
"lora_dropout": 0.1,
|
| 15 |
+
"megatron_config": null,
|
| 16 |
+
"megatron_core": "megatron.core",
|
| 17 |
+
"modules_to_save": null,
|
| 18 |
+
"peft_type": "LORA",
|
| 19 |
+
"r": 16,
|
| 20 |
+
"rank_pattern": {},
|
| 21 |
+
"revision": null,
|
| 22 |
+
"target_modules": [
|
| 23 |
+
"down_proj",
|
| 24 |
+
"gate_proj",
|
| 25 |
+
"q_proj",
|
| 26 |
+
"lm_head",
|
| 27 |
+
"o_proj",
|
| 28 |
+
"v_proj",
|
| 29 |
+
"up_proj",
|
| 30 |
+
"k_proj"
|
| 31 |
+
],
|
| 32 |
+
"task_type": "CAUSAL_LM",
|
| 33 |
+
"use_dora": false,
|
| 34 |
+
"use_rslora": false
|
| 35 |
+
}
|
checkpoint-100/adapter_model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:baca06cc42b3510c3c3bf848f73f79b72b1baea5a4f963941e134e37b3ab777d
|
| 3 |
+
size 1138856856
|
checkpoint-100/global_step100/bf16_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 sha256:808ab758e2efff532198a64b20a7f9c499ac3216290ea2598de18360f006e24c
|
| 3 |
+
size 528781328
|
checkpoint-100/global_step100/mp_rank_00_model_states.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:cba7f894e1d634e3bd3acbb30c43a7817dcd9bbc30549de0142cf5b4970bdd25
|
| 3 |
+
size 199905337
|
checkpoint-100/latest
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
global_step100
|
checkpoint-100/rng_state.pth
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:1a2d9af3fcfc54434e74034cabf1cd700dd64663d32e5f7d97b79b71b2284409
|
| 3 |
+
size 14244
|
checkpoint-100/trainer_state.json
ADDED
|
@@ -0,0 +1,733 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"best_metric": null,
|
| 3 |
+
"best_model_checkpoint": null,
|
| 4 |
+
"epoch": 0.8533333333333334,
|
| 5 |
+
"eval_steps": 500,
|
| 6 |
+
"global_step": 100,
|
| 7 |
+
"is_hyper_param_search": false,
|
| 8 |
+
"is_local_process_zero": true,
|
| 9 |
+
"is_world_process_zero": true,
|
| 10 |
+
"log_history": [
|
| 11 |
+
{
|
| 12 |
+
"epoch": 0.008533333333333334,
|
| 13 |
+
"grad_norm": 100.21848203113535,
|
| 14 |
+
"learning_rate": 0.0,
|
| 15 |
+
"loss": 7.1962,
|
| 16 |
+
"step": 1
|
| 17 |
+
},
|
| 18 |
+
{
|
| 19 |
+
"epoch": 0.017066666666666667,
|
| 20 |
+
"grad_norm": 96.43006188910957,
|
| 21 |
+
"learning_rate": 3.0102999566398115e-06,
|
| 22 |
+
"loss": 6.9414,
|
| 23 |
+
"step": 2
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"epoch": 0.0256,
|
| 27 |
+
"grad_norm": 97.35803466618715,
|
| 28 |
+
"learning_rate": 4.771212547196624e-06,
|
| 29 |
+
"loss": 7.0102,
|
| 30 |
+
"step": 3
|
| 31 |
+
},
|
| 32 |
+
{
|
| 33 |
+
"epoch": 0.034133333333333335,
|
| 34 |
+
"grad_norm": 95.14837816372646,
|
| 35 |
+
"learning_rate": 6.020599913279623e-06,
|
| 36 |
+
"loss": 6.5295,
|
| 37 |
+
"step": 4
|
| 38 |
+
},
|
| 39 |
+
{
|
| 40 |
+
"epoch": 0.042666666666666665,
|
| 41 |
+
"grad_norm": 91.76544275692784,
|
| 42 |
+
"learning_rate": 6.989700043360187e-06,
|
| 43 |
+
"loss": 6.4806,
|
| 44 |
+
"step": 5
|
| 45 |
+
},
|
| 46 |
+
{
|
| 47 |
+
"epoch": 0.0512,
|
| 48 |
+
"grad_norm": 84.4494318688335,
|
| 49 |
+
"learning_rate": 7.781512503836437e-06,
|
| 50 |
+
"loss": 6.4194,
|
| 51 |
+
"step": 6
|
| 52 |
+
},
|
| 53 |
+
{
|
| 54 |
+
"epoch": 0.05973333333333333,
|
| 55 |
+
"grad_norm": 71.37977490595638,
|
| 56 |
+
"learning_rate": 8.450980400142568e-06,
|
| 57 |
+
"loss": 5.4953,
|
| 58 |
+
"step": 7
|
| 59 |
+
},
|
| 60 |
+
{
|
| 61 |
+
"epoch": 0.06826666666666667,
|
| 62 |
+
"grad_norm": 49.31153456754566,
|
| 63 |
+
"learning_rate": 9.030899869919434e-06,
|
| 64 |
+
"loss": 5.4123,
|
| 65 |
+
"step": 8
|
| 66 |
+
},
|
| 67 |
+
{
|
| 68 |
+
"epoch": 0.0768,
|
| 69 |
+
"grad_norm": 20.37296364560341,
|
| 70 |
+
"learning_rate": 9.542425094393249e-06,
|
| 71 |
+
"loss": 5.2334,
|
| 72 |
+
"step": 9
|
| 73 |
+
},
|
| 74 |
+
{
|
| 75 |
+
"epoch": 0.08533333333333333,
|
| 76 |
+
"grad_norm": 7.968467079076881,
|
| 77 |
+
"learning_rate": 9.999999999999999e-06,
|
| 78 |
+
"loss": 5.0282,
|
| 79 |
+
"step": 10
|
| 80 |
+
},
|
| 81 |
+
{
|
| 82 |
+
"epoch": 0.09386666666666667,
|
| 83 |
+
"grad_norm": 3.559446532055649,
|
| 84 |
+
"learning_rate": 1.041392685158225e-05,
|
| 85 |
+
"loss": 4.612,
|
| 86 |
+
"step": 11
|
| 87 |
+
},
|
| 88 |
+
{
|
| 89 |
+
"epoch": 0.1024,
|
| 90 |
+
"grad_norm": 3.5528846947995674,
|
| 91 |
+
"learning_rate": 1.0791812460476248e-05,
|
| 92 |
+
"loss": 4.9475,
|
| 93 |
+
"step": 12
|
| 94 |
+
},
|
| 95 |
+
{
|
| 96 |
+
"epoch": 0.11093333333333333,
|
| 97 |
+
"grad_norm": 3.541968897471334,
|
| 98 |
+
"learning_rate": 1.1139433523068365e-05,
|
| 99 |
+
"loss": 4.2777,
|
| 100 |
+
"step": 13
|
| 101 |
+
},
|
| 102 |
+
{
|
| 103 |
+
"epoch": 0.11946666666666667,
|
| 104 |
+
"grad_norm": 3.54718070036198,
|
| 105 |
+
"learning_rate": 1.1461280356782378e-05,
|
| 106 |
+
"loss": 4.3507,
|
| 107 |
+
"step": 14
|
| 108 |
+
},
|
| 109 |
+
{
|
| 110 |
+
"epoch": 0.128,
|
| 111 |
+
"grad_norm": 3.8632334830606747,
|
| 112 |
+
"learning_rate": 1.1760912590556813e-05,
|
| 113 |
+
"loss": 4.5364,
|
| 114 |
+
"step": 15
|
| 115 |
+
},
|
| 116 |
+
{
|
| 117 |
+
"epoch": 0.13653333333333334,
|
| 118 |
+
"grad_norm": 3.6637424744054004,
|
| 119 |
+
"learning_rate": 1.2041199826559246e-05,
|
| 120 |
+
"loss": 3.9672,
|
| 121 |
+
"step": 16
|
| 122 |
+
},
|
| 123 |
+
{
|
| 124 |
+
"epoch": 0.14506666666666668,
|
| 125 |
+
"grad_norm": 3.919802904818311,
|
| 126 |
+
"learning_rate": 1.230448921378274e-05,
|
| 127 |
+
"loss": 4.0618,
|
| 128 |
+
"step": 17
|
| 129 |
+
},
|
| 130 |
+
{
|
| 131 |
+
"epoch": 0.1536,
|
| 132 |
+
"grad_norm": 4.71904950738746,
|
| 133 |
+
"learning_rate": 1.2552725051033058e-05,
|
| 134 |
+
"loss": 4.6656,
|
| 135 |
+
"step": 18
|
| 136 |
+
},
|
| 137 |
+
{
|
| 138 |
+
"epoch": 0.16213333333333332,
|
| 139 |
+
"grad_norm": 4.6656317698690835,
|
| 140 |
+
"learning_rate": 1.2787536009528288e-05,
|
| 141 |
+
"loss": 4.1131,
|
| 142 |
+
"step": 19
|
| 143 |
+
},
|
| 144 |
+
{
|
| 145 |
+
"epoch": 0.17066666666666666,
|
| 146 |
+
"grad_norm": 5.145138692367417,
|
| 147 |
+
"learning_rate": 1.301029995663981e-05,
|
| 148 |
+
"loss": 4.0989,
|
| 149 |
+
"step": 20
|
| 150 |
+
},
|
| 151 |
+
{
|
| 152 |
+
"epoch": 0.1792,
|
| 153 |
+
"grad_norm": 5.923538014759818,
|
| 154 |
+
"learning_rate": 1.3222192947339192e-05,
|
| 155 |
+
"loss": 4.4991,
|
| 156 |
+
"step": 21
|
| 157 |
+
},
|
| 158 |
+
{
|
| 159 |
+
"epoch": 0.18773333333333334,
|
| 160 |
+
"grad_norm": 5.941056962941364,
|
| 161 |
+
"learning_rate": 1.3424226808222062e-05,
|
| 162 |
+
"loss": 4.0836,
|
| 163 |
+
"step": 22
|
| 164 |
+
},
|
| 165 |
+
{
|
| 166 |
+
"epoch": 0.19626666666666667,
|
| 167 |
+
"grad_norm": 6.171026012117947,
|
| 168 |
+
"learning_rate": 1.3617278360175927e-05,
|
| 169 |
+
"loss": 3.6861,
|
| 170 |
+
"step": 23
|
| 171 |
+
},
|
| 172 |
+
{
|
| 173 |
+
"epoch": 0.2048,
|
| 174 |
+
"grad_norm": 7.130542138930838,
|
| 175 |
+
"learning_rate": 1.380211241711606e-05,
|
| 176 |
+
"loss": 4.0958,
|
| 177 |
+
"step": 24
|
| 178 |
+
},
|
| 179 |
+
{
|
| 180 |
+
"epoch": 0.21333333333333335,
|
| 181 |
+
"grad_norm": 7.328837606110418,
|
| 182 |
+
"learning_rate": 1.3979400086720374e-05,
|
| 183 |
+
"loss": 3.9524,
|
| 184 |
+
"step": 25
|
| 185 |
+
},
|
| 186 |
+
{
|
| 187 |
+
"epoch": 0.22186666666666666,
|
| 188 |
+
"grad_norm": 6.923489005711429,
|
| 189 |
+
"learning_rate": 1.4149733479708178e-05,
|
| 190 |
+
"loss": 3.6062,
|
| 191 |
+
"step": 26
|
| 192 |
+
},
|
| 193 |
+
{
|
| 194 |
+
"epoch": 0.2304,
|
| 195 |
+
"grad_norm": 6.409498926059221,
|
| 196 |
+
"learning_rate": 1.4313637641589872e-05,
|
| 197 |
+
"loss": 3.2034,
|
| 198 |
+
"step": 27
|
| 199 |
+
},
|
| 200 |
+
{
|
| 201 |
+
"epoch": 0.23893333333333333,
|
| 202 |
+
"grad_norm": 5.781628405584682,
|
| 203 |
+
"learning_rate": 1.4471580313422191e-05,
|
| 204 |
+
"loss": 2.8158,
|
| 205 |
+
"step": 28
|
| 206 |
+
},
|
| 207 |
+
{
|
| 208 |
+
"epoch": 0.24746666666666667,
|
| 209 |
+
"grad_norm": 6.2927590068243315,
|
| 210 |
+
"learning_rate": 1.4623979978989559e-05,
|
| 211 |
+
"loss": 2.9803,
|
| 212 |
+
"step": 29
|
| 213 |
+
},
|
| 214 |
+
{
|
| 215 |
+
"epoch": 0.256,
|
| 216 |
+
"grad_norm": 6.103844678630006,
|
| 217 |
+
"learning_rate": 1.4771212547196623e-05,
|
| 218 |
+
"loss": 2.847,
|
| 219 |
+
"step": 30
|
| 220 |
+
},
|
| 221 |
+
{
|
| 222 |
+
"epoch": 0.26453333333333334,
|
| 223 |
+
"grad_norm": 7.656341925867144,
|
| 224 |
+
"learning_rate": 1.4913616938342726e-05,
|
| 225 |
+
"loss": 3.0907,
|
| 226 |
+
"step": 31
|
| 227 |
+
},
|
| 228 |
+
{
|
| 229 |
+
"epoch": 0.2730666666666667,
|
| 230 |
+
"grad_norm": 6.324242877501844,
|
| 231 |
+
"learning_rate": 1.5051499783199059e-05,
|
| 232 |
+
"loss": 2.3467,
|
| 233 |
+
"step": 32
|
| 234 |
+
},
|
| 235 |
+
{
|
| 236 |
+
"epoch": 0.2816,
|
| 237 |
+
"grad_norm": 7.606313101162739,
|
| 238 |
+
"learning_rate": 1.5185139398778874e-05,
|
| 239 |
+
"loss": 2.5292,
|
| 240 |
+
"step": 33
|
| 241 |
+
},
|
| 242 |
+
{
|
| 243 |
+
"epoch": 0.29013333333333335,
|
| 244 |
+
"grad_norm": 8.553792493849265,
|
| 245 |
+
"learning_rate": 1.531478917042255e-05,
|
| 246 |
+
"loss": 2.4547,
|
| 247 |
+
"step": 34
|
| 248 |
+
},
|
| 249 |
+
{
|
| 250 |
+
"epoch": 0.2986666666666667,
|
| 251 |
+
"grad_norm": 8.483368703272543,
|
| 252 |
+
"learning_rate": 1.5440680443502753e-05,
|
| 253 |
+
"loss": 2.1956,
|
| 254 |
+
"step": 35
|
| 255 |
+
},
|
| 256 |
+
{
|
| 257 |
+
"epoch": 0.3072,
|
| 258 |
+
"grad_norm": 8.364739352838077,
|
| 259 |
+
"learning_rate": 1.5563025007672873e-05,
|
| 260 |
+
"loss": 1.8552,
|
| 261 |
+
"step": 36
|
| 262 |
+
},
|
| 263 |
+
{
|
| 264 |
+
"epoch": 0.3157333333333333,
|
| 265 |
+
"grad_norm": 9.37663682000104,
|
| 266 |
+
"learning_rate": 1.5682017240669948e-05,
|
| 267 |
+
"loss": 1.9228,
|
| 268 |
+
"step": 37
|
| 269 |
+
},
|
| 270 |
+
{
|
| 271 |
+
"epoch": 0.32426666666666665,
|
| 272 |
+
"grad_norm": 9.823047193440066,
|
| 273 |
+
"learning_rate": 1.57978359661681e-05,
|
| 274 |
+
"loss": 1.7033,
|
| 275 |
+
"step": 38
|
| 276 |
+
},
|
| 277 |
+
{
|
| 278 |
+
"epoch": 0.3328,
|
| 279 |
+
"grad_norm": 9.692618512955894,
|
| 280 |
+
"learning_rate": 1.591064607026499e-05,
|
| 281 |
+
"loss": 1.3768,
|
| 282 |
+
"step": 39
|
| 283 |
+
},
|
| 284 |
+
{
|
| 285 |
+
"epoch": 0.3413333333333333,
|
| 286 |
+
"grad_norm": 9.08889021911031,
|
| 287 |
+
"learning_rate": 1.6020599913279622e-05,
|
| 288 |
+
"loss": 1.3015,
|
| 289 |
+
"step": 40
|
| 290 |
+
},
|
| 291 |
+
{
|
| 292 |
+
"epoch": 0.34986666666666666,
|
| 293 |
+
"grad_norm": 8.081534221516058,
|
| 294 |
+
"learning_rate": 1.6127838567197353e-05,
|
| 295 |
+
"loss": 0.9228,
|
| 296 |
+
"step": 41
|
| 297 |
+
},
|
| 298 |
+
{
|
| 299 |
+
"epoch": 0.3584,
|
| 300 |
+
"grad_norm": 6.238638048950311,
|
| 301 |
+
"learning_rate": 1.6232492903979005e-05,
|
| 302 |
+
"loss": 0.7267,
|
| 303 |
+
"step": 42
|
| 304 |
+
},
|
| 305 |
+
{
|
| 306 |
+
"epoch": 0.36693333333333333,
|
| 307 |
+
"grad_norm": 3.4058036861773604,
|
| 308 |
+
"learning_rate": 1.6334684555795865e-05,
|
| 309 |
+
"loss": 0.5875,
|
| 310 |
+
"step": 43
|
| 311 |
+
},
|
| 312 |
+
{
|
| 313 |
+
"epoch": 0.37546666666666667,
|
| 314 |
+
"grad_norm": 2.079163829467713,
|
| 315 |
+
"learning_rate": 1.6434526764861872e-05,
|
| 316 |
+
"loss": 0.6355,
|
| 317 |
+
"step": 44
|
| 318 |
+
},
|
| 319 |
+
{
|
| 320 |
+
"epoch": 0.384,
|
| 321 |
+
"grad_norm": 1.5597487833024746,
|
| 322 |
+
"learning_rate": 1.6532125137753435e-05,
|
| 323 |
+
"loss": 0.5106,
|
| 324 |
+
"step": 45
|
| 325 |
+
},
|
| 326 |
+
{
|
| 327 |
+
"epoch": 0.39253333333333335,
|
| 328 |
+
"grad_norm": 2.491689602375256,
|
| 329 |
+
"learning_rate": 1.662757831681574e-05,
|
| 330 |
+
"loss": 0.6454,
|
| 331 |
+
"step": 46
|
| 332 |
+
},
|
| 333 |
+
{
|
| 334 |
+
"epoch": 0.4010666666666667,
|
| 335 |
+
"grad_norm": 2.010880438195854,
|
| 336 |
+
"learning_rate": 1.672097857935717e-05,
|
| 337 |
+
"loss": 0.4757,
|
| 338 |
+
"step": 47
|
| 339 |
+
},
|
| 340 |
+
{
|
| 341 |
+
"epoch": 0.4096,
|
| 342 |
+
"grad_norm": 1.9452805114322096,
|
| 343 |
+
"learning_rate": 1.681241237375587e-05,
|
| 344 |
+
"loss": 0.4133,
|
| 345 |
+
"step": 48
|
| 346 |
+
},
|
| 347 |
+
{
|
| 348 |
+
"epoch": 0.41813333333333336,
|
| 349 |
+
"grad_norm": 1.7620848552306103,
|
| 350 |
+
"learning_rate": 1.6901960800285137e-05,
|
| 351 |
+
"loss": 0.4004,
|
| 352 |
+
"step": 49
|
| 353 |
+
},
|
| 354 |
+
{
|
| 355 |
+
"epoch": 0.4266666666666667,
|
| 356 |
+
"grad_norm": 1.278224489774809,
|
| 357 |
+
"learning_rate": 1.6989700043360187e-05,
|
| 358 |
+
"loss": 0.3523,
|
| 359 |
+
"step": 50
|
| 360 |
+
},
|
| 361 |
+
{
|
| 362 |
+
"epoch": 0.4352,
|
| 363 |
+
"grad_norm": 1.6151354758303231,
|
| 364 |
+
"learning_rate": 1.7075701760979363e-05,
|
| 365 |
+
"loss": 0.4317,
|
| 366 |
+
"step": 51
|
| 367 |
+
},
|
| 368 |
+
{
|
| 369 |
+
"epoch": 0.4437333333333333,
|
| 370 |
+
"grad_norm": 1.3451396055695035,
|
| 371 |
+
"learning_rate": 1.716003343634799e-05,
|
| 372 |
+
"loss": 0.3474,
|
| 373 |
+
"step": 52
|
| 374 |
+
},
|
| 375 |
+
{
|
| 376 |
+
"epoch": 0.45226666666666665,
|
| 377 |
+
"grad_norm": 1.6814977782362666,
|
| 378 |
+
"learning_rate": 1.724275869600789e-05,
|
| 379 |
+
"loss": 0.3706,
|
| 380 |
+
"step": 53
|
| 381 |
+
},
|
| 382 |
+
{
|
| 383 |
+
"epoch": 0.4608,
|
| 384 |
+
"grad_norm": 1.1682442432688667,
|
| 385 |
+
"learning_rate": 1.7323937598229687e-05,
|
| 386 |
+
"loss": 0.3488,
|
| 387 |
+
"step": 54
|
| 388 |
+
},
|
| 389 |
+
{
|
| 390 |
+
"epoch": 0.4693333333333333,
|
| 391 |
+
"grad_norm": 0.8839814540462471,
|
| 392 |
+
"learning_rate": 1.7403626894942437e-05,
|
| 393 |
+
"loss": 0.293,
|
| 394 |
+
"step": 55
|
| 395 |
+
},
|
| 396 |
+
{
|
| 397 |
+
"epoch": 0.47786666666666666,
|
| 398 |
+
"grad_norm": 0.7974304806295485,
|
| 399 |
+
"learning_rate": 1.7481880270062003e-05,
|
| 400 |
+
"loss": 0.2717,
|
| 401 |
+
"step": 56
|
| 402 |
+
},
|
| 403 |
+
{
|
| 404 |
+
"epoch": 0.4864,
|
| 405 |
+
"grad_norm": 0.6232203657585239,
|
| 406 |
+
"learning_rate": 1.7558748556724913e-05,
|
| 407 |
+
"loss": 0.1741,
|
| 408 |
+
"step": 57
|
| 409 |
+
},
|
| 410 |
+
{
|
| 411 |
+
"epoch": 0.49493333333333334,
|
| 412 |
+
"grad_norm": 0.6850688604563008,
|
| 413 |
+
"learning_rate": 1.763427993562937e-05,
|
| 414 |
+
"loss": 0.228,
|
| 415 |
+
"step": 58
|
| 416 |
+
},
|
| 417 |
+
{
|
| 418 |
+
"epoch": 0.5034666666666666,
|
| 419 |
+
"grad_norm": 0.5923826384300431,
|
| 420 |
+
"learning_rate": 1.7708520116421443e-05,
|
| 421 |
+
"loss": 0.2131,
|
| 422 |
+
"step": 59
|
| 423 |
+
},
|
| 424 |
+
{
|
| 425 |
+
"epoch": 0.512,
|
| 426 |
+
"grad_norm": 0.489754430485032,
|
| 427 |
+
"learning_rate": 1.7781512503836432e-05,
|
| 428 |
+
"loss": 0.165,
|
| 429 |
+
"step": 60
|
| 430 |
+
},
|
| 431 |
+
{
|
| 432 |
+
"epoch": 0.5205333333333333,
|
| 433 |
+
"grad_norm": 0.5280940052395061,
|
| 434 |
+
"learning_rate": 1.7853298350107667e-05,
|
| 435 |
+
"loss": 0.1658,
|
| 436 |
+
"step": 61
|
| 437 |
+
},
|
| 438 |
+
{
|
| 439 |
+
"epoch": 0.5290666666666667,
|
| 440 |
+
"grad_norm": 0.4750905992036739,
|
| 441 |
+
"learning_rate": 1.7923916894982537e-05,
|
| 442 |
+
"loss": 0.1438,
|
| 443 |
+
"step": 62
|
| 444 |
+
},
|
| 445 |
+
{
|
| 446 |
+
"epoch": 0.5376,
|
| 447 |
+
"grad_norm": 0.4162998392722401,
|
| 448 |
+
"learning_rate": 1.7993405494535815e-05,
|
| 449 |
+
"loss": 0.1555,
|
| 450 |
+
"step": 63
|
| 451 |
+
},
|
| 452 |
+
{
|
| 453 |
+
"epoch": 0.5461333333333334,
|
| 454 |
+
"grad_norm": 0.26863266246370443,
|
| 455 |
+
"learning_rate": 1.806179973983887e-05,
|
| 456 |
+
"loss": 0.1323,
|
| 457 |
+
"step": 64
|
| 458 |
+
},
|
| 459 |
+
{
|
| 460 |
+
"epoch": 0.5546666666666666,
|
| 461 |
+
"grad_norm": 0.26534221125601215,
|
| 462 |
+
"learning_rate": 1.8129133566428553e-05,
|
| 463 |
+
"loss": 0.1671,
|
| 464 |
+
"step": 65
|
| 465 |
+
},
|
| 466 |
+
{
|
| 467 |
+
"epoch": 0.5632,
|
| 468 |
+
"grad_norm": 0.2548662962257576,
|
| 469 |
+
"learning_rate": 1.8195439355418686e-05,
|
| 470 |
+
"loss": 0.1308,
|
| 471 |
+
"step": 66
|
| 472 |
+
},
|
| 473 |
+
{
|
| 474 |
+
"epoch": 0.5717333333333333,
|
| 475 |
+
"grad_norm": 0.18045394638578796,
|
| 476 |
+
"learning_rate": 1.8260748027008263e-05,
|
| 477 |
+
"loss": 0.1262,
|
| 478 |
+
"step": 67
|
| 479 |
+
},
|
| 480 |
+
{
|
| 481 |
+
"epoch": 0.5802666666666667,
|
| 482 |
+
"grad_norm": 0.17070388073612064,
|
| 483 |
+
"learning_rate": 1.8325089127062364e-05,
|
| 484 |
+
"loss": 0.1192,
|
| 485 |
+
"step": 68
|
| 486 |
+
},
|
| 487 |
+
{
|
| 488 |
+
"epoch": 0.5888,
|
| 489 |
+
"grad_norm": 0.1531381679244776,
|
| 490 |
+
"learning_rate": 1.8388490907372553e-05,
|
| 491 |
+
"loss": 0.1274,
|
| 492 |
+
"step": 69
|
| 493 |
+
},
|
| 494 |
+
{
|
| 495 |
+
"epoch": 0.5973333333333334,
|
| 496 |
+
"grad_norm": 0.18196434993687946,
|
| 497 |
+
"learning_rate": 1.8450980400142568e-05,
|
| 498 |
+
"loss": 0.1375,
|
| 499 |
+
"step": 70
|
| 500 |
+
},
|
| 501 |
+
{
|
| 502 |
+
"epoch": 0.6058666666666667,
|
| 503 |
+
"grad_norm": 0.15324416972951205,
|
| 504 |
+
"learning_rate": 1.8512583487190752e-05,
|
| 505 |
+
"loss": 0.1599,
|
| 506 |
+
"step": 71
|
| 507 |
+
},
|
| 508 |
+
{
|
| 509 |
+
"epoch": 0.6144,
|
| 510 |
+
"grad_norm": 0.10884462064503801,
|
| 511 |
+
"learning_rate": 1.857332496431268e-05,
|
| 512 |
+
"loss": 0.1041,
|
| 513 |
+
"step": 72
|
| 514 |
+
},
|
| 515 |
+
{
|
| 516 |
+
"epoch": 0.6229333333333333,
|
| 517 |
+
"grad_norm": 0.12915133528192668,
|
| 518 |
+
"learning_rate": 1.8633228601204555e-05,
|
| 519 |
+
"loss": 0.1406,
|
| 520 |
+
"step": 73
|
| 521 |
+
},
|
| 522 |
+
{
|
| 523 |
+
"epoch": 0.6314666666666666,
|
| 524 |
+
"grad_norm": 0.12553425699952878,
|
| 525 |
+
"learning_rate": 1.8692317197309763e-05,
|
| 526 |
+
"loss": 0.1256,
|
| 527 |
+
"step": 74
|
| 528 |
+
},
|
| 529 |
+
{
|
| 530 |
+
"epoch": 0.64,
|
| 531 |
+
"grad_norm": 0.11976960918968543,
|
| 532 |
+
"learning_rate": 1.8750612633916997e-05,
|
| 533 |
+
"loss": 0.1144,
|
| 534 |
+
"step": 75
|
| 535 |
+
},
|
| 536 |
+
{
|
| 537 |
+
"epoch": 0.6485333333333333,
|
| 538 |
+
"grad_norm": 0.115805998298789,
|
| 539 |
+
"learning_rate": 1.8808135922807914e-05,
|
| 540 |
+
"loss": 0.1528,
|
| 541 |
+
"step": 76
|
| 542 |
+
},
|
| 543 |
+
{
|
| 544 |
+
"epoch": 0.6570666666666667,
|
| 545 |
+
"grad_norm": 0.10325948496697443,
|
| 546 |
+
"learning_rate": 1.8864907251724818e-05,
|
| 547 |
+
"loss": 0.1044,
|
| 548 |
+
"step": 77
|
| 549 |
+
},
|
| 550 |
+
{
|
| 551 |
+
"epoch": 0.6656,
|
| 552 |
+
"grad_norm": 0.09595064346541006,
|
| 553 |
+
"learning_rate": 1.8920946026904802e-05,
|
| 554 |
+
"loss": 0.1534,
|
| 555 |
+
"step": 78
|
| 556 |
+
},
|
| 557 |
+
{
|
| 558 |
+
"epoch": 0.6741333333333334,
|
| 559 |
+
"grad_norm": 0.08796742845240496,
|
| 560 |
+
"learning_rate": 1.8976270912904414e-05,
|
| 561 |
+
"loss": 0.1155,
|
| 562 |
+
"step": 79
|
| 563 |
+
},
|
| 564 |
+
{
|
| 565 |
+
"epoch": 0.6826666666666666,
|
| 566 |
+
"grad_norm": 0.08218991738379527,
|
| 567 |
+
"learning_rate": 1.9030899869919434e-05,
|
| 568 |
+
"loss": 0.1311,
|
| 569 |
+
"step": 80
|
| 570 |
+
},
|
| 571 |
+
{
|
| 572 |
+
"epoch": 0.6912,
|
| 573 |
+
"grad_norm": 0.08290815261109215,
|
| 574 |
+
"learning_rate": 1.9084850188786497e-05,
|
| 575 |
+
"loss": 0.11,
|
| 576 |
+
"step": 81
|
| 577 |
+
},
|
| 578 |
+
{
|
| 579 |
+
"epoch": 0.6997333333333333,
|
| 580 |
+
"grad_norm": 0.0794132180176064,
|
| 581 |
+
"learning_rate": 1.9138138523837165e-05,
|
| 582 |
+
"loss": 0.1135,
|
| 583 |
+
"step": 82
|
| 584 |
+
},
|
| 585 |
+
{
|
| 586 |
+
"epoch": 0.7082666666666667,
|
| 587 |
+
"grad_norm": 0.06934410705255296,
|
| 588 |
+
"learning_rate": 1.919078092376074e-05,
|
| 589 |
+
"loss": 0.109,
|
| 590 |
+
"step": 83
|
| 591 |
+
},
|
| 592 |
+
{
|
| 593 |
+
"epoch": 0.7168,
|
| 594 |
+
"grad_norm": 0.09000563031870593,
|
| 595 |
+
"learning_rate": 1.9242792860618813e-05,
|
| 596 |
+
"loss": 0.12,
|
| 597 |
+
"step": 84
|
| 598 |
+
},
|
| 599 |
+
{
|
| 600 |
+
"epoch": 0.7253333333333334,
|
| 601 |
+
"grad_norm": 0.1134042277879818,
|
| 602 |
+
"learning_rate": 1.929418925714293e-05,
|
| 603 |
+
"loss": 0.1223,
|
| 604 |
+
"step": 85
|
| 605 |
+
},
|
| 606 |
+
{
|
| 607 |
+
"epoch": 0.7338666666666667,
|
| 608 |
+
"grad_norm": 0.09118764690233076,
|
| 609 |
+
"learning_rate": 1.9344984512435673e-05,
|
| 610 |
+
"loss": 0.1459,
|
| 611 |
+
"step": 86
|
| 612 |
+
},
|
| 613 |
+
{
|
| 614 |
+
"epoch": 0.7424,
|
| 615 |
+
"grad_norm": 0.07873016754353963,
|
| 616 |
+
"learning_rate": 1.9395192526186183e-05,
|
| 617 |
+
"loss": 0.1422,
|
| 618 |
+
"step": 87
|
| 619 |
+
},
|
| 620 |
+
{
|
| 621 |
+
"epoch": 0.7509333333333333,
|
| 622 |
+
"grad_norm": 0.1796495874463076,
|
| 623 |
+
"learning_rate": 1.9444826721501687e-05,
|
| 624 |
+
"loss": 0.1291,
|
| 625 |
+
"step": 88
|
| 626 |
+
},
|
| 627 |
+
{
|
| 628 |
+
"epoch": 0.7594666666666666,
|
| 629 |
+
"grad_norm": 0.0679589944174269,
|
| 630 |
+
"learning_rate": 1.9493900066449125e-05,
|
| 631 |
+
"loss": 0.108,
|
| 632 |
+
"step": 89
|
| 633 |
+
},
|
| 634 |
+
{
|
| 635 |
+
"epoch": 0.768,
|
| 636 |
+
"grad_norm": 0.08174688574235538,
|
| 637 |
+
"learning_rate": 1.9542425094393246e-05,
|
| 638 |
+
"loss": 0.1081,
|
| 639 |
+
"step": 90
|
| 640 |
+
},
|
| 641 |
+
{
|
| 642 |
+
"epoch": 0.7765333333333333,
|
| 643 |
+
"grad_norm": 0.057137370501406756,
|
| 644 |
+
"learning_rate": 1.9590413923210934e-05,
|
| 645 |
+
"loss": 0.0934,
|
| 646 |
+
"step": 91
|
| 647 |
+
},
|
| 648 |
+
{
|
| 649 |
+
"epoch": 0.7850666666666667,
|
| 650 |
+
"grad_norm": 0.06578111924908255,
|
| 651 |
+
"learning_rate": 1.9637878273455555e-05,
|
| 652 |
+
"loss": 0.1085,
|
| 653 |
+
"step": 92
|
| 654 |
+
},
|
| 655 |
+
{
|
| 656 |
+
"epoch": 0.7936,
|
| 657 |
+
"grad_norm": 0.08945990540906254,
|
| 658 |
+
"learning_rate": 1.968482948553935e-05,
|
| 659 |
+
"loss": 0.1747,
|
| 660 |
+
"step": 93
|
| 661 |
+
},
|
| 662 |
+
{
|
| 663 |
+
"epoch": 0.8021333333333334,
|
| 664 |
+
"grad_norm": 0.06183863311044229,
|
| 665 |
+
"learning_rate": 1.9731278535996986e-05,
|
| 666 |
+
"loss": 0.1136,
|
| 667 |
+
"step": 94
|
| 668 |
+
},
|
| 669 |
+
{
|
| 670 |
+
"epoch": 0.8106666666666666,
|
| 671 |
+
"grad_norm": 0.05777899602544702,
|
| 672 |
+
"learning_rate": 1.9777236052888476e-05,
|
| 673 |
+
"loss": 0.0984,
|
| 674 |
+
"step": 95
|
| 675 |
+
},
|
| 676 |
+
{
|
| 677 |
+
"epoch": 0.8192,
|
| 678 |
+
"grad_norm": 0.08130851607693534,
|
| 679 |
+
"learning_rate": 1.9822712330395683e-05,
|
| 680 |
+
"loss": 0.187,
|
| 681 |
+
"step": 96
|
| 682 |
+
},
|
| 683 |
+
{
|
| 684 |
+
"epoch": 0.8277333333333333,
|
| 685 |
+
"grad_norm": 0.06426546202002927,
|
| 686 |
+
"learning_rate": 1.986771734266245e-05,
|
| 687 |
+
"loss": 0.1296,
|
| 688 |
+
"step": 97
|
| 689 |
+
},
|
| 690 |
+
{
|
| 691 |
+
"epoch": 0.8362666666666667,
|
| 692 |
+
"grad_norm": 0.069692313707994,
|
| 693 |
+
"learning_rate": 1.991226075692495e-05,
|
| 694 |
+
"loss": 0.1404,
|
| 695 |
+
"step": 98
|
| 696 |
+
},
|
| 697 |
+
{
|
| 698 |
+
"epoch": 0.8448,
|
| 699 |
+
"grad_norm": 0.05494542266886729,
|
| 700 |
+
"learning_rate": 1.9956351945975496e-05,
|
| 701 |
+
"loss": 0.116,
|
| 702 |
+
"step": 99
|
| 703 |
+
},
|
| 704 |
+
{
|
| 705 |
+
"epoch": 0.8533333333333334,
|
| 706 |
+
"grad_norm": 0.07571686966840627,
|
| 707 |
+
"learning_rate": 1.9999999999999998e-05,
|
| 708 |
+
"loss": 0.1539,
|
| 709 |
+
"step": 100
|
| 710 |
+
}
|
| 711 |
+
],
|
| 712 |
+
"logging_steps": 1,
|
| 713 |
+
"max_steps": 351,
|
| 714 |
+
"num_input_tokens_seen": 0,
|
| 715 |
+
"num_train_epochs": 3,
|
| 716 |
+
"save_steps": 20,
|
| 717 |
+
"stateful_callbacks": {
|
| 718 |
+
"TrainerControl": {
|
| 719 |
+
"args": {
|
| 720 |
+
"should_epoch_stop": false,
|
| 721 |
+
"should_evaluate": false,
|
| 722 |
+
"should_log": false,
|
| 723 |
+
"should_save": true,
|
| 724 |
+
"should_training_stop": false
|
| 725 |
+
},
|
| 726 |
+
"attributes": {}
|
| 727 |
+
}
|
| 728 |
+
},
|
| 729 |
+
"total_flos": 1.4183197916633498e+18,
|
| 730 |
+
"train_batch_size": 16,
|
| 731 |
+
"trial_name": null,
|
| 732 |
+
"trial_params": null
|
| 733 |
+
}
|
checkpoint-100/training_args.bin
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:aa4edad1da5557fcd52a6da980443588016cead6f0444a3562cfa68029c66a04
|
| 3 |
+
size 6840
|
checkpoint-100/zero_to_fp32.py
ADDED
|
@@ -0,0 +1,592 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 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):
|
| 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 |
+
elif zero_stage == 3:
|
| 216 |
+
return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
|
| 217 |
+
|
| 218 |
+
|
| 219 |
+
def _zero2_merge_frozen_params(state_dict, zero_model_states):
|
| 220 |
+
if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
|
| 221 |
+
return
|
| 222 |
+
|
| 223 |
+
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
|
| 224 |
+
frozen_param_fragments = zero_model_states[0].frozen_param_fragments
|
| 225 |
+
|
| 226 |
+
if debug:
|
| 227 |
+
num_elem = sum(s.numel() for s in frozen_param_shapes.values())
|
| 228 |
+
print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
|
| 229 |
+
|
| 230 |
+
wanted_params = len(frozen_param_shapes)
|
| 231 |
+
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
| 232 |
+
avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
|
| 233 |
+
print(f'Frozen params: Have {avail_numel} numels to process.')
|
| 234 |
+
print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
|
| 235 |
+
|
| 236 |
+
total_params = 0
|
| 237 |
+
total_numel = 0
|
| 238 |
+
for name, shape in frozen_param_shapes.items():
|
| 239 |
+
total_params += 1
|
| 240 |
+
unpartitioned_numel = shape.numel()
|
| 241 |
+
total_numel += unpartitioned_numel
|
| 242 |
+
|
| 243 |
+
state_dict[name] = frozen_param_fragments[name]
|
| 244 |
+
|
| 245 |
+
if debug:
|
| 246 |
+
print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
|
| 247 |
+
|
| 248 |
+
print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
|
| 249 |
+
|
| 250 |
+
|
| 251 |
+
def _has_callable(obj, fn):
|
| 252 |
+
attr = getattr(obj, fn, None)
|
| 253 |
+
return callable(attr)
|
| 254 |
+
|
| 255 |
+
|
| 256 |
+
def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
|
| 257 |
+
param_shapes = zero_model_states[0].param_shapes
|
| 258 |
+
|
| 259 |
+
# Reconstruction protocol:
|
| 260 |
+
#
|
| 261 |
+
# XXX: document this
|
| 262 |
+
|
| 263 |
+
if debug:
|
| 264 |
+
for i in range(world_size):
|
| 265 |
+
for j in range(len(fp32_flat_groups[0])):
|
| 266 |
+
print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
|
| 267 |
+
|
| 268 |
+
# XXX: memory usage doubles here (zero2)
|
| 269 |
+
num_param_groups = len(fp32_flat_groups[0])
|
| 270 |
+
merged_single_partition_of_fp32_groups = []
|
| 271 |
+
for i in range(num_param_groups):
|
| 272 |
+
merged_partitions = [sd[i] for sd in fp32_flat_groups]
|
| 273 |
+
full_single_fp32_vector = torch.cat(merged_partitions, 0)
|
| 274 |
+
merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
|
| 275 |
+
avail_numel = sum(
|
| 276 |
+
[full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
|
| 277 |
+
|
| 278 |
+
if debug:
|
| 279 |
+
wanted_params = sum([len(shapes) for shapes in param_shapes])
|
| 280 |
+
wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
|
| 281 |
+
# not asserting if there is a mismatch due to possible padding
|
| 282 |
+
print(f"Have {avail_numel} numels to process.")
|
| 283 |
+
print(f"Need {wanted_numel} numels in {wanted_params} params.")
|
| 284 |
+
|
| 285 |
+
# params
|
| 286 |
+
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
| 287 |
+
# out-of-core computing solution
|
| 288 |
+
total_numel = 0
|
| 289 |
+
total_params = 0
|
| 290 |
+
for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
|
| 291 |
+
offset = 0
|
| 292 |
+
avail_numel = full_single_fp32_vector.numel()
|
| 293 |
+
for name, shape in shapes.items():
|
| 294 |
+
|
| 295 |
+
unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
|
| 296 |
+
total_numel += unpartitioned_numel
|
| 297 |
+
total_params += 1
|
| 298 |
+
|
| 299 |
+
if debug:
|
| 300 |
+
print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
|
| 301 |
+
state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
|
| 302 |
+
offset += unpartitioned_numel
|
| 303 |
+
|
| 304 |
+
# Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
|
| 305 |
+
# avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
|
| 306 |
+
# paddings performed in the code it's almost impossible to predict the exact numbers w/o the
|
| 307 |
+
# live optimizer object, so we are checking that the numbers are within the right range
|
| 308 |
+
align_to = 2 * world_size
|
| 309 |
+
|
| 310 |
+
def zero2_align(x):
|
| 311 |
+
return align_to * math.ceil(x / align_to)
|
| 312 |
+
|
| 313 |
+
if debug:
|
| 314 |
+
print(f"original offset={offset}, avail_numel={avail_numel}")
|
| 315 |
+
|
| 316 |
+
offset = zero2_align(offset)
|
| 317 |
+
avail_numel = zero2_align(avail_numel)
|
| 318 |
+
|
| 319 |
+
if debug:
|
| 320 |
+
print(f"aligned offset={offset}, avail_numel={avail_numel}")
|
| 321 |
+
|
| 322 |
+
# Sanity check
|
| 323 |
+
if offset != avail_numel:
|
| 324 |
+
raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
| 325 |
+
|
| 326 |
+
print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
|
| 327 |
+
|
| 328 |
+
|
| 329 |
+
def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):
|
| 330 |
+
state_dict = OrderedDict()
|
| 331 |
+
|
| 332 |
+
# buffers
|
| 333 |
+
buffers = zero_model_states[0].buffers
|
| 334 |
+
state_dict.update(buffers)
|
| 335 |
+
if debug:
|
| 336 |
+
print(f"added {len(buffers)} buffers")
|
| 337 |
+
|
| 338 |
+
_zero2_merge_frozen_params(state_dict, zero_model_states)
|
| 339 |
+
|
| 340 |
+
_zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
|
| 341 |
+
|
| 342 |
+
# recover shared parameters
|
| 343 |
+
for pair in zero_model_states[0].shared_params:
|
| 344 |
+
if pair[1] in state_dict:
|
| 345 |
+
state_dict[pair[0]] = state_dict[pair[1]]
|
| 346 |
+
|
| 347 |
+
return state_dict
|
| 348 |
+
|
| 349 |
+
|
| 350 |
+
def zero3_partitioned_param_info(unpartitioned_numel, world_size):
|
| 351 |
+
remainder = unpartitioned_numel % world_size
|
| 352 |
+
padding_numel = (world_size - remainder) if remainder else 0
|
| 353 |
+
partitioned_numel = math.ceil(unpartitioned_numel / world_size)
|
| 354 |
+
return partitioned_numel, padding_numel
|
| 355 |
+
|
| 356 |
+
|
| 357 |
+
def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
|
| 358 |
+
if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
|
| 359 |
+
return
|
| 360 |
+
|
| 361 |
+
if debug:
|
| 362 |
+
for i in range(world_size):
|
| 363 |
+
num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
|
| 364 |
+
print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
|
| 365 |
+
|
| 366 |
+
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
|
| 367 |
+
wanted_params = len(frozen_param_shapes)
|
| 368 |
+
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
| 369 |
+
avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
|
| 370 |
+
print(f'Frozen params: Have {avail_numel} numels to process.')
|
| 371 |
+
print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
|
| 372 |
+
|
| 373 |
+
total_params = 0
|
| 374 |
+
total_numel = 0
|
| 375 |
+
for name, shape in zero_model_states[0].frozen_param_shapes.items():
|
| 376 |
+
total_params += 1
|
| 377 |
+
unpartitioned_numel = shape.numel()
|
| 378 |
+
total_numel += unpartitioned_numel
|
| 379 |
+
|
| 380 |
+
param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
|
| 381 |
+
state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
|
| 382 |
+
|
| 383 |
+
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
| 384 |
+
|
| 385 |
+
if debug:
|
| 386 |
+
print(
|
| 387 |
+
f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
| 388 |
+
)
|
| 389 |
+
|
| 390 |
+
print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
|
| 391 |
+
|
| 392 |
+
|
| 393 |
+
def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
|
| 394 |
+
param_shapes = zero_model_states[0].param_shapes
|
| 395 |
+
avail_numel = fp32_flat_groups[0].numel() * world_size
|
| 396 |
+
# Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
|
| 397 |
+
# param, re-consolidating each param, while dealing with padding if any
|
| 398 |
+
|
| 399 |
+
# merge list of dicts, preserving order
|
| 400 |
+
param_shapes = {k: v for d in param_shapes for k, v in d.items()}
|
| 401 |
+
|
| 402 |
+
if debug:
|
| 403 |
+
for i in range(world_size):
|
| 404 |
+
print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
|
| 405 |
+
|
| 406 |
+
wanted_params = len(param_shapes)
|
| 407 |
+
wanted_numel = sum(shape.numel() for shape in param_shapes.values())
|
| 408 |
+
# not asserting if there is a mismatch due to possible padding
|
| 409 |
+
avail_numel = fp32_flat_groups[0].numel() * world_size
|
| 410 |
+
print(f"Trainable params: Have {avail_numel} numels to process.")
|
| 411 |
+
print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
|
| 412 |
+
|
| 413 |
+
# params
|
| 414 |
+
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
| 415 |
+
# out-of-core computing solution
|
| 416 |
+
offset = 0
|
| 417 |
+
total_numel = 0
|
| 418 |
+
total_params = 0
|
| 419 |
+
for name, shape in param_shapes.items():
|
| 420 |
+
|
| 421 |
+
unpartitioned_numel = shape.numel()
|
| 422 |
+
total_numel += unpartitioned_numel
|
| 423 |
+
total_params += 1
|
| 424 |
+
|
| 425 |
+
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
| 426 |
+
|
| 427 |
+
if debug:
|
| 428 |
+
print(
|
| 429 |
+
f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
| 430 |
+
)
|
| 431 |
+
|
| 432 |
+
# XXX: memory usage doubles here
|
| 433 |
+
state_dict[name] = torch.cat(
|
| 434 |
+
tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
|
| 435 |
+
0).narrow(0, 0, unpartitioned_numel).view(shape)
|
| 436 |
+
offset += partitioned_numel
|
| 437 |
+
|
| 438 |
+
offset *= world_size
|
| 439 |
+
|
| 440 |
+
# Sanity check
|
| 441 |
+
if offset != avail_numel:
|
| 442 |
+
raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
| 443 |
+
|
| 444 |
+
print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
|
| 445 |
+
|
| 446 |
+
|
| 447 |
+
def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):
|
| 448 |
+
state_dict = OrderedDict()
|
| 449 |
+
|
| 450 |
+
# buffers
|
| 451 |
+
buffers = zero_model_states[0].buffers
|
| 452 |
+
state_dict.update(buffers)
|
| 453 |
+
if debug:
|
| 454 |
+
print(f"added {len(buffers)} buffers")
|
| 455 |
+
|
| 456 |
+
_zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
|
| 457 |
+
|
| 458 |
+
_zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
|
| 459 |
+
|
| 460 |
+
# recover shared parameters
|
| 461 |
+
for pair in zero_model_states[0].shared_params:
|
| 462 |
+
if pair[1] in state_dict:
|
| 463 |
+
state_dict[pair[0]] = state_dict[pair[1]]
|
| 464 |
+
|
| 465 |
+
return state_dict
|
| 466 |
+
|
| 467 |
+
|
| 468 |
+
def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
|
| 469 |
+
"""
|
| 470 |
+
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
|
| 471 |
+
``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
|
| 472 |
+
via a model hub.
|
| 473 |
+
|
| 474 |
+
Args:
|
| 475 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder
|
| 476 |
+
- ``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``
|
| 477 |
+
|
| 478 |
+
Returns:
|
| 479 |
+
- pytorch ``state_dict``
|
| 480 |
+
|
| 481 |
+
Note: this approach may not work if your application doesn't have sufficient free CPU memory and
|
| 482 |
+
you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
|
| 483 |
+
the checkpoint.
|
| 484 |
+
|
| 485 |
+
A typical usage might be ::
|
| 486 |
+
|
| 487 |
+
from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
|
| 488 |
+
# do the training and checkpoint saving
|
| 489 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
|
| 490 |
+
model = model.cpu() # move to cpu
|
| 491 |
+
model.load_state_dict(state_dict)
|
| 492 |
+
# submit to model hub or save the model to share with others
|
| 493 |
+
|
| 494 |
+
In this example the ``model`` will no longer be usable in the deepspeed context of the same
|
| 495 |
+
application. i.e. you will need to re-initialize the deepspeed engine, since
|
| 496 |
+
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
| 497 |
+
|
| 498 |
+
If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
|
| 499 |
+
|
| 500 |
+
"""
|
| 501 |
+
if tag is None:
|
| 502 |
+
latest_path = os.path.join(checkpoint_dir, 'latest')
|
| 503 |
+
if os.path.isfile(latest_path):
|
| 504 |
+
with open(latest_path, 'r') as fd:
|
| 505 |
+
tag = fd.read().strip()
|
| 506 |
+
else:
|
| 507 |
+
raise ValueError(f"Unable to find 'latest' file at {latest_path}")
|
| 508 |
+
|
| 509 |
+
ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
|
| 510 |
+
|
| 511 |
+
if not os.path.isdir(ds_checkpoint_dir):
|
| 512 |
+
raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
|
| 513 |
+
|
| 514 |
+
return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
|
| 515 |
+
|
| 516 |
+
|
| 517 |
+
def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
|
| 518 |
+
"""
|
| 519 |
+
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
|
| 520 |
+
loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
|
| 521 |
+
|
| 522 |
+
Args:
|
| 523 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
| 524 |
+
- ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
|
| 525 |
+
- ``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``
|
| 526 |
+
"""
|
| 527 |
+
|
| 528 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
| 529 |
+
print(f"Saving fp32 state dict to {output_file}")
|
| 530 |
+
torch.save(state_dict, output_file)
|
| 531 |
+
|
| 532 |
+
|
| 533 |
+
def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
|
| 534 |
+
"""
|
| 535 |
+
1. Put the provided model to cpu
|
| 536 |
+
2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
|
| 537 |
+
3. Load it into the provided model
|
| 538 |
+
|
| 539 |
+
Args:
|
| 540 |
+
- ``model``: the model object to update
|
| 541 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
| 542 |
+
- ``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``
|
| 543 |
+
|
| 544 |
+
Returns:
|
| 545 |
+
- ``model`: modified model
|
| 546 |
+
|
| 547 |
+
Make sure you have plenty of CPU memory available before you call this function. If you don't
|
| 548 |
+
have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
|
| 549 |
+
conveniently placed for you in the checkpoint folder.
|
| 550 |
+
|
| 551 |
+
A typical usage might be ::
|
| 552 |
+
|
| 553 |
+
from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
|
| 554 |
+
model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
|
| 555 |
+
# submit to model hub or save the model to share with others
|
| 556 |
+
|
| 557 |
+
Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
|
| 558 |
+
of the same application. i.e. you will need to re-initialize the deepspeed engine, since
|
| 559 |
+
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
| 560 |
+
|
| 561 |
+
"""
|
| 562 |
+
logger.info(f"Extracting fp32 weights")
|
| 563 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
| 564 |
+
|
| 565 |
+
logger.info(f"Overwriting model with fp32 weights")
|
| 566 |
+
model = model.cpu()
|
| 567 |
+
model.load_state_dict(state_dict, strict=False)
|
| 568 |
+
|
| 569 |
+
return model
|
| 570 |
+
|
| 571 |
+
|
| 572 |
+
if __name__ == "__main__":
|
| 573 |
+
|
| 574 |
+
parser = argparse.ArgumentParser()
|
| 575 |
+
parser.add_argument("checkpoint_dir",
|
| 576 |
+
type=str,
|
| 577 |
+
help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
|
| 578 |
+
parser.add_argument(
|
| 579 |
+
"output_file",
|
| 580 |
+
type=str,
|
| 581 |
+
help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
|
| 582 |
+
parser.add_argument("-t",
|
| 583 |
+
"--tag",
|
| 584 |
+
type=str,
|
| 585 |
+
default=None,
|
| 586 |
+
help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
|
| 587 |
+
parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
|
| 588 |
+
args = parser.parse_args()
|
| 589 |
+
|
| 590 |
+
debug = args.debug
|
| 591 |
+
|
| 592 |
+
convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file, tag=args.tag)
|
checkpoint-120/README.md
ADDED
|
@@ -0,0 +1,202 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: peft
|
| 3 |
+
base_model: ../ckpts/Meta-Llama-3-8B-Instruct
|
| 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 0.11.1
|
checkpoint-120/adapter_config.json
ADDED
|
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"alpha_pattern": {},
|
| 3 |
+
"auto_mapping": null,
|
| 4 |
+
"base_model_name_or_path": "../ckpts/Meta-Llama-3-8B-Instruct",
|
| 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": 32,
|
| 14 |
+
"lora_dropout": 0.1,
|
| 15 |
+
"megatron_config": null,
|
| 16 |
+
"megatron_core": "megatron.core",
|
| 17 |
+
"modules_to_save": null,
|
| 18 |
+
"peft_type": "LORA",
|
| 19 |
+
"r": 16,
|
| 20 |
+
"rank_pattern": {},
|
| 21 |
+
"revision": null,
|
| 22 |
+
"target_modules": [
|
| 23 |
+
"down_proj",
|
| 24 |
+
"gate_proj",
|
| 25 |
+
"q_proj",
|
| 26 |
+
"lm_head",
|
| 27 |
+
"o_proj",
|
| 28 |
+
"v_proj",
|
| 29 |
+
"up_proj",
|
| 30 |
+
"k_proj"
|
| 31 |
+
],
|
| 32 |
+
"task_type": "CAUSAL_LM",
|
| 33 |
+
"use_dora": false,
|
| 34 |
+
"use_rslora": false
|
| 35 |
+
}
|
checkpoint-120/adapter_model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:33e6839258702615da06d8aa30c9fd67b2657cce43937f62cf3927f85b0a8362
|
| 3 |
+
size 1138856856
|
checkpoint-120/global_step120/bf16_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 sha256:11b436ebd244ff8f8485f00d23b9c863c009715ce5d24e64a356edf305f17b6e
|
| 3 |
+
size 528781328
|
checkpoint-120/global_step120/mp_rank_00_model_states.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2c17e1cbdbbc34d79755357fd83899cf1a03c70d31c45b9c5104634809569283
|
| 3 |
+
size 199905337
|
checkpoint-120/latest
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
global_step120
|
checkpoint-120/rng_state.pth
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:9ae17b05257a556f53e7168228629d1bb3ae5e5c9000a8d7dd9d2cc833ef3948
|
| 3 |
+
size 14244
|
checkpoint-120/trainer_state.json
ADDED
|
@@ -0,0 +1,873 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"best_metric": null,
|
| 3 |
+
"best_model_checkpoint": null,
|
| 4 |
+
"epoch": 1.024,
|
| 5 |
+
"eval_steps": 500,
|
| 6 |
+
"global_step": 120,
|
| 7 |
+
"is_hyper_param_search": false,
|
| 8 |
+
"is_local_process_zero": true,
|
| 9 |
+
"is_world_process_zero": true,
|
| 10 |
+
"log_history": [
|
| 11 |
+
{
|
| 12 |
+
"epoch": 0.008533333333333334,
|
| 13 |
+
"grad_norm": 100.21848203113535,
|
| 14 |
+
"learning_rate": 0.0,
|
| 15 |
+
"loss": 7.1962,
|
| 16 |
+
"step": 1
|
| 17 |
+
},
|
| 18 |
+
{
|
| 19 |
+
"epoch": 0.017066666666666667,
|
| 20 |
+
"grad_norm": 96.43006188910957,
|
| 21 |
+
"learning_rate": 3.0102999566398115e-06,
|
| 22 |
+
"loss": 6.9414,
|
| 23 |
+
"step": 2
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"epoch": 0.0256,
|
| 27 |
+
"grad_norm": 97.35803466618715,
|
| 28 |
+
"learning_rate": 4.771212547196624e-06,
|
| 29 |
+
"loss": 7.0102,
|
| 30 |
+
"step": 3
|
| 31 |
+
},
|
| 32 |
+
{
|
| 33 |
+
"epoch": 0.034133333333333335,
|
| 34 |
+
"grad_norm": 95.14837816372646,
|
| 35 |
+
"learning_rate": 6.020599913279623e-06,
|
| 36 |
+
"loss": 6.5295,
|
| 37 |
+
"step": 4
|
| 38 |
+
},
|
| 39 |
+
{
|
| 40 |
+
"epoch": 0.042666666666666665,
|
| 41 |
+
"grad_norm": 91.76544275692784,
|
| 42 |
+
"learning_rate": 6.989700043360187e-06,
|
| 43 |
+
"loss": 6.4806,
|
| 44 |
+
"step": 5
|
| 45 |
+
},
|
| 46 |
+
{
|
| 47 |
+
"epoch": 0.0512,
|
| 48 |
+
"grad_norm": 84.4494318688335,
|
| 49 |
+
"learning_rate": 7.781512503836437e-06,
|
| 50 |
+
"loss": 6.4194,
|
| 51 |
+
"step": 6
|
| 52 |
+
},
|
| 53 |
+
{
|
| 54 |
+
"epoch": 0.05973333333333333,
|
| 55 |
+
"grad_norm": 71.37977490595638,
|
| 56 |
+
"learning_rate": 8.450980400142568e-06,
|
| 57 |
+
"loss": 5.4953,
|
| 58 |
+
"step": 7
|
| 59 |
+
},
|
| 60 |
+
{
|
| 61 |
+
"epoch": 0.06826666666666667,
|
| 62 |
+
"grad_norm": 49.31153456754566,
|
| 63 |
+
"learning_rate": 9.030899869919434e-06,
|
| 64 |
+
"loss": 5.4123,
|
| 65 |
+
"step": 8
|
| 66 |
+
},
|
| 67 |
+
{
|
| 68 |
+
"epoch": 0.0768,
|
| 69 |
+
"grad_norm": 20.37296364560341,
|
| 70 |
+
"learning_rate": 9.542425094393249e-06,
|
| 71 |
+
"loss": 5.2334,
|
| 72 |
+
"step": 9
|
| 73 |
+
},
|
| 74 |
+
{
|
| 75 |
+
"epoch": 0.08533333333333333,
|
| 76 |
+
"grad_norm": 7.968467079076881,
|
| 77 |
+
"learning_rate": 9.999999999999999e-06,
|
| 78 |
+
"loss": 5.0282,
|
| 79 |
+
"step": 10
|
| 80 |
+
},
|
| 81 |
+
{
|
| 82 |
+
"epoch": 0.09386666666666667,
|
| 83 |
+
"grad_norm": 3.559446532055649,
|
| 84 |
+
"learning_rate": 1.041392685158225e-05,
|
| 85 |
+
"loss": 4.612,
|
| 86 |
+
"step": 11
|
| 87 |
+
},
|
| 88 |
+
{
|
| 89 |
+
"epoch": 0.1024,
|
| 90 |
+
"grad_norm": 3.5528846947995674,
|
| 91 |
+
"learning_rate": 1.0791812460476248e-05,
|
| 92 |
+
"loss": 4.9475,
|
| 93 |
+
"step": 12
|
| 94 |
+
},
|
| 95 |
+
{
|
| 96 |
+
"epoch": 0.11093333333333333,
|
| 97 |
+
"grad_norm": 3.541968897471334,
|
| 98 |
+
"learning_rate": 1.1139433523068365e-05,
|
| 99 |
+
"loss": 4.2777,
|
| 100 |
+
"step": 13
|
| 101 |
+
},
|
| 102 |
+
{
|
| 103 |
+
"epoch": 0.11946666666666667,
|
| 104 |
+
"grad_norm": 3.54718070036198,
|
| 105 |
+
"learning_rate": 1.1461280356782378e-05,
|
| 106 |
+
"loss": 4.3507,
|
| 107 |
+
"step": 14
|
| 108 |
+
},
|
| 109 |
+
{
|
| 110 |
+
"epoch": 0.128,
|
| 111 |
+
"grad_norm": 3.8632334830606747,
|
| 112 |
+
"learning_rate": 1.1760912590556813e-05,
|
| 113 |
+
"loss": 4.5364,
|
| 114 |
+
"step": 15
|
| 115 |
+
},
|
| 116 |
+
{
|
| 117 |
+
"epoch": 0.13653333333333334,
|
| 118 |
+
"grad_norm": 3.6637424744054004,
|
| 119 |
+
"learning_rate": 1.2041199826559246e-05,
|
| 120 |
+
"loss": 3.9672,
|
| 121 |
+
"step": 16
|
| 122 |
+
},
|
| 123 |
+
{
|
| 124 |
+
"epoch": 0.14506666666666668,
|
| 125 |
+
"grad_norm": 3.919802904818311,
|
| 126 |
+
"learning_rate": 1.230448921378274e-05,
|
| 127 |
+
"loss": 4.0618,
|
| 128 |
+
"step": 17
|
| 129 |
+
},
|
| 130 |
+
{
|
| 131 |
+
"epoch": 0.1536,
|
| 132 |
+
"grad_norm": 4.71904950738746,
|
| 133 |
+
"learning_rate": 1.2552725051033058e-05,
|
| 134 |
+
"loss": 4.6656,
|
| 135 |
+
"step": 18
|
| 136 |
+
},
|
| 137 |
+
{
|
| 138 |
+
"epoch": 0.16213333333333332,
|
| 139 |
+
"grad_norm": 4.6656317698690835,
|
| 140 |
+
"learning_rate": 1.2787536009528288e-05,
|
| 141 |
+
"loss": 4.1131,
|
| 142 |
+
"step": 19
|
| 143 |
+
},
|
| 144 |
+
{
|
| 145 |
+
"epoch": 0.17066666666666666,
|
| 146 |
+
"grad_norm": 5.145138692367417,
|
| 147 |
+
"learning_rate": 1.301029995663981e-05,
|
| 148 |
+
"loss": 4.0989,
|
| 149 |
+
"step": 20
|
| 150 |
+
},
|
| 151 |
+
{
|
| 152 |
+
"epoch": 0.1792,
|
| 153 |
+
"grad_norm": 5.923538014759818,
|
| 154 |
+
"learning_rate": 1.3222192947339192e-05,
|
| 155 |
+
"loss": 4.4991,
|
| 156 |
+
"step": 21
|
| 157 |
+
},
|
| 158 |
+
{
|
| 159 |
+
"epoch": 0.18773333333333334,
|
| 160 |
+
"grad_norm": 5.941056962941364,
|
| 161 |
+
"learning_rate": 1.3424226808222062e-05,
|
| 162 |
+
"loss": 4.0836,
|
| 163 |
+
"step": 22
|
| 164 |
+
},
|
| 165 |
+
{
|
| 166 |
+
"epoch": 0.19626666666666667,
|
| 167 |
+
"grad_norm": 6.171026012117947,
|
| 168 |
+
"learning_rate": 1.3617278360175927e-05,
|
| 169 |
+
"loss": 3.6861,
|
| 170 |
+
"step": 23
|
| 171 |
+
},
|
| 172 |
+
{
|
| 173 |
+
"epoch": 0.2048,
|
| 174 |
+
"grad_norm": 7.130542138930838,
|
| 175 |
+
"learning_rate": 1.380211241711606e-05,
|
| 176 |
+
"loss": 4.0958,
|
| 177 |
+
"step": 24
|
| 178 |
+
},
|
| 179 |
+
{
|
| 180 |
+
"epoch": 0.21333333333333335,
|
| 181 |
+
"grad_norm": 7.328837606110418,
|
| 182 |
+
"learning_rate": 1.3979400086720374e-05,
|
| 183 |
+
"loss": 3.9524,
|
| 184 |
+
"step": 25
|
| 185 |
+
},
|
| 186 |
+
{
|
| 187 |
+
"epoch": 0.22186666666666666,
|
| 188 |
+
"grad_norm": 6.923489005711429,
|
| 189 |
+
"learning_rate": 1.4149733479708178e-05,
|
| 190 |
+
"loss": 3.6062,
|
| 191 |
+
"step": 26
|
| 192 |
+
},
|
| 193 |
+
{
|
| 194 |
+
"epoch": 0.2304,
|
| 195 |
+
"grad_norm": 6.409498926059221,
|
| 196 |
+
"learning_rate": 1.4313637641589872e-05,
|
| 197 |
+
"loss": 3.2034,
|
| 198 |
+
"step": 27
|
| 199 |
+
},
|
| 200 |
+
{
|
| 201 |
+
"epoch": 0.23893333333333333,
|
| 202 |
+
"grad_norm": 5.781628405584682,
|
| 203 |
+
"learning_rate": 1.4471580313422191e-05,
|
| 204 |
+
"loss": 2.8158,
|
| 205 |
+
"step": 28
|
| 206 |
+
},
|
| 207 |
+
{
|
| 208 |
+
"epoch": 0.24746666666666667,
|
| 209 |
+
"grad_norm": 6.2927590068243315,
|
| 210 |
+
"learning_rate": 1.4623979978989559e-05,
|
| 211 |
+
"loss": 2.9803,
|
| 212 |
+
"step": 29
|
| 213 |
+
},
|
| 214 |
+
{
|
| 215 |
+
"epoch": 0.256,
|
| 216 |
+
"grad_norm": 6.103844678630006,
|
| 217 |
+
"learning_rate": 1.4771212547196623e-05,
|
| 218 |
+
"loss": 2.847,
|
| 219 |
+
"step": 30
|
| 220 |
+
},
|
| 221 |
+
{
|
| 222 |
+
"epoch": 0.26453333333333334,
|
| 223 |
+
"grad_norm": 7.656341925867144,
|
| 224 |
+
"learning_rate": 1.4913616938342726e-05,
|
| 225 |
+
"loss": 3.0907,
|
| 226 |
+
"step": 31
|
| 227 |
+
},
|
| 228 |
+
{
|
| 229 |
+
"epoch": 0.2730666666666667,
|
| 230 |
+
"grad_norm": 6.324242877501844,
|
| 231 |
+
"learning_rate": 1.5051499783199059e-05,
|
| 232 |
+
"loss": 2.3467,
|
| 233 |
+
"step": 32
|
| 234 |
+
},
|
| 235 |
+
{
|
| 236 |
+
"epoch": 0.2816,
|
| 237 |
+
"grad_norm": 7.606313101162739,
|
| 238 |
+
"learning_rate": 1.5185139398778874e-05,
|
| 239 |
+
"loss": 2.5292,
|
| 240 |
+
"step": 33
|
| 241 |
+
},
|
| 242 |
+
{
|
| 243 |
+
"epoch": 0.29013333333333335,
|
| 244 |
+
"grad_norm": 8.553792493849265,
|
| 245 |
+
"learning_rate": 1.531478917042255e-05,
|
| 246 |
+
"loss": 2.4547,
|
| 247 |
+
"step": 34
|
| 248 |
+
},
|
| 249 |
+
{
|
| 250 |
+
"epoch": 0.2986666666666667,
|
| 251 |
+
"grad_norm": 8.483368703272543,
|
| 252 |
+
"learning_rate": 1.5440680443502753e-05,
|
| 253 |
+
"loss": 2.1956,
|
| 254 |
+
"step": 35
|
| 255 |
+
},
|
| 256 |
+
{
|
| 257 |
+
"epoch": 0.3072,
|
| 258 |
+
"grad_norm": 8.364739352838077,
|
| 259 |
+
"learning_rate": 1.5563025007672873e-05,
|
| 260 |
+
"loss": 1.8552,
|
| 261 |
+
"step": 36
|
| 262 |
+
},
|
| 263 |
+
{
|
| 264 |
+
"epoch": 0.3157333333333333,
|
| 265 |
+
"grad_norm": 9.37663682000104,
|
| 266 |
+
"learning_rate": 1.5682017240669948e-05,
|
| 267 |
+
"loss": 1.9228,
|
| 268 |
+
"step": 37
|
| 269 |
+
},
|
| 270 |
+
{
|
| 271 |
+
"epoch": 0.32426666666666665,
|
| 272 |
+
"grad_norm": 9.823047193440066,
|
| 273 |
+
"learning_rate": 1.57978359661681e-05,
|
| 274 |
+
"loss": 1.7033,
|
| 275 |
+
"step": 38
|
| 276 |
+
},
|
| 277 |
+
{
|
| 278 |
+
"epoch": 0.3328,
|
| 279 |
+
"grad_norm": 9.692618512955894,
|
| 280 |
+
"learning_rate": 1.591064607026499e-05,
|
| 281 |
+
"loss": 1.3768,
|
| 282 |
+
"step": 39
|
| 283 |
+
},
|
| 284 |
+
{
|
| 285 |
+
"epoch": 0.3413333333333333,
|
| 286 |
+
"grad_norm": 9.08889021911031,
|
| 287 |
+
"learning_rate": 1.6020599913279622e-05,
|
| 288 |
+
"loss": 1.3015,
|
| 289 |
+
"step": 40
|
| 290 |
+
},
|
| 291 |
+
{
|
| 292 |
+
"epoch": 0.34986666666666666,
|
| 293 |
+
"grad_norm": 8.081534221516058,
|
| 294 |
+
"learning_rate": 1.6127838567197353e-05,
|
| 295 |
+
"loss": 0.9228,
|
| 296 |
+
"step": 41
|
| 297 |
+
},
|
| 298 |
+
{
|
| 299 |
+
"epoch": 0.3584,
|
| 300 |
+
"grad_norm": 6.238638048950311,
|
| 301 |
+
"learning_rate": 1.6232492903979005e-05,
|
| 302 |
+
"loss": 0.7267,
|
| 303 |
+
"step": 42
|
| 304 |
+
},
|
| 305 |
+
{
|
| 306 |
+
"epoch": 0.36693333333333333,
|
| 307 |
+
"grad_norm": 3.4058036861773604,
|
| 308 |
+
"learning_rate": 1.6334684555795865e-05,
|
| 309 |
+
"loss": 0.5875,
|
| 310 |
+
"step": 43
|
| 311 |
+
},
|
| 312 |
+
{
|
| 313 |
+
"epoch": 0.37546666666666667,
|
| 314 |
+
"grad_norm": 2.079163829467713,
|
| 315 |
+
"learning_rate": 1.6434526764861872e-05,
|
| 316 |
+
"loss": 0.6355,
|
| 317 |
+
"step": 44
|
| 318 |
+
},
|
| 319 |
+
{
|
| 320 |
+
"epoch": 0.384,
|
| 321 |
+
"grad_norm": 1.5597487833024746,
|
| 322 |
+
"learning_rate": 1.6532125137753435e-05,
|
| 323 |
+
"loss": 0.5106,
|
| 324 |
+
"step": 45
|
| 325 |
+
},
|
| 326 |
+
{
|
| 327 |
+
"epoch": 0.39253333333333335,
|
| 328 |
+
"grad_norm": 2.491689602375256,
|
| 329 |
+
"learning_rate": 1.662757831681574e-05,
|
| 330 |
+
"loss": 0.6454,
|
| 331 |
+
"step": 46
|
| 332 |
+
},
|
| 333 |
+
{
|
| 334 |
+
"epoch": 0.4010666666666667,
|
| 335 |
+
"grad_norm": 2.010880438195854,
|
| 336 |
+
"learning_rate": 1.672097857935717e-05,
|
| 337 |
+
"loss": 0.4757,
|
| 338 |
+
"step": 47
|
| 339 |
+
},
|
| 340 |
+
{
|
| 341 |
+
"epoch": 0.4096,
|
| 342 |
+
"grad_norm": 1.9452805114322096,
|
| 343 |
+
"learning_rate": 1.681241237375587e-05,
|
| 344 |
+
"loss": 0.4133,
|
| 345 |
+
"step": 48
|
| 346 |
+
},
|
| 347 |
+
{
|
| 348 |
+
"epoch": 0.41813333333333336,
|
| 349 |
+
"grad_norm": 1.7620848552306103,
|
| 350 |
+
"learning_rate": 1.6901960800285137e-05,
|
| 351 |
+
"loss": 0.4004,
|
| 352 |
+
"step": 49
|
| 353 |
+
},
|
| 354 |
+
{
|
| 355 |
+
"epoch": 0.4266666666666667,
|
| 356 |
+
"grad_norm": 1.278224489774809,
|
| 357 |
+
"learning_rate": 1.6989700043360187e-05,
|
| 358 |
+
"loss": 0.3523,
|
| 359 |
+
"step": 50
|
| 360 |
+
},
|
| 361 |
+
{
|
| 362 |
+
"epoch": 0.4352,
|
| 363 |
+
"grad_norm": 1.6151354758303231,
|
| 364 |
+
"learning_rate": 1.7075701760979363e-05,
|
| 365 |
+
"loss": 0.4317,
|
| 366 |
+
"step": 51
|
| 367 |
+
},
|
| 368 |
+
{
|
| 369 |
+
"epoch": 0.4437333333333333,
|
| 370 |
+
"grad_norm": 1.3451396055695035,
|
| 371 |
+
"learning_rate": 1.716003343634799e-05,
|
| 372 |
+
"loss": 0.3474,
|
| 373 |
+
"step": 52
|
| 374 |
+
},
|
| 375 |
+
{
|
| 376 |
+
"epoch": 0.45226666666666665,
|
| 377 |
+
"grad_norm": 1.6814977782362666,
|
| 378 |
+
"learning_rate": 1.724275869600789e-05,
|
| 379 |
+
"loss": 0.3706,
|
| 380 |
+
"step": 53
|
| 381 |
+
},
|
| 382 |
+
{
|
| 383 |
+
"epoch": 0.4608,
|
| 384 |
+
"grad_norm": 1.1682442432688667,
|
| 385 |
+
"learning_rate": 1.7323937598229687e-05,
|
| 386 |
+
"loss": 0.3488,
|
| 387 |
+
"step": 54
|
| 388 |
+
},
|
| 389 |
+
{
|
| 390 |
+
"epoch": 0.4693333333333333,
|
| 391 |
+
"grad_norm": 0.8839814540462471,
|
| 392 |
+
"learning_rate": 1.7403626894942437e-05,
|
| 393 |
+
"loss": 0.293,
|
| 394 |
+
"step": 55
|
| 395 |
+
},
|
| 396 |
+
{
|
| 397 |
+
"epoch": 0.47786666666666666,
|
| 398 |
+
"grad_norm": 0.7974304806295485,
|
| 399 |
+
"learning_rate": 1.7481880270062003e-05,
|
| 400 |
+
"loss": 0.2717,
|
| 401 |
+
"step": 56
|
| 402 |
+
},
|
| 403 |
+
{
|
| 404 |
+
"epoch": 0.4864,
|
| 405 |
+
"grad_norm": 0.6232203657585239,
|
| 406 |
+
"learning_rate": 1.7558748556724913e-05,
|
| 407 |
+
"loss": 0.1741,
|
| 408 |
+
"step": 57
|
| 409 |
+
},
|
| 410 |
+
{
|
| 411 |
+
"epoch": 0.49493333333333334,
|
| 412 |
+
"grad_norm": 0.6850688604563008,
|
| 413 |
+
"learning_rate": 1.763427993562937e-05,
|
| 414 |
+
"loss": 0.228,
|
| 415 |
+
"step": 58
|
| 416 |
+
},
|
| 417 |
+
{
|
| 418 |
+
"epoch": 0.5034666666666666,
|
| 419 |
+
"grad_norm": 0.5923826384300431,
|
| 420 |
+
"learning_rate": 1.7708520116421443e-05,
|
| 421 |
+
"loss": 0.2131,
|
| 422 |
+
"step": 59
|
| 423 |
+
},
|
| 424 |
+
{
|
| 425 |
+
"epoch": 0.512,
|
| 426 |
+
"grad_norm": 0.489754430485032,
|
| 427 |
+
"learning_rate": 1.7781512503836432e-05,
|
| 428 |
+
"loss": 0.165,
|
| 429 |
+
"step": 60
|
| 430 |
+
},
|
| 431 |
+
{
|
| 432 |
+
"epoch": 0.5205333333333333,
|
| 433 |
+
"grad_norm": 0.5280940052395061,
|
| 434 |
+
"learning_rate": 1.7853298350107667e-05,
|
| 435 |
+
"loss": 0.1658,
|
| 436 |
+
"step": 61
|
| 437 |
+
},
|
| 438 |
+
{
|
| 439 |
+
"epoch": 0.5290666666666667,
|
| 440 |
+
"grad_norm": 0.4750905992036739,
|
| 441 |
+
"learning_rate": 1.7923916894982537e-05,
|
| 442 |
+
"loss": 0.1438,
|
| 443 |
+
"step": 62
|
| 444 |
+
},
|
| 445 |
+
{
|
| 446 |
+
"epoch": 0.5376,
|
| 447 |
+
"grad_norm": 0.4162998392722401,
|
| 448 |
+
"learning_rate": 1.7993405494535815e-05,
|
| 449 |
+
"loss": 0.1555,
|
| 450 |
+
"step": 63
|
| 451 |
+
},
|
| 452 |
+
{
|
| 453 |
+
"epoch": 0.5461333333333334,
|
| 454 |
+
"grad_norm": 0.26863266246370443,
|
| 455 |
+
"learning_rate": 1.806179973983887e-05,
|
| 456 |
+
"loss": 0.1323,
|
| 457 |
+
"step": 64
|
| 458 |
+
},
|
| 459 |
+
{
|
| 460 |
+
"epoch": 0.5546666666666666,
|
| 461 |
+
"grad_norm": 0.26534221125601215,
|
| 462 |
+
"learning_rate": 1.8129133566428553e-05,
|
| 463 |
+
"loss": 0.1671,
|
| 464 |
+
"step": 65
|
| 465 |
+
},
|
| 466 |
+
{
|
| 467 |
+
"epoch": 0.5632,
|
| 468 |
+
"grad_norm": 0.2548662962257576,
|
| 469 |
+
"learning_rate": 1.8195439355418686e-05,
|
| 470 |
+
"loss": 0.1308,
|
| 471 |
+
"step": 66
|
| 472 |
+
},
|
| 473 |
+
{
|
| 474 |
+
"epoch": 0.5717333333333333,
|
| 475 |
+
"grad_norm": 0.18045394638578796,
|
| 476 |
+
"learning_rate": 1.8260748027008263e-05,
|
| 477 |
+
"loss": 0.1262,
|
| 478 |
+
"step": 67
|
| 479 |
+
},
|
| 480 |
+
{
|
| 481 |
+
"epoch": 0.5802666666666667,
|
| 482 |
+
"grad_norm": 0.17070388073612064,
|
| 483 |
+
"learning_rate": 1.8325089127062364e-05,
|
| 484 |
+
"loss": 0.1192,
|
| 485 |
+
"step": 68
|
| 486 |
+
},
|
| 487 |
+
{
|
| 488 |
+
"epoch": 0.5888,
|
| 489 |
+
"grad_norm": 0.1531381679244776,
|
| 490 |
+
"learning_rate": 1.8388490907372553e-05,
|
| 491 |
+
"loss": 0.1274,
|
| 492 |
+
"step": 69
|
| 493 |
+
},
|
| 494 |
+
{
|
| 495 |
+
"epoch": 0.5973333333333334,
|
| 496 |
+
"grad_norm": 0.18196434993687946,
|
| 497 |
+
"learning_rate": 1.8450980400142568e-05,
|
| 498 |
+
"loss": 0.1375,
|
| 499 |
+
"step": 70
|
| 500 |
+
},
|
| 501 |
+
{
|
| 502 |
+
"epoch": 0.6058666666666667,
|
| 503 |
+
"grad_norm": 0.15324416972951205,
|
| 504 |
+
"learning_rate": 1.8512583487190752e-05,
|
| 505 |
+
"loss": 0.1599,
|
| 506 |
+
"step": 71
|
| 507 |
+
},
|
| 508 |
+
{
|
| 509 |
+
"epoch": 0.6144,
|
| 510 |
+
"grad_norm": 0.10884462064503801,
|
| 511 |
+
"learning_rate": 1.857332496431268e-05,
|
| 512 |
+
"loss": 0.1041,
|
| 513 |
+
"step": 72
|
| 514 |
+
},
|
| 515 |
+
{
|
| 516 |
+
"epoch": 0.6229333333333333,
|
| 517 |
+
"grad_norm": 0.12915133528192668,
|
| 518 |
+
"learning_rate": 1.8633228601204555e-05,
|
| 519 |
+
"loss": 0.1406,
|
| 520 |
+
"step": 73
|
| 521 |
+
},
|
| 522 |
+
{
|
| 523 |
+
"epoch": 0.6314666666666666,
|
| 524 |
+
"grad_norm": 0.12553425699952878,
|
| 525 |
+
"learning_rate": 1.8692317197309763e-05,
|
| 526 |
+
"loss": 0.1256,
|
| 527 |
+
"step": 74
|
| 528 |
+
},
|
| 529 |
+
{
|
| 530 |
+
"epoch": 0.64,
|
| 531 |
+
"grad_norm": 0.11976960918968543,
|
| 532 |
+
"learning_rate": 1.8750612633916997e-05,
|
| 533 |
+
"loss": 0.1144,
|
| 534 |
+
"step": 75
|
| 535 |
+
},
|
| 536 |
+
{
|
| 537 |
+
"epoch": 0.6485333333333333,
|
| 538 |
+
"grad_norm": 0.115805998298789,
|
| 539 |
+
"learning_rate": 1.8808135922807914e-05,
|
| 540 |
+
"loss": 0.1528,
|
| 541 |
+
"step": 76
|
| 542 |
+
},
|
| 543 |
+
{
|
| 544 |
+
"epoch": 0.6570666666666667,
|
| 545 |
+
"grad_norm": 0.10325948496697443,
|
| 546 |
+
"learning_rate": 1.8864907251724818e-05,
|
| 547 |
+
"loss": 0.1044,
|
| 548 |
+
"step": 77
|
| 549 |
+
},
|
| 550 |
+
{
|
| 551 |
+
"epoch": 0.6656,
|
| 552 |
+
"grad_norm": 0.09595064346541006,
|
| 553 |
+
"learning_rate": 1.8920946026904802e-05,
|
| 554 |
+
"loss": 0.1534,
|
| 555 |
+
"step": 78
|
| 556 |
+
},
|
| 557 |
+
{
|
| 558 |
+
"epoch": 0.6741333333333334,
|
| 559 |
+
"grad_norm": 0.08796742845240496,
|
| 560 |
+
"learning_rate": 1.8976270912904414e-05,
|
| 561 |
+
"loss": 0.1155,
|
| 562 |
+
"step": 79
|
| 563 |
+
},
|
| 564 |
+
{
|
| 565 |
+
"epoch": 0.6826666666666666,
|
| 566 |
+
"grad_norm": 0.08218991738379527,
|
| 567 |
+
"learning_rate": 1.9030899869919434e-05,
|
| 568 |
+
"loss": 0.1311,
|
| 569 |
+
"step": 80
|
| 570 |
+
},
|
| 571 |
+
{
|
| 572 |
+
"epoch": 0.6912,
|
| 573 |
+
"grad_norm": 0.08290815261109215,
|
| 574 |
+
"learning_rate": 1.9084850188786497e-05,
|
| 575 |
+
"loss": 0.11,
|
| 576 |
+
"step": 81
|
| 577 |
+
},
|
| 578 |
+
{
|
| 579 |
+
"epoch": 0.6997333333333333,
|
| 580 |
+
"grad_norm": 0.0794132180176064,
|
| 581 |
+
"learning_rate": 1.9138138523837165e-05,
|
| 582 |
+
"loss": 0.1135,
|
| 583 |
+
"step": 82
|
| 584 |
+
},
|
| 585 |
+
{
|
| 586 |
+
"epoch": 0.7082666666666667,
|
| 587 |
+
"grad_norm": 0.06934410705255296,
|
| 588 |
+
"learning_rate": 1.919078092376074e-05,
|
| 589 |
+
"loss": 0.109,
|
| 590 |
+
"step": 83
|
| 591 |
+
},
|
| 592 |
+
{
|
| 593 |
+
"epoch": 0.7168,
|
| 594 |
+
"grad_norm": 0.09000563031870593,
|
| 595 |
+
"learning_rate": 1.9242792860618813e-05,
|
| 596 |
+
"loss": 0.12,
|
| 597 |
+
"step": 84
|
| 598 |
+
},
|
| 599 |
+
{
|
| 600 |
+
"epoch": 0.7253333333333334,
|
| 601 |
+
"grad_norm": 0.1134042277879818,
|
| 602 |
+
"learning_rate": 1.929418925714293e-05,
|
| 603 |
+
"loss": 0.1223,
|
| 604 |
+
"step": 85
|
| 605 |
+
},
|
| 606 |
+
{
|
| 607 |
+
"epoch": 0.7338666666666667,
|
| 608 |
+
"grad_norm": 0.09118764690233076,
|
| 609 |
+
"learning_rate": 1.9344984512435673e-05,
|
| 610 |
+
"loss": 0.1459,
|
| 611 |
+
"step": 86
|
| 612 |
+
},
|
| 613 |
+
{
|
| 614 |
+
"epoch": 0.7424,
|
| 615 |
+
"grad_norm": 0.07873016754353963,
|
| 616 |
+
"learning_rate": 1.9395192526186183e-05,
|
| 617 |
+
"loss": 0.1422,
|
| 618 |
+
"step": 87
|
| 619 |
+
},
|
| 620 |
+
{
|
| 621 |
+
"epoch": 0.7509333333333333,
|
| 622 |
+
"grad_norm": 0.1796495874463076,
|
| 623 |
+
"learning_rate": 1.9444826721501687e-05,
|
| 624 |
+
"loss": 0.1291,
|
| 625 |
+
"step": 88
|
| 626 |
+
},
|
| 627 |
+
{
|
| 628 |
+
"epoch": 0.7594666666666666,
|
| 629 |
+
"grad_norm": 0.0679589944174269,
|
| 630 |
+
"learning_rate": 1.9493900066449125e-05,
|
| 631 |
+
"loss": 0.108,
|
| 632 |
+
"step": 89
|
| 633 |
+
},
|
| 634 |
+
{
|
| 635 |
+
"epoch": 0.768,
|
| 636 |
+
"grad_norm": 0.08174688574235538,
|
| 637 |
+
"learning_rate": 1.9542425094393246e-05,
|
| 638 |
+
"loss": 0.1081,
|
| 639 |
+
"step": 90
|
| 640 |
+
},
|
| 641 |
+
{
|
| 642 |
+
"epoch": 0.7765333333333333,
|
| 643 |
+
"grad_norm": 0.057137370501406756,
|
| 644 |
+
"learning_rate": 1.9590413923210934e-05,
|
| 645 |
+
"loss": 0.0934,
|
| 646 |
+
"step": 91
|
| 647 |
+
},
|
| 648 |
+
{
|
| 649 |
+
"epoch": 0.7850666666666667,
|
| 650 |
+
"grad_norm": 0.06578111924908255,
|
| 651 |
+
"learning_rate": 1.9637878273455555e-05,
|
| 652 |
+
"loss": 0.1085,
|
| 653 |
+
"step": 92
|
| 654 |
+
},
|
| 655 |
+
{
|
| 656 |
+
"epoch": 0.7936,
|
| 657 |
+
"grad_norm": 0.08945990540906254,
|
| 658 |
+
"learning_rate": 1.968482948553935e-05,
|
| 659 |
+
"loss": 0.1747,
|
| 660 |
+
"step": 93
|
| 661 |
+
},
|
| 662 |
+
{
|
| 663 |
+
"epoch": 0.8021333333333334,
|
| 664 |
+
"grad_norm": 0.06183863311044229,
|
| 665 |
+
"learning_rate": 1.9731278535996986e-05,
|
| 666 |
+
"loss": 0.1136,
|
| 667 |
+
"step": 94
|
| 668 |
+
},
|
| 669 |
+
{
|
| 670 |
+
"epoch": 0.8106666666666666,
|
| 671 |
+
"grad_norm": 0.05777899602544702,
|
| 672 |
+
"learning_rate": 1.9777236052888476e-05,
|
| 673 |
+
"loss": 0.0984,
|
| 674 |
+
"step": 95
|
| 675 |
+
},
|
| 676 |
+
{
|
| 677 |
+
"epoch": 0.8192,
|
| 678 |
+
"grad_norm": 0.08130851607693534,
|
| 679 |
+
"learning_rate": 1.9822712330395683e-05,
|
| 680 |
+
"loss": 0.187,
|
| 681 |
+
"step": 96
|
| 682 |
+
},
|
| 683 |
+
{
|
| 684 |
+
"epoch": 0.8277333333333333,
|
| 685 |
+
"grad_norm": 0.06426546202002927,
|
| 686 |
+
"learning_rate": 1.986771734266245e-05,
|
| 687 |
+
"loss": 0.1296,
|
| 688 |
+
"step": 97
|
| 689 |
+
},
|
| 690 |
+
{
|
| 691 |
+
"epoch": 0.8362666666666667,
|
| 692 |
+
"grad_norm": 0.069692313707994,
|
| 693 |
+
"learning_rate": 1.991226075692495e-05,
|
| 694 |
+
"loss": 0.1404,
|
| 695 |
+
"step": 98
|
| 696 |
+
},
|
| 697 |
+
{
|
| 698 |
+
"epoch": 0.8448,
|
| 699 |
+
"grad_norm": 0.05494542266886729,
|
| 700 |
+
"learning_rate": 1.9956351945975496e-05,
|
| 701 |
+
"loss": 0.116,
|
| 702 |
+
"step": 99
|
| 703 |
+
},
|
| 704 |
+
{
|
| 705 |
+
"epoch": 0.8533333333333334,
|
| 706 |
+
"grad_norm": 0.07571686966840627,
|
| 707 |
+
"learning_rate": 1.9999999999999998e-05,
|
| 708 |
+
"loss": 0.1539,
|
| 709 |
+
"step": 100
|
| 710 |
+
},
|
| 711 |
+
{
|
| 712 |
+
"epoch": 0.8618666666666667,
|
| 713 |
+
"grad_norm": 0.054351059117603705,
|
| 714 |
+
"learning_rate": 2e-05,
|
| 715 |
+
"loss": 0.1037,
|
| 716 |
+
"step": 101
|
| 717 |
+
},
|
| 718 |
+
{
|
| 719 |
+
"epoch": 0.8704,
|
| 720 |
+
"grad_norm": 0.06531899611551092,
|
| 721 |
+
"learning_rate": 2e-05,
|
| 722 |
+
"loss": 0.0827,
|
| 723 |
+
"step": 102
|
| 724 |
+
},
|
| 725 |
+
{
|
| 726 |
+
"epoch": 0.8789333333333333,
|
| 727 |
+
"grad_norm": 0.06131678646504652,
|
| 728 |
+
"learning_rate": 2e-05,
|
| 729 |
+
"loss": 0.1266,
|
| 730 |
+
"step": 103
|
| 731 |
+
},
|
| 732 |
+
{
|
| 733 |
+
"epoch": 0.8874666666666666,
|
| 734 |
+
"grad_norm": 0.06850220540661824,
|
| 735 |
+
"learning_rate": 2e-05,
|
| 736 |
+
"loss": 0.1456,
|
| 737 |
+
"step": 104
|
| 738 |
+
},
|
| 739 |
+
{
|
| 740 |
+
"epoch": 0.896,
|
| 741 |
+
"grad_norm": 0.05806908951252483,
|
| 742 |
+
"learning_rate": 2e-05,
|
| 743 |
+
"loss": 0.0954,
|
| 744 |
+
"step": 105
|
| 745 |
+
},
|
| 746 |
+
{
|
| 747 |
+
"epoch": 0.9045333333333333,
|
| 748 |
+
"grad_norm": 0.06503642452033717,
|
| 749 |
+
"learning_rate": 2e-05,
|
| 750 |
+
"loss": 0.1417,
|
| 751 |
+
"step": 106
|
| 752 |
+
},
|
| 753 |
+
{
|
| 754 |
+
"epoch": 0.9130666666666667,
|
| 755 |
+
"grad_norm": 0.050486271853277066,
|
| 756 |
+
"learning_rate": 2e-05,
|
| 757 |
+
"loss": 0.0959,
|
| 758 |
+
"step": 107
|
| 759 |
+
},
|
| 760 |
+
{
|
| 761 |
+
"epoch": 0.9216,
|
| 762 |
+
"grad_norm": 0.07746063813802379,
|
| 763 |
+
"learning_rate": 2e-05,
|
| 764 |
+
"loss": 0.1256,
|
| 765 |
+
"step": 108
|
| 766 |
+
},
|
| 767 |
+
{
|
| 768 |
+
"epoch": 0.9301333333333334,
|
| 769 |
+
"grad_norm": 0.051231172380840004,
|
| 770 |
+
"learning_rate": 2e-05,
|
| 771 |
+
"loss": 0.1116,
|
| 772 |
+
"step": 109
|
| 773 |
+
},
|
| 774 |
+
{
|
| 775 |
+
"epoch": 0.9386666666666666,
|
| 776 |
+
"grad_norm": 0.056296443557859455,
|
| 777 |
+
"learning_rate": 2e-05,
|
| 778 |
+
"loss": 0.1056,
|
| 779 |
+
"step": 110
|
| 780 |
+
},
|
| 781 |
+
{
|
| 782 |
+
"epoch": 0.9472,
|
| 783 |
+
"grad_norm": 0.05058663240713958,
|
| 784 |
+
"learning_rate": 2e-05,
|
| 785 |
+
"loss": 0.0971,
|
| 786 |
+
"step": 111
|
| 787 |
+
},
|
| 788 |
+
{
|
| 789 |
+
"epoch": 0.9557333333333333,
|
| 790 |
+
"grad_norm": 0.05532886570130611,
|
| 791 |
+
"learning_rate": 2e-05,
|
| 792 |
+
"loss": 0.1086,
|
| 793 |
+
"step": 112
|
| 794 |
+
},
|
| 795 |
+
{
|
| 796 |
+
"epoch": 0.9642666666666667,
|
| 797 |
+
"grad_norm": 0.05327811326654907,
|
| 798 |
+
"learning_rate": 2e-05,
|
| 799 |
+
"loss": 0.0989,
|
| 800 |
+
"step": 113
|
| 801 |
+
},
|
| 802 |
+
{
|
| 803 |
+
"epoch": 0.9728,
|
| 804 |
+
"grad_norm": 0.05663279364147864,
|
| 805 |
+
"learning_rate": 2e-05,
|
| 806 |
+
"loss": 0.0958,
|
| 807 |
+
"step": 114
|
| 808 |
+
},
|
| 809 |
+
{
|
| 810 |
+
"epoch": 0.9813333333333333,
|
| 811 |
+
"grad_norm": 0.04930904541225805,
|
| 812 |
+
"learning_rate": 2e-05,
|
| 813 |
+
"loss": 0.0887,
|
| 814 |
+
"step": 115
|
| 815 |
+
},
|
| 816 |
+
{
|
| 817 |
+
"epoch": 0.9898666666666667,
|
| 818 |
+
"grad_norm": 0.06096947951115022,
|
| 819 |
+
"learning_rate": 2e-05,
|
| 820 |
+
"loss": 0.106,
|
| 821 |
+
"step": 116
|
| 822 |
+
},
|
| 823 |
+
{
|
| 824 |
+
"epoch": 0.9984,
|
| 825 |
+
"grad_norm": 0.050092322361182495,
|
| 826 |
+
"learning_rate": 2e-05,
|
| 827 |
+
"loss": 0.0931,
|
| 828 |
+
"step": 117
|
| 829 |
+
},
|
| 830 |
+
{
|
| 831 |
+
"epoch": 1.0069333333333332,
|
| 832 |
+
"grad_norm": 0.04980408443758999,
|
| 833 |
+
"learning_rate": 2e-05,
|
| 834 |
+
"loss": 0.0955,
|
| 835 |
+
"step": 118
|
| 836 |
+
},
|
| 837 |
+
{
|
| 838 |
+
"epoch": 1.0154666666666667,
|
| 839 |
+
"grad_norm": 0.051183082721834305,
|
| 840 |
+
"learning_rate": 2e-05,
|
| 841 |
+
"loss": 0.1049,
|
| 842 |
+
"step": 119
|
| 843 |
+
},
|
| 844 |
+
{
|
| 845 |
+
"epoch": 1.024,
|
| 846 |
+
"grad_norm": 0.04332220265802814,
|
| 847 |
+
"learning_rate": 2e-05,
|
| 848 |
+
"loss": 0.0983,
|
| 849 |
+
"step": 120
|
| 850 |
+
}
|
| 851 |
+
],
|
| 852 |
+
"logging_steps": 1,
|
| 853 |
+
"max_steps": 351,
|
| 854 |
+
"num_input_tokens_seen": 0,
|
| 855 |
+
"num_train_epochs": 3,
|
| 856 |
+
"save_steps": 20,
|
| 857 |
+
"stateful_callbacks": {
|
| 858 |
+
"TrainerControl": {
|
| 859 |
+
"args": {
|
| 860 |
+
"should_epoch_stop": false,
|
| 861 |
+
"should_evaluate": false,
|
| 862 |
+
"should_log": false,
|
| 863 |
+
"should_save": true,
|
| 864 |
+
"should_training_stop": false
|
| 865 |
+
},
|
| 866 |
+
"attributes": {}
|
| 867 |
+
}
|
| 868 |
+
},
|
| 869 |
+
"total_flos": 1.7047845829953454e+18,
|
| 870 |
+
"train_batch_size": 16,
|
| 871 |
+
"trial_name": null,
|
| 872 |
+
"trial_params": null
|
| 873 |
+
}
|
checkpoint-120/training_args.bin
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:aa4edad1da5557fcd52a6da980443588016cead6f0444a3562cfa68029c66a04
|
| 3 |
+
size 6840
|
checkpoint-120/zero_to_fp32.py
ADDED
|
@@ -0,0 +1,592 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 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):
|
| 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 |
+
elif zero_stage == 3:
|
| 216 |
+
return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
|
| 217 |
+
|
| 218 |
+
|
| 219 |
+
def _zero2_merge_frozen_params(state_dict, zero_model_states):
|
| 220 |
+
if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
|
| 221 |
+
return
|
| 222 |
+
|
| 223 |
+
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
|
| 224 |
+
frozen_param_fragments = zero_model_states[0].frozen_param_fragments
|
| 225 |
+
|
| 226 |
+
if debug:
|
| 227 |
+
num_elem = sum(s.numel() for s in frozen_param_shapes.values())
|
| 228 |
+
print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
|
| 229 |
+
|
| 230 |
+
wanted_params = len(frozen_param_shapes)
|
| 231 |
+
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
| 232 |
+
avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
|
| 233 |
+
print(f'Frozen params: Have {avail_numel} numels to process.')
|
| 234 |
+
print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
|
| 235 |
+
|
| 236 |
+
total_params = 0
|
| 237 |
+
total_numel = 0
|
| 238 |
+
for name, shape in frozen_param_shapes.items():
|
| 239 |
+
total_params += 1
|
| 240 |
+
unpartitioned_numel = shape.numel()
|
| 241 |
+
total_numel += unpartitioned_numel
|
| 242 |
+
|
| 243 |
+
state_dict[name] = frozen_param_fragments[name]
|
| 244 |
+
|
| 245 |
+
if debug:
|
| 246 |
+
print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
|
| 247 |
+
|
| 248 |
+
print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
|
| 249 |
+
|
| 250 |
+
|
| 251 |
+
def _has_callable(obj, fn):
|
| 252 |
+
attr = getattr(obj, fn, None)
|
| 253 |
+
return callable(attr)
|
| 254 |
+
|
| 255 |
+
|
| 256 |
+
def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
|
| 257 |
+
param_shapes = zero_model_states[0].param_shapes
|
| 258 |
+
|
| 259 |
+
# Reconstruction protocol:
|
| 260 |
+
#
|
| 261 |
+
# XXX: document this
|
| 262 |
+
|
| 263 |
+
if debug:
|
| 264 |
+
for i in range(world_size):
|
| 265 |
+
for j in range(len(fp32_flat_groups[0])):
|
| 266 |
+
print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
|
| 267 |
+
|
| 268 |
+
# XXX: memory usage doubles here (zero2)
|
| 269 |
+
num_param_groups = len(fp32_flat_groups[0])
|
| 270 |
+
merged_single_partition_of_fp32_groups = []
|
| 271 |
+
for i in range(num_param_groups):
|
| 272 |
+
merged_partitions = [sd[i] for sd in fp32_flat_groups]
|
| 273 |
+
full_single_fp32_vector = torch.cat(merged_partitions, 0)
|
| 274 |
+
merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
|
| 275 |
+
avail_numel = sum(
|
| 276 |
+
[full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
|
| 277 |
+
|
| 278 |
+
if debug:
|
| 279 |
+
wanted_params = sum([len(shapes) for shapes in param_shapes])
|
| 280 |
+
wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
|
| 281 |
+
# not asserting if there is a mismatch due to possible padding
|
| 282 |
+
print(f"Have {avail_numel} numels to process.")
|
| 283 |
+
print(f"Need {wanted_numel} numels in {wanted_params} params.")
|
| 284 |
+
|
| 285 |
+
# params
|
| 286 |
+
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
| 287 |
+
# out-of-core computing solution
|
| 288 |
+
total_numel = 0
|
| 289 |
+
total_params = 0
|
| 290 |
+
for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
|
| 291 |
+
offset = 0
|
| 292 |
+
avail_numel = full_single_fp32_vector.numel()
|
| 293 |
+
for name, shape in shapes.items():
|
| 294 |
+
|
| 295 |
+
unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
|
| 296 |
+
total_numel += unpartitioned_numel
|
| 297 |
+
total_params += 1
|
| 298 |
+
|
| 299 |
+
if debug:
|
| 300 |
+
print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
|
| 301 |
+
state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
|
| 302 |
+
offset += unpartitioned_numel
|
| 303 |
+
|
| 304 |
+
# Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
|
| 305 |
+
# avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
|
| 306 |
+
# paddings performed in the code it's almost impossible to predict the exact numbers w/o the
|
| 307 |
+
# live optimizer object, so we are checking that the numbers are within the right range
|
| 308 |
+
align_to = 2 * world_size
|
| 309 |
+
|
| 310 |
+
def zero2_align(x):
|
| 311 |
+
return align_to * math.ceil(x / align_to)
|
| 312 |
+
|
| 313 |
+
if debug:
|
| 314 |
+
print(f"original offset={offset}, avail_numel={avail_numel}")
|
| 315 |
+
|
| 316 |
+
offset = zero2_align(offset)
|
| 317 |
+
avail_numel = zero2_align(avail_numel)
|
| 318 |
+
|
| 319 |
+
if debug:
|
| 320 |
+
print(f"aligned offset={offset}, avail_numel={avail_numel}")
|
| 321 |
+
|
| 322 |
+
# Sanity check
|
| 323 |
+
if offset != avail_numel:
|
| 324 |
+
raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
| 325 |
+
|
| 326 |
+
print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
|
| 327 |
+
|
| 328 |
+
|
| 329 |
+
def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):
|
| 330 |
+
state_dict = OrderedDict()
|
| 331 |
+
|
| 332 |
+
# buffers
|
| 333 |
+
buffers = zero_model_states[0].buffers
|
| 334 |
+
state_dict.update(buffers)
|
| 335 |
+
if debug:
|
| 336 |
+
print(f"added {len(buffers)} buffers")
|
| 337 |
+
|
| 338 |
+
_zero2_merge_frozen_params(state_dict, zero_model_states)
|
| 339 |
+
|
| 340 |
+
_zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
|
| 341 |
+
|
| 342 |
+
# recover shared parameters
|
| 343 |
+
for pair in zero_model_states[0].shared_params:
|
| 344 |
+
if pair[1] in state_dict:
|
| 345 |
+
state_dict[pair[0]] = state_dict[pair[1]]
|
| 346 |
+
|
| 347 |
+
return state_dict
|
| 348 |
+
|
| 349 |
+
|
| 350 |
+
def zero3_partitioned_param_info(unpartitioned_numel, world_size):
|
| 351 |
+
remainder = unpartitioned_numel % world_size
|
| 352 |
+
padding_numel = (world_size - remainder) if remainder else 0
|
| 353 |
+
partitioned_numel = math.ceil(unpartitioned_numel / world_size)
|
| 354 |
+
return partitioned_numel, padding_numel
|
| 355 |
+
|
| 356 |
+
|
| 357 |
+
def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
|
| 358 |
+
if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
|
| 359 |
+
return
|
| 360 |
+
|
| 361 |
+
if debug:
|
| 362 |
+
for i in range(world_size):
|
| 363 |
+
num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
|
| 364 |
+
print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
|
| 365 |
+
|
| 366 |
+
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
|
| 367 |
+
wanted_params = len(frozen_param_shapes)
|
| 368 |
+
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
| 369 |
+
avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
|
| 370 |
+
print(f'Frozen params: Have {avail_numel} numels to process.')
|
| 371 |
+
print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
|
| 372 |
+
|
| 373 |
+
total_params = 0
|
| 374 |
+
total_numel = 0
|
| 375 |
+
for name, shape in zero_model_states[0].frozen_param_shapes.items():
|
| 376 |
+
total_params += 1
|
| 377 |
+
unpartitioned_numel = shape.numel()
|
| 378 |
+
total_numel += unpartitioned_numel
|
| 379 |
+
|
| 380 |
+
param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
|
| 381 |
+
state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
|
| 382 |
+
|
| 383 |
+
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
| 384 |
+
|
| 385 |
+
if debug:
|
| 386 |
+
print(
|
| 387 |
+
f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
| 388 |
+
)
|
| 389 |
+
|
| 390 |
+
print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
|
| 391 |
+
|
| 392 |
+
|
| 393 |
+
def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
|
| 394 |
+
param_shapes = zero_model_states[0].param_shapes
|
| 395 |
+
avail_numel = fp32_flat_groups[0].numel() * world_size
|
| 396 |
+
# Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
|
| 397 |
+
# param, re-consolidating each param, while dealing with padding if any
|
| 398 |
+
|
| 399 |
+
# merge list of dicts, preserving order
|
| 400 |
+
param_shapes = {k: v for d in param_shapes for k, v in d.items()}
|
| 401 |
+
|
| 402 |
+
if debug:
|
| 403 |
+
for i in range(world_size):
|
| 404 |
+
print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
|
| 405 |
+
|
| 406 |
+
wanted_params = len(param_shapes)
|
| 407 |
+
wanted_numel = sum(shape.numel() for shape in param_shapes.values())
|
| 408 |
+
# not asserting if there is a mismatch due to possible padding
|
| 409 |
+
avail_numel = fp32_flat_groups[0].numel() * world_size
|
| 410 |
+
print(f"Trainable params: Have {avail_numel} numels to process.")
|
| 411 |
+
print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
|
| 412 |
+
|
| 413 |
+
# params
|
| 414 |
+
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
| 415 |
+
# out-of-core computing solution
|
| 416 |
+
offset = 0
|
| 417 |
+
total_numel = 0
|
| 418 |
+
total_params = 0
|
| 419 |
+
for name, shape in param_shapes.items():
|
| 420 |
+
|
| 421 |
+
unpartitioned_numel = shape.numel()
|
| 422 |
+
total_numel += unpartitioned_numel
|
| 423 |
+
total_params += 1
|
| 424 |
+
|
| 425 |
+
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
| 426 |
+
|
| 427 |
+
if debug:
|
| 428 |
+
print(
|
| 429 |
+
f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
| 430 |
+
)
|
| 431 |
+
|
| 432 |
+
# XXX: memory usage doubles here
|
| 433 |
+
state_dict[name] = torch.cat(
|
| 434 |
+
tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
|
| 435 |
+
0).narrow(0, 0, unpartitioned_numel).view(shape)
|
| 436 |
+
offset += partitioned_numel
|
| 437 |
+
|
| 438 |
+
offset *= world_size
|
| 439 |
+
|
| 440 |
+
# Sanity check
|
| 441 |
+
if offset != avail_numel:
|
| 442 |
+
raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
| 443 |
+
|
| 444 |
+
print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
|
| 445 |
+
|
| 446 |
+
|
| 447 |
+
def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):
|
| 448 |
+
state_dict = OrderedDict()
|
| 449 |
+
|
| 450 |
+
# buffers
|
| 451 |
+
buffers = zero_model_states[0].buffers
|
| 452 |
+
state_dict.update(buffers)
|
| 453 |
+
if debug:
|
| 454 |
+
print(f"added {len(buffers)} buffers")
|
| 455 |
+
|
| 456 |
+
_zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
|
| 457 |
+
|
| 458 |
+
_zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
|
| 459 |
+
|
| 460 |
+
# recover shared parameters
|
| 461 |
+
for pair in zero_model_states[0].shared_params:
|
| 462 |
+
if pair[1] in state_dict:
|
| 463 |
+
state_dict[pair[0]] = state_dict[pair[1]]
|
| 464 |
+
|
| 465 |
+
return state_dict
|
| 466 |
+
|
| 467 |
+
|
| 468 |
+
def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
|
| 469 |
+
"""
|
| 470 |
+
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
|
| 471 |
+
``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
|
| 472 |
+
via a model hub.
|
| 473 |
+
|
| 474 |
+
Args:
|
| 475 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder
|
| 476 |
+
- ``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``
|
| 477 |
+
|
| 478 |
+
Returns:
|
| 479 |
+
- pytorch ``state_dict``
|
| 480 |
+
|
| 481 |
+
Note: this approach may not work if your application doesn't have sufficient free CPU memory and
|
| 482 |
+
you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
|
| 483 |
+
the checkpoint.
|
| 484 |
+
|
| 485 |
+
A typical usage might be ::
|
| 486 |
+
|
| 487 |
+
from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
|
| 488 |
+
# do the training and checkpoint saving
|
| 489 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
|
| 490 |
+
model = model.cpu() # move to cpu
|
| 491 |
+
model.load_state_dict(state_dict)
|
| 492 |
+
# submit to model hub or save the model to share with others
|
| 493 |
+
|
| 494 |
+
In this example the ``model`` will no longer be usable in the deepspeed context of the same
|
| 495 |
+
application. i.e. you will need to re-initialize the deepspeed engine, since
|
| 496 |
+
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
| 497 |
+
|
| 498 |
+
If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
|
| 499 |
+
|
| 500 |
+
"""
|
| 501 |
+
if tag is None:
|
| 502 |
+
latest_path = os.path.join(checkpoint_dir, 'latest')
|
| 503 |
+
if os.path.isfile(latest_path):
|
| 504 |
+
with open(latest_path, 'r') as fd:
|
| 505 |
+
tag = fd.read().strip()
|
| 506 |
+
else:
|
| 507 |
+
raise ValueError(f"Unable to find 'latest' file at {latest_path}")
|
| 508 |
+
|
| 509 |
+
ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
|
| 510 |
+
|
| 511 |
+
if not os.path.isdir(ds_checkpoint_dir):
|
| 512 |
+
raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
|
| 513 |
+
|
| 514 |
+
return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
|
| 515 |
+
|
| 516 |
+
|
| 517 |
+
def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
|
| 518 |
+
"""
|
| 519 |
+
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
|
| 520 |
+
loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
|
| 521 |
+
|
| 522 |
+
Args:
|
| 523 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
| 524 |
+
- ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
|
| 525 |
+
- ``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``
|
| 526 |
+
"""
|
| 527 |
+
|
| 528 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
| 529 |
+
print(f"Saving fp32 state dict to {output_file}")
|
| 530 |
+
torch.save(state_dict, output_file)
|
| 531 |
+
|
| 532 |
+
|
| 533 |
+
def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
|
| 534 |
+
"""
|
| 535 |
+
1. Put the provided model to cpu
|
| 536 |
+
2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
|
| 537 |
+
3. Load it into the provided model
|
| 538 |
+
|
| 539 |
+
Args:
|
| 540 |
+
- ``model``: the model object to update
|
| 541 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
| 542 |
+
- ``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``
|
| 543 |
+
|
| 544 |
+
Returns:
|
| 545 |
+
- ``model`: modified model
|
| 546 |
+
|
| 547 |
+
Make sure you have plenty of CPU memory available before you call this function. If you don't
|
| 548 |
+
have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
|
| 549 |
+
conveniently placed for you in the checkpoint folder.
|
| 550 |
+
|
| 551 |
+
A typical usage might be ::
|
| 552 |
+
|
| 553 |
+
from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
|
| 554 |
+
model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
|
| 555 |
+
# submit to model hub or save the model to share with others
|
| 556 |
+
|
| 557 |
+
Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
|
| 558 |
+
of the same application. i.e. you will need to re-initialize the deepspeed engine, since
|
| 559 |
+
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
| 560 |
+
|
| 561 |
+
"""
|
| 562 |
+
logger.info(f"Extracting fp32 weights")
|
| 563 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
| 564 |
+
|
| 565 |
+
logger.info(f"Overwriting model with fp32 weights")
|
| 566 |
+
model = model.cpu()
|
| 567 |
+
model.load_state_dict(state_dict, strict=False)
|
| 568 |
+
|
| 569 |
+
return model
|
| 570 |
+
|
| 571 |
+
|
| 572 |
+
if __name__ == "__main__":
|
| 573 |
+
|
| 574 |
+
parser = argparse.ArgumentParser()
|
| 575 |
+
parser.add_argument("checkpoint_dir",
|
| 576 |
+
type=str,
|
| 577 |
+
help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
|
| 578 |
+
parser.add_argument(
|
| 579 |
+
"output_file",
|
| 580 |
+
type=str,
|
| 581 |
+
help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
|
| 582 |
+
parser.add_argument("-t",
|
| 583 |
+
"--tag",
|
| 584 |
+
type=str,
|
| 585 |
+
default=None,
|
| 586 |
+
help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
|
| 587 |
+
parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
|
| 588 |
+
args = parser.parse_args()
|
| 589 |
+
|
| 590 |
+
debug = args.debug
|
| 591 |
+
|
| 592 |
+
convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file, tag=args.tag)
|
checkpoint-140/README.md
ADDED
|
@@ -0,0 +1,202 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: peft
|
| 3 |
+
base_model: ../ckpts/Meta-Llama-3-8B-Instruct
|
| 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 0.11.1
|
checkpoint-140/adapter_config.json
ADDED
|
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"alpha_pattern": {},
|
| 3 |
+
"auto_mapping": null,
|
| 4 |
+
"base_model_name_or_path": "../ckpts/Meta-Llama-3-8B-Instruct",
|
| 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": 32,
|
| 14 |
+
"lora_dropout": 0.1,
|
| 15 |
+
"megatron_config": null,
|
| 16 |
+
"megatron_core": "megatron.core",
|
| 17 |
+
"modules_to_save": null,
|
| 18 |
+
"peft_type": "LORA",
|
| 19 |
+
"r": 16,
|
| 20 |
+
"rank_pattern": {},
|
| 21 |
+
"revision": null,
|
| 22 |
+
"target_modules": [
|
| 23 |
+
"down_proj",
|
| 24 |
+
"gate_proj",
|
| 25 |
+
"q_proj",
|
| 26 |
+
"lm_head",
|
| 27 |
+
"o_proj",
|
| 28 |
+
"v_proj",
|
| 29 |
+
"up_proj",
|
| 30 |
+
"k_proj"
|
| 31 |
+
],
|
| 32 |
+
"task_type": "CAUSAL_LM",
|
| 33 |
+
"use_dora": false,
|
| 34 |
+
"use_rslora": false
|
| 35 |
+
}
|
checkpoint-140/adapter_model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b12f85af7122bcce5f6997bec8d9a8a845cf9ab8d2c4f241d37562051e34c216
|
| 3 |
+
size 1138856856
|
checkpoint-140/global_step140/bf16_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 sha256:8993927bbec291df7a8a041058bc3762b88d9ca9fcfbdccdabc686717c9c5f5d
|
| 3 |
+
size 528781328
|
checkpoint-140/global_step140/mp_rank_00_model_states.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a4dca3099b75369d573830bab987eb6f984e58440ba145f2f47df24a9212b7c4
|
| 3 |
+
size 199905337
|
checkpoint-140/latest
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
global_step140
|
checkpoint-140/rng_state.pth
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:cb5b9188bbc87bb2ff7dba4844a90c5c236962fc66f3da23e5af9ee0eda34a1b
|
| 3 |
+
size 14244
|
checkpoint-140/trainer_state.json
ADDED
|
@@ -0,0 +1,1013 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"best_metric": null,
|
| 3 |
+
"best_model_checkpoint": null,
|
| 4 |
+
"epoch": 1.1946666666666665,
|
| 5 |
+
"eval_steps": 500,
|
| 6 |
+
"global_step": 140,
|
| 7 |
+
"is_hyper_param_search": false,
|
| 8 |
+
"is_local_process_zero": true,
|
| 9 |
+
"is_world_process_zero": true,
|
| 10 |
+
"log_history": [
|
| 11 |
+
{
|
| 12 |
+
"epoch": 0.008533333333333334,
|
| 13 |
+
"grad_norm": 100.21848203113535,
|
| 14 |
+
"learning_rate": 0.0,
|
| 15 |
+
"loss": 7.1962,
|
| 16 |
+
"step": 1
|
| 17 |
+
},
|
| 18 |
+
{
|
| 19 |
+
"epoch": 0.017066666666666667,
|
| 20 |
+
"grad_norm": 96.43006188910957,
|
| 21 |
+
"learning_rate": 3.0102999566398115e-06,
|
| 22 |
+
"loss": 6.9414,
|
| 23 |
+
"step": 2
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"epoch": 0.0256,
|
| 27 |
+
"grad_norm": 97.35803466618715,
|
| 28 |
+
"learning_rate": 4.771212547196624e-06,
|
| 29 |
+
"loss": 7.0102,
|
| 30 |
+
"step": 3
|
| 31 |
+
},
|
| 32 |
+
{
|
| 33 |
+
"epoch": 0.034133333333333335,
|
| 34 |
+
"grad_norm": 95.14837816372646,
|
| 35 |
+
"learning_rate": 6.020599913279623e-06,
|
| 36 |
+
"loss": 6.5295,
|
| 37 |
+
"step": 4
|
| 38 |
+
},
|
| 39 |
+
{
|
| 40 |
+
"epoch": 0.042666666666666665,
|
| 41 |
+
"grad_norm": 91.76544275692784,
|
| 42 |
+
"learning_rate": 6.989700043360187e-06,
|
| 43 |
+
"loss": 6.4806,
|
| 44 |
+
"step": 5
|
| 45 |
+
},
|
| 46 |
+
{
|
| 47 |
+
"epoch": 0.0512,
|
| 48 |
+
"grad_norm": 84.4494318688335,
|
| 49 |
+
"learning_rate": 7.781512503836437e-06,
|
| 50 |
+
"loss": 6.4194,
|
| 51 |
+
"step": 6
|
| 52 |
+
},
|
| 53 |
+
{
|
| 54 |
+
"epoch": 0.05973333333333333,
|
| 55 |
+
"grad_norm": 71.37977490595638,
|
| 56 |
+
"learning_rate": 8.450980400142568e-06,
|
| 57 |
+
"loss": 5.4953,
|
| 58 |
+
"step": 7
|
| 59 |
+
},
|
| 60 |
+
{
|
| 61 |
+
"epoch": 0.06826666666666667,
|
| 62 |
+
"grad_norm": 49.31153456754566,
|
| 63 |
+
"learning_rate": 9.030899869919434e-06,
|
| 64 |
+
"loss": 5.4123,
|
| 65 |
+
"step": 8
|
| 66 |
+
},
|
| 67 |
+
{
|
| 68 |
+
"epoch": 0.0768,
|
| 69 |
+
"grad_norm": 20.37296364560341,
|
| 70 |
+
"learning_rate": 9.542425094393249e-06,
|
| 71 |
+
"loss": 5.2334,
|
| 72 |
+
"step": 9
|
| 73 |
+
},
|
| 74 |
+
{
|
| 75 |
+
"epoch": 0.08533333333333333,
|
| 76 |
+
"grad_norm": 7.968467079076881,
|
| 77 |
+
"learning_rate": 9.999999999999999e-06,
|
| 78 |
+
"loss": 5.0282,
|
| 79 |
+
"step": 10
|
| 80 |
+
},
|
| 81 |
+
{
|
| 82 |
+
"epoch": 0.09386666666666667,
|
| 83 |
+
"grad_norm": 3.559446532055649,
|
| 84 |
+
"learning_rate": 1.041392685158225e-05,
|
| 85 |
+
"loss": 4.612,
|
| 86 |
+
"step": 11
|
| 87 |
+
},
|
| 88 |
+
{
|
| 89 |
+
"epoch": 0.1024,
|
| 90 |
+
"grad_norm": 3.5528846947995674,
|
| 91 |
+
"learning_rate": 1.0791812460476248e-05,
|
| 92 |
+
"loss": 4.9475,
|
| 93 |
+
"step": 12
|
| 94 |
+
},
|
| 95 |
+
{
|
| 96 |
+
"epoch": 0.11093333333333333,
|
| 97 |
+
"grad_norm": 3.541968897471334,
|
| 98 |
+
"learning_rate": 1.1139433523068365e-05,
|
| 99 |
+
"loss": 4.2777,
|
| 100 |
+
"step": 13
|
| 101 |
+
},
|
| 102 |
+
{
|
| 103 |
+
"epoch": 0.11946666666666667,
|
| 104 |
+
"grad_norm": 3.54718070036198,
|
| 105 |
+
"learning_rate": 1.1461280356782378e-05,
|
| 106 |
+
"loss": 4.3507,
|
| 107 |
+
"step": 14
|
| 108 |
+
},
|
| 109 |
+
{
|
| 110 |
+
"epoch": 0.128,
|
| 111 |
+
"grad_norm": 3.8632334830606747,
|
| 112 |
+
"learning_rate": 1.1760912590556813e-05,
|
| 113 |
+
"loss": 4.5364,
|
| 114 |
+
"step": 15
|
| 115 |
+
},
|
| 116 |
+
{
|
| 117 |
+
"epoch": 0.13653333333333334,
|
| 118 |
+
"grad_norm": 3.6637424744054004,
|
| 119 |
+
"learning_rate": 1.2041199826559246e-05,
|
| 120 |
+
"loss": 3.9672,
|
| 121 |
+
"step": 16
|
| 122 |
+
},
|
| 123 |
+
{
|
| 124 |
+
"epoch": 0.14506666666666668,
|
| 125 |
+
"grad_norm": 3.919802904818311,
|
| 126 |
+
"learning_rate": 1.230448921378274e-05,
|
| 127 |
+
"loss": 4.0618,
|
| 128 |
+
"step": 17
|
| 129 |
+
},
|
| 130 |
+
{
|
| 131 |
+
"epoch": 0.1536,
|
| 132 |
+
"grad_norm": 4.71904950738746,
|
| 133 |
+
"learning_rate": 1.2552725051033058e-05,
|
| 134 |
+
"loss": 4.6656,
|
| 135 |
+
"step": 18
|
| 136 |
+
},
|
| 137 |
+
{
|
| 138 |
+
"epoch": 0.16213333333333332,
|
| 139 |
+
"grad_norm": 4.6656317698690835,
|
| 140 |
+
"learning_rate": 1.2787536009528288e-05,
|
| 141 |
+
"loss": 4.1131,
|
| 142 |
+
"step": 19
|
| 143 |
+
},
|
| 144 |
+
{
|
| 145 |
+
"epoch": 0.17066666666666666,
|
| 146 |
+
"grad_norm": 5.145138692367417,
|
| 147 |
+
"learning_rate": 1.301029995663981e-05,
|
| 148 |
+
"loss": 4.0989,
|
| 149 |
+
"step": 20
|
| 150 |
+
},
|
| 151 |
+
{
|
| 152 |
+
"epoch": 0.1792,
|
| 153 |
+
"grad_norm": 5.923538014759818,
|
| 154 |
+
"learning_rate": 1.3222192947339192e-05,
|
| 155 |
+
"loss": 4.4991,
|
| 156 |
+
"step": 21
|
| 157 |
+
},
|
| 158 |
+
{
|
| 159 |
+
"epoch": 0.18773333333333334,
|
| 160 |
+
"grad_norm": 5.941056962941364,
|
| 161 |
+
"learning_rate": 1.3424226808222062e-05,
|
| 162 |
+
"loss": 4.0836,
|
| 163 |
+
"step": 22
|
| 164 |
+
},
|
| 165 |
+
{
|
| 166 |
+
"epoch": 0.19626666666666667,
|
| 167 |
+
"grad_norm": 6.171026012117947,
|
| 168 |
+
"learning_rate": 1.3617278360175927e-05,
|
| 169 |
+
"loss": 3.6861,
|
| 170 |
+
"step": 23
|
| 171 |
+
},
|
| 172 |
+
{
|
| 173 |
+
"epoch": 0.2048,
|
| 174 |
+
"grad_norm": 7.130542138930838,
|
| 175 |
+
"learning_rate": 1.380211241711606e-05,
|
| 176 |
+
"loss": 4.0958,
|
| 177 |
+
"step": 24
|
| 178 |
+
},
|
| 179 |
+
{
|
| 180 |
+
"epoch": 0.21333333333333335,
|
| 181 |
+
"grad_norm": 7.328837606110418,
|
| 182 |
+
"learning_rate": 1.3979400086720374e-05,
|
| 183 |
+
"loss": 3.9524,
|
| 184 |
+
"step": 25
|
| 185 |
+
},
|
| 186 |
+
{
|
| 187 |
+
"epoch": 0.22186666666666666,
|
| 188 |
+
"grad_norm": 6.923489005711429,
|
| 189 |
+
"learning_rate": 1.4149733479708178e-05,
|
| 190 |
+
"loss": 3.6062,
|
| 191 |
+
"step": 26
|
| 192 |
+
},
|
| 193 |
+
{
|
| 194 |
+
"epoch": 0.2304,
|
| 195 |
+
"grad_norm": 6.409498926059221,
|
| 196 |
+
"learning_rate": 1.4313637641589872e-05,
|
| 197 |
+
"loss": 3.2034,
|
| 198 |
+
"step": 27
|
| 199 |
+
},
|
| 200 |
+
{
|
| 201 |
+
"epoch": 0.23893333333333333,
|
| 202 |
+
"grad_norm": 5.781628405584682,
|
| 203 |
+
"learning_rate": 1.4471580313422191e-05,
|
| 204 |
+
"loss": 2.8158,
|
| 205 |
+
"step": 28
|
| 206 |
+
},
|
| 207 |
+
{
|
| 208 |
+
"epoch": 0.24746666666666667,
|
| 209 |
+
"grad_norm": 6.2927590068243315,
|
| 210 |
+
"learning_rate": 1.4623979978989559e-05,
|
| 211 |
+
"loss": 2.9803,
|
| 212 |
+
"step": 29
|
| 213 |
+
},
|
| 214 |
+
{
|
| 215 |
+
"epoch": 0.256,
|
| 216 |
+
"grad_norm": 6.103844678630006,
|
| 217 |
+
"learning_rate": 1.4771212547196623e-05,
|
| 218 |
+
"loss": 2.847,
|
| 219 |
+
"step": 30
|
| 220 |
+
},
|
| 221 |
+
{
|
| 222 |
+
"epoch": 0.26453333333333334,
|
| 223 |
+
"grad_norm": 7.656341925867144,
|
| 224 |
+
"learning_rate": 1.4913616938342726e-05,
|
| 225 |
+
"loss": 3.0907,
|
| 226 |
+
"step": 31
|
| 227 |
+
},
|
| 228 |
+
{
|
| 229 |
+
"epoch": 0.2730666666666667,
|
| 230 |
+
"grad_norm": 6.324242877501844,
|
| 231 |
+
"learning_rate": 1.5051499783199059e-05,
|
| 232 |
+
"loss": 2.3467,
|
| 233 |
+
"step": 32
|
| 234 |
+
},
|
| 235 |
+
{
|
| 236 |
+
"epoch": 0.2816,
|
| 237 |
+
"grad_norm": 7.606313101162739,
|
| 238 |
+
"learning_rate": 1.5185139398778874e-05,
|
| 239 |
+
"loss": 2.5292,
|
| 240 |
+
"step": 33
|
| 241 |
+
},
|
| 242 |
+
{
|
| 243 |
+
"epoch": 0.29013333333333335,
|
| 244 |
+
"grad_norm": 8.553792493849265,
|
| 245 |
+
"learning_rate": 1.531478917042255e-05,
|
| 246 |
+
"loss": 2.4547,
|
| 247 |
+
"step": 34
|
| 248 |
+
},
|
| 249 |
+
{
|
| 250 |
+
"epoch": 0.2986666666666667,
|
| 251 |
+
"grad_norm": 8.483368703272543,
|
| 252 |
+
"learning_rate": 1.5440680443502753e-05,
|
| 253 |
+
"loss": 2.1956,
|
| 254 |
+
"step": 35
|
| 255 |
+
},
|
| 256 |
+
{
|
| 257 |
+
"epoch": 0.3072,
|
| 258 |
+
"grad_norm": 8.364739352838077,
|
| 259 |
+
"learning_rate": 1.5563025007672873e-05,
|
| 260 |
+
"loss": 1.8552,
|
| 261 |
+
"step": 36
|
| 262 |
+
},
|
| 263 |
+
{
|
| 264 |
+
"epoch": 0.3157333333333333,
|
| 265 |
+
"grad_norm": 9.37663682000104,
|
| 266 |
+
"learning_rate": 1.5682017240669948e-05,
|
| 267 |
+
"loss": 1.9228,
|
| 268 |
+
"step": 37
|
| 269 |
+
},
|
| 270 |
+
{
|
| 271 |
+
"epoch": 0.32426666666666665,
|
| 272 |
+
"grad_norm": 9.823047193440066,
|
| 273 |
+
"learning_rate": 1.57978359661681e-05,
|
| 274 |
+
"loss": 1.7033,
|
| 275 |
+
"step": 38
|
| 276 |
+
},
|
| 277 |
+
{
|
| 278 |
+
"epoch": 0.3328,
|
| 279 |
+
"grad_norm": 9.692618512955894,
|
| 280 |
+
"learning_rate": 1.591064607026499e-05,
|
| 281 |
+
"loss": 1.3768,
|
| 282 |
+
"step": 39
|
| 283 |
+
},
|
| 284 |
+
{
|
| 285 |
+
"epoch": 0.3413333333333333,
|
| 286 |
+
"grad_norm": 9.08889021911031,
|
| 287 |
+
"learning_rate": 1.6020599913279622e-05,
|
| 288 |
+
"loss": 1.3015,
|
| 289 |
+
"step": 40
|
| 290 |
+
},
|
| 291 |
+
{
|
| 292 |
+
"epoch": 0.34986666666666666,
|
| 293 |
+
"grad_norm": 8.081534221516058,
|
| 294 |
+
"learning_rate": 1.6127838567197353e-05,
|
| 295 |
+
"loss": 0.9228,
|
| 296 |
+
"step": 41
|
| 297 |
+
},
|
| 298 |
+
{
|
| 299 |
+
"epoch": 0.3584,
|
| 300 |
+
"grad_norm": 6.238638048950311,
|
| 301 |
+
"learning_rate": 1.6232492903979005e-05,
|
| 302 |
+
"loss": 0.7267,
|
| 303 |
+
"step": 42
|
| 304 |
+
},
|
| 305 |
+
{
|
| 306 |
+
"epoch": 0.36693333333333333,
|
| 307 |
+
"grad_norm": 3.4058036861773604,
|
| 308 |
+
"learning_rate": 1.6334684555795865e-05,
|
| 309 |
+
"loss": 0.5875,
|
| 310 |
+
"step": 43
|
| 311 |
+
},
|
| 312 |
+
{
|
| 313 |
+
"epoch": 0.37546666666666667,
|
| 314 |
+
"grad_norm": 2.079163829467713,
|
| 315 |
+
"learning_rate": 1.6434526764861872e-05,
|
| 316 |
+
"loss": 0.6355,
|
| 317 |
+
"step": 44
|
| 318 |
+
},
|
| 319 |
+
{
|
| 320 |
+
"epoch": 0.384,
|
| 321 |
+
"grad_norm": 1.5597487833024746,
|
| 322 |
+
"learning_rate": 1.6532125137753435e-05,
|
| 323 |
+
"loss": 0.5106,
|
| 324 |
+
"step": 45
|
| 325 |
+
},
|
| 326 |
+
{
|
| 327 |
+
"epoch": 0.39253333333333335,
|
| 328 |
+
"grad_norm": 2.491689602375256,
|
| 329 |
+
"learning_rate": 1.662757831681574e-05,
|
| 330 |
+
"loss": 0.6454,
|
| 331 |
+
"step": 46
|
| 332 |
+
},
|
| 333 |
+
{
|
| 334 |
+
"epoch": 0.4010666666666667,
|
| 335 |
+
"grad_norm": 2.010880438195854,
|
| 336 |
+
"learning_rate": 1.672097857935717e-05,
|
| 337 |
+
"loss": 0.4757,
|
| 338 |
+
"step": 47
|
| 339 |
+
},
|
| 340 |
+
{
|
| 341 |
+
"epoch": 0.4096,
|
| 342 |
+
"grad_norm": 1.9452805114322096,
|
| 343 |
+
"learning_rate": 1.681241237375587e-05,
|
| 344 |
+
"loss": 0.4133,
|
| 345 |
+
"step": 48
|
| 346 |
+
},
|
| 347 |
+
{
|
| 348 |
+
"epoch": 0.41813333333333336,
|
| 349 |
+
"grad_norm": 1.7620848552306103,
|
| 350 |
+
"learning_rate": 1.6901960800285137e-05,
|
| 351 |
+
"loss": 0.4004,
|
| 352 |
+
"step": 49
|
| 353 |
+
},
|
| 354 |
+
{
|
| 355 |
+
"epoch": 0.4266666666666667,
|
| 356 |
+
"grad_norm": 1.278224489774809,
|
| 357 |
+
"learning_rate": 1.6989700043360187e-05,
|
| 358 |
+
"loss": 0.3523,
|
| 359 |
+
"step": 50
|
| 360 |
+
},
|
| 361 |
+
{
|
| 362 |
+
"epoch": 0.4352,
|
| 363 |
+
"grad_norm": 1.6151354758303231,
|
| 364 |
+
"learning_rate": 1.7075701760979363e-05,
|
| 365 |
+
"loss": 0.4317,
|
| 366 |
+
"step": 51
|
| 367 |
+
},
|
| 368 |
+
{
|
| 369 |
+
"epoch": 0.4437333333333333,
|
| 370 |
+
"grad_norm": 1.3451396055695035,
|
| 371 |
+
"learning_rate": 1.716003343634799e-05,
|
| 372 |
+
"loss": 0.3474,
|
| 373 |
+
"step": 52
|
| 374 |
+
},
|
| 375 |
+
{
|
| 376 |
+
"epoch": 0.45226666666666665,
|
| 377 |
+
"grad_norm": 1.6814977782362666,
|
| 378 |
+
"learning_rate": 1.724275869600789e-05,
|
| 379 |
+
"loss": 0.3706,
|
| 380 |
+
"step": 53
|
| 381 |
+
},
|
| 382 |
+
{
|
| 383 |
+
"epoch": 0.4608,
|
| 384 |
+
"grad_norm": 1.1682442432688667,
|
| 385 |
+
"learning_rate": 1.7323937598229687e-05,
|
| 386 |
+
"loss": 0.3488,
|
| 387 |
+
"step": 54
|
| 388 |
+
},
|
| 389 |
+
{
|
| 390 |
+
"epoch": 0.4693333333333333,
|
| 391 |
+
"grad_norm": 0.8839814540462471,
|
| 392 |
+
"learning_rate": 1.7403626894942437e-05,
|
| 393 |
+
"loss": 0.293,
|
| 394 |
+
"step": 55
|
| 395 |
+
},
|
| 396 |
+
{
|
| 397 |
+
"epoch": 0.47786666666666666,
|
| 398 |
+
"grad_norm": 0.7974304806295485,
|
| 399 |
+
"learning_rate": 1.7481880270062003e-05,
|
| 400 |
+
"loss": 0.2717,
|
| 401 |
+
"step": 56
|
| 402 |
+
},
|
| 403 |
+
{
|
| 404 |
+
"epoch": 0.4864,
|
| 405 |
+
"grad_norm": 0.6232203657585239,
|
| 406 |
+
"learning_rate": 1.7558748556724913e-05,
|
| 407 |
+
"loss": 0.1741,
|
| 408 |
+
"step": 57
|
| 409 |
+
},
|
| 410 |
+
{
|
| 411 |
+
"epoch": 0.49493333333333334,
|
| 412 |
+
"grad_norm": 0.6850688604563008,
|
| 413 |
+
"learning_rate": 1.763427993562937e-05,
|
| 414 |
+
"loss": 0.228,
|
| 415 |
+
"step": 58
|
| 416 |
+
},
|
| 417 |
+
{
|
| 418 |
+
"epoch": 0.5034666666666666,
|
| 419 |
+
"grad_norm": 0.5923826384300431,
|
| 420 |
+
"learning_rate": 1.7708520116421443e-05,
|
| 421 |
+
"loss": 0.2131,
|
| 422 |
+
"step": 59
|
| 423 |
+
},
|
| 424 |
+
{
|
| 425 |
+
"epoch": 0.512,
|
| 426 |
+
"grad_norm": 0.489754430485032,
|
| 427 |
+
"learning_rate": 1.7781512503836432e-05,
|
| 428 |
+
"loss": 0.165,
|
| 429 |
+
"step": 60
|
| 430 |
+
},
|
| 431 |
+
{
|
| 432 |
+
"epoch": 0.5205333333333333,
|
| 433 |
+
"grad_norm": 0.5280940052395061,
|
| 434 |
+
"learning_rate": 1.7853298350107667e-05,
|
| 435 |
+
"loss": 0.1658,
|
| 436 |
+
"step": 61
|
| 437 |
+
},
|
| 438 |
+
{
|
| 439 |
+
"epoch": 0.5290666666666667,
|
| 440 |
+
"grad_norm": 0.4750905992036739,
|
| 441 |
+
"learning_rate": 1.7923916894982537e-05,
|
| 442 |
+
"loss": 0.1438,
|
| 443 |
+
"step": 62
|
| 444 |
+
},
|
| 445 |
+
{
|
| 446 |
+
"epoch": 0.5376,
|
| 447 |
+
"grad_norm": 0.4162998392722401,
|
| 448 |
+
"learning_rate": 1.7993405494535815e-05,
|
| 449 |
+
"loss": 0.1555,
|
| 450 |
+
"step": 63
|
| 451 |
+
},
|
| 452 |
+
{
|
| 453 |
+
"epoch": 0.5461333333333334,
|
| 454 |
+
"grad_norm": 0.26863266246370443,
|
| 455 |
+
"learning_rate": 1.806179973983887e-05,
|
| 456 |
+
"loss": 0.1323,
|
| 457 |
+
"step": 64
|
| 458 |
+
},
|
| 459 |
+
{
|
| 460 |
+
"epoch": 0.5546666666666666,
|
| 461 |
+
"grad_norm": 0.26534221125601215,
|
| 462 |
+
"learning_rate": 1.8129133566428553e-05,
|
| 463 |
+
"loss": 0.1671,
|
| 464 |
+
"step": 65
|
| 465 |
+
},
|
| 466 |
+
{
|
| 467 |
+
"epoch": 0.5632,
|
| 468 |
+
"grad_norm": 0.2548662962257576,
|
| 469 |
+
"learning_rate": 1.8195439355418686e-05,
|
| 470 |
+
"loss": 0.1308,
|
| 471 |
+
"step": 66
|
| 472 |
+
},
|
| 473 |
+
{
|
| 474 |
+
"epoch": 0.5717333333333333,
|
| 475 |
+
"grad_norm": 0.18045394638578796,
|
| 476 |
+
"learning_rate": 1.8260748027008263e-05,
|
| 477 |
+
"loss": 0.1262,
|
| 478 |
+
"step": 67
|
| 479 |
+
},
|
| 480 |
+
{
|
| 481 |
+
"epoch": 0.5802666666666667,
|
| 482 |
+
"grad_norm": 0.17070388073612064,
|
| 483 |
+
"learning_rate": 1.8325089127062364e-05,
|
| 484 |
+
"loss": 0.1192,
|
| 485 |
+
"step": 68
|
| 486 |
+
},
|
| 487 |
+
{
|
| 488 |
+
"epoch": 0.5888,
|
| 489 |
+
"grad_norm": 0.1531381679244776,
|
| 490 |
+
"learning_rate": 1.8388490907372553e-05,
|
| 491 |
+
"loss": 0.1274,
|
| 492 |
+
"step": 69
|
| 493 |
+
},
|
| 494 |
+
{
|
| 495 |
+
"epoch": 0.5973333333333334,
|
| 496 |
+
"grad_norm": 0.18196434993687946,
|
| 497 |
+
"learning_rate": 1.8450980400142568e-05,
|
| 498 |
+
"loss": 0.1375,
|
| 499 |
+
"step": 70
|
| 500 |
+
},
|
| 501 |
+
{
|
| 502 |
+
"epoch": 0.6058666666666667,
|
| 503 |
+
"grad_norm": 0.15324416972951205,
|
| 504 |
+
"learning_rate": 1.8512583487190752e-05,
|
| 505 |
+
"loss": 0.1599,
|
| 506 |
+
"step": 71
|
| 507 |
+
},
|
| 508 |
+
{
|
| 509 |
+
"epoch": 0.6144,
|
| 510 |
+
"grad_norm": 0.10884462064503801,
|
| 511 |
+
"learning_rate": 1.857332496431268e-05,
|
| 512 |
+
"loss": 0.1041,
|
| 513 |
+
"step": 72
|
| 514 |
+
},
|
| 515 |
+
{
|
| 516 |
+
"epoch": 0.6229333333333333,
|
| 517 |
+
"grad_norm": 0.12915133528192668,
|
| 518 |
+
"learning_rate": 1.8633228601204555e-05,
|
| 519 |
+
"loss": 0.1406,
|
| 520 |
+
"step": 73
|
| 521 |
+
},
|
| 522 |
+
{
|
| 523 |
+
"epoch": 0.6314666666666666,
|
| 524 |
+
"grad_norm": 0.12553425699952878,
|
| 525 |
+
"learning_rate": 1.8692317197309763e-05,
|
| 526 |
+
"loss": 0.1256,
|
| 527 |
+
"step": 74
|
| 528 |
+
},
|
| 529 |
+
{
|
| 530 |
+
"epoch": 0.64,
|
| 531 |
+
"grad_norm": 0.11976960918968543,
|
| 532 |
+
"learning_rate": 1.8750612633916997e-05,
|
| 533 |
+
"loss": 0.1144,
|
| 534 |
+
"step": 75
|
| 535 |
+
},
|
| 536 |
+
{
|
| 537 |
+
"epoch": 0.6485333333333333,
|
| 538 |
+
"grad_norm": 0.115805998298789,
|
| 539 |
+
"learning_rate": 1.8808135922807914e-05,
|
| 540 |
+
"loss": 0.1528,
|
| 541 |
+
"step": 76
|
| 542 |
+
},
|
| 543 |
+
{
|
| 544 |
+
"epoch": 0.6570666666666667,
|
| 545 |
+
"grad_norm": 0.10325948496697443,
|
| 546 |
+
"learning_rate": 1.8864907251724818e-05,
|
| 547 |
+
"loss": 0.1044,
|
| 548 |
+
"step": 77
|
| 549 |
+
},
|
| 550 |
+
{
|
| 551 |
+
"epoch": 0.6656,
|
| 552 |
+
"grad_norm": 0.09595064346541006,
|
| 553 |
+
"learning_rate": 1.8920946026904802e-05,
|
| 554 |
+
"loss": 0.1534,
|
| 555 |
+
"step": 78
|
| 556 |
+
},
|
| 557 |
+
{
|
| 558 |
+
"epoch": 0.6741333333333334,
|
| 559 |
+
"grad_norm": 0.08796742845240496,
|
| 560 |
+
"learning_rate": 1.8976270912904414e-05,
|
| 561 |
+
"loss": 0.1155,
|
| 562 |
+
"step": 79
|
| 563 |
+
},
|
| 564 |
+
{
|
| 565 |
+
"epoch": 0.6826666666666666,
|
| 566 |
+
"grad_norm": 0.08218991738379527,
|
| 567 |
+
"learning_rate": 1.9030899869919434e-05,
|
| 568 |
+
"loss": 0.1311,
|
| 569 |
+
"step": 80
|
| 570 |
+
},
|
| 571 |
+
{
|
| 572 |
+
"epoch": 0.6912,
|
| 573 |
+
"grad_norm": 0.08290815261109215,
|
| 574 |
+
"learning_rate": 1.9084850188786497e-05,
|
| 575 |
+
"loss": 0.11,
|
| 576 |
+
"step": 81
|
| 577 |
+
},
|
| 578 |
+
{
|
| 579 |
+
"epoch": 0.6997333333333333,
|
| 580 |
+
"grad_norm": 0.0794132180176064,
|
| 581 |
+
"learning_rate": 1.9138138523837165e-05,
|
| 582 |
+
"loss": 0.1135,
|
| 583 |
+
"step": 82
|
| 584 |
+
},
|
| 585 |
+
{
|
| 586 |
+
"epoch": 0.7082666666666667,
|
| 587 |
+
"grad_norm": 0.06934410705255296,
|
| 588 |
+
"learning_rate": 1.919078092376074e-05,
|
| 589 |
+
"loss": 0.109,
|
| 590 |
+
"step": 83
|
| 591 |
+
},
|
| 592 |
+
{
|
| 593 |
+
"epoch": 0.7168,
|
| 594 |
+
"grad_norm": 0.09000563031870593,
|
| 595 |
+
"learning_rate": 1.9242792860618813e-05,
|
| 596 |
+
"loss": 0.12,
|
| 597 |
+
"step": 84
|
| 598 |
+
},
|
| 599 |
+
{
|
| 600 |
+
"epoch": 0.7253333333333334,
|
| 601 |
+
"grad_norm": 0.1134042277879818,
|
| 602 |
+
"learning_rate": 1.929418925714293e-05,
|
| 603 |
+
"loss": 0.1223,
|
| 604 |
+
"step": 85
|
| 605 |
+
},
|
| 606 |
+
{
|
| 607 |
+
"epoch": 0.7338666666666667,
|
| 608 |
+
"grad_norm": 0.09118764690233076,
|
| 609 |
+
"learning_rate": 1.9344984512435673e-05,
|
| 610 |
+
"loss": 0.1459,
|
| 611 |
+
"step": 86
|
| 612 |
+
},
|
| 613 |
+
{
|
| 614 |
+
"epoch": 0.7424,
|
| 615 |
+
"grad_norm": 0.07873016754353963,
|
| 616 |
+
"learning_rate": 1.9395192526186183e-05,
|
| 617 |
+
"loss": 0.1422,
|
| 618 |
+
"step": 87
|
| 619 |
+
},
|
| 620 |
+
{
|
| 621 |
+
"epoch": 0.7509333333333333,
|
| 622 |
+
"grad_norm": 0.1796495874463076,
|
| 623 |
+
"learning_rate": 1.9444826721501687e-05,
|
| 624 |
+
"loss": 0.1291,
|
| 625 |
+
"step": 88
|
| 626 |
+
},
|
| 627 |
+
{
|
| 628 |
+
"epoch": 0.7594666666666666,
|
| 629 |
+
"grad_norm": 0.0679589944174269,
|
| 630 |
+
"learning_rate": 1.9493900066449125e-05,
|
| 631 |
+
"loss": 0.108,
|
| 632 |
+
"step": 89
|
| 633 |
+
},
|
| 634 |
+
{
|
| 635 |
+
"epoch": 0.768,
|
| 636 |
+
"grad_norm": 0.08174688574235538,
|
| 637 |
+
"learning_rate": 1.9542425094393246e-05,
|
| 638 |
+
"loss": 0.1081,
|
| 639 |
+
"step": 90
|
| 640 |
+
},
|
| 641 |
+
{
|
| 642 |
+
"epoch": 0.7765333333333333,
|
| 643 |
+
"grad_norm": 0.057137370501406756,
|
| 644 |
+
"learning_rate": 1.9590413923210934e-05,
|
| 645 |
+
"loss": 0.0934,
|
| 646 |
+
"step": 91
|
| 647 |
+
},
|
| 648 |
+
{
|
| 649 |
+
"epoch": 0.7850666666666667,
|
| 650 |
+
"grad_norm": 0.06578111924908255,
|
| 651 |
+
"learning_rate": 1.9637878273455555e-05,
|
| 652 |
+
"loss": 0.1085,
|
| 653 |
+
"step": 92
|
| 654 |
+
},
|
| 655 |
+
{
|
| 656 |
+
"epoch": 0.7936,
|
| 657 |
+
"grad_norm": 0.08945990540906254,
|
| 658 |
+
"learning_rate": 1.968482948553935e-05,
|
| 659 |
+
"loss": 0.1747,
|
| 660 |
+
"step": 93
|
| 661 |
+
},
|
| 662 |
+
{
|
| 663 |
+
"epoch": 0.8021333333333334,
|
| 664 |
+
"grad_norm": 0.06183863311044229,
|
| 665 |
+
"learning_rate": 1.9731278535996986e-05,
|
| 666 |
+
"loss": 0.1136,
|
| 667 |
+
"step": 94
|
| 668 |
+
},
|
| 669 |
+
{
|
| 670 |
+
"epoch": 0.8106666666666666,
|
| 671 |
+
"grad_norm": 0.05777899602544702,
|
| 672 |
+
"learning_rate": 1.9777236052888476e-05,
|
| 673 |
+
"loss": 0.0984,
|
| 674 |
+
"step": 95
|
| 675 |
+
},
|
| 676 |
+
{
|
| 677 |
+
"epoch": 0.8192,
|
| 678 |
+
"grad_norm": 0.08130851607693534,
|
| 679 |
+
"learning_rate": 1.9822712330395683e-05,
|
| 680 |
+
"loss": 0.187,
|
| 681 |
+
"step": 96
|
| 682 |
+
},
|
| 683 |
+
{
|
| 684 |
+
"epoch": 0.8277333333333333,
|
| 685 |
+
"grad_norm": 0.06426546202002927,
|
| 686 |
+
"learning_rate": 1.986771734266245e-05,
|
| 687 |
+
"loss": 0.1296,
|
| 688 |
+
"step": 97
|
| 689 |
+
},
|
| 690 |
+
{
|
| 691 |
+
"epoch": 0.8362666666666667,
|
| 692 |
+
"grad_norm": 0.069692313707994,
|
| 693 |
+
"learning_rate": 1.991226075692495e-05,
|
| 694 |
+
"loss": 0.1404,
|
| 695 |
+
"step": 98
|
| 696 |
+
},
|
| 697 |
+
{
|
| 698 |
+
"epoch": 0.8448,
|
| 699 |
+
"grad_norm": 0.05494542266886729,
|
| 700 |
+
"learning_rate": 1.9956351945975496e-05,
|
| 701 |
+
"loss": 0.116,
|
| 702 |
+
"step": 99
|
| 703 |
+
},
|
| 704 |
+
{
|
| 705 |
+
"epoch": 0.8533333333333334,
|
| 706 |
+
"grad_norm": 0.07571686966840627,
|
| 707 |
+
"learning_rate": 1.9999999999999998e-05,
|
| 708 |
+
"loss": 0.1539,
|
| 709 |
+
"step": 100
|
| 710 |
+
},
|
| 711 |
+
{
|
| 712 |
+
"epoch": 0.8618666666666667,
|
| 713 |
+
"grad_norm": 0.054351059117603705,
|
| 714 |
+
"learning_rate": 2e-05,
|
| 715 |
+
"loss": 0.1037,
|
| 716 |
+
"step": 101
|
| 717 |
+
},
|
| 718 |
+
{
|
| 719 |
+
"epoch": 0.8704,
|
| 720 |
+
"grad_norm": 0.06531899611551092,
|
| 721 |
+
"learning_rate": 2e-05,
|
| 722 |
+
"loss": 0.0827,
|
| 723 |
+
"step": 102
|
| 724 |
+
},
|
| 725 |
+
{
|
| 726 |
+
"epoch": 0.8789333333333333,
|
| 727 |
+
"grad_norm": 0.06131678646504652,
|
| 728 |
+
"learning_rate": 2e-05,
|
| 729 |
+
"loss": 0.1266,
|
| 730 |
+
"step": 103
|
| 731 |
+
},
|
| 732 |
+
{
|
| 733 |
+
"epoch": 0.8874666666666666,
|
| 734 |
+
"grad_norm": 0.06850220540661824,
|
| 735 |
+
"learning_rate": 2e-05,
|
| 736 |
+
"loss": 0.1456,
|
| 737 |
+
"step": 104
|
| 738 |
+
},
|
| 739 |
+
{
|
| 740 |
+
"epoch": 0.896,
|
| 741 |
+
"grad_norm": 0.05806908951252483,
|
| 742 |
+
"learning_rate": 2e-05,
|
| 743 |
+
"loss": 0.0954,
|
| 744 |
+
"step": 105
|
| 745 |
+
},
|
| 746 |
+
{
|
| 747 |
+
"epoch": 0.9045333333333333,
|
| 748 |
+
"grad_norm": 0.06503642452033717,
|
| 749 |
+
"learning_rate": 2e-05,
|
| 750 |
+
"loss": 0.1417,
|
| 751 |
+
"step": 106
|
| 752 |
+
},
|
| 753 |
+
{
|
| 754 |
+
"epoch": 0.9130666666666667,
|
| 755 |
+
"grad_norm": 0.050486271853277066,
|
| 756 |
+
"learning_rate": 2e-05,
|
| 757 |
+
"loss": 0.0959,
|
| 758 |
+
"step": 107
|
| 759 |
+
},
|
| 760 |
+
{
|
| 761 |
+
"epoch": 0.9216,
|
| 762 |
+
"grad_norm": 0.07746063813802379,
|
| 763 |
+
"learning_rate": 2e-05,
|
| 764 |
+
"loss": 0.1256,
|
| 765 |
+
"step": 108
|
| 766 |
+
},
|
| 767 |
+
{
|
| 768 |
+
"epoch": 0.9301333333333334,
|
| 769 |
+
"grad_norm": 0.051231172380840004,
|
| 770 |
+
"learning_rate": 2e-05,
|
| 771 |
+
"loss": 0.1116,
|
| 772 |
+
"step": 109
|
| 773 |
+
},
|
| 774 |
+
{
|
| 775 |
+
"epoch": 0.9386666666666666,
|
| 776 |
+
"grad_norm": 0.056296443557859455,
|
| 777 |
+
"learning_rate": 2e-05,
|
| 778 |
+
"loss": 0.1056,
|
| 779 |
+
"step": 110
|
| 780 |
+
},
|
| 781 |
+
{
|
| 782 |
+
"epoch": 0.9472,
|
| 783 |
+
"grad_norm": 0.05058663240713958,
|
| 784 |
+
"learning_rate": 2e-05,
|
| 785 |
+
"loss": 0.0971,
|
| 786 |
+
"step": 111
|
| 787 |
+
},
|
| 788 |
+
{
|
| 789 |
+
"epoch": 0.9557333333333333,
|
| 790 |
+
"grad_norm": 0.05532886570130611,
|
| 791 |
+
"learning_rate": 2e-05,
|
| 792 |
+
"loss": 0.1086,
|
| 793 |
+
"step": 112
|
| 794 |
+
},
|
| 795 |
+
{
|
| 796 |
+
"epoch": 0.9642666666666667,
|
| 797 |
+
"grad_norm": 0.05327811326654907,
|
| 798 |
+
"learning_rate": 2e-05,
|
| 799 |
+
"loss": 0.0989,
|
| 800 |
+
"step": 113
|
| 801 |
+
},
|
| 802 |
+
{
|
| 803 |
+
"epoch": 0.9728,
|
| 804 |
+
"grad_norm": 0.05663279364147864,
|
| 805 |
+
"learning_rate": 2e-05,
|
| 806 |
+
"loss": 0.0958,
|
| 807 |
+
"step": 114
|
| 808 |
+
},
|
| 809 |
+
{
|
| 810 |
+
"epoch": 0.9813333333333333,
|
| 811 |
+
"grad_norm": 0.04930904541225805,
|
| 812 |
+
"learning_rate": 2e-05,
|
| 813 |
+
"loss": 0.0887,
|
| 814 |
+
"step": 115
|
| 815 |
+
},
|
| 816 |
+
{
|
| 817 |
+
"epoch": 0.9898666666666667,
|
| 818 |
+
"grad_norm": 0.06096947951115022,
|
| 819 |
+
"learning_rate": 2e-05,
|
| 820 |
+
"loss": 0.106,
|
| 821 |
+
"step": 116
|
| 822 |
+
},
|
| 823 |
+
{
|
| 824 |
+
"epoch": 0.9984,
|
| 825 |
+
"grad_norm": 0.050092322361182495,
|
| 826 |
+
"learning_rate": 2e-05,
|
| 827 |
+
"loss": 0.0931,
|
| 828 |
+
"step": 117
|
| 829 |
+
},
|
| 830 |
+
{
|
| 831 |
+
"epoch": 1.0069333333333332,
|
| 832 |
+
"grad_norm": 0.04980408443758999,
|
| 833 |
+
"learning_rate": 2e-05,
|
| 834 |
+
"loss": 0.0955,
|
| 835 |
+
"step": 118
|
| 836 |
+
},
|
| 837 |
+
{
|
| 838 |
+
"epoch": 1.0154666666666667,
|
| 839 |
+
"grad_norm": 0.051183082721834305,
|
| 840 |
+
"learning_rate": 2e-05,
|
| 841 |
+
"loss": 0.1049,
|
| 842 |
+
"step": 119
|
| 843 |
+
},
|
| 844 |
+
{
|
| 845 |
+
"epoch": 1.024,
|
| 846 |
+
"grad_norm": 0.04332220265802814,
|
| 847 |
+
"learning_rate": 2e-05,
|
| 848 |
+
"loss": 0.0983,
|
| 849 |
+
"step": 120
|
| 850 |
+
},
|
| 851 |
+
{
|
| 852 |
+
"epoch": 1.0325333333333333,
|
| 853 |
+
"grad_norm": 0.07211731499677299,
|
| 854 |
+
"learning_rate": 2e-05,
|
| 855 |
+
"loss": 0.1386,
|
| 856 |
+
"step": 121
|
| 857 |
+
},
|
| 858 |
+
{
|
| 859 |
+
"epoch": 1.0410666666666666,
|
| 860 |
+
"grad_norm": 0.06550870223740553,
|
| 861 |
+
"learning_rate": 2e-05,
|
| 862 |
+
"loss": 0.1334,
|
| 863 |
+
"step": 122
|
| 864 |
+
},
|
| 865 |
+
{
|
| 866 |
+
"epoch": 1.0496,
|
| 867 |
+
"grad_norm": 0.05331839690767287,
|
| 868 |
+
"learning_rate": 2e-05,
|
| 869 |
+
"loss": 0.1014,
|
| 870 |
+
"step": 123
|
| 871 |
+
},
|
| 872 |
+
{
|
| 873 |
+
"epoch": 1.0581333333333334,
|
| 874 |
+
"grad_norm": 0.05227685628767905,
|
| 875 |
+
"learning_rate": 2e-05,
|
| 876 |
+
"loss": 0.1098,
|
| 877 |
+
"step": 124
|
| 878 |
+
},
|
| 879 |
+
{
|
| 880 |
+
"epoch": 1.0666666666666667,
|
| 881 |
+
"grad_norm": 0.07641693882491171,
|
| 882 |
+
"learning_rate": 2e-05,
|
| 883 |
+
"loss": 0.1127,
|
| 884 |
+
"step": 125
|
| 885 |
+
},
|
| 886 |
+
{
|
| 887 |
+
"epoch": 1.0752,
|
| 888 |
+
"grad_norm": 0.052835367770791786,
|
| 889 |
+
"learning_rate": 2e-05,
|
| 890 |
+
"loss": 0.1231,
|
| 891 |
+
"step": 126
|
| 892 |
+
},
|
| 893 |
+
{
|
| 894 |
+
"epoch": 1.0837333333333334,
|
| 895 |
+
"grad_norm": 0.07520329755025788,
|
| 896 |
+
"learning_rate": 2e-05,
|
| 897 |
+
"loss": 0.085,
|
| 898 |
+
"step": 127
|
| 899 |
+
},
|
| 900 |
+
{
|
| 901 |
+
"epoch": 1.0922666666666667,
|
| 902 |
+
"grad_norm": 0.07670066152157425,
|
| 903 |
+
"learning_rate": 2e-05,
|
| 904 |
+
"loss": 0.1071,
|
| 905 |
+
"step": 128
|
| 906 |
+
},
|
| 907 |
+
{
|
| 908 |
+
"epoch": 1.1008,
|
| 909 |
+
"grad_norm": 0.052832906560645154,
|
| 910 |
+
"learning_rate": 2e-05,
|
| 911 |
+
"loss": 0.1093,
|
| 912 |
+
"step": 129
|
| 913 |
+
},
|
| 914 |
+
{
|
| 915 |
+
"epoch": 1.1093333333333333,
|
| 916 |
+
"grad_norm": 0.06573889037311398,
|
| 917 |
+
"learning_rate": 2e-05,
|
| 918 |
+
"loss": 0.1193,
|
| 919 |
+
"step": 130
|
| 920 |
+
},
|
| 921 |
+
{
|
| 922 |
+
"epoch": 1.1178666666666666,
|
| 923 |
+
"grad_norm": 0.05175471296566334,
|
| 924 |
+
"learning_rate": 2e-05,
|
| 925 |
+
"loss": 0.1184,
|
| 926 |
+
"step": 131
|
| 927 |
+
},
|
| 928 |
+
{
|
| 929 |
+
"epoch": 1.1264,
|
| 930 |
+
"grad_norm": 0.05912231419793496,
|
| 931 |
+
"learning_rate": 2e-05,
|
| 932 |
+
"loss": 0.1154,
|
| 933 |
+
"step": 132
|
| 934 |
+
},
|
| 935 |
+
{
|
| 936 |
+
"epoch": 1.1349333333333333,
|
| 937 |
+
"grad_norm": 0.04899140475981105,
|
| 938 |
+
"learning_rate": 2e-05,
|
| 939 |
+
"loss": 0.0957,
|
| 940 |
+
"step": 133
|
| 941 |
+
},
|
| 942 |
+
{
|
| 943 |
+
"epoch": 1.1434666666666666,
|
| 944 |
+
"grad_norm": 0.05939916939142137,
|
| 945 |
+
"learning_rate": 2e-05,
|
| 946 |
+
"loss": 0.0979,
|
| 947 |
+
"step": 134
|
| 948 |
+
},
|
| 949 |
+
{
|
| 950 |
+
"epoch": 1.152,
|
| 951 |
+
"grad_norm": 0.0516819217599706,
|
| 952 |
+
"learning_rate": 2e-05,
|
| 953 |
+
"loss": 0.0834,
|
| 954 |
+
"step": 135
|
| 955 |
+
},
|
| 956 |
+
{
|
| 957 |
+
"epoch": 1.1605333333333334,
|
| 958 |
+
"grad_norm": 0.05456440346454737,
|
| 959 |
+
"learning_rate": 2e-05,
|
| 960 |
+
"loss": 0.1183,
|
| 961 |
+
"step": 136
|
| 962 |
+
},
|
| 963 |
+
{
|
| 964 |
+
"epoch": 1.1690666666666667,
|
| 965 |
+
"grad_norm": 0.059906464476343235,
|
| 966 |
+
"learning_rate": 2e-05,
|
| 967 |
+
"loss": 0.1048,
|
| 968 |
+
"step": 137
|
| 969 |
+
},
|
| 970 |
+
{
|
| 971 |
+
"epoch": 1.1776,
|
| 972 |
+
"grad_norm": 0.0720112680204319,
|
| 973 |
+
"learning_rate": 2e-05,
|
| 974 |
+
"loss": 0.1168,
|
| 975 |
+
"step": 138
|
| 976 |
+
},
|
| 977 |
+
{
|
| 978 |
+
"epoch": 1.1861333333333333,
|
| 979 |
+
"grad_norm": 0.04940202805828527,
|
| 980 |
+
"learning_rate": 2e-05,
|
| 981 |
+
"loss": 0.0948,
|
| 982 |
+
"step": 139
|
| 983 |
+
},
|
| 984 |
+
{
|
| 985 |
+
"epoch": 1.1946666666666665,
|
| 986 |
+
"grad_norm": 0.060088609545130046,
|
| 987 |
+
"learning_rate": 2e-05,
|
| 988 |
+
"loss": 0.0952,
|
| 989 |
+
"step": 140
|
| 990 |
+
}
|
| 991 |
+
],
|
| 992 |
+
"logging_steps": 1,
|
| 993 |
+
"max_steps": 351,
|
| 994 |
+
"num_input_tokens_seen": 0,
|
| 995 |
+
"num_train_epochs": 3,
|
| 996 |
+
"save_steps": 20,
|
| 997 |
+
"stateful_callbacks": {
|
| 998 |
+
"TrainerControl": {
|
| 999 |
+
"args": {
|
| 1000 |
+
"should_epoch_stop": false,
|
| 1001 |
+
"should_evaluate": false,
|
| 1002 |
+
"should_log": false,
|
| 1003 |
+
"should_save": true,
|
| 1004 |
+
"should_training_stop": false
|
| 1005 |
+
},
|
| 1006 |
+
"attributes": {}
|
| 1007 |
+
}
|
| 1008 |
+
},
|
| 1009 |
+
"total_flos": 1.9872207486907843e+18,
|
| 1010 |
+
"train_batch_size": 16,
|
| 1011 |
+
"trial_name": null,
|
| 1012 |
+
"trial_params": null
|
| 1013 |
+
}
|
checkpoint-140/training_args.bin
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:aa4edad1da5557fcd52a6da980443588016cead6f0444a3562cfa68029c66a04
|
| 3 |
+
size 6840
|
checkpoint-140/zero_to_fp32.py
ADDED
|
@@ -0,0 +1,592 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 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):
|
| 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 |
+
elif zero_stage == 3:
|
| 216 |
+
return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
|
| 217 |
+
|
| 218 |
+
|
| 219 |
+
def _zero2_merge_frozen_params(state_dict, zero_model_states):
|
| 220 |
+
if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
|
| 221 |
+
return
|
| 222 |
+
|
| 223 |
+
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
|
| 224 |
+
frozen_param_fragments = zero_model_states[0].frozen_param_fragments
|
| 225 |
+
|
| 226 |
+
if debug:
|
| 227 |
+
num_elem = sum(s.numel() for s in frozen_param_shapes.values())
|
| 228 |
+
print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
|
| 229 |
+
|
| 230 |
+
wanted_params = len(frozen_param_shapes)
|
| 231 |
+
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
| 232 |
+
avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
|
| 233 |
+
print(f'Frozen params: Have {avail_numel} numels to process.')
|
| 234 |
+
print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
|
| 235 |
+
|
| 236 |
+
total_params = 0
|
| 237 |
+
total_numel = 0
|
| 238 |
+
for name, shape in frozen_param_shapes.items():
|
| 239 |
+
total_params += 1
|
| 240 |
+
unpartitioned_numel = shape.numel()
|
| 241 |
+
total_numel += unpartitioned_numel
|
| 242 |
+
|
| 243 |
+
state_dict[name] = frozen_param_fragments[name]
|
| 244 |
+
|
| 245 |
+
if debug:
|
| 246 |
+
print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
|
| 247 |
+
|
| 248 |
+
print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
|
| 249 |
+
|
| 250 |
+
|
| 251 |
+
def _has_callable(obj, fn):
|
| 252 |
+
attr = getattr(obj, fn, None)
|
| 253 |
+
return callable(attr)
|
| 254 |
+
|
| 255 |
+
|
| 256 |
+
def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
|
| 257 |
+
param_shapes = zero_model_states[0].param_shapes
|
| 258 |
+
|
| 259 |
+
# Reconstruction protocol:
|
| 260 |
+
#
|
| 261 |
+
# XXX: document this
|
| 262 |
+
|
| 263 |
+
if debug:
|
| 264 |
+
for i in range(world_size):
|
| 265 |
+
for j in range(len(fp32_flat_groups[0])):
|
| 266 |
+
print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
|
| 267 |
+
|
| 268 |
+
# XXX: memory usage doubles here (zero2)
|
| 269 |
+
num_param_groups = len(fp32_flat_groups[0])
|
| 270 |
+
merged_single_partition_of_fp32_groups = []
|
| 271 |
+
for i in range(num_param_groups):
|
| 272 |
+
merged_partitions = [sd[i] for sd in fp32_flat_groups]
|
| 273 |
+
full_single_fp32_vector = torch.cat(merged_partitions, 0)
|
| 274 |
+
merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
|
| 275 |
+
avail_numel = sum(
|
| 276 |
+
[full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
|
| 277 |
+
|
| 278 |
+
if debug:
|
| 279 |
+
wanted_params = sum([len(shapes) for shapes in param_shapes])
|
| 280 |
+
wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
|
| 281 |
+
# not asserting if there is a mismatch due to possible padding
|
| 282 |
+
print(f"Have {avail_numel} numels to process.")
|
| 283 |
+
print(f"Need {wanted_numel} numels in {wanted_params} params.")
|
| 284 |
+
|
| 285 |
+
# params
|
| 286 |
+
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
| 287 |
+
# out-of-core computing solution
|
| 288 |
+
total_numel = 0
|
| 289 |
+
total_params = 0
|
| 290 |
+
for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
|
| 291 |
+
offset = 0
|
| 292 |
+
avail_numel = full_single_fp32_vector.numel()
|
| 293 |
+
for name, shape in shapes.items():
|
| 294 |
+
|
| 295 |
+
unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
|
| 296 |
+
total_numel += unpartitioned_numel
|
| 297 |
+
total_params += 1
|
| 298 |
+
|
| 299 |
+
if debug:
|
| 300 |
+
print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
|
| 301 |
+
state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
|
| 302 |
+
offset += unpartitioned_numel
|
| 303 |
+
|
| 304 |
+
# Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
|
| 305 |
+
# avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
|
| 306 |
+
# paddings performed in the code it's almost impossible to predict the exact numbers w/o the
|
| 307 |
+
# live optimizer object, so we are checking that the numbers are within the right range
|
| 308 |
+
align_to = 2 * world_size
|
| 309 |
+
|
| 310 |
+
def zero2_align(x):
|
| 311 |
+
return align_to * math.ceil(x / align_to)
|
| 312 |
+
|
| 313 |
+
if debug:
|
| 314 |
+
print(f"original offset={offset}, avail_numel={avail_numel}")
|
| 315 |
+
|
| 316 |
+
offset = zero2_align(offset)
|
| 317 |
+
avail_numel = zero2_align(avail_numel)
|
| 318 |
+
|
| 319 |
+
if debug:
|
| 320 |
+
print(f"aligned offset={offset}, avail_numel={avail_numel}")
|
| 321 |
+
|
| 322 |
+
# Sanity check
|
| 323 |
+
if offset != avail_numel:
|
| 324 |
+
raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
| 325 |
+
|
| 326 |
+
print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
|
| 327 |
+
|
| 328 |
+
|
| 329 |
+
def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):
|
| 330 |
+
state_dict = OrderedDict()
|
| 331 |
+
|
| 332 |
+
# buffers
|
| 333 |
+
buffers = zero_model_states[0].buffers
|
| 334 |
+
state_dict.update(buffers)
|
| 335 |
+
if debug:
|
| 336 |
+
print(f"added {len(buffers)} buffers")
|
| 337 |
+
|
| 338 |
+
_zero2_merge_frozen_params(state_dict, zero_model_states)
|
| 339 |
+
|
| 340 |
+
_zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
|
| 341 |
+
|
| 342 |
+
# recover shared parameters
|
| 343 |
+
for pair in zero_model_states[0].shared_params:
|
| 344 |
+
if pair[1] in state_dict:
|
| 345 |
+
state_dict[pair[0]] = state_dict[pair[1]]
|
| 346 |
+
|
| 347 |
+
return state_dict
|
| 348 |
+
|
| 349 |
+
|
| 350 |
+
def zero3_partitioned_param_info(unpartitioned_numel, world_size):
|
| 351 |
+
remainder = unpartitioned_numel % world_size
|
| 352 |
+
padding_numel = (world_size - remainder) if remainder else 0
|
| 353 |
+
partitioned_numel = math.ceil(unpartitioned_numel / world_size)
|
| 354 |
+
return partitioned_numel, padding_numel
|
| 355 |
+
|
| 356 |
+
|
| 357 |
+
def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
|
| 358 |
+
if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
|
| 359 |
+
return
|
| 360 |
+
|
| 361 |
+
if debug:
|
| 362 |
+
for i in range(world_size):
|
| 363 |
+
num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
|
| 364 |
+
print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
|
| 365 |
+
|
| 366 |
+
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
|
| 367 |
+
wanted_params = len(frozen_param_shapes)
|
| 368 |
+
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
| 369 |
+
avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
|
| 370 |
+
print(f'Frozen params: Have {avail_numel} numels to process.')
|
| 371 |
+
print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
|
| 372 |
+
|
| 373 |
+
total_params = 0
|
| 374 |
+
total_numel = 0
|
| 375 |
+
for name, shape in zero_model_states[0].frozen_param_shapes.items():
|
| 376 |
+
total_params += 1
|
| 377 |
+
unpartitioned_numel = shape.numel()
|
| 378 |
+
total_numel += unpartitioned_numel
|
| 379 |
+
|
| 380 |
+
param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
|
| 381 |
+
state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
|
| 382 |
+
|
| 383 |
+
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
| 384 |
+
|
| 385 |
+
if debug:
|
| 386 |
+
print(
|
| 387 |
+
f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
| 388 |
+
)
|
| 389 |
+
|
| 390 |
+
print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
|
| 391 |
+
|
| 392 |
+
|
| 393 |
+
def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
|
| 394 |
+
param_shapes = zero_model_states[0].param_shapes
|
| 395 |
+
avail_numel = fp32_flat_groups[0].numel() * world_size
|
| 396 |
+
# Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
|
| 397 |
+
# param, re-consolidating each param, while dealing with padding if any
|
| 398 |
+
|
| 399 |
+
# merge list of dicts, preserving order
|
| 400 |
+
param_shapes = {k: v for d in param_shapes for k, v in d.items()}
|
| 401 |
+
|
| 402 |
+
if debug:
|
| 403 |
+
for i in range(world_size):
|
| 404 |
+
print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
|
| 405 |
+
|
| 406 |
+
wanted_params = len(param_shapes)
|
| 407 |
+
wanted_numel = sum(shape.numel() for shape in param_shapes.values())
|
| 408 |
+
# not asserting if there is a mismatch due to possible padding
|
| 409 |
+
avail_numel = fp32_flat_groups[0].numel() * world_size
|
| 410 |
+
print(f"Trainable params: Have {avail_numel} numels to process.")
|
| 411 |
+
print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
|
| 412 |
+
|
| 413 |
+
# params
|
| 414 |
+
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
| 415 |
+
# out-of-core computing solution
|
| 416 |
+
offset = 0
|
| 417 |
+
total_numel = 0
|
| 418 |
+
total_params = 0
|
| 419 |
+
for name, shape in param_shapes.items():
|
| 420 |
+
|
| 421 |
+
unpartitioned_numel = shape.numel()
|
| 422 |
+
total_numel += unpartitioned_numel
|
| 423 |
+
total_params += 1
|
| 424 |
+
|
| 425 |
+
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
| 426 |
+
|
| 427 |
+
if debug:
|
| 428 |
+
print(
|
| 429 |
+
f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
| 430 |
+
)
|
| 431 |
+
|
| 432 |
+
# XXX: memory usage doubles here
|
| 433 |
+
state_dict[name] = torch.cat(
|
| 434 |
+
tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
|
| 435 |
+
0).narrow(0, 0, unpartitioned_numel).view(shape)
|
| 436 |
+
offset += partitioned_numel
|
| 437 |
+
|
| 438 |
+
offset *= world_size
|
| 439 |
+
|
| 440 |
+
# Sanity check
|
| 441 |
+
if offset != avail_numel:
|
| 442 |
+
raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
| 443 |
+
|
| 444 |
+
print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
|
| 445 |
+
|
| 446 |
+
|
| 447 |
+
def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):
|
| 448 |
+
state_dict = OrderedDict()
|
| 449 |
+
|
| 450 |
+
# buffers
|
| 451 |
+
buffers = zero_model_states[0].buffers
|
| 452 |
+
state_dict.update(buffers)
|
| 453 |
+
if debug:
|
| 454 |
+
print(f"added {len(buffers)} buffers")
|
| 455 |
+
|
| 456 |
+
_zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
|
| 457 |
+
|
| 458 |
+
_zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
|
| 459 |
+
|
| 460 |
+
# recover shared parameters
|
| 461 |
+
for pair in zero_model_states[0].shared_params:
|
| 462 |
+
if pair[1] in state_dict:
|
| 463 |
+
state_dict[pair[0]] = state_dict[pair[1]]
|
| 464 |
+
|
| 465 |
+
return state_dict
|
| 466 |
+
|
| 467 |
+
|
| 468 |
+
def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
|
| 469 |
+
"""
|
| 470 |
+
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
|
| 471 |
+
``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
|
| 472 |
+
via a model hub.
|
| 473 |
+
|
| 474 |
+
Args:
|
| 475 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder
|
| 476 |
+
- ``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``
|
| 477 |
+
|
| 478 |
+
Returns:
|
| 479 |
+
- pytorch ``state_dict``
|
| 480 |
+
|
| 481 |
+
Note: this approach may not work if your application doesn't have sufficient free CPU memory and
|
| 482 |
+
you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
|
| 483 |
+
the checkpoint.
|
| 484 |
+
|
| 485 |
+
A typical usage might be ::
|
| 486 |
+
|
| 487 |
+
from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
|
| 488 |
+
# do the training and checkpoint saving
|
| 489 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
|
| 490 |
+
model = model.cpu() # move to cpu
|
| 491 |
+
model.load_state_dict(state_dict)
|
| 492 |
+
# submit to model hub or save the model to share with others
|
| 493 |
+
|
| 494 |
+
In this example the ``model`` will no longer be usable in the deepspeed context of the same
|
| 495 |
+
application. i.e. you will need to re-initialize the deepspeed engine, since
|
| 496 |
+
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
| 497 |
+
|
| 498 |
+
If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
|
| 499 |
+
|
| 500 |
+
"""
|
| 501 |
+
if tag is None:
|
| 502 |
+
latest_path = os.path.join(checkpoint_dir, 'latest')
|
| 503 |
+
if os.path.isfile(latest_path):
|
| 504 |
+
with open(latest_path, 'r') as fd:
|
| 505 |
+
tag = fd.read().strip()
|
| 506 |
+
else:
|
| 507 |
+
raise ValueError(f"Unable to find 'latest' file at {latest_path}")
|
| 508 |
+
|
| 509 |
+
ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
|
| 510 |
+
|
| 511 |
+
if not os.path.isdir(ds_checkpoint_dir):
|
| 512 |
+
raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
|
| 513 |
+
|
| 514 |
+
return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
|
| 515 |
+
|
| 516 |
+
|
| 517 |
+
def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
|
| 518 |
+
"""
|
| 519 |
+
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
|
| 520 |
+
loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
|
| 521 |
+
|
| 522 |
+
Args:
|
| 523 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
| 524 |
+
- ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
|
| 525 |
+
- ``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``
|
| 526 |
+
"""
|
| 527 |
+
|
| 528 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
| 529 |
+
print(f"Saving fp32 state dict to {output_file}")
|
| 530 |
+
torch.save(state_dict, output_file)
|
| 531 |
+
|
| 532 |
+
|
| 533 |
+
def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
|
| 534 |
+
"""
|
| 535 |
+
1. Put the provided model to cpu
|
| 536 |
+
2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
|
| 537 |
+
3. Load it into the provided model
|
| 538 |
+
|
| 539 |
+
Args:
|
| 540 |
+
- ``model``: the model object to update
|
| 541 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
| 542 |
+
- ``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``
|
| 543 |
+
|
| 544 |
+
Returns:
|
| 545 |
+
- ``model`: modified model
|
| 546 |
+
|
| 547 |
+
Make sure you have plenty of CPU memory available before you call this function. If you don't
|
| 548 |
+
have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
|
| 549 |
+
conveniently placed for you in the checkpoint folder.
|
| 550 |
+
|
| 551 |
+
A typical usage might be ::
|
| 552 |
+
|
| 553 |
+
from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
|
| 554 |
+
model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
|
| 555 |
+
# submit to model hub or save the model to share with others
|
| 556 |
+
|
| 557 |
+
Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
|
| 558 |
+
of the same application. i.e. you will need to re-initialize the deepspeed engine, since
|
| 559 |
+
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
| 560 |
+
|
| 561 |
+
"""
|
| 562 |
+
logger.info(f"Extracting fp32 weights")
|
| 563 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
| 564 |
+
|
| 565 |
+
logger.info(f"Overwriting model with fp32 weights")
|
| 566 |
+
model = model.cpu()
|
| 567 |
+
model.load_state_dict(state_dict, strict=False)
|
| 568 |
+
|
| 569 |
+
return model
|
| 570 |
+
|
| 571 |
+
|
| 572 |
+
if __name__ == "__main__":
|
| 573 |
+
|
| 574 |
+
parser = argparse.ArgumentParser()
|
| 575 |
+
parser.add_argument("checkpoint_dir",
|
| 576 |
+
type=str,
|
| 577 |
+
help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
|
| 578 |
+
parser.add_argument(
|
| 579 |
+
"output_file",
|
| 580 |
+
type=str,
|
| 581 |
+
help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
|
| 582 |
+
parser.add_argument("-t",
|
| 583 |
+
"--tag",
|
| 584 |
+
type=str,
|
| 585 |
+
default=None,
|
| 586 |
+
help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
|
| 587 |
+
parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
|
| 588 |
+
args = parser.parse_args()
|
| 589 |
+
|
| 590 |
+
debug = args.debug
|
| 591 |
+
|
| 592 |
+
convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file, tag=args.tag)
|
checkpoint-160/README.md
ADDED
|
@@ -0,0 +1,202 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: peft
|
| 3 |
+
base_model: ../ckpts/Meta-Llama-3-8B-Instruct
|
| 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 0.11.1
|
checkpoint-160/adapter_config.json
ADDED
|
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"alpha_pattern": {},
|
| 3 |
+
"auto_mapping": null,
|
| 4 |
+
"base_model_name_or_path": "../ckpts/Meta-Llama-3-8B-Instruct",
|
| 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": 32,
|
| 14 |
+
"lora_dropout": 0.1,
|
| 15 |
+
"megatron_config": null,
|
| 16 |
+
"megatron_core": "megatron.core",
|
| 17 |
+
"modules_to_save": null,
|
| 18 |
+
"peft_type": "LORA",
|
| 19 |
+
"r": 16,
|
| 20 |
+
"rank_pattern": {},
|
| 21 |
+
"revision": null,
|
| 22 |
+
"target_modules": [
|
| 23 |
+
"down_proj",
|
| 24 |
+
"gate_proj",
|
| 25 |
+
"q_proj",
|
| 26 |
+
"lm_head",
|
| 27 |
+
"o_proj",
|
| 28 |
+
"v_proj",
|
| 29 |
+
"up_proj",
|
| 30 |
+
"k_proj"
|
| 31 |
+
],
|
| 32 |
+
"task_type": "CAUSAL_LM",
|
| 33 |
+
"use_dora": false,
|
| 34 |
+
"use_rslora": false
|
| 35 |
+
}
|
checkpoint-160/adapter_model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:6b6b5a540a8357ce98d65b484037f08c6e0295863490398dfd6671ec06a2916b
|
| 3 |
+
size 1138856856
|
checkpoint-160/global_step160/bf16_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:f1a136bd5515088baf044edf425835b86e73f43dc41097181483574d0a3ec311
|
| 3 |
+
size 528781328
|
checkpoint-160/global_step160/mp_rank_00_model_states.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:e64b47cf0673bd0721e3b682a0e4fd592936923b3a1fc6966a1236c9b9f1a4ec
|
| 3 |
+
size 199905337
|
checkpoint-160/latest
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
global_step160
|
checkpoint-160/rng_state.pth
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:cefbcec9b3679fcdb9b96cdb9ed85fb37b575748c126094c451cc80787e18e73
|
| 3 |
+
size 14244
|
checkpoint-160/trainer_state.json
ADDED
|
@@ -0,0 +1,1153 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"best_metric": null,
|
| 3 |
+
"best_model_checkpoint": null,
|
| 4 |
+
"epoch": 1.3653333333333333,
|
| 5 |
+
"eval_steps": 500,
|
| 6 |
+
"global_step": 160,
|
| 7 |
+
"is_hyper_param_search": false,
|
| 8 |
+
"is_local_process_zero": true,
|
| 9 |
+
"is_world_process_zero": true,
|
| 10 |
+
"log_history": [
|
| 11 |
+
{
|
| 12 |
+
"epoch": 0.008533333333333334,
|
| 13 |
+
"grad_norm": 100.21848203113535,
|
| 14 |
+
"learning_rate": 0.0,
|
| 15 |
+
"loss": 7.1962,
|
| 16 |
+
"step": 1
|
| 17 |
+
},
|
| 18 |
+
{
|
| 19 |
+
"epoch": 0.017066666666666667,
|
| 20 |
+
"grad_norm": 96.43006188910957,
|
| 21 |
+
"learning_rate": 3.0102999566398115e-06,
|
| 22 |
+
"loss": 6.9414,
|
| 23 |
+
"step": 2
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"epoch": 0.0256,
|
| 27 |
+
"grad_norm": 97.35803466618715,
|
| 28 |
+
"learning_rate": 4.771212547196624e-06,
|
| 29 |
+
"loss": 7.0102,
|
| 30 |
+
"step": 3
|
| 31 |
+
},
|
| 32 |
+
{
|
| 33 |
+
"epoch": 0.034133333333333335,
|
| 34 |
+
"grad_norm": 95.14837816372646,
|
| 35 |
+
"learning_rate": 6.020599913279623e-06,
|
| 36 |
+
"loss": 6.5295,
|
| 37 |
+
"step": 4
|
| 38 |
+
},
|
| 39 |
+
{
|
| 40 |
+
"epoch": 0.042666666666666665,
|
| 41 |
+
"grad_norm": 91.76544275692784,
|
| 42 |
+
"learning_rate": 6.989700043360187e-06,
|
| 43 |
+
"loss": 6.4806,
|
| 44 |
+
"step": 5
|
| 45 |
+
},
|
| 46 |
+
{
|
| 47 |
+
"epoch": 0.0512,
|
| 48 |
+
"grad_norm": 84.4494318688335,
|
| 49 |
+
"learning_rate": 7.781512503836437e-06,
|
| 50 |
+
"loss": 6.4194,
|
| 51 |
+
"step": 6
|
| 52 |
+
},
|
| 53 |
+
{
|
| 54 |
+
"epoch": 0.05973333333333333,
|
| 55 |
+
"grad_norm": 71.37977490595638,
|
| 56 |
+
"learning_rate": 8.450980400142568e-06,
|
| 57 |
+
"loss": 5.4953,
|
| 58 |
+
"step": 7
|
| 59 |
+
},
|
| 60 |
+
{
|
| 61 |
+
"epoch": 0.06826666666666667,
|
| 62 |
+
"grad_norm": 49.31153456754566,
|
| 63 |
+
"learning_rate": 9.030899869919434e-06,
|
| 64 |
+
"loss": 5.4123,
|
| 65 |
+
"step": 8
|
| 66 |
+
},
|
| 67 |
+
{
|
| 68 |
+
"epoch": 0.0768,
|
| 69 |
+
"grad_norm": 20.37296364560341,
|
| 70 |
+
"learning_rate": 9.542425094393249e-06,
|
| 71 |
+
"loss": 5.2334,
|
| 72 |
+
"step": 9
|
| 73 |
+
},
|
| 74 |
+
{
|
| 75 |
+
"epoch": 0.08533333333333333,
|
| 76 |
+
"grad_norm": 7.968467079076881,
|
| 77 |
+
"learning_rate": 9.999999999999999e-06,
|
| 78 |
+
"loss": 5.0282,
|
| 79 |
+
"step": 10
|
| 80 |
+
},
|
| 81 |
+
{
|
| 82 |
+
"epoch": 0.09386666666666667,
|
| 83 |
+
"grad_norm": 3.559446532055649,
|
| 84 |
+
"learning_rate": 1.041392685158225e-05,
|
| 85 |
+
"loss": 4.612,
|
| 86 |
+
"step": 11
|
| 87 |
+
},
|
| 88 |
+
{
|
| 89 |
+
"epoch": 0.1024,
|
| 90 |
+
"grad_norm": 3.5528846947995674,
|
| 91 |
+
"learning_rate": 1.0791812460476248e-05,
|
| 92 |
+
"loss": 4.9475,
|
| 93 |
+
"step": 12
|
| 94 |
+
},
|
| 95 |
+
{
|
| 96 |
+
"epoch": 0.11093333333333333,
|
| 97 |
+
"grad_norm": 3.541968897471334,
|
| 98 |
+
"learning_rate": 1.1139433523068365e-05,
|
| 99 |
+
"loss": 4.2777,
|
| 100 |
+
"step": 13
|
| 101 |
+
},
|
| 102 |
+
{
|
| 103 |
+
"epoch": 0.11946666666666667,
|
| 104 |
+
"grad_norm": 3.54718070036198,
|
| 105 |
+
"learning_rate": 1.1461280356782378e-05,
|
| 106 |
+
"loss": 4.3507,
|
| 107 |
+
"step": 14
|
| 108 |
+
},
|
| 109 |
+
{
|
| 110 |
+
"epoch": 0.128,
|
| 111 |
+
"grad_norm": 3.8632334830606747,
|
| 112 |
+
"learning_rate": 1.1760912590556813e-05,
|
| 113 |
+
"loss": 4.5364,
|
| 114 |
+
"step": 15
|
| 115 |
+
},
|
| 116 |
+
{
|
| 117 |
+
"epoch": 0.13653333333333334,
|
| 118 |
+
"grad_norm": 3.6637424744054004,
|
| 119 |
+
"learning_rate": 1.2041199826559246e-05,
|
| 120 |
+
"loss": 3.9672,
|
| 121 |
+
"step": 16
|
| 122 |
+
},
|
| 123 |
+
{
|
| 124 |
+
"epoch": 0.14506666666666668,
|
| 125 |
+
"grad_norm": 3.919802904818311,
|
| 126 |
+
"learning_rate": 1.230448921378274e-05,
|
| 127 |
+
"loss": 4.0618,
|
| 128 |
+
"step": 17
|
| 129 |
+
},
|
| 130 |
+
{
|
| 131 |
+
"epoch": 0.1536,
|
| 132 |
+
"grad_norm": 4.71904950738746,
|
| 133 |
+
"learning_rate": 1.2552725051033058e-05,
|
| 134 |
+
"loss": 4.6656,
|
| 135 |
+
"step": 18
|
| 136 |
+
},
|
| 137 |
+
{
|
| 138 |
+
"epoch": 0.16213333333333332,
|
| 139 |
+
"grad_norm": 4.6656317698690835,
|
| 140 |
+
"learning_rate": 1.2787536009528288e-05,
|
| 141 |
+
"loss": 4.1131,
|
| 142 |
+
"step": 19
|
| 143 |
+
},
|
| 144 |
+
{
|
| 145 |
+
"epoch": 0.17066666666666666,
|
| 146 |
+
"grad_norm": 5.145138692367417,
|
| 147 |
+
"learning_rate": 1.301029995663981e-05,
|
| 148 |
+
"loss": 4.0989,
|
| 149 |
+
"step": 20
|
| 150 |
+
},
|
| 151 |
+
{
|
| 152 |
+
"epoch": 0.1792,
|
| 153 |
+
"grad_norm": 5.923538014759818,
|
| 154 |
+
"learning_rate": 1.3222192947339192e-05,
|
| 155 |
+
"loss": 4.4991,
|
| 156 |
+
"step": 21
|
| 157 |
+
},
|
| 158 |
+
{
|
| 159 |
+
"epoch": 0.18773333333333334,
|
| 160 |
+
"grad_norm": 5.941056962941364,
|
| 161 |
+
"learning_rate": 1.3424226808222062e-05,
|
| 162 |
+
"loss": 4.0836,
|
| 163 |
+
"step": 22
|
| 164 |
+
},
|
| 165 |
+
{
|
| 166 |
+
"epoch": 0.19626666666666667,
|
| 167 |
+
"grad_norm": 6.171026012117947,
|
| 168 |
+
"learning_rate": 1.3617278360175927e-05,
|
| 169 |
+
"loss": 3.6861,
|
| 170 |
+
"step": 23
|
| 171 |
+
},
|
| 172 |
+
{
|
| 173 |
+
"epoch": 0.2048,
|
| 174 |
+
"grad_norm": 7.130542138930838,
|
| 175 |
+
"learning_rate": 1.380211241711606e-05,
|
| 176 |
+
"loss": 4.0958,
|
| 177 |
+
"step": 24
|
| 178 |
+
},
|
| 179 |
+
{
|
| 180 |
+
"epoch": 0.21333333333333335,
|
| 181 |
+
"grad_norm": 7.328837606110418,
|
| 182 |
+
"learning_rate": 1.3979400086720374e-05,
|
| 183 |
+
"loss": 3.9524,
|
| 184 |
+
"step": 25
|
| 185 |
+
},
|
| 186 |
+
{
|
| 187 |
+
"epoch": 0.22186666666666666,
|
| 188 |
+
"grad_norm": 6.923489005711429,
|
| 189 |
+
"learning_rate": 1.4149733479708178e-05,
|
| 190 |
+
"loss": 3.6062,
|
| 191 |
+
"step": 26
|
| 192 |
+
},
|
| 193 |
+
{
|
| 194 |
+
"epoch": 0.2304,
|
| 195 |
+
"grad_norm": 6.409498926059221,
|
| 196 |
+
"learning_rate": 1.4313637641589872e-05,
|
| 197 |
+
"loss": 3.2034,
|
| 198 |
+
"step": 27
|
| 199 |
+
},
|
| 200 |
+
{
|
| 201 |
+
"epoch": 0.23893333333333333,
|
| 202 |
+
"grad_norm": 5.781628405584682,
|
| 203 |
+
"learning_rate": 1.4471580313422191e-05,
|
| 204 |
+
"loss": 2.8158,
|
| 205 |
+
"step": 28
|
| 206 |
+
},
|
| 207 |
+
{
|
| 208 |
+
"epoch": 0.24746666666666667,
|
| 209 |
+
"grad_norm": 6.2927590068243315,
|
| 210 |
+
"learning_rate": 1.4623979978989559e-05,
|
| 211 |
+
"loss": 2.9803,
|
| 212 |
+
"step": 29
|
| 213 |
+
},
|
| 214 |
+
{
|
| 215 |
+
"epoch": 0.256,
|
| 216 |
+
"grad_norm": 6.103844678630006,
|
| 217 |
+
"learning_rate": 1.4771212547196623e-05,
|
| 218 |
+
"loss": 2.847,
|
| 219 |
+
"step": 30
|
| 220 |
+
},
|
| 221 |
+
{
|
| 222 |
+
"epoch": 0.26453333333333334,
|
| 223 |
+
"grad_norm": 7.656341925867144,
|
| 224 |
+
"learning_rate": 1.4913616938342726e-05,
|
| 225 |
+
"loss": 3.0907,
|
| 226 |
+
"step": 31
|
| 227 |
+
},
|
| 228 |
+
{
|
| 229 |
+
"epoch": 0.2730666666666667,
|
| 230 |
+
"grad_norm": 6.324242877501844,
|
| 231 |
+
"learning_rate": 1.5051499783199059e-05,
|
| 232 |
+
"loss": 2.3467,
|
| 233 |
+
"step": 32
|
| 234 |
+
},
|
| 235 |
+
{
|
| 236 |
+
"epoch": 0.2816,
|
| 237 |
+
"grad_norm": 7.606313101162739,
|
| 238 |
+
"learning_rate": 1.5185139398778874e-05,
|
| 239 |
+
"loss": 2.5292,
|
| 240 |
+
"step": 33
|
| 241 |
+
},
|
| 242 |
+
{
|
| 243 |
+
"epoch": 0.29013333333333335,
|
| 244 |
+
"grad_norm": 8.553792493849265,
|
| 245 |
+
"learning_rate": 1.531478917042255e-05,
|
| 246 |
+
"loss": 2.4547,
|
| 247 |
+
"step": 34
|
| 248 |
+
},
|
| 249 |
+
{
|
| 250 |
+
"epoch": 0.2986666666666667,
|
| 251 |
+
"grad_norm": 8.483368703272543,
|
| 252 |
+
"learning_rate": 1.5440680443502753e-05,
|
| 253 |
+
"loss": 2.1956,
|
| 254 |
+
"step": 35
|
| 255 |
+
},
|
| 256 |
+
{
|
| 257 |
+
"epoch": 0.3072,
|
| 258 |
+
"grad_norm": 8.364739352838077,
|
| 259 |
+
"learning_rate": 1.5563025007672873e-05,
|
| 260 |
+
"loss": 1.8552,
|
| 261 |
+
"step": 36
|
| 262 |
+
},
|
| 263 |
+
{
|
| 264 |
+
"epoch": 0.3157333333333333,
|
| 265 |
+
"grad_norm": 9.37663682000104,
|
| 266 |
+
"learning_rate": 1.5682017240669948e-05,
|
| 267 |
+
"loss": 1.9228,
|
| 268 |
+
"step": 37
|
| 269 |
+
},
|
| 270 |
+
{
|
| 271 |
+
"epoch": 0.32426666666666665,
|
| 272 |
+
"grad_norm": 9.823047193440066,
|
| 273 |
+
"learning_rate": 1.57978359661681e-05,
|
| 274 |
+
"loss": 1.7033,
|
| 275 |
+
"step": 38
|
| 276 |
+
},
|
| 277 |
+
{
|
| 278 |
+
"epoch": 0.3328,
|
| 279 |
+
"grad_norm": 9.692618512955894,
|
| 280 |
+
"learning_rate": 1.591064607026499e-05,
|
| 281 |
+
"loss": 1.3768,
|
| 282 |
+
"step": 39
|
| 283 |
+
},
|
| 284 |
+
{
|
| 285 |
+
"epoch": 0.3413333333333333,
|
| 286 |
+
"grad_norm": 9.08889021911031,
|
| 287 |
+
"learning_rate": 1.6020599913279622e-05,
|
| 288 |
+
"loss": 1.3015,
|
| 289 |
+
"step": 40
|
| 290 |
+
},
|
| 291 |
+
{
|
| 292 |
+
"epoch": 0.34986666666666666,
|
| 293 |
+
"grad_norm": 8.081534221516058,
|
| 294 |
+
"learning_rate": 1.6127838567197353e-05,
|
| 295 |
+
"loss": 0.9228,
|
| 296 |
+
"step": 41
|
| 297 |
+
},
|
| 298 |
+
{
|
| 299 |
+
"epoch": 0.3584,
|
| 300 |
+
"grad_norm": 6.238638048950311,
|
| 301 |
+
"learning_rate": 1.6232492903979005e-05,
|
| 302 |
+
"loss": 0.7267,
|
| 303 |
+
"step": 42
|
| 304 |
+
},
|
| 305 |
+
{
|
| 306 |
+
"epoch": 0.36693333333333333,
|
| 307 |
+
"grad_norm": 3.4058036861773604,
|
| 308 |
+
"learning_rate": 1.6334684555795865e-05,
|
| 309 |
+
"loss": 0.5875,
|
| 310 |
+
"step": 43
|
| 311 |
+
},
|
| 312 |
+
{
|
| 313 |
+
"epoch": 0.37546666666666667,
|
| 314 |
+
"grad_norm": 2.079163829467713,
|
| 315 |
+
"learning_rate": 1.6434526764861872e-05,
|
| 316 |
+
"loss": 0.6355,
|
| 317 |
+
"step": 44
|
| 318 |
+
},
|
| 319 |
+
{
|
| 320 |
+
"epoch": 0.384,
|
| 321 |
+
"grad_norm": 1.5597487833024746,
|
| 322 |
+
"learning_rate": 1.6532125137753435e-05,
|
| 323 |
+
"loss": 0.5106,
|
| 324 |
+
"step": 45
|
| 325 |
+
},
|
| 326 |
+
{
|
| 327 |
+
"epoch": 0.39253333333333335,
|
| 328 |
+
"grad_norm": 2.491689602375256,
|
| 329 |
+
"learning_rate": 1.662757831681574e-05,
|
| 330 |
+
"loss": 0.6454,
|
| 331 |
+
"step": 46
|
| 332 |
+
},
|
| 333 |
+
{
|
| 334 |
+
"epoch": 0.4010666666666667,
|
| 335 |
+
"grad_norm": 2.010880438195854,
|
| 336 |
+
"learning_rate": 1.672097857935717e-05,
|
| 337 |
+
"loss": 0.4757,
|
| 338 |
+
"step": 47
|
| 339 |
+
},
|
| 340 |
+
{
|
| 341 |
+
"epoch": 0.4096,
|
| 342 |
+
"grad_norm": 1.9452805114322096,
|
| 343 |
+
"learning_rate": 1.681241237375587e-05,
|
| 344 |
+
"loss": 0.4133,
|
| 345 |
+
"step": 48
|
| 346 |
+
},
|
| 347 |
+
{
|
| 348 |
+
"epoch": 0.41813333333333336,
|
| 349 |
+
"grad_norm": 1.7620848552306103,
|
| 350 |
+
"learning_rate": 1.6901960800285137e-05,
|
| 351 |
+
"loss": 0.4004,
|
| 352 |
+
"step": 49
|
| 353 |
+
},
|
| 354 |
+
{
|
| 355 |
+
"epoch": 0.4266666666666667,
|
| 356 |
+
"grad_norm": 1.278224489774809,
|
| 357 |
+
"learning_rate": 1.6989700043360187e-05,
|
| 358 |
+
"loss": 0.3523,
|
| 359 |
+
"step": 50
|
| 360 |
+
},
|
| 361 |
+
{
|
| 362 |
+
"epoch": 0.4352,
|
| 363 |
+
"grad_norm": 1.6151354758303231,
|
| 364 |
+
"learning_rate": 1.7075701760979363e-05,
|
| 365 |
+
"loss": 0.4317,
|
| 366 |
+
"step": 51
|
| 367 |
+
},
|
| 368 |
+
{
|
| 369 |
+
"epoch": 0.4437333333333333,
|
| 370 |
+
"grad_norm": 1.3451396055695035,
|
| 371 |
+
"learning_rate": 1.716003343634799e-05,
|
| 372 |
+
"loss": 0.3474,
|
| 373 |
+
"step": 52
|
| 374 |
+
},
|
| 375 |
+
{
|
| 376 |
+
"epoch": 0.45226666666666665,
|
| 377 |
+
"grad_norm": 1.6814977782362666,
|
| 378 |
+
"learning_rate": 1.724275869600789e-05,
|
| 379 |
+
"loss": 0.3706,
|
| 380 |
+
"step": 53
|
| 381 |
+
},
|
| 382 |
+
{
|
| 383 |
+
"epoch": 0.4608,
|
| 384 |
+
"grad_norm": 1.1682442432688667,
|
| 385 |
+
"learning_rate": 1.7323937598229687e-05,
|
| 386 |
+
"loss": 0.3488,
|
| 387 |
+
"step": 54
|
| 388 |
+
},
|
| 389 |
+
{
|
| 390 |
+
"epoch": 0.4693333333333333,
|
| 391 |
+
"grad_norm": 0.8839814540462471,
|
| 392 |
+
"learning_rate": 1.7403626894942437e-05,
|
| 393 |
+
"loss": 0.293,
|
| 394 |
+
"step": 55
|
| 395 |
+
},
|
| 396 |
+
{
|
| 397 |
+
"epoch": 0.47786666666666666,
|
| 398 |
+
"grad_norm": 0.7974304806295485,
|
| 399 |
+
"learning_rate": 1.7481880270062003e-05,
|
| 400 |
+
"loss": 0.2717,
|
| 401 |
+
"step": 56
|
| 402 |
+
},
|
| 403 |
+
{
|
| 404 |
+
"epoch": 0.4864,
|
| 405 |
+
"grad_norm": 0.6232203657585239,
|
| 406 |
+
"learning_rate": 1.7558748556724913e-05,
|
| 407 |
+
"loss": 0.1741,
|
| 408 |
+
"step": 57
|
| 409 |
+
},
|
| 410 |
+
{
|
| 411 |
+
"epoch": 0.49493333333333334,
|
| 412 |
+
"grad_norm": 0.6850688604563008,
|
| 413 |
+
"learning_rate": 1.763427993562937e-05,
|
| 414 |
+
"loss": 0.228,
|
| 415 |
+
"step": 58
|
| 416 |
+
},
|
| 417 |
+
{
|
| 418 |
+
"epoch": 0.5034666666666666,
|
| 419 |
+
"grad_norm": 0.5923826384300431,
|
| 420 |
+
"learning_rate": 1.7708520116421443e-05,
|
| 421 |
+
"loss": 0.2131,
|
| 422 |
+
"step": 59
|
| 423 |
+
},
|
| 424 |
+
{
|
| 425 |
+
"epoch": 0.512,
|
| 426 |
+
"grad_norm": 0.489754430485032,
|
| 427 |
+
"learning_rate": 1.7781512503836432e-05,
|
| 428 |
+
"loss": 0.165,
|
| 429 |
+
"step": 60
|
| 430 |
+
},
|
| 431 |
+
{
|
| 432 |
+
"epoch": 0.5205333333333333,
|
| 433 |
+
"grad_norm": 0.5280940052395061,
|
| 434 |
+
"learning_rate": 1.7853298350107667e-05,
|
| 435 |
+
"loss": 0.1658,
|
| 436 |
+
"step": 61
|
| 437 |
+
},
|
| 438 |
+
{
|
| 439 |
+
"epoch": 0.5290666666666667,
|
| 440 |
+
"grad_norm": 0.4750905992036739,
|
| 441 |
+
"learning_rate": 1.7923916894982537e-05,
|
| 442 |
+
"loss": 0.1438,
|
| 443 |
+
"step": 62
|
| 444 |
+
},
|
| 445 |
+
{
|
| 446 |
+
"epoch": 0.5376,
|
| 447 |
+
"grad_norm": 0.4162998392722401,
|
| 448 |
+
"learning_rate": 1.7993405494535815e-05,
|
| 449 |
+
"loss": 0.1555,
|
| 450 |
+
"step": 63
|
| 451 |
+
},
|
| 452 |
+
{
|
| 453 |
+
"epoch": 0.5461333333333334,
|
| 454 |
+
"grad_norm": 0.26863266246370443,
|
| 455 |
+
"learning_rate": 1.806179973983887e-05,
|
| 456 |
+
"loss": 0.1323,
|
| 457 |
+
"step": 64
|
| 458 |
+
},
|
| 459 |
+
{
|
| 460 |
+
"epoch": 0.5546666666666666,
|
| 461 |
+
"grad_norm": 0.26534221125601215,
|
| 462 |
+
"learning_rate": 1.8129133566428553e-05,
|
| 463 |
+
"loss": 0.1671,
|
| 464 |
+
"step": 65
|
| 465 |
+
},
|
| 466 |
+
{
|
| 467 |
+
"epoch": 0.5632,
|
| 468 |
+
"grad_norm": 0.2548662962257576,
|
| 469 |
+
"learning_rate": 1.8195439355418686e-05,
|
| 470 |
+
"loss": 0.1308,
|
| 471 |
+
"step": 66
|
| 472 |
+
},
|
| 473 |
+
{
|
| 474 |
+
"epoch": 0.5717333333333333,
|
| 475 |
+
"grad_norm": 0.18045394638578796,
|
| 476 |
+
"learning_rate": 1.8260748027008263e-05,
|
| 477 |
+
"loss": 0.1262,
|
| 478 |
+
"step": 67
|
| 479 |
+
},
|
| 480 |
+
{
|
| 481 |
+
"epoch": 0.5802666666666667,
|
| 482 |
+
"grad_norm": 0.17070388073612064,
|
| 483 |
+
"learning_rate": 1.8325089127062364e-05,
|
| 484 |
+
"loss": 0.1192,
|
| 485 |
+
"step": 68
|
| 486 |
+
},
|
| 487 |
+
{
|
| 488 |
+
"epoch": 0.5888,
|
| 489 |
+
"grad_norm": 0.1531381679244776,
|
| 490 |
+
"learning_rate": 1.8388490907372553e-05,
|
| 491 |
+
"loss": 0.1274,
|
| 492 |
+
"step": 69
|
| 493 |
+
},
|
| 494 |
+
{
|
| 495 |
+
"epoch": 0.5973333333333334,
|
| 496 |
+
"grad_norm": 0.18196434993687946,
|
| 497 |
+
"learning_rate": 1.8450980400142568e-05,
|
| 498 |
+
"loss": 0.1375,
|
| 499 |
+
"step": 70
|
| 500 |
+
},
|
| 501 |
+
{
|
| 502 |
+
"epoch": 0.6058666666666667,
|
| 503 |
+
"grad_norm": 0.15324416972951205,
|
| 504 |
+
"learning_rate": 1.8512583487190752e-05,
|
| 505 |
+
"loss": 0.1599,
|
| 506 |
+
"step": 71
|
| 507 |
+
},
|
| 508 |
+
{
|
| 509 |
+
"epoch": 0.6144,
|
| 510 |
+
"grad_norm": 0.10884462064503801,
|
| 511 |
+
"learning_rate": 1.857332496431268e-05,
|
| 512 |
+
"loss": 0.1041,
|
| 513 |
+
"step": 72
|
| 514 |
+
},
|
| 515 |
+
{
|
| 516 |
+
"epoch": 0.6229333333333333,
|
| 517 |
+
"grad_norm": 0.12915133528192668,
|
| 518 |
+
"learning_rate": 1.8633228601204555e-05,
|
| 519 |
+
"loss": 0.1406,
|
| 520 |
+
"step": 73
|
| 521 |
+
},
|
| 522 |
+
{
|
| 523 |
+
"epoch": 0.6314666666666666,
|
| 524 |
+
"grad_norm": 0.12553425699952878,
|
| 525 |
+
"learning_rate": 1.8692317197309763e-05,
|
| 526 |
+
"loss": 0.1256,
|
| 527 |
+
"step": 74
|
| 528 |
+
},
|
| 529 |
+
{
|
| 530 |
+
"epoch": 0.64,
|
| 531 |
+
"grad_norm": 0.11976960918968543,
|
| 532 |
+
"learning_rate": 1.8750612633916997e-05,
|
| 533 |
+
"loss": 0.1144,
|
| 534 |
+
"step": 75
|
| 535 |
+
},
|
| 536 |
+
{
|
| 537 |
+
"epoch": 0.6485333333333333,
|
| 538 |
+
"grad_norm": 0.115805998298789,
|
| 539 |
+
"learning_rate": 1.8808135922807914e-05,
|
| 540 |
+
"loss": 0.1528,
|
| 541 |
+
"step": 76
|
| 542 |
+
},
|
| 543 |
+
{
|
| 544 |
+
"epoch": 0.6570666666666667,
|
| 545 |
+
"grad_norm": 0.10325948496697443,
|
| 546 |
+
"learning_rate": 1.8864907251724818e-05,
|
| 547 |
+
"loss": 0.1044,
|
| 548 |
+
"step": 77
|
| 549 |
+
},
|
| 550 |
+
{
|
| 551 |
+
"epoch": 0.6656,
|
| 552 |
+
"grad_norm": 0.09595064346541006,
|
| 553 |
+
"learning_rate": 1.8920946026904802e-05,
|
| 554 |
+
"loss": 0.1534,
|
| 555 |
+
"step": 78
|
| 556 |
+
},
|
| 557 |
+
{
|
| 558 |
+
"epoch": 0.6741333333333334,
|
| 559 |
+
"grad_norm": 0.08796742845240496,
|
| 560 |
+
"learning_rate": 1.8976270912904414e-05,
|
| 561 |
+
"loss": 0.1155,
|
| 562 |
+
"step": 79
|
| 563 |
+
},
|
| 564 |
+
{
|
| 565 |
+
"epoch": 0.6826666666666666,
|
| 566 |
+
"grad_norm": 0.08218991738379527,
|
| 567 |
+
"learning_rate": 1.9030899869919434e-05,
|
| 568 |
+
"loss": 0.1311,
|
| 569 |
+
"step": 80
|
| 570 |
+
},
|
| 571 |
+
{
|
| 572 |
+
"epoch": 0.6912,
|
| 573 |
+
"grad_norm": 0.08290815261109215,
|
| 574 |
+
"learning_rate": 1.9084850188786497e-05,
|
| 575 |
+
"loss": 0.11,
|
| 576 |
+
"step": 81
|
| 577 |
+
},
|
| 578 |
+
{
|
| 579 |
+
"epoch": 0.6997333333333333,
|
| 580 |
+
"grad_norm": 0.0794132180176064,
|
| 581 |
+
"learning_rate": 1.9138138523837165e-05,
|
| 582 |
+
"loss": 0.1135,
|
| 583 |
+
"step": 82
|
| 584 |
+
},
|
| 585 |
+
{
|
| 586 |
+
"epoch": 0.7082666666666667,
|
| 587 |
+
"grad_norm": 0.06934410705255296,
|
| 588 |
+
"learning_rate": 1.919078092376074e-05,
|
| 589 |
+
"loss": 0.109,
|
| 590 |
+
"step": 83
|
| 591 |
+
},
|
| 592 |
+
{
|
| 593 |
+
"epoch": 0.7168,
|
| 594 |
+
"grad_norm": 0.09000563031870593,
|
| 595 |
+
"learning_rate": 1.9242792860618813e-05,
|
| 596 |
+
"loss": 0.12,
|
| 597 |
+
"step": 84
|
| 598 |
+
},
|
| 599 |
+
{
|
| 600 |
+
"epoch": 0.7253333333333334,
|
| 601 |
+
"grad_norm": 0.1134042277879818,
|
| 602 |
+
"learning_rate": 1.929418925714293e-05,
|
| 603 |
+
"loss": 0.1223,
|
| 604 |
+
"step": 85
|
| 605 |
+
},
|
| 606 |
+
{
|
| 607 |
+
"epoch": 0.7338666666666667,
|
| 608 |
+
"grad_norm": 0.09118764690233076,
|
| 609 |
+
"learning_rate": 1.9344984512435673e-05,
|
| 610 |
+
"loss": 0.1459,
|
| 611 |
+
"step": 86
|
| 612 |
+
},
|
| 613 |
+
{
|
| 614 |
+
"epoch": 0.7424,
|
| 615 |
+
"grad_norm": 0.07873016754353963,
|
| 616 |
+
"learning_rate": 1.9395192526186183e-05,
|
| 617 |
+
"loss": 0.1422,
|
| 618 |
+
"step": 87
|
| 619 |
+
},
|
| 620 |
+
{
|
| 621 |
+
"epoch": 0.7509333333333333,
|
| 622 |
+
"grad_norm": 0.1796495874463076,
|
| 623 |
+
"learning_rate": 1.9444826721501687e-05,
|
| 624 |
+
"loss": 0.1291,
|
| 625 |
+
"step": 88
|
| 626 |
+
},
|
| 627 |
+
{
|
| 628 |
+
"epoch": 0.7594666666666666,
|
| 629 |
+
"grad_norm": 0.0679589944174269,
|
| 630 |
+
"learning_rate": 1.9493900066449125e-05,
|
| 631 |
+
"loss": 0.108,
|
| 632 |
+
"step": 89
|
| 633 |
+
},
|
| 634 |
+
{
|
| 635 |
+
"epoch": 0.768,
|
| 636 |
+
"grad_norm": 0.08174688574235538,
|
| 637 |
+
"learning_rate": 1.9542425094393246e-05,
|
| 638 |
+
"loss": 0.1081,
|
| 639 |
+
"step": 90
|
| 640 |
+
},
|
| 641 |
+
{
|
| 642 |
+
"epoch": 0.7765333333333333,
|
| 643 |
+
"grad_norm": 0.057137370501406756,
|
| 644 |
+
"learning_rate": 1.9590413923210934e-05,
|
| 645 |
+
"loss": 0.0934,
|
| 646 |
+
"step": 91
|
| 647 |
+
},
|
| 648 |
+
{
|
| 649 |
+
"epoch": 0.7850666666666667,
|
| 650 |
+
"grad_norm": 0.06578111924908255,
|
| 651 |
+
"learning_rate": 1.9637878273455555e-05,
|
| 652 |
+
"loss": 0.1085,
|
| 653 |
+
"step": 92
|
| 654 |
+
},
|
| 655 |
+
{
|
| 656 |
+
"epoch": 0.7936,
|
| 657 |
+
"grad_norm": 0.08945990540906254,
|
| 658 |
+
"learning_rate": 1.968482948553935e-05,
|
| 659 |
+
"loss": 0.1747,
|
| 660 |
+
"step": 93
|
| 661 |
+
},
|
| 662 |
+
{
|
| 663 |
+
"epoch": 0.8021333333333334,
|
| 664 |
+
"grad_norm": 0.06183863311044229,
|
| 665 |
+
"learning_rate": 1.9731278535996986e-05,
|
| 666 |
+
"loss": 0.1136,
|
| 667 |
+
"step": 94
|
| 668 |
+
},
|
| 669 |
+
{
|
| 670 |
+
"epoch": 0.8106666666666666,
|
| 671 |
+
"grad_norm": 0.05777899602544702,
|
| 672 |
+
"learning_rate": 1.9777236052888476e-05,
|
| 673 |
+
"loss": 0.0984,
|
| 674 |
+
"step": 95
|
| 675 |
+
},
|
| 676 |
+
{
|
| 677 |
+
"epoch": 0.8192,
|
| 678 |
+
"grad_norm": 0.08130851607693534,
|
| 679 |
+
"learning_rate": 1.9822712330395683e-05,
|
| 680 |
+
"loss": 0.187,
|
| 681 |
+
"step": 96
|
| 682 |
+
},
|
| 683 |
+
{
|
| 684 |
+
"epoch": 0.8277333333333333,
|
| 685 |
+
"grad_norm": 0.06426546202002927,
|
| 686 |
+
"learning_rate": 1.986771734266245e-05,
|
| 687 |
+
"loss": 0.1296,
|
| 688 |
+
"step": 97
|
| 689 |
+
},
|
| 690 |
+
{
|
| 691 |
+
"epoch": 0.8362666666666667,
|
| 692 |
+
"grad_norm": 0.069692313707994,
|
| 693 |
+
"learning_rate": 1.991226075692495e-05,
|
| 694 |
+
"loss": 0.1404,
|
| 695 |
+
"step": 98
|
| 696 |
+
},
|
| 697 |
+
{
|
| 698 |
+
"epoch": 0.8448,
|
| 699 |
+
"grad_norm": 0.05494542266886729,
|
| 700 |
+
"learning_rate": 1.9956351945975496e-05,
|
| 701 |
+
"loss": 0.116,
|
| 702 |
+
"step": 99
|
| 703 |
+
},
|
| 704 |
+
{
|
| 705 |
+
"epoch": 0.8533333333333334,
|
| 706 |
+
"grad_norm": 0.07571686966840627,
|
| 707 |
+
"learning_rate": 1.9999999999999998e-05,
|
| 708 |
+
"loss": 0.1539,
|
| 709 |
+
"step": 100
|
| 710 |
+
},
|
| 711 |
+
{
|
| 712 |
+
"epoch": 0.8618666666666667,
|
| 713 |
+
"grad_norm": 0.054351059117603705,
|
| 714 |
+
"learning_rate": 2e-05,
|
| 715 |
+
"loss": 0.1037,
|
| 716 |
+
"step": 101
|
| 717 |
+
},
|
| 718 |
+
{
|
| 719 |
+
"epoch": 0.8704,
|
| 720 |
+
"grad_norm": 0.06531899611551092,
|
| 721 |
+
"learning_rate": 2e-05,
|
| 722 |
+
"loss": 0.0827,
|
| 723 |
+
"step": 102
|
| 724 |
+
},
|
| 725 |
+
{
|
| 726 |
+
"epoch": 0.8789333333333333,
|
| 727 |
+
"grad_norm": 0.06131678646504652,
|
| 728 |
+
"learning_rate": 2e-05,
|
| 729 |
+
"loss": 0.1266,
|
| 730 |
+
"step": 103
|
| 731 |
+
},
|
| 732 |
+
{
|
| 733 |
+
"epoch": 0.8874666666666666,
|
| 734 |
+
"grad_norm": 0.06850220540661824,
|
| 735 |
+
"learning_rate": 2e-05,
|
| 736 |
+
"loss": 0.1456,
|
| 737 |
+
"step": 104
|
| 738 |
+
},
|
| 739 |
+
{
|
| 740 |
+
"epoch": 0.896,
|
| 741 |
+
"grad_norm": 0.05806908951252483,
|
| 742 |
+
"learning_rate": 2e-05,
|
| 743 |
+
"loss": 0.0954,
|
| 744 |
+
"step": 105
|
| 745 |
+
},
|
| 746 |
+
{
|
| 747 |
+
"epoch": 0.9045333333333333,
|
| 748 |
+
"grad_norm": 0.06503642452033717,
|
| 749 |
+
"learning_rate": 2e-05,
|
| 750 |
+
"loss": 0.1417,
|
| 751 |
+
"step": 106
|
| 752 |
+
},
|
| 753 |
+
{
|
| 754 |
+
"epoch": 0.9130666666666667,
|
| 755 |
+
"grad_norm": 0.050486271853277066,
|
| 756 |
+
"learning_rate": 2e-05,
|
| 757 |
+
"loss": 0.0959,
|
| 758 |
+
"step": 107
|
| 759 |
+
},
|
| 760 |
+
{
|
| 761 |
+
"epoch": 0.9216,
|
| 762 |
+
"grad_norm": 0.07746063813802379,
|
| 763 |
+
"learning_rate": 2e-05,
|
| 764 |
+
"loss": 0.1256,
|
| 765 |
+
"step": 108
|
| 766 |
+
},
|
| 767 |
+
{
|
| 768 |
+
"epoch": 0.9301333333333334,
|
| 769 |
+
"grad_norm": 0.051231172380840004,
|
| 770 |
+
"learning_rate": 2e-05,
|
| 771 |
+
"loss": 0.1116,
|
| 772 |
+
"step": 109
|
| 773 |
+
},
|
| 774 |
+
{
|
| 775 |
+
"epoch": 0.9386666666666666,
|
| 776 |
+
"grad_norm": 0.056296443557859455,
|
| 777 |
+
"learning_rate": 2e-05,
|
| 778 |
+
"loss": 0.1056,
|
| 779 |
+
"step": 110
|
| 780 |
+
},
|
| 781 |
+
{
|
| 782 |
+
"epoch": 0.9472,
|
| 783 |
+
"grad_norm": 0.05058663240713958,
|
| 784 |
+
"learning_rate": 2e-05,
|
| 785 |
+
"loss": 0.0971,
|
| 786 |
+
"step": 111
|
| 787 |
+
},
|
| 788 |
+
{
|
| 789 |
+
"epoch": 0.9557333333333333,
|
| 790 |
+
"grad_norm": 0.05532886570130611,
|
| 791 |
+
"learning_rate": 2e-05,
|
| 792 |
+
"loss": 0.1086,
|
| 793 |
+
"step": 112
|
| 794 |
+
},
|
| 795 |
+
{
|
| 796 |
+
"epoch": 0.9642666666666667,
|
| 797 |
+
"grad_norm": 0.05327811326654907,
|
| 798 |
+
"learning_rate": 2e-05,
|
| 799 |
+
"loss": 0.0989,
|
| 800 |
+
"step": 113
|
| 801 |
+
},
|
| 802 |
+
{
|
| 803 |
+
"epoch": 0.9728,
|
| 804 |
+
"grad_norm": 0.05663279364147864,
|
| 805 |
+
"learning_rate": 2e-05,
|
| 806 |
+
"loss": 0.0958,
|
| 807 |
+
"step": 114
|
| 808 |
+
},
|
| 809 |
+
{
|
| 810 |
+
"epoch": 0.9813333333333333,
|
| 811 |
+
"grad_norm": 0.04930904541225805,
|
| 812 |
+
"learning_rate": 2e-05,
|
| 813 |
+
"loss": 0.0887,
|
| 814 |
+
"step": 115
|
| 815 |
+
},
|
| 816 |
+
{
|
| 817 |
+
"epoch": 0.9898666666666667,
|
| 818 |
+
"grad_norm": 0.06096947951115022,
|
| 819 |
+
"learning_rate": 2e-05,
|
| 820 |
+
"loss": 0.106,
|
| 821 |
+
"step": 116
|
| 822 |
+
},
|
| 823 |
+
{
|
| 824 |
+
"epoch": 0.9984,
|
| 825 |
+
"grad_norm": 0.050092322361182495,
|
| 826 |
+
"learning_rate": 2e-05,
|
| 827 |
+
"loss": 0.0931,
|
| 828 |
+
"step": 117
|
| 829 |
+
},
|
| 830 |
+
{
|
| 831 |
+
"epoch": 1.0069333333333332,
|
| 832 |
+
"grad_norm": 0.04980408443758999,
|
| 833 |
+
"learning_rate": 2e-05,
|
| 834 |
+
"loss": 0.0955,
|
| 835 |
+
"step": 118
|
| 836 |
+
},
|
| 837 |
+
{
|
| 838 |
+
"epoch": 1.0154666666666667,
|
| 839 |
+
"grad_norm": 0.051183082721834305,
|
| 840 |
+
"learning_rate": 2e-05,
|
| 841 |
+
"loss": 0.1049,
|
| 842 |
+
"step": 119
|
| 843 |
+
},
|
| 844 |
+
{
|
| 845 |
+
"epoch": 1.024,
|
| 846 |
+
"grad_norm": 0.04332220265802814,
|
| 847 |
+
"learning_rate": 2e-05,
|
| 848 |
+
"loss": 0.0983,
|
| 849 |
+
"step": 120
|
| 850 |
+
},
|
| 851 |
+
{
|
| 852 |
+
"epoch": 1.0325333333333333,
|
| 853 |
+
"grad_norm": 0.07211731499677299,
|
| 854 |
+
"learning_rate": 2e-05,
|
| 855 |
+
"loss": 0.1386,
|
| 856 |
+
"step": 121
|
| 857 |
+
},
|
| 858 |
+
{
|
| 859 |
+
"epoch": 1.0410666666666666,
|
| 860 |
+
"grad_norm": 0.06550870223740553,
|
| 861 |
+
"learning_rate": 2e-05,
|
| 862 |
+
"loss": 0.1334,
|
| 863 |
+
"step": 122
|
| 864 |
+
},
|
| 865 |
+
{
|
| 866 |
+
"epoch": 1.0496,
|
| 867 |
+
"grad_norm": 0.05331839690767287,
|
| 868 |
+
"learning_rate": 2e-05,
|
| 869 |
+
"loss": 0.1014,
|
| 870 |
+
"step": 123
|
| 871 |
+
},
|
| 872 |
+
{
|
| 873 |
+
"epoch": 1.0581333333333334,
|
| 874 |
+
"grad_norm": 0.05227685628767905,
|
| 875 |
+
"learning_rate": 2e-05,
|
| 876 |
+
"loss": 0.1098,
|
| 877 |
+
"step": 124
|
| 878 |
+
},
|
| 879 |
+
{
|
| 880 |
+
"epoch": 1.0666666666666667,
|
| 881 |
+
"grad_norm": 0.07641693882491171,
|
| 882 |
+
"learning_rate": 2e-05,
|
| 883 |
+
"loss": 0.1127,
|
| 884 |
+
"step": 125
|
| 885 |
+
},
|
| 886 |
+
{
|
| 887 |
+
"epoch": 1.0752,
|
| 888 |
+
"grad_norm": 0.052835367770791786,
|
| 889 |
+
"learning_rate": 2e-05,
|
| 890 |
+
"loss": 0.1231,
|
| 891 |
+
"step": 126
|
| 892 |
+
},
|
| 893 |
+
{
|
| 894 |
+
"epoch": 1.0837333333333334,
|
| 895 |
+
"grad_norm": 0.07520329755025788,
|
| 896 |
+
"learning_rate": 2e-05,
|
| 897 |
+
"loss": 0.085,
|
| 898 |
+
"step": 127
|
| 899 |
+
},
|
| 900 |
+
{
|
| 901 |
+
"epoch": 1.0922666666666667,
|
| 902 |
+
"grad_norm": 0.07670066152157425,
|
| 903 |
+
"learning_rate": 2e-05,
|
| 904 |
+
"loss": 0.1071,
|
| 905 |
+
"step": 128
|
| 906 |
+
},
|
| 907 |
+
{
|
| 908 |
+
"epoch": 1.1008,
|
| 909 |
+
"grad_norm": 0.052832906560645154,
|
| 910 |
+
"learning_rate": 2e-05,
|
| 911 |
+
"loss": 0.1093,
|
| 912 |
+
"step": 129
|
| 913 |
+
},
|
| 914 |
+
{
|
| 915 |
+
"epoch": 1.1093333333333333,
|
| 916 |
+
"grad_norm": 0.06573889037311398,
|
| 917 |
+
"learning_rate": 2e-05,
|
| 918 |
+
"loss": 0.1193,
|
| 919 |
+
"step": 130
|
| 920 |
+
},
|
| 921 |
+
{
|
| 922 |
+
"epoch": 1.1178666666666666,
|
| 923 |
+
"grad_norm": 0.05175471296566334,
|
| 924 |
+
"learning_rate": 2e-05,
|
| 925 |
+
"loss": 0.1184,
|
| 926 |
+
"step": 131
|
| 927 |
+
},
|
| 928 |
+
{
|
| 929 |
+
"epoch": 1.1264,
|
| 930 |
+
"grad_norm": 0.05912231419793496,
|
| 931 |
+
"learning_rate": 2e-05,
|
| 932 |
+
"loss": 0.1154,
|
| 933 |
+
"step": 132
|
| 934 |
+
},
|
| 935 |
+
{
|
| 936 |
+
"epoch": 1.1349333333333333,
|
| 937 |
+
"grad_norm": 0.04899140475981105,
|
| 938 |
+
"learning_rate": 2e-05,
|
| 939 |
+
"loss": 0.0957,
|
| 940 |
+
"step": 133
|
| 941 |
+
},
|
| 942 |
+
{
|
| 943 |
+
"epoch": 1.1434666666666666,
|
| 944 |
+
"grad_norm": 0.05939916939142137,
|
| 945 |
+
"learning_rate": 2e-05,
|
| 946 |
+
"loss": 0.0979,
|
| 947 |
+
"step": 134
|
| 948 |
+
},
|
| 949 |
+
{
|
| 950 |
+
"epoch": 1.152,
|
| 951 |
+
"grad_norm": 0.0516819217599706,
|
| 952 |
+
"learning_rate": 2e-05,
|
| 953 |
+
"loss": 0.0834,
|
| 954 |
+
"step": 135
|
| 955 |
+
},
|
| 956 |
+
{
|
| 957 |
+
"epoch": 1.1605333333333334,
|
| 958 |
+
"grad_norm": 0.05456440346454737,
|
| 959 |
+
"learning_rate": 2e-05,
|
| 960 |
+
"loss": 0.1183,
|
| 961 |
+
"step": 136
|
| 962 |
+
},
|
| 963 |
+
{
|
| 964 |
+
"epoch": 1.1690666666666667,
|
| 965 |
+
"grad_norm": 0.059906464476343235,
|
| 966 |
+
"learning_rate": 2e-05,
|
| 967 |
+
"loss": 0.1048,
|
| 968 |
+
"step": 137
|
| 969 |
+
},
|
| 970 |
+
{
|
| 971 |
+
"epoch": 1.1776,
|
| 972 |
+
"grad_norm": 0.0720112680204319,
|
| 973 |
+
"learning_rate": 2e-05,
|
| 974 |
+
"loss": 0.1168,
|
| 975 |
+
"step": 138
|
| 976 |
+
},
|
| 977 |
+
{
|
| 978 |
+
"epoch": 1.1861333333333333,
|
| 979 |
+
"grad_norm": 0.04940202805828527,
|
| 980 |
+
"learning_rate": 2e-05,
|
| 981 |
+
"loss": 0.0948,
|
| 982 |
+
"step": 139
|
| 983 |
+
},
|
| 984 |
+
{
|
| 985 |
+
"epoch": 1.1946666666666665,
|
| 986 |
+
"grad_norm": 0.060088609545130046,
|
| 987 |
+
"learning_rate": 2e-05,
|
| 988 |
+
"loss": 0.0952,
|
| 989 |
+
"step": 140
|
| 990 |
+
},
|
| 991 |
+
{
|
| 992 |
+
"epoch": 1.2032,
|
| 993 |
+
"grad_norm": 0.04694761423612446,
|
| 994 |
+
"learning_rate": 2e-05,
|
| 995 |
+
"loss": 0.0717,
|
| 996 |
+
"step": 141
|
| 997 |
+
},
|
| 998 |
+
{
|
| 999 |
+
"epoch": 1.2117333333333333,
|
| 1000 |
+
"grad_norm": 0.05628581562512457,
|
| 1001 |
+
"learning_rate": 2e-05,
|
| 1002 |
+
"loss": 0.1062,
|
| 1003 |
+
"step": 142
|
| 1004 |
+
},
|
| 1005 |
+
{
|
| 1006 |
+
"epoch": 1.2202666666666666,
|
| 1007 |
+
"grad_norm": 0.06876420990437652,
|
| 1008 |
+
"learning_rate": 2e-05,
|
| 1009 |
+
"loss": 0.1218,
|
| 1010 |
+
"step": 143
|
| 1011 |
+
},
|
| 1012 |
+
{
|
| 1013 |
+
"epoch": 1.2288000000000001,
|
| 1014 |
+
"grad_norm": 0.058774700501610655,
|
| 1015 |
+
"learning_rate": 2e-05,
|
| 1016 |
+
"loss": 0.1125,
|
| 1017 |
+
"step": 144
|
| 1018 |
+
},
|
| 1019 |
+
{
|
| 1020 |
+
"epoch": 1.2373333333333334,
|
| 1021 |
+
"grad_norm": 0.061855922064341186,
|
| 1022 |
+
"learning_rate": 2e-05,
|
| 1023 |
+
"loss": 0.1295,
|
| 1024 |
+
"step": 145
|
| 1025 |
+
},
|
| 1026 |
+
{
|
| 1027 |
+
"epoch": 1.2458666666666667,
|
| 1028 |
+
"grad_norm": 0.0813047704730138,
|
| 1029 |
+
"learning_rate": 2e-05,
|
| 1030 |
+
"loss": 0.1165,
|
| 1031 |
+
"step": 146
|
| 1032 |
+
},
|
| 1033 |
+
{
|
| 1034 |
+
"epoch": 1.2544,
|
| 1035 |
+
"grad_norm": 0.061374000305305446,
|
| 1036 |
+
"learning_rate": 2e-05,
|
| 1037 |
+
"loss": 0.1094,
|
| 1038 |
+
"step": 147
|
| 1039 |
+
},
|
| 1040 |
+
{
|
| 1041 |
+
"epoch": 1.2629333333333332,
|
| 1042 |
+
"grad_norm": 0.055537169110833,
|
| 1043 |
+
"learning_rate": 2e-05,
|
| 1044 |
+
"loss": 0.1054,
|
| 1045 |
+
"step": 148
|
| 1046 |
+
},
|
| 1047 |
+
{
|
| 1048 |
+
"epoch": 1.2714666666666667,
|
| 1049 |
+
"grad_norm": 0.04423248714119304,
|
| 1050 |
+
"learning_rate": 2e-05,
|
| 1051 |
+
"loss": 0.0841,
|
| 1052 |
+
"step": 149
|
| 1053 |
+
},
|
| 1054 |
+
{
|
| 1055 |
+
"epoch": 1.28,
|
| 1056 |
+
"grad_norm": 0.049931966607835034,
|
| 1057 |
+
"learning_rate": 2e-05,
|
| 1058 |
+
"loss": 0.0961,
|
| 1059 |
+
"step": 150
|
| 1060 |
+
},
|
| 1061 |
+
{
|
| 1062 |
+
"epoch": 1.2885333333333333,
|
| 1063 |
+
"grad_norm": 0.06178656953298769,
|
| 1064 |
+
"learning_rate": 2e-05,
|
| 1065 |
+
"loss": 0.0854,
|
| 1066 |
+
"step": 151
|
| 1067 |
+
},
|
| 1068 |
+
{
|
| 1069 |
+
"epoch": 1.2970666666666666,
|
| 1070 |
+
"grad_norm": 0.05783812343287897,
|
| 1071 |
+
"learning_rate": 2e-05,
|
| 1072 |
+
"loss": 0.1141,
|
| 1073 |
+
"step": 152
|
| 1074 |
+
},
|
| 1075 |
+
{
|
| 1076 |
+
"epoch": 1.3056,
|
| 1077 |
+
"grad_norm": 0.048955120400167584,
|
| 1078 |
+
"learning_rate": 2e-05,
|
| 1079 |
+
"loss": 0.0947,
|
| 1080 |
+
"step": 153
|
| 1081 |
+
},
|
| 1082 |
+
{
|
| 1083 |
+
"epoch": 1.3141333333333334,
|
| 1084 |
+
"grad_norm": 0.12270174746806978,
|
| 1085 |
+
"learning_rate": 2e-05,
|
| 1086 |
+
"loss": 0.1553,
|
| 1087 |
+
"step": 154
|
| 1088 |
+
},
|
| 1089 |
+
{
|
| 1090 |
+
"epoch": 1.3226666666666667,
|
| 1091 |
+
"grad_norm": 0.06928026959973474,
|
| 1092 |
+
"learning_rate": 2e-05,
|
| 1093 |
+
"loss": 0.1274,
|
| 1094 |
+
"step": 155
|
| 1095 |
+
},
|
| 1096 |
+
{
|
| 1097 |
+
"epoch": 1.3312,
|
| 1098 |
+
"grad_norm": 0.04756100666105405,
|
| 1099 |
+
"learning_rate": 2e-05,
|
| 1100 |
+
"loss": 0.0893,
|
| 1101 |
+
"step": 156
|
| 1102 |
+
},
|
| 1103 |
+
{
|
| 1104 |
+
"epoch": 1.3397333333333332,
|
| 1105 |
+
"grad_norm": 0.056054951338196934,
|
| 1106 |
+
"learning_rate": 2e-05,
|
| 1107 |
+
"loss": 0.0831,
|
| 1108 |
+
"step": 157
|
| 1109 |
+
},
|
| 1110 |
+
{
|
| 1111 |
+
"epoch": 1.3482666666666667,
|
| 1112 |
+
"grad_norm": 0.0516990471964239,
|
| 1113 |
+
"learning_rate": 2e-05,
|
| 1114 |
+
"loss": 0.0883,
|
| 1115 |
+
"step": 158
|
| 1116 |
+
},
|
| 1117 |
+
{
|
| 1118 |
+
"epoch": 1.3568,
|
| 1119 |
+
"grad_norm": 0.06011650542069954,
|
| 1120 |
+
"learning_rate": 2e-05,
|
| 1121 |
+
"loss": 0.0938,
|
| 1122 |
+
"step": 159
|
| 1123 |
+
},
|
| 1124 |
+
{
|
| 1125 |
+
"epoch": 1.3653333333333333,
|
| 1126 |
+
"grad_norm": 0.051831307951873976,
|
| 1127 |
+
"learning_rate": 2e-05,
|
| 1128 |
+
"loss": 0.1019,
|
| 1129 |
+
"step": 160
|
| 1130 |
+
}
|
| 1131 |
+
],
|
| 1132 |
+
"logging_steps": 1,
|
| 1133 |
+
"max_steps": 351,
|
| 1134 |
+
"num_input_tokens_seen": 0,
|
| 1135 |
+
"num_train_epochs": 3,
|
| 1136 |
+
"save_steps": 20,
|
| 1137 |
+
"stateful_callbacks": {
|
| 1138 |
+
"TrainerControl": {
|
| 1139 |
+
"args": {
|
| 1140 |
+
"should_epoch_stop": false,
|
| 1141 |
+
"should_evaluate": false,
|
| 1142 |
+
"should_log": false,
|
| 1143 |
+
"should_save": true,
|
| 1144 |
+
"should_training_stop": false
|
| 1145 |
+
},
|
| 1146 |
+
"attributes": {}
|
| 1147 |
+
}
|
| 1148 |
+
},
|
| 1149 |
+
"total_flos": 2.273910921652863e+18,
|
| 1150 |
+
"train_batch_size": 16,
|
| 1151 |
+
"trial_name": null,
|
| 1152 |
+
"trial_params": null
|
| 1153 |
+
}
|
checkpoint-160/training_args.bin
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:aa4edad1da5557fcd52a6da980443588016cead6f0444a3562cfa68029c66a04
|
| 3 |
+
size 6840
|
checkpoint-160/zero_to_fp32.py
ADDED
|
@@ -0,0 +1,592 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 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):
|
| 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 |
+
elif zero_stage == 3:
|
| 216 |
+
return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
|
| 217 |
+
|
| 218 |
+
|
| 219 |
+
def _zero2_merge_frozen_params(state_dict, zero_model_states):
|
| 220 |
+
if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
|
| 221 |
+
return
|
| 222 |
+
|
| 223 |
+
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
|
| 224 |
+
frozen_param_fragments = zero_model_states[0].frozen_param_fragments
|
| 225 |
+
|
| 226 |
+
if debug:
|
| 227 |
+
num_elem = sum(s.numel() for s in frozen_param_shapes.values())
|
| 228 |
+
print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
|
| 229 |
+
|
| 230 |
+
wanted_params = len(frozen_param_shapes)
|
| 231 |
+
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
| 232 |
+
avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
|
| 233 |
+
print(f'Frozen params: Have {avail_numel} numels to process.')
|
| 234 |
+
print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
|
| 235 |
+
|
| 236 |
+
total_params = 0
|
| 237 |
+
total_numel = 0
|
| 238 |
+
for name, shape in frozen_param_shapes.items():
|
| 239 |
+
total_params += 1
|
| 240 |
+
unpartitioned_numel = shape.numel()
|
| 241 |
+
total_numel += unpartitioned_numel
|
| 242 |
+
|
| 243 |
+
state_dict[name] = frozen_param_fragments[name]
|
| 244 |
+
|
| 245 |
+
if debug:
|
| 246 |
+
print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
|
| 247 |
+
|
| 248 |
+
print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
|
| 249 |
+
|
| 250 |
+
|
| 251 |
+
def _has_callable(obj, fn):
|
| 252 |
+
attr = getattr(obj, fn, None)
|
| 253 |
+
return callable(attr)
|
| 254 |
+
|
| 255 |
+
|
| 256 |
+
def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
|
| 257 |
+
param_shapes = zero_model_states[0].param_shapes
|
| 258 |
+
|
| 259 |
+
# Reconstruction protocol:
|
| 260 |
+
#
|
| 261 |
+
# XXX: document this
|
| 262 |
+
|
| 263 |
+
if debug:
|
| 264 |
+
for i in range(world_size):
|
| 265 |
+
for j in range(len(fp32_flat_groups[0])):
|
| 266 |
+
print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
|
| 267 |
+
|
| 268 |
+
# XXX: memory usage doubles here (zero2)
|
| 269 |
+
num_param_groups = len(fp32_flat_groups[0])
|
| 270 |
+
merged_single_partition_of_fp32_groups = []
|
| 271 |
+
for i in range(num_param_groups):
|
| 272 |
+
merged_partitions = [sd[i] for sd in fp32_flat_groups]
|
| 273 |
+
full_single_fp32_vector = torch.cat(merged_partitions, 0)
|
| 274 |
+
merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
|
| 275 |
+
avail_numel = sum(
|
| 276 |
+
[full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
|
| 277 |
+
|
| 278 |
+
if debug:
|
| 279 |
+
wanted_params = sum([len(shapes) for shapes in param_shapes])
|
| 280 |
+
wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
|
| 281 |
+
# not asserting if there is a mismatch due to possible padding
|
| 282 |
+
print(f"Have {avail_numel} numels to process.")
|
| 283 |
+
print(f"Need {wanted_numel} numels in {wanted_params} params.")
|
| 284 |
+
|
| 285 |
+
# params
|
| 286 |
+
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
| 287 |
+
# out-of-core computing solution
|
| 288 |
+
total_numel = 0
|
| 289 |
+
total_params = 0
|
| 290 |
+
for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
|
| 291 |
+
offset = 0
|
| 292 |
+
avail_numel = full_single_fp32_vector.numel()
|
| 293 |
+
for name, shape in shapes.items():
|
| 294 |
+
|
| 295 |
+
unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
|
| 296 |
+
total_numel += unpartitioned_numel
|
| 297 |
+
total_params += 1
|
| 298 |
+
|
| 299 |
+
if debug:
|
| 300 |
+
print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
|
| 301 |
+
state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
|
| 302 |
+
offset += unpartitioned_numel
|
| 303 |
+
|
| 304 |
+
# Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
|
| 305 |
+
# avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
|
| 306 |
+
# paddings performed in the code it's almost impossible to predict the exact numbers w/o the
|
| 307 |
+
# live optimizer object, so we are checking that the numbers are within the right range
|
| 308 |
+
align_to = 2 * world_size
|
| 309 |
+
|
| 310 |
+
def zero2_align(x):
|
| 311 |
+
return align_to * math.ceil(x / align_to)
|
| 312 |
+
|
| 313 |
+
if debug:
|
| 314 |
+
print(f"original offset={offset}, avail_numel={avail_numel}")
|
| 315 |
+
|
| 316 |
+
offset = zero2_align(offset)
|
| 317 |
+
avail_numel = zero2_align(avail_numel)
|
| 318 |
+
|
| 319 |
+
if debug:
|
| 320 |
+
print(f"aligned offset={offset}, avail_numel={avail_numel}")
|
| 321 |
+
|
| 322 |
+
# Sanity check
|
| 323 |
+
if offset != avail_numel:
|
| 324 |
+
raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
| 325 |
+
|
| 326 |
+
print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
|
| 327 |
+
|
| 328 |
+
|
| 329 |
+
def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):
|
| 330 |
+
state_dict = OrderedDict()
|
| 331 |
+
|
| 332 |
+
# buffers
|
| 333 |
+
buffers = zero_model_states[0].buffers
|
| 334 |
+
state_dict.update(buffers)
|
| 335 |
+
if debug:
|
| 336 |
+
print(f"added {len(buffers)} buffers")
|
| 337 |
+
|
| 338 |
+
_zero2_merge_frozen_params(state_dict, zero_model_states)
|
| 339 |
+
|
| 340 |
+
_zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
|
| 341 |
+
|
| 342 |
+
# recover shared parameters
|
| 343 |
+
for pair in zero_model_states[0].shared_params:
|
| 344 |
+
if pair[1] in state_dict:
|
| 345 |
+
state_dict[pair[0]] = state_dict[pair[1]]
|
| 346 |
+
|
| 347 |
+
return state_dict
|
| 348 |
+
|
| 349 |
+
|
| 350 |
+
def zero3_partitioned_param_info(unpartitioned_numel, world_size):
|
| 351 |
+
remainder = unpartitioned_numel % world_size
|
| 352 |
+
padding_numel = (world_size - remainder) if remainder else 0
|
| 353 |
+
partitioned_numel = math.ceil(unpartitioned_numel / world_size)
|
| 354 |
+
return partitioned_numel, padding_numel
|
| 355 |
+
|
| 356 |
+
|
| 357 |
+
def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
|
| 358 |
+
if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
|
| 359 |
+
return
|
| 360 |
+
|
| 361 |
+
if debug:
|
| 362 |
+
for i in range(world_size):
|
| 363 |
+
num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
|
| 364 |
+
print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
|
| 365 |
+
|
| 366 |
+
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
|
| 367 |
+
wanted_params = len(frozen_param_shapes)
|
| 368 |
+
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
| 369 |
+
avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
|
| 370 |
+
print(f'Frozen params: Have {avail_numel} numels to process.')
|
| 371 |
+
print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
|
| 372 |
+
|
| 373 |
+
total_params = 0
|
| 374 |
+
total_numel = 0
|
| 375 |
+
for name, shape in zero_model_states[0].frozen_param_shapes.items():
|
| 376 |
+
total_params += 1
|
| 377 |
+
unpartitioned_numel = shape.numel()
|
| 378 |
+
total_numel += unpartitioned_numel
|
| 379 |
+
|
| 380 |
+
param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
|
| 381 |
+
state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
|
| 382 |
+
|
| 383 |
+
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
| 384 |
+
|
| 385 |
+
if debug:
|
| 386 |
+
print(
|
| 387 |
+
f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
| 388 |
+
)
|
| 389 |
+
|
| 390 |
+
print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
|
| 391 |
+
|
| 392 |
+
|
| 393 |
+
def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
|
| 394 |
+
param_shapes = zero_model_states[0].param_shapes
|
| 395 |
+
avail_numel = fp32_flat_groups[0].numel() * world_size
|
| 396 |
+
# Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
|
| 397 |
+
# param, re-consolidating each param, while dealing with padding if any
|
| 398 |
+
|
| 399 |
+
# merge list of dicts, preserving order
|
| 400 |
+
param_shapes = {k: v for d in param_shapes for k, v in d.items()}
|
| 401 |
+
|
| 402 |
+
if debug:
|
| 403 |
+
for i in range(world_size):
|
| 404 |
+
print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
|
| 405 |
+
|
| 406 |
+
wanted_params = len(param_shapes)
|
| 407 |
+
wanted_numel = sum(shape.numel() for shape in param_shapes.values())
|
| 408 |
+
# not asserting if there is a mismatch due to possible padding
|
| 409 |
+
avail_numel = fp32_flat_groups[0].numel() * world_size
|
| 410 |
+
print(f"Trainable params: Have {avail_numel} numels to process.")
|
| 411 |
+
print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
|
| 412 |
+
|
| 413 |
+
# params
|
| 414 |
+
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
| 415 |
+
# out-of-core computing solution
|
| 416 |
+
offset = 0
|
| 417 |
+
total_numel = 0
|
| 418 |
+
total_params = 0
|
| 419 |
+
for name, shape in param_shapes.items():
|
| 420 |
+
|
| 421 |
+
unpartitioned_numel = shape.numel()
|
| 422 |
+
total_numel += unpartitioned_numel
|
| 423 |
+
total_params += 1
|
| 424 |
+
|
| 425 |
+
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
| 426 |
+
|
| 427 |
+
if debug:
|
| 428 |
+
print(
|
| 429 |
+
f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
| 430 |
+
)
|
| 431 |
+
|
| 432 |
+
# XXX: memory usage doubles here
|
| 433 |
+
state_dict[name] = torch.cat(
|
| 434 |
+
tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
|
| 435 |
+
0).narrow(0, 0, unpartitioned_numel).view(shape)
|
| 436 |
+
offset += partitioned_numel
|
| 437 |
+
|
| 438 |
+
offset *= world_size
|
| 439 |
+
|
| 440 |
+
# Sanity check
|
| 441 |
+
if offset != avail_numel:
|
| 442 |
+
raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
| 443 |
+
|
| 444 |
+
print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
|
| 445 |
+
|
| 446 |
+
|
| 447 |
+
def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):
|
| 448 |
+
state_dict = OrderedDict()
|
| 449 |
+
|
| 450 |
+
# buffers
|
| 451 |
+
buffers = zero_model_states[0].buffers
|
| 452 |
+
state_dict.update(buffers)
|
| 453 |
+
if debug:
|
| 454 |
+
print(f"added {len(buffers)} buffers")
|
| 455 |
+
|
| 456 |
+
_zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
|
| 457 |
+
|
| 458 |
+
_zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
|
| 459 |
+
|
| 460 |
+
# recover shared parameters
|
| 461 |
+
for pair in zero_model_states[0].shared_params:
|
| 462 |
+
if pair[1] in state_dict:
|
| 463 |
+
state_dict[pair[0]] = state_dict[pair[1]]
|
| 464 |
+
|
| 465 |
+
return state_dict
|
| 466 |
+
|
| 467 |
+
|
| 468 |
+
def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
|
| 469 |
+
"""
|
| 470 |
+
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
|
| 471 |
+
``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
|
| 472 |
+
via a model hub.
|
| 473 |
+
|
| 474 |
+
Args:
|
| 475 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder
|
| 476 |
+
- ``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``
|
| 477 |
+
|
| 478 |
+
Returns:
|
| 479 |
+
- pytorch ``state_dict``
|
| 480 |
+
|
| 481 |
+
Note: this approach may not work if your application doesn't have sufficient free CPU memory and
|
| 482 |
+
you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
|
| 483 |
+
the checkpoint.
|
| 484 |
+
|
| 485 |
+
A typical usage might be ::
|
| 486 |
+
|
| 487 |
+
from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
|
| 488 |
+
# do the training and checkpoint saving
|
| 489 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
|
| 490 |
+
model = model.cpu() # move to cpu
|
| 491 |
+
model.load_state_dict(state_dict)
|
| 492 |
+
# submit to model hub or save the model to share with others
|
| 493 |
+
|
| 494 |
+
In this example the ``model`` will no longer be usable in the deepspeed context of the same
|
| 495 |
+
application. i.e. you will need to re-initialize the deepspeed engine, since
|
| 496 |
+
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
| 497 |
+
|
| 498 |
+
If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
|
| 499 |
+
|
| 500 |
+
"""
|
| 501 |
+
if tag is None:
|
| 502 |
+
latest_path = os.path.join(checkpoint_dir, 'latest')
|
| 503 |
+
if os.path.isfile(latest_path):
|
| 504 |
+
with open(latest_path, 'r') as fd:
|
| 505 |
+
tag = fd.read().strip()
|
| 506 |
+
else:
|
| 507 |
+
raise ValueError(f"Unable to find 'latest' file at {latest_path}")
|
| 508 |
+
|
| 509 |
+
ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
|
| 510 |
+
|
| 511 |
+
if not os.path.isdir(ds_checkpoint_dir):
|
| 512 |
+
raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
|
| 513 |
+
|
| 514 |
+
return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
|
| 515 |
+
|
| 516 |
+
|
| 517 |
+
def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
|
| 518 |
+
"""
|
| 519 |
+
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
|
| 520 |
+
loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
|
| 521 |
+
|
| 522 |
+
Args:
|
| 523 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
| 524 |
+
- ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
|
| 525 |
+
- ``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``
|
| 526 |
+
"""
|
| 527 |
+
|
| 528 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
| 529 |
+
print(f"Saving fp32 state dict to {output_file}")
|
| 530 |
+
torch.save(state_dict, output_file)
|
| 531 |
+
|
| 532 |
+
|
| 533 |
+
def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
|
| 534 |
+
"""
|
| 535 |
+
1. Put the provided model to cpu
|
| 536 |
+
2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
|
| 537 |
+
3. Load it into the provided model
|
| 538 |
+
|
| 539 |
+
Args:
|
| 540 |
+
- ``model``: the model object to update
|
| 541 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
| 542 |
+
- ``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``
|
| 543 |
+
|
| 544 |
+
Returns:
|
| 545 |
+
- ``model`: modified model
|
| 546 |
+
|
| 547 |
+
Make sure you have plenty of CPU memory available before you call this function. If you don't
|
| 548 |
+
have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
|
| 549 |
+
conveniently placed for you in the checkpoint folder.
|
| 550 |
+
|
| 551 |
+
A typical usage might be ::
|
| 552 |
+
|
| 553 |
+
from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
|
| 554 |
+
model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
|
| 555 |
+
# submit to model hub or save the model to share with others
|
| 556 |
+
|
| 557 |
+
Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
|
| 558 |
+
of the same application. i.e. you will need to re-initialize the deepspeed engine, since
|
| 559 |
+
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
| 560 |
+
|
| 561 |
+
"""
|
| 562 |
+
logger.info(f"Extracting fp32 weights")
|
| 563 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
| 564 |
+
|
| 565 |
+
logger.info(f"Overwriting model with fp32 weights")
|
| 566 |
+
model = model.cpu()
|
| 567 |
+
model.load_state_dict(state_dict, strict=False)
|
| 568 |
+
|
| 569 |
+
return model
|
| 570 |
+
|
| 571 |
+
|
| 572 |
+
if __name__ == "__main__":
|
| 573 |
+
|
| 574 |
+
parser = argparse.ArgumentParser()
|
| 575 |
+
parser.add_argument("checkpoint_dir",
|
| 576 |
+
type=str,
|
| 577 |
+
help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
|
| 578 |
+
parser.add_argument(
|
| 579 |
+
"output_file",
|
| 580 |
+
type=str,
|
| 581 |
+
help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
|
| 582 |
+
parser.add_argument("-t",
|
| 583 |
+
"--tag",
|
| 584 |
+
type=str,
|
| 585 |
+
default=None,
|
| 586 |
+
help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
|
| 587 |
+
parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
|
| 588 |
+
args = parser.parse_args()
|
| 589 |
+
|
| 590 |
+
debug = args.debug
|
| 591 |
+
|
| 592 |
+
convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file, tag=args.tag)
|
checkpoint-20/README.md
ADDED
|
@@ -0,0 +1,202 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: peft
|
| 3 |
+
base_model: ../ckpts/Meta-Llama-3-8B-Instruct
|
| 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 0.11.1
|
checkpoint-20/adapter_config.json
ADDED
|
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"alpha_pattern": {},
|
| 3 |
+
"auto_mapping": null,
|
| 4 |
+
"base_model_name_or_path": "../ckpts/Meta-Llama-3-8B-Instruct",
|
| 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": 32,
|
| 14 |
+
"lora_dropout": 0.1,
|
| 15 |
+
"megatron_config": null,
|
| 16 |
+
"megatron_core": "megatron.core",
|
| 17 |
+
"modules_to_save": null,
|
| 18 |
+
"peft_type": "LORA",
|
| 19 |
+
"r": 16,
|
| 20 |
+
"rank_pattern": {},
|
| 21 |
+
"revision": null,
|
| 22 |
+
"target_modules": [
|
| 23 |
+
"down_proj",
|
| 24 |
+
"gate_proj",
|
| 25 |
+
"q_proj",
|
| 26 |
+
"lm_head",
|
| 27 |
+
"o_proj",
|
| 28 |
+
"v_proj",
|
| 29 |
+
"up_proj",
|
| 30 |
+
"k_proj"
|
| 31 |
+
],
|
| 32 |
+
"task_type": "CAUSAL_LM",
|
| 33 |
+
"use_dora": false,
|
| 34 |
+
"use_rslora": false
|
| 35 |
+
}
|
checkpoint-20/adapter_model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:181727761583eeba365bc29814e3196c50f90d02ae039c8a529d15c6b58813d4
|
| 3 |
+
size 1138856856
|
checkpoint-20/global_step20/bf16_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:fc02bac7488ab5bf8abdc7c2b5344d7b8dfb0afbb5b445938a1d48306aa2f091
|
| 3 |
+
size 528781328
|
checkpoint-20/global_step20/mp_rank_00_model_states.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:92ae839a372c83084c52ed7acb3f7c30fafafaf5b2d3a11194adddf87b919f14
|
| 3 |
+
size 199905337
|
checkpoint-20/latest
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
global_step20
|
checkpoint-20/rng_state.pth
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:926d305e96380bc96ae95066ad252aa5ad89a958f613cc36a658033c5409657d
|
| 3 |
+
size 14244
|
checkpoint-20/trainer_state.json
ADDED
|
@@ -0,0 +1,173 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"best_metric": null,
|
| 3 |
+
"best_model_checkpoint": null,
|
| 4 |
+
"epoch": 0.17066666666666666,
|
| 5 |
+
"eval_steps": 500,
|
| 6 |
+
"global_step": 20,
|
| 7 |
+
"is_hyper_param_search": false,
|
| 8 |
+
"is_local_process_zero": true,
|
| 9 |
+
"is_world_process_zero": true,
|
| 10 |
+
"log_history": [
|
| 11 |
+
{
|
| 12 |
+
"epoch": 0.008533333333333334,
|
| 13 |
+
"grad_norm": 100.21848203113535,
|
| 14 |
+
"learning_rate": 0.0,
|
| 15 |
+
"loss": 7.1962,
|
| 16 |
+
"step": 1
|
| 17 |
+
},
|
| 18 |
+
{
|
| 19 |
+
"epoch": 0.017066666666666667,
|
| 20 |
+
"grad_norm": 96.43006188910957,
|
| 21 |
+
"learning_rate": 3.0102999566398115e-06,
|
| 22 |
+
"loss": 6.9414,
|
| 23 |
+
"step": 2
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"epoch": 0.0256,
|
| 27 |
+
"grad_norm": 97.35803466618715,
|
| 28 |
+
"learning_rate": 4.771212547196624e-06,
|
| 29 |
+
"loss": 7.0102,
|
| 30 |
+
"step": 3
|
| 31 |
+
},
|
| 32 |
+
{
|
| 33 |
+
"epoch": 0.034133333333333335,
|
| 34 |
+
"grad_norm": 95.14837816372646,
|
| 35 |
+
"learning_rate": 6.020599913279623e-06,
|
| 36 |
+
"loss": 6.5295,
|
| 37 |
+
"step": 4
|
| 38 |
+
},
|
| 39 |
+
{
|
| 40 |
+
"epoch": 0.042666666666666665,
|
| 41 |
+
"grad_norm": 91.76544275692784,
|
| 42 |
+
"learning_rate": 6.989700043360187e-06,
|
| 43 |
+
"loss": 6.4806,
|
| 44 |
+
"step": 5
|
| 45 |
+
},
|
| 46 |
+
{
|
| 47 |
+
"epoch": 0.0512,
|
| 48 |
+
"grad_norm": 84.4494318688335,
|
| 49 |
+
"learning_rate": 7.781512503836437e-06,
|
| 50 |
+
"loss": 6.4194,
|
| 51 |
+
"step": 6
|
| 52 |
+
},
|
| 53 |
+
{
|
| 54 |
+
"epoch": 0.05973333333333333,
|
| 55 |
+
"grad_norm": 71.37977490595638,
|
| 56 |
+
"learning_rate": 8.450980400142568e-06,
|
| 57 |
+
"loss": 5.4953,
|
| 58 |
+
"step": 7
|
| 59 |
+
},
|
| 60 |
+
{
|
| 61 |
+
"epoch": 0.06826666666666667,
|
| 62 |
+
"grad_norm": 49.31153456754566,
|
| 63 |
+
"learning_rate": 9.030899869919434e-06,
|
| 64 |
+
"loss": 5.4123,
|
| 65 |
+
"step": 8
|
| 66 |
+
},
|
| 67 |
+
{
|
| 68 |
+
"epoch": 0.0768,
|
| 69 |
+
"grad_norm": 20.37296364560341,
|
| 70 |
+
"learning_rate": 9.542425094393249e-06,
|
| 71 |
+
"loss": 5.2334,
|
| 72 |
+
"step": 9
|
| 73 |
+
},
|
| 74 |
+
{
|
| 75 |
+
"epoch": 0.08533333333333333,
|
| 76 |
+
"grad_norm": 7.968467079076881,
|
| 77 |
+
"learning_rate": 9.999999999999999e-06,
|
| 78 |
+
"loss": 5.0282,
|
| 79 |
+
"step": 10
|
| 80 |
+
},
|
| 81 |
+
{
|
| 82 |
+
"epoch": 0.09386666666666667,
|
| 83 |
+
"grad_norm": 3.559446532055649,
|
| 84 |
+
"learning_rate": 1.041392685158225e-05,
|
| 85 |
+
"loss": 4.612,
|
| 86 |
+
"step": 11
|
| 87 |
+
},
|
| 88 |
+
{
|
| 89 |
+
"epoch": 0.1024,
|
| 90 |
+
"grad_norm": 3.5528846947995674,
|
| 91 |
+
"learning_rate": 1.0791812460476248e-05,
|
| 92 |
+
"loss": 4.9475,
|
| 93 |
+
"step": 12
|
| 94 |
+
},
|
| 95 |
+
{
|
| 96 |
+
"epoch": 0.11093333333333333,
|
| 97 |
+
"grad_norm": 3.541968897471334,
|
| 98 |
+
"learning_rate": 1.1139433523068365e-05,
|
| 99 |
+
"loss": 4.2777,
|
| 100 |
+
"step": 13
|
| 101 |
+
},
|
| 102 |
+
{
|
| 103 |
+
"epoch": 0.11946666666666667,
|
| 104 |
+
"grad_norm": 3.54718070036198,
|
| 105 |
+
"learning_rate": 1.1461280356782378e-05,
|
| 106 |
+
"loss": 4.3507,
|
| 107 |
+
"step": 14
|
| 108 |
+
},
|
| 109 |
+
{
|
| 110 |
+
"epoch": 0.128,
|
| 111 |
+
"grad_norm": 3.8632334830606747,
|
| 112 |
+
"learning_rate": 1.1760912590556813e-05,
|
| 113 |
+
"loss": 4.5364,
|
| 114 |
+
"step": 15
|
| 115 |
+
},
|
| 116 |
+
{
|
| 117 |
+
"epoch": 0.13653333333333334,
|
| 118 |
+
"grad_norm": 3.6637424744054004,
|
| 119 |
+
"learning_rate": 1.2041199826559246e-05,
|
| 120 |
+
"loss": 3.9672,
|
| 121 |
+
"step": 16
|
| 122 |
+
},
|
| 123 |
+
{
|
| 124 |
+
"epoch": 0.14506666666666668,
|
| 125 |
+
"grad_norm": 3.919802904818311,
|
| 126 |
+
"learning_rate": 1.230448921378274e-05,
|
| 127 |
+
"loss": 4.0618,
|
| 128 |
+
"step": 17
|
| 129 |
+
},
|
| 130 |
+
{
|
| 131 |
+
"epoch": 0.1536,
|
| 132 |
+
"grad_norm": 4.71904950738746,
|
| 133 |
+
"learning_rate": 1.2552725051033058e-05,
|
| 134 |
+
"loss": 4.6656,
|
| 135 |
+
"step": 18
|
| 136 |
+
},
|
| 137 |
+
{
|
| 138 |
+
"epoch": 0.16213333333333332,
|
| 139 |
+
"grad_norm": 4.6656317698690835,
|
| 140 |
+
"learning_rate": 1.2787536009528288e-05,
|
| 141 |
+
"loss": 4.1131,
|
| 142 |
+
"step": 19
|
| 143 |
+
},
|
| 144 |
+
{
|
| 145 |
+
"epoch": 0.17066666666666666,
|
| 146 |
+
"grad_norm": 5.145138692367417,
|
| 147 |
+
"learning_rate": 1.301029995663981e-05,
|
| 148 |
+
"loss": 4.0989,
|
| 149 |
+
"step": 20
|
| 150 |
+
}
|
| 151 |
+
],
|
| 152 |
+
"logging_steps": 1,
|
| 153 |
+
"max_steps": 351,
|
| 154 |
+
"num_input_tokens_seen": 0,
|
| 155 |
+
"num_train_epochs": 3,
|
| 156 |
+
"save_steps": 20,
|
| 157 |
+
"stateful_callbacks": {
|
| 158 |
+
"TrainerControl": {
|
| 159 |
+
"args": {
|
| 160 |
+
"should_epoch_stop": false,
|
| 161 |
+
"should_evaluate": false,
|
| 162 |
+
"should_log": false,
|
| 163 |
+
"should_save": true,
|
| 164 |
+
"should_training_stop": false
|
| 165 |
+
},
|
| 166 |
+
"attributes": {}
|
| 167 |
+
}
|
| 168 |
+
},
|
| 169 |
+
"total_flos": 2.924957640079442e+17,
|
| 170 |
+
"train_batch_size": 16,
|
| 171 |
+
"trial_name": null,
|
| 172 |
+
"trial_params": null
|
| 173 |
+
}
|
checkpoint-20/training_args.bin
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:aa4edad1da5557fcd52a6da980443588016cead6f0444a3562cfa68029c66a04
|
| 3 |
+
size 6840
|
checkpoint-20/zero_to_fp32.py
ADDED
|
@@ -0,0 +1,592 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 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):
|
| 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 |
+
elif zero_stage == 3:
|
| 216 |
+
return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
|
| 217 |
+
|
| 218 |
+
|
| 219 |
+
def _zero2_merge_frozen_params(state_dict, zero_model_states):
|
| 220 |
+
if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
|
| 221 |
+
return
|
| 222 |
+
|
| 223 |
+
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
|
| 224 |
+
frozen_param_fragments = zero_model_states[0].frozen_param_fragments
|
| 225 |
+
|
| 226 |
+
if debug:
|
| 227 |
+
num_elem = sum(s.numel() for s in frozen_param_shapes.values())
|
| 228 |
+
print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
|
| 229 |
+
|
| 230 |
+
wanted_params = len(frozen_param_shapes)
|
| 231 |
+
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
| 232 |
+
avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
|
| 233 |
+
print(f'Frozen params: Have {avail_numel} numels to process.')
|
| 234 |
+
print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
|
| 235 |
+
|
| 236 |
+
total_params = 0
|
| 237 |
+
total_numel = 0
|
| 238 |
+
for name, shape in frozen_param_shapes.items():
|
| 239 |
+
total_params += 1
|
| 240 |
+
unpartitioned_numel = shape.numel()
|
| 241 |
+
total_numel += unpartitioned_numel
|
| 242 |
+
|
| 243 |
+
state_dict[name] = frozen_param_fragments[name]
|
| 244 |
+
|
| 245 |
+
if debug:
|
| 246 |
+
print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
|
| 247 |
+
|
| 248 |
+
print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
|
| 249 |
+
|
| 250 |
+
|
| 251 |
+
def _has_callable(obj, fn):
|
| 252 |
+
attr = getattr(obj, fn, None)
|
| 253 |
+
return callable(attr)
|
| 254 |
+
|
| 255 |
+
|
| 256 |
+
def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
|
| 257 |
+
param_shapes = zero_model_states[0].param_shapes
|
| 258 |
+
|
| 259 |
+
# Reconstruction protocol:
|
| 260 |
+
#
|
| 261 |
+
# XXX: document this
|
| 262 |
+
|
| 263 |
+
if debug:
|
| 264 |
+
for i in range(world_size):
|
| 265 |
+
for j in range(len(fp32_flat_groups[0])):
|
| 266 |
+
print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
|
| 267 |
+
|
| 268 |
+
# XXX: memory usage doubles here (zero2)
|
| 269 |
+
num_param_groups = len(fp32_flat_groups[0])
|
| 270 |
+
merged_single_partition_of_fp32_groups = []
|
| 271 |
+
for i in range(num_param_groups):
|
| 272 |
+
merged_partitions = [sd[i] for sd in fp32_flat_groups]
|
| 273 |
+
full_single_fp32_vector = torch.cat(merged_partitions, 0)
|
| 274 |
+
merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
|
| 275 |
+
avail_numel = sum(
|
| 276 |
+
[full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
|
| 277 |
+
|
| 278 |
+
if debug:
|
| 279 |
+
wanted_params = sum([len(shapes) for shapes in param_shapes])
|
| 280 |
+
wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
|
| 281 |
+
# not asserting if there is a mismatch due to possible padding
|
| 282 |
+
print(f"Have {avail_numel} numels to process.")
|
| 283 |
+
print(f"Need {wanted_numel} numels in {wanted_params} params.")
|
| 284 |
+
|
| 285 |
+
# params
|
| 286 |
+
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
| 287 |
+
# out-of-core computing solution
|
| 288 |
+
total_numel = 0
|
| 289 |
+
total_params = 0
|
| 290 |
+
for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
|
| 291 |
+
offset = 0
|
| 292 |
+
avail_numel = full_single_fp32_vector.numel()
|
| 293 |
+
for name, shape in shapes.items():
|
| 294 |
+
|
| 295 |
+
unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
|
| 296 |
+
total_numel += unpartitioned_numel
|
| 297 |
+
total_params += 1
|
| 298 |
+
|
| 299 |
+
if debug:
|
| 300 |
+
print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
|
| 301 |
+
state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
|
| 302 |
+
offset += unpartitioned_numel
|
| 303 |
+
|
| 304 |
+
# Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
|
| 305 |
+
# avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
|
| 306 |
+
# paddings performed in the code it's almost impossible to predict the exact numbers w/o the
|
| 307 |
+
# live optimizer object, so we are checking that the numbers are within the right range
|
| 308 |
+
align_to = 2 * world_size
|
| 309 |
+
|
| 310 |
+
def zero2_align(x):
|
| 311 |
+
return align_to * math.ceil(x / align_to)
|
| 312 |
+
|
| 313 |
+
if debug:
|
| 314 |
+
print(f"original offset={offset}, avail_numel={avail_numel}")
|
| 315 |
+
|
| 316 |
+
offset = zero2_align(offset)
|
| 317 |
+
avail_numel = zero2_align(avail_numel)
|
| 318 |
+
|
| 319 |
+
if debug:
|
| 320 |
+
print(f"aligned offset={offset}, avail_numel={avail_numel}")
|
| 321 |
+
|
| 322 |
+
# Sanity check
|
| 323 |
+
if offset != avail_numel:
|
| 324 |
+
raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
| 325 |
+
|
| 326 |
+
print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
|
| 327 |
+
|
| 328 |
+
|
| 329 |
+
def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):
|
| 330 |
+
state_dict = OrderedDict()
|
| 331 |
+
|
| 332 |
+
# buffers
|
| 333 |
+
buffers = zero_model_states[0].buffers
|
| 334 |
+
state_dict.update(buffers)
|
| 335 |
+
if debug:
|
| 336 |
+
print(f"added {len(buffers)} buffers")
|
| 337 |
+
|
| 338 |
+
_zero2_merge_frozen_params(state_dict, zero_model_states)
|
| 339 |
+
|
| 340 |
+
_zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
|
| 341 |
+
|
| 342 |
+
# recover shared parameters
|
| 343 |
+
for pair in zero_model_states[0].shared_params:
|
| 344 |
+
if pair[1] in state_dict:
|
| 345 |
+
state_dict[pair[0]] = state_dict[pair[1]]
|
| 346 |
+
|
| 347 |
+
return state_dict
|
| 348 |
+
|
| 349 |
+
|
| 350 |
+
def zero3_partitioned_param_info(unpartitioned_numel, world_size):
|
| 351 |
+
remainder = unpartitioned_numel % world_size
|
| 352 |
+
padding_numel = (world_size - remainder) if remainder else 0
|
| 353 |
+
partitioned_numel = math.ceil(unpartitioned_numel / world_size)
|
| 354 |
+
return partitioned_numel, padding_numel
|
| 355 |
+
|
| 356 |
+
|
| 357 |
+
def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
|
| 358 |
+
if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
|
| 359 |
+
return
|
| 360 |
+
|
| 361 |
+
if debug:
|
| 362 |
+
for i in range(world_size):
|
| 363 |
+
num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
|
| 364 |
+
print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
|
| 365 |
+
|
| 366 |
+
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
|
| 367 |
+
wanted_params = len(frozen_param_shapes)
|
| 368 |
+
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
| 369 |
+
avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
|
| 370 |
+
print(f'Frozen params: Have {avail_numel} numels to process.')
|
| 371 |
+
print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
|
| 372 |
+
|
| 373 |
+
total_params = 0
|
| 374 |
+
total_numel = 0
|
| 375 |
+
for name, shape in zero_model_states[0].frozen_param_shapes.items():
|
| 376 |
+
total_params += 1
|
| 377 |
+
unpartitioned_numel = shape.numel()
|
| 378 |
+
total_numel += unpartitioned_numel
|
| 379 |
+
|
| 380 |
+
param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
|
| 381 |
+
state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
|
| 382 |
+
|
| 383 |
+
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
| 384 |
+
|
| 385 |
+
if debug:
|
| 386 |
+
print(
|
| 387 |
+
f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
| 388 |
+
)
|
| 389 |
+
|
| 390 |
+
print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
|
| 391 |
+
|
| 392 |
+
|
| 393 |
+
def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
|
| 394 |
+
param_shapes = zero_model_states[0].param_shapes
|
| 395 |
+
avail_numel = fp32_flat_groups[0].numel() * world_size
|
| 396 |
+
# Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
|
| 397 |
+
# param, re-consolidating each param, while dealing with padding if any
|
| 398 |
+
|
| 399 |
+
# merge list of dicts, preserving order
|
| 400 |
+
param_shapes = {k: v for d in param_shapes for k, v in d.items()}
|
| 401 |
+
|
| 402 |
+
if debug:
|
| 403 |
+
for i in range(world_size):
|
| 404 |
+
print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
|
| 405 |
+
|
| 406 |
+
wanted_params = len(param_shapes)
|
| 407 |
+
wanted_numel = sum(shape.numel() for shape in param_shapes.values())
|
| 408 |
+
# not asserting if there is a mismatch due to possible padding
|
| 409 |
+
avail_numel = fp32_flat_groups[0].numel() * world_size
|
| 410 |
+
print(f"Trainable params: Have {avail_numel} numels to process.")
|
| 411 |
+
print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
|
| 412 |
+
|
| 413 |
+
# params
|
| 414 |
+
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
| 415 |
+
# out-of-core computing solution
|
| 416 |
+
offset = 0
|
| 417 |
+
total_numel = 0
|
| 418 |
+
total_params = 0
|
| 419 |
+
for name, shape in param_shapes.items():
|
| 420 |
+
|
| 421 |
+
unpartitioned_numel = shape.numel()
|
| 422 |
+
total_numel += unpartitioned_numel
|
| 423 |
+
total_params += 1
|
| 424 |
+
|
| 425 |
+
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
| 426 |
+
|
| 427 |
+
if debug:
|
| 428 |
+
print(
|
| 429 |
+
f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
| 430 |
+
)
|
| 431 |
+
|
| 432 |
+
# XXX: memory usage doubles here
|
| 433 |
+
state_dict[name] = torch.cat(
|
| 434 |
+
tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
|
| 435 |
+
0).narrow(0, 0, unpartitioned_numel).view(shape)
|
| 436 |
+
offset += partitioned_numel
|
| 437 |
+
|
| 438 |
+
offset *= world_size
|
| 439 |
+
|
| 440 |
+
# Sanity check
|
| 441 |
+
if offset != avail_numel:
|
| 442 |
+
raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
| 443 |
+
|
| 444 |
+
print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
|
| 445 |
+
|
| 446 |
+
|
| 447 |
+
def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):
|
| 448 |
+
state_dict = OrderedDict()
|
| 449 |
+
|
| 450 |
+
# buffers
|
| 451 |
+
buffers = zero_model_states[0].buffers
|
| 452 |
+
state_dict.update(buffers)
|
| 453 |
+
if debug:
|
| 454 |
+
print(f"added {len(buffers)} buffers")
|
| 455 |
+
|
| 456 |
+
_zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
|
| 457 |
+
|
| 458 |
+
_zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
|
| 459 |
+
|
| 460 |
+
# recover shared parameters
|
| 461 |
+
for pair in zero_model_states[0].shared_params:
|
| 462 |
+
if pair[1] in state_dict:
|
| 463 |
+
state_dict[pair[0]] = state_dict[pair[1]]
|
| 464 |
+
|
| 465 |
+
return state_dict
|
| 466 |
+
|
| 467 |
+
|
| 468 |
+
def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
|
| 469 |
+
"""
|
| 470 |
+
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
|
| 471 |
+
``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
|
| 472 |
+
via a model hub.
|
| 473 |
+
|
| 474 |
+
Args:
|
| 475 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder
|
| 476 |
+
- ``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``
|
| 477 |
+
|
| 478 |
+
Returns:
|
| 479 |
+
- pytorch ``state_dict``
|
| 480 |
+
|
| 481 |
+
Note: this approach may not work if your application doesn't have sufficient free CPU memory and
|
| 482 |
+
you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
|
| 483 |
+
the checkpoint.
|
| 484 |
+
|
| 485 |
+
A typical usage might be ::
|
| 486 |
+
|
| 487 |
+
from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
|
| 488 |
+
# do the training and checkpoint saving
|
| 489 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
|
| 490 |
+
model = model.cpu() # move to cpu
|
| 491 |
+
model.load_state_dict(state_dict)
|
| 492 |
+
# submit to model hub or save the model to share with others
|
| 493 |
+
|
| 494 |
+
In this example the ``model`` will no longer be usable in the deepspeed context of the same
|
| 495 |
+
application. i.e. you will need to re-initialize the deepspeed engine, since
|
| 496 |
+
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
| 497 |
+
|
| 498 |
+
If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
|
| 499 |
+
|
| 500 |
+
"""
|
| 501 |
+
if tag is None:
|
| 502 |
+
latest_path = os.path.join(checkpoint_dir, 'latest')
|
| 503 |
+
if os.path.isfile(latest_path):
|
| 504 |
+
with open(latest_path, 'r') as fd:
|
| 505 |
+
tag = fd.read().strip()
|
| 506 |
+
else:
|
| 507 |
+
raise ValueError(f"Unable to find 'latest' file at {latest_path}")
|
| 508 |
+
|
| 509 |
+
ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
|
| 510 |
+
|
| 511 |
+
if not os.path.isdir(ds_checkpoint_dir):
|
| 512 |
+
raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
|
| 513 |
+
|
| 514 |
+
return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
|
| 515 |
+
|
| 516 |
+
|
| 517 |
+
def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
|
| 518 |
+
"""
|
| 519 |
+
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
|
| 520 |
+
loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
|
| 521 |
+
|
| 522 |
+
Args:
|
| 523 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
| 524 |
+
- ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
|
| 525 |
+
- ``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``
|
| 526 |
+
"""
|
| 527 |
+
|
| 528 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
| 529 |
+
print(f"Saving fp32 state dict to {output_file}")
|
| 530 |
+
torch.save(state_dict, output_file)
|
| 531 |
+
|
| 532 |
+
|
| 533 |
+
def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
|
| 534 |
+
"""
|
| 535 |
+
1. Put the provided model to cpu
|
| 536 |
+
2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
|
| 537 |
+
3. Load it into the provided model
|
| 538 |
+
|
| 539 |
+
Args:
|
| 540 |
+
- ``model``: the model object to update
|
| 541 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
| 542 |
+
- ``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``
|
| 543 |
+
|
| 544 |
+
Returns:
|
| 545 |
+
- ``model`: modified model
|
| 546 |
+
|
| 547 |
+
Make sure you have plenty of CPU memory available before you call this function. If you don't
|
| 548 |
+
have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
|
| 549 |
+
conveniently placed for you in the checkpoint folder.
|
| 550 |
+
|
| 551 |
+
A typical usage might be ::
|
| 552 |
+
|
| 553 |
+
from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
|
| 554 |
+
model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
|
| 555 |
+
# submit to model hub or save the model to share with others
|
| 556 |
+
|
| 557 |
+
Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
|
| 558 |
+
of the same application. i.e. you will need to re-initialize the deepspeed engine, since
|
| 559 |
+
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
| 560 |
+
|
| 561 |
+
"""
|
| 562 |
+
logger.info(f"Extracting fp32 weights")
|
| 563 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
| 564 |
+
|
| 565 |
+
logger.info(f"Overwriting model with fp32 weights")
|
| 566 |
+
model = model.cpu()
|
| 567 |
+
model.load_state_dict(state_dict, strict=False)
|
| 568 |
+
|
| 569 |
+
return model
|
| 570 |
+
|
| 571 |
+
|
| 572 |
+
if __name__ == "__main__":
|
| 573 |
+
|
| 574 |
+
parser = argparse.ArgumentParser()
|
| 575 |
+
parser.add_argument("checkpoint_dir",
|
| 576 |
+
type=str,
|
| 577 |
+
help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
|
| 578 |
+
parser.add_argument(
|
| 579 |
+
"output_file",
|
| 580 |
+
type=str,
|
| 581 |
+
help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
|
| 582 |
+
parser.add_argument("-t",
|
| 583 |
+
"--tag",
|
| 584 |
+
type=str,
|
| 585 |
+
default=None,
|
| 586 |
+
help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
|
| 587 |
+
parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
|
| 588 |
+
args = parser.parse_args()
|
| 589 |
+
|
| 590 |
+
debug = args.debug
|
| 591 |
+
|
| 592 |
+
convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file, tag=args.tag)
|