Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+merged/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,155 @@
+---
+license: other
+library_name: peft
+tags:
+- generated_from_trainer
+base_model: google/gemma-7b-it
+model-index:
+- name: out
+  results: []
+---
+
+<!-- This model card has been generated automatically according to the information the Trainer had access to. You
+should probably proofread and complete it, then remove this comment. -->
+
+[<img src="https://raw.githubusercontent.com/OpenAccess-AI-Collective/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/OpenAccess-AI-Collective/axolotl)
+<details><summary>See axolotl config</summary>
+
+axolotl version: `0.4.0`
+```yaml
+# use google/gemma-7b if you have access
+base_model: google/gemma-7b-it
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+
+load_in_8bit: false
+load_in_4bit: true
+strict: false
+
+# huggingface repo
+datasets:
+  - path: ./python-oasst/chunk_1.jsonl
+    type: oasst
+val_set_size: 0.1
+output_dir: ./out
+
+adapter: qlora
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+sequence_len: 4096
+sample_packing: false
+pad_to_sequence_len: true
+
+wandb_project: gemma-7b-it
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+
+gradient_accumulation_steps: 6
+micro_batch_size: 4
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: true
+group_by_length: false
+bf16: auto
+fp16:
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 4
+eval_table_size:
+eval_max_new_tokens: 128
+saves_per_epoch: 1
+debug:
+deepspeed: deepspeed_configs/zero1.json
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+
+```
+
+</details><br>
+
+# out
+
+This model is a fine-tuned version of [google/gemma-7b-it](https://huggingface.co/google/gemma-7b-it) on the None dataset.
+It achieves the following results on the evaluation set:
+- Loss: 1.1911
+
+## Model description
+
+More information needed
+
+## Intended uses & limitations
+
+More information needed
+
+## Training and evaluation data
+
+More information needed
+
+## Training procedure
+
+### Training hyperparameters
+
+The following hyperparameters were used during training:
+- learning_rate: 0.0002
+- train_batch_size: 4
+- eval_batch_size: 4
+- seed: 42
+- distributed_type: multi-GPU
+- num_devices: 4
+- gradient_accumulation_steps: 6
+- total_train_batch_size: 96
+- total_eval_batch_size: 16
+- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
+- lr_scheduler_type: cosine
+- lr_scheduler_warmup_steps: 9
+- num_epochs: 4
+
+### Training results
+
+| Training Loss | Epoch | Step | Validation Loss |
+|:-------------:|:-----:|:----:|:---------------:|
+| 5.0474        | 0.01  | 1    | 5.9279          |
+| 1.2191        | 0.26  | 24   | 1.2947          |
+| 1.1165        | 0.51  | 48   | 1.1679          |
+| 1.0711        | 0.77  | 72   | 1.1377          |
+| 0.9546        | 1.02  | 96   | 1.1303          |
+| 0.9309        | 1.28  | 120  | 1.1298          |
+| 0.9588        | 1.54  | 144  | 1.1242          |
+| 0.8553        | 1.79  | 168  | 1.1259          |
+| 0.8231        | 2.05  | 192  | 1.1449          |
+| 0.8154        | 2.31  | 216  | 1.1514          |
+| 0.7354        | 2.56  | 240  | 1.1471          |
+| 0.7577        | 2.82  | 264  | 1.1479          |
+| 0.6647        | 3.07  | 288  | 1.1923          |
+| 0.6928        | 3.33  | 312  | 1.1856          |
+| 0.731         | 3.59  | 336  | 1.1890          |
+| 0.7193        | 3.84  | 360  | 1.1911          |
+
+
+### Framework versions
+
+- PEFT 0.9.0
+- Transformers 4.39.0.dev0
+- Pytorch 2.1.2+cu118
+- Datasets 2.18.0
+- Tokenizers 0.15.0

README1.md

@@ -0,0 +1,153 @@
+---
+license: other
+library_name: peft
+tags:
+- generated_from_trainer
+base_model: google/gemma-7b-it
+model-index:
+- name: out
+  results: []
+---
+
+
+[<img src="https://raw.githubusercontent.com/OpenAccess-AI-Collective/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/OpenAccess-AI-Collective/axolotl)
+<details><summary>See axolotl config</summary>
+
+axolotl version: `0.4.0`
+```yaml
+# use google/gemma-7b if you have access
+base_model: google/gemma-7b-it
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+
+load_in_8bit: false
+load_in_4bit: true
+strict: false
+
+# huggingface repo
+datasets:
+  - path: ./python-oasst/chunk_1.jsonl
+    type: oasst
+val_set_size: 0.1
+output_dir: ./out
+
+adapter: qlora
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+sequence_len: 4096
+sample_packing: false
+pad_to_sequence_len: true
+
+wandb_project: gemma-7b-it
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+
+gradient_accumulation_steps: 6
+micro_batch_size: 4
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: true
+group_by_length: false
+bf16: auto
+fp16:
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 4
+eval_table_size:
+eval_max_new_tokens: 128
+saves_per_epoch: 1
+debug:
+deepspeed: deepspeed_configs/zero1.json
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+
+```
+
+</details><br>
+
+# out
+
+This model is a fine-tuned version of [google/gemma-7b-it](https://huggingface.co/google/gemma-7b-it) on the None dataset.
+It achieves the following results on the evaluation set:
+- Loss: 1.1911
+
+## Model description
+
+More information needed
+
+## Intended uses & limitations
+
+More information needed
+
+## Training and evaluation data
+
+More information needed
+
+## Training procedure
+
+### Training hyperparameters
+
+The following hyperparameters were used during training:
+- learning_rate: 0.0002
+- train_batch_size: 4
+- eval_batch_size: 4
+- seed: 42
+- distributed_type: multi-GPU
+- num_devices: 4
+- gradient_accumulation_steps: 6
+- total_train_batch_size: 96
+- total_eval_batch_size: 16
+- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
+- lr_scheduler_type: cosine
+- lr_scheduler_warmup_steps: 9
+- num_epochs: 4
+
+### Training results
+
+| Training Loss | Epoch | Step | Validation Loss |
+|:-------------:|:-----:|:----:|:---------------:|
+| 5.0474        | 0.01  | 1    | 5.9279          |
+| 1.2191        | 0.26  | 24   | 1.2947          |
+| 1.1165        | 0.51  | 48   | 1.1679          |
+| 1.0711        | 0.77  | 72   | 1.1377          |
+| 0.9546        | 1.02  | 96   | 1.1303          |
+| 0.9309        | 1.28  | 120  | 1.1298          |
+| 0.9588        | 1.54  | 144  | 1.1242          |
+| 0.8553        | 1.79  | 168  | 1.1259          |
+| 0.8231        | 2.05  | 192  | 1.1449          |
+| 0.8154        | 2.31  | 216  | 1.1514          |
+| 0.7354        | 2.56  | 240  | 1.1471          |
+| 0.7577        | 2.82  | 264  | 1.1479          |
+| 0.6647        | 3.07  | 288  | 1.1923          |
+| 0.6928        | 3.33  | 312  | 1.1856          |
+| 0.731         | 3.59  | 336  | 1.1890          |
+| 0.7193        | 3.84  | 360  | 1.1911          |
+
+
+### Framework versions
+
+- PEFT 0.9.0
+- Transformers 4.39.0.dev0
+- Pytorch 2.1.2+cu118
+- Datasets 2.18.0
+- Tokenizers 0.15.0

adapter_config.json

@@ -0,0 +1,33 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "google/gemma-7b-it",
+  "bias": "none",
+  "fan_in_fan_out": null,
+  "inference_mode": false,
+  "init_lora_weights": true,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 16,
+  "lora_dropout": 0.05,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "gate_proj",
+    "up_proj",
+    "k_proj",
+    "q_proj",
+    "down_proj",
+    "v_proj",
+    "o_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "use_dora": false,
+  "use_rslora": false
+}

adapter_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:25b9fe36c6655d98214b7521f5d8d9f662fc0c4007a06eeefea535ccfec3dc1e
+size 200078074

checkpoint-279/README.md

@@ -0,0 +1,202 @@
+---
+library_name: peft
+base_model: google/gemma-7b-it
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- 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. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+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).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.9.0

checkpoint-279/adapter_config.json

@@ -0,0 +1,33 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "google/gemma-7b-it",
+  "bias": "none",
+  "fan_in_fan_out": null,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 16,
+  "lora_dropout": 0.05,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "down_proj",
+    "o_proj",
+    "k_proj",
+    "q_proj",
+    "gate_proj",
+    "up_proj",
+    "v_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "use_dora": false,
+  "use_rslora": false
+}

checkpoint-279/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0831f70d185dae9ca69f58be3eab596067ac52e75e3e97b46d23ecd486b83942
+size 200068904

checkpoint-279/global_step279/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:90f2db91b1ca035dfa781beb0567b7cfaaf6646de04cc9a82d8e80069e7a5b09
+size 150126608

checkpoint-279/global_step279/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b9ec41ba7f5c3131e00c854ec2bbfca98e6a3321e5f2ddf6efdc6056fa008c5a
+size 150126672

checkpoint-279/global_step279/bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:41f227ca1d3c19b4cd53567e28a2d395c2e804bd38dfd9bb3c937adab1daf5a3
+size 150126736

checkpoint-279/global_step279/bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7e486ddf3459c4f6befb004a9374e7e4fb9bd64bba72dd2e6f7051ee89939988
+size 150126736

checkpoint-279/global_step279/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:feed7b7a8694c54651374fb581d67d60790a016e23023446231557add62ffc80
+size 1896781286

checkpoint-279/latest

@@ -0,0 +1 @@
+global_step279

checkpoint-279/rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a84c3f9fa55e23a5c4d93b108c705b57ba9a5ed816191e6dfbb6e72ad2857e6d
+size 15024

checkpoint-279/rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fbb1da31ff41578c72556d0a8b9b94abf6be26bf16b6456ecd87d2b611f5b9bd
+size 15024

checkpoint-279/rng_state_2.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:11a7b38529914886a43976df69af7f331315329e1d38788c57003ca4cd1a849f
+size 15024

checkpoint-279/rng_state_3.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7d65b4248464f467db8226c5cc4ba4aa32e06af0bf915b61ea8a2db71d16b5ce
+size 15024

checkpoint-279/scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:099f524a0aa9353b01bf7d70e5a899c6e8ee8efc46e982213631888df6e5111b
+size 1064

checkpoint-279/trainer_state.json

@@ -0,0 +1,2070 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 2.97864768683274,
+  "eval_steps": 24,
+  "global_step": 279,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.01,
+      "grad_norm": 1.8206765789002874,
+      "learning_rate": 2.2222222222222223e-05,
+      "loss": 5.0474,
+      "step": 1
+    },
+    {
+      "epoch": 0.01,
+      "eval_loss": 5.927858829498291,
+      "eval_runtime": 117.3665,
+      "eval_samples_per_second": 8.512,
+      "eval_steps_per_second": 0.537,
+      "step": 1
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 1.9889295079554647,
+      "learning_rate": 4.4444444444444447e-05,
+      "loss": 5.5569,
+      "step": 2
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 1.8931443004310682,
+      "learning_rate": 6.666666666666667e-05,
+      "loss": 5.2383,
+      "step": 3
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 2.195266234429632,
+      "learning_rate": 8.888888888888889e-05,
+      "loss": 5.4943,
+      "step": 4
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 2.6001064132041503,
+      "learning_rate": 0.00011111111111111112,
+      "loss": 5.2602,
+      "step": 5
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 3.26301463076567,
+      "learning_rate": 0.00013333333333333334,
+      "loss": 4.8182,
+      "step": 6
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 3.476044691292363,
+      "learning_rate": 0.00015555555555555556,
+      "loss": 4.0432,
+      "step": 7
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 3.378803229553045,
+      "learning_rate": 0.00017777777777777779,
+      "loss": 3.5212,
+      "step": 8
+    },
+    {
+      "epoch": 0.1,
+      "grad_norm": 3.9419449437137017,
+      "learning_rate": 0.0002,
+      "loss": 3.2239,
+      "step": 9
+    },
+    {
+      "epoch": 0.11,
+      "grad_norm": 5.8833082175146485,
+      "learning_rate": 0.00019999625498303932,
+      "loss": 3.4319,
+      "step": 10
+    },
+    {
+      "epoch": 0.12,
+      "grad_norm": 5.4690223843996515,
+      "learning_rate": 0.0001999850202126604,
+      "loss": 2.8167,
+      "step": 11
+    },
+    {
+      "epoch": 0.13,
+      "grad_norm": 7.009614336449043,
+      "learning_rate": 0.00019996629653035126,
+      "loss": 2.7966,
+      "step": 12
+    },
+    {
+      "epoch": 0.14,
+      "grad_norm": 6.254841874500106,
+      "learning_rate": 0.0001999400853385221,
+      "loss": 2.1336,
+      "step": 13
+    },
+    {
+      "epoch": 0.15,
+      "grad_norm": 6.037710889841169,
+      "learning_rate": 0.00019990638860040006,
+      "loss": 1.85,
+      "step": 14
+    },
+    {
+      "epoch": 0.16,
+      "grad_norm": 1.0500019118881985,
+      "learning_rate": 0.00019986520883988232,
+      "loss": 1.5964,
+      "step": 15
+    },
+    {
+      "epoch": 0.17,
+      "grad_norm": 0.6169710624824223,
+      "learning_rate": 0.00019981654914134686,
+      "loss": 1.4307,
+      "step": 16
+    },
+    {
+      "epoch": 0.18,
+      "grad_norm": 1.86114059095932,
+      "learning_rate": 0.00019976041314942155,
+      "loss": 1.4285,
+      "step": 17
+    },
+    {
+      "epoch": 0.19,
+      "grad_norm": 1.6513877610200167,
+      "learning_rate": 0.00019969680506871137,
+      "loss": 1.4621,
+      "step": 18
+    },
+    {
+      "epoch": 0.2,
+      "grad_norm": 1.4395882738454628,
+      "learning_rate": 0.000199625729663483,
+      "loss": 1.3561,
+      "step": 19
+    },
+    {
+      "epoch": 0.21,
+      "grad_norm": 0.70847060238536,
+      "learning_rate": 0.00019954719225730847,
+      "loss": 1.3565,
+      "step": 20
+    },
+    {
+      "epoch": 0.22,
+      "grad_norm": 0.4331630595385925,
+      "learning_rate": 0.00019946119873266613,
+      "loss": 1.3374,
+      "step": 21
+    },
+    {
+      "epoch": 0.23,
+      "grad_norm": 0.5580281682185451,
+      "learning_rate": 0.0001993677555305002,
+      "loss": 1.313,
+      "step": 22
+    },
+    {
+      "epoch": 0.25,
+      "grad_norm": 0.5217443953771937,
+      "learning_rate": 0.00019926686964973813,
+      "loss": 1.2541,
+      "step": 23
+    },
+    {
+      "epoch": 0.26,
+      "grad_norm": 0.36823120314463453,
+      "learning_rate": 0.00019915854864676664,
+      "loss": 1.2191,
+      "step": 24
+    },
+    {
+      "epoch": 0.26,
+      "eval_loss": 1.2946609258651733,
+      "eval_runtime": 118.9039,
+      "eval_samples_per_second": 8.402,
+      "eval_steps_per_second": 0.53,
+      "step": 24
+    },
+    {
+      "epoch": 0.27,
+      "grad_norm": 0.5797477063688413,
+      "learning_rate": 0.0001990428006348656,
+      "loss": 1.24,
+      "step": 25
+    },
+    {
+      "epoch": 0.28,
+      "grad_norm": 0.41369538857234545,
+      "learning_rate": 0.00019891963428360043,
+      "loss": 1.209,
+      "step": 26
+    },
+    {
+      "epoch": 0.29,
+      "grad_norm": 0.36666008426797836,
+      "learning_rate": 0.00019878905881817252,
+      "loss": 1.2543,
+      "step": 27
+    },
+    {
+      "epoch": 0.3,
+      "grad_norm": 0.3976779691989045,
+      "learning_rate": 0.00019865108401872857,
+      "loss": 1.2431,
+      "step": 28
+    },
+    {
+      "epoch": 0.31,
+      "grad_norm": 0.4992861718630414,
+      "learning_rate": 0.00019850572021962788,
+      "loss": 1.2471,
+      "step": 29
+    },
+    {
+      "epoch": 0.32,
+      "grad_norm": 0.33729072192890136,
+      "learning_rate": 0.00019835297830866826,
+      "loss": 1.1933,
+      "step": 30
+    },
+    {
+      "epoch": 0.33,
+      "grad_norm": 0.29373457949318904,
+      "learning_rate": 0.00019819286972627066,
+      "loss": 1.1761,
+      "step": 31
+    },
+    {
+      "epoch": 0.34,
+      "grad_norm": 0.5339184947140588,
+      "learning_rate": 0.0001980254064646223,
+      "loss": 1.165,
+      "step": 32
+    },
+    {
+      "epoch": 0.35,
+      "grad_norm": 0.38755069216510263,
+      "learning_rate": 0.00019785060106677818,
+      "loss": 1.1236,
+      "step": 33
+    },
+    {
+      "epoch": 0.36,
+      "grad_norm": 0.338373181403367,
+      "learning_rate": 0.00019766846662572191,
+      "loss": 1.2102,
+      "step": 34
+    },
+    {
+      "epoch": 0.37,
+      "grad_norm": 0.39237714718744304,
+      "learning_rate": 0.00019747901678338496,
+      "loss": 1.1642,
+      "step": 35
+    },
+    {
+      "epoch": 0.38,
+      "grad_norm": 0.3614249847081747,
+      "learning_rate": 0.00019728226572962473,
+      "loss": 1.1387,
+      "step": 36
+    },
+    {
+      "epoch": 0.4,
+      "grad_norm": 0.28278007479509987,
+      "learning_rate": 0.00019707822820116193,
+      "loss": 1.0939,
+      "step": 37
+    },
+    {
+      "epoch": 0.41,
+      "grad_norm": 0.3008254873268798,
+      "learning_rate": 0.00019686691948047664,
+      "loss": 1.1346,
+      "step": 38
+    },
+    {
+      "epoch": 0.42,
+      "grad_norm": 0.4263010439416343,
+      "learning_rate": 0.0001966483553946637,
+      "loss": 1.1015,
+      "step": 39
+    },
+    {
+      "epoch": 0.43,
+      "grad_norm": 0.32725448028464205,
+      "learning_rate": 0.00019642255231424729,
+      "loss": 1.1324,
+      "step": 40
+    },
+    {
+      "epoch": 0.44,
+      "grad_norm": 0.3028242900588441,
+      "learning_rate": 0.00019618952715195475,
+      "loss": 1.1147,
+      "step": 41
+    },
+    {
+      "epoch": 0.45,
+      "grad_norm": 0.33893311928252234,
+      "learning_rate": 0.00019594929736144976,
+      "loss": 1.0978,
+      "step": 42
+    },
+    {
+      "epoch": 0.46,
+      "grad_norm": 0.2786082334492372,
+      "learning_rate": 0.0001957018809360251,
+      "loss": 1.0933,
+      "step": 43
+    },
+    {
+      "epoch": 0.47,
+      "grad_norm": 0.2732185168098956,
+      "learning_rate": 0.00019544729640725498,
+      "loss": 1.084,
+      "step": 44
+    },
+    {
+      "epoch": 0.48,
+      "grad_norm": 0.33386436894143035,
+      "learning_rate": 0.00019518556284360696,
+      "loss": 1.0673,
+      "step": 45
+    },
+    {
+      "epoch": 0.49,
+      "grad_norm": 0.2761688734050621,
+      "learning_rate": 0.00019491669984901379,
+      "loss": 1.0523,
+      "step": 46
+    },
+    {
+      "epoch": 0.5,
+      "grad_norm": 0.3346957388610895,
+      "learning_rate": 0.00019464072756140486,
+      "loss": 1.0913,
+      "step": 47
+    },
+    {
+      "epoch": 0.51,
+      "grad_norm": 0.30196058996924285,
+      "learning_rate": 0.0001943576666511982,
+      "loss": 1.1165,
+      "step": 48
+    },
+    {
+      "epoch": 0.51,
+      "eval_loss": 1.167867660522461,
+      "eval_runtime": 119.1485,
+      "eval_samples_per_second": 8.384,
+      "eval_steps_per_second": 0.529,
+      "step": 48
+    },
+    {
+      "epoch": 0.52,
+      "grad_norm": 0.27445390350987153,
+      "learning_rate": 0.00019406753831975203,
+      "loss": 1.1069,
+      "step": 49
+    },
+    {
+      "epoch": 0.53,
+      "grad_norm": 0.34729097228771255,
+      "learning_rate": 0.00019377036429777672,
+      "loss": 1.0567,
+      "step": 50
+    },
+    {
+      "epoch": 0.54,
+      "grad_norm": 0.31314016575739406,
+      "learning_rate": 0.0001934661668437073,
+      "loss": 1.0875,
+      "step": 51
+    },
+    {
+      "epoch": 0.56,
+      "grad_norm": 0.29140014335226905,
+      "learning_rate": 0.0001931549687420364,
+      "loss": 1.0929,
+      "step": 52
+    },
+    {
+      "epoch": 0.57,
+      "grad_norm": 0.2638104110161505,
+      "learning_rate": 0.00019283679330160726,
+      "loss": 1.0963,
+      "step": 53
+    },
+    {
+      "epoch": 0.58,
+      "grad_norm": 0.2833945318119855,
+      "learning_rate": 0.0001925116643538684,
+      "loss": 1.0535,
+      "step": 54
+    },
+    {
+      "epoch": 0.59,
+      "grad_norm": 0.28672689795285417,
+      "learning_rate": 0.0001921796062510882,
+      "loss": 1.0699,
+      "step": 55
+    },
+    {
+      "epoch": 0.6,
+      "grad_norm": 0.261255409262294,
+      "learning_rate": 0.00019184064386453128,
+      "loss": 1.0658,
+      "step": 56
+    },
+    {
+      "epoch": 0.61,
+      "grad_norm": 0.24304864434604007,
+      "learning_rate": 0.00019149480258259533,
+      "loss": 1.0441,
+      "step": 57
+    },
+    {
+      "epoch": 0.62,
+      "grad_norm": 0.2987107937915846,
+      "learning_rate": 0.00019114210830890969,
+      "loss": 1.0061,
+      "step": 58
+    },
+    {
+      "epoch": 0.63,
+      "grad_norm": 0.2617045441373282,
+      "learning_rate": 0.00019078258746039507,
+      "loss": 1.0578,
+      "step": 59
+    },
+    {
+      "epoch": 0.64,
+      "grad_norm": 0.2577955355987167,
+      "learning_rate": 0.00019041626696528503,
+      "loss": 1.0333,
+      "step": 60
+    },
+    {
+      "epoch": 0.65,
+      "grad_norm": 0.2823058812174375,
+      "learning_rate": 0.0001900431742611089,
+      "loss": 1.0837,
+      "step": 61
+    },
+    {
+      "epoch": 0.66,
+      "grad_norm": 0.30425238718712166,
+      "learning_rate": 0.00018966333729263674,
+      "loss": 1.0619,
+      "step": 62
+    },
+    {
+      "epoch": 0.67,
+      "grad_norm": 0.29826831116146957,
+      "learning_rate": 0.0001892767845097864,
+      "loss": 1.056,
+      "step": 63
+    },
+    {
+      "epoch": 0.68,
+      "grad_norm": 0.22990267950533677,
+      "learning_rate": 0.00018888354486549237,
+      "loss": 1.061,
+      "step": 64
+    },
+    {
+      "epoch": 0.69,
+      "grad_norm": 0.27604852373975236,
+      "learning_rate": 0.00018848364781353744,
+      "loss": 1.0624,
+      "step": 65
+    },
+    {
+      "epoch": 0.7,
+      "grad_norm": 0.302101014156969,
+      "learning_rate": 0.00018807712330634642,
+      "loss": 1.0965,
+      "step": 66
+    },
+    {
+      "epoch": 0.72,
+      "grad_norm": 0.2532153192142023,
+      "learning_rate": 0.00018766400179274286,
+      "loss": 1.0972,
+      "step": 67
+    },
+    {
+      "epoch": 0.73,
+      "grad_norm": 0.23803088057755897,
+      "learning_rate": 0.00018724431421566823,
+      "loss": 1.0823,
+      "step": 68
+    },
+    {
+      "epoch": 0.74,
+      "grad_norm": 0.2200041903156331,
+      "learning_rate": 0.0001868180920098644,
+      "loss": 1.037,
+      "step": 69
+    },
+    {
+      "epoch": 0.75,
+      "grad_norm": 0.31123761066229655,
+      "learning_rate": 0.00018638536709951917,
+      "loss": 1.0689,
+      "step": 70
+    },
+    {
+      "epoch": 0.76,
+      "grad_norm": 0.2760757149384919,
+      "learning_rate": 0.00018594617189587512,
+      "loss": 1.0071,
+      "step": 71
+    },
+    {
+      "epoch": 0.77,
+      "grad_norm": 0.2452672521810973,
+      "learning_rate": 0.00018550053929480202,
+      "loss": 1.0711,
+      "step": 72
+    },
+    {
+      "epoch": 0.77,
+      "eval_loss": 1.1377497911453247,
+      "eval_runtime": 119.461,
+      "eval_samples_per_second": 8.363,
+      "eval_steps_per_second": 0.527,
+      "step": 72
+    },
+    {
+      "epoch": 0.78,
+      "grad_norm": 0.30897216290479246,
+      "learning_rate": 0.0001850485026743328,
+      "loss": 1.0508,
+      "step": 73
+    },
+    {
+      "epoch": 0.79,
+      "grad_norm": 0.24165903393157925,
+      "learning_rate": 0.00018459009589216364,
+      "loss": 1.046,
+      "step": 74
+    },
+    {
+      "epoch": 0.8,
+      "grad_norm": 0.2509819208307879,
+      "learning_rate": 0.00018412535328311814,
+      "loss": 1.0726,
+      "step": 75
+    },
+    {
+      "epoch": 0.81,
+      "grad_norm": 0.26145395006758515,
+      "learning_rate": 0.00018365430965657526,
+      "loss": 0.9998,
+      "step": 76
+    },
+    {
+      "epoch": 0.82,
+      "grad_norm": 0.26920709605794424,
+      "learning_rate": 0.00018317700029386245,
+      "loss": 1.065,
+      "step": 77
+    },
+    {
+      "epoch": 0.83,
+      "grad_norm": 0.24226754926786417,
+      "learning_rate": 0.0001826934609456129,
+      "loss": 1.0489,
+      "step": 78
+    },
+    {
+      "epoch": 0.84,
+      "grad_norm": 0.3022365661006827,
+      "learning_rate": 0.00018220372782908777,
+      "loss": 1.0372,
+      "step": 79
+    },
+    {
+      "epoch": 0.85,
+      "grad_norm": 0.25795710005352673,
+      "learning_rate": 0.00018170783762546365,
+      "loss": 1.0128,
+      "step": 80
+    },
+    {
+      "epoch": 0.86,
+      "grad_norm": 0.3490748875058354,
+      "learning_rate": 0.00018120582747708502,
+      "loss": 1.0168,
+      "step": 81
+    },
+    {
+      "epoch": 0.88,
+      "grad_norm": 0.24938209735120945,
+      "learning_rate": 0.00018069773498468223,
+      "loss": 0.9586,
+      "step": 82
+    },
+    {
+      "epoch": 0.89,
+      "grad_norm": 0.2527612545099894,
+      "learning_rate": 0.00018018359820455536,
+      "loss": 1.0385,
+      "step": 83
+    },
+    {
+      "epoch": 0.9,
+      "grad_norm": 0.27528879975094916,
+      "learning_rate": 0.0001796634556457236,
+      "loss": 1.0328,
+      "step": 84
+    },
+    {
+      "epoch": 0.91,
+      "grad_norm": 0.2605002777661913,
+      "learning_rate": 0.0001791373462670411,
+      "loss": 0.9966,
+      "step": 85
+    },
+    {
+      "epoch": 0.92,
+      "grad_norm": 0.3117107796665858,
+      "learning_rate": 0.00017860530947427875,
+      "loss": 0.9772,
+      "step": 86
+    },
+    {
+      "epoch": 0.93,
+      "grad_norm": 0.28336227154677734,
+      "learning_rate": 0.0001780673851171728,
+      "loss": 1.0724,
+      "step": 87
+    },
+    {
+      "epoch": 0.94,
+      "grad_norm": 0.42707817919652674,
+      "learning_rate": 0.0001775236134864401,
+      "loss": 1.0038,
+      "step": 88
+    },
+    {
+      "epoch": 0.95,
+      "grad_norm": 0.29236016959846456,
+      "learning_rate": 0.0001769740353107602,
+      "loss": 1.0083,
+      "step": 89
+    },
+    {
+      "epoch": 0.96,
+      "grad_norm": 0.43295063403530637,
+      "learning_rate": 0.00017641869175372493,
+      "loss": 1.022,
+      "step": 90
+    },
+    {
+      "epoch": 0.97,
+      "grad_norm": 0.3086663897043129,
+      "learning_rate": 0.00017585762441075503,
+      "loss": 1.0303,
+      "step": 91
+    },
+    {
+      "epoch": 0.98,
+      "grad_norm": 0.2783768981163154,
+      "learning_rate": 0.0001752908753059849,
+      "loss": 1.061,
+      "step": 92
+    },
+    {
+      "epoch": 0.99,
+      "grad_norm": 0.43168501819843275,
+      "learning_rate": 0.00017471848688911464,
+      "loss": 1.0631,
+      "step": 93
+    },
+    {
+      "epoch": 1.0,
+      "grad_norm": 0.25487494913299935,
+      "learning_rate": 0.0001741405020322309,
+      "loss": 0.9858,
+      "step": 94
+    },
+    {
+      "epoch": 1.01,
+      "grad_norm": 0.3229761094582219,
+      "learning_rate": 0.00017355696402659548,
+      "loss": 0.9495,
+      "step": 95
+    },
+    {
+      "epoch": 1.02,
+      "grad_norm": 0.3178464701266748,
+      "learning_rate": 0.000172967916579403,
+      "loss": 0.9546,
+      "step": 96
+    },
+    {
+      "epoch": 1.02,
+      "eval_loss": 1.1303094625473022,
+      "eval_runtime": 119.6761,
+      "eval_samples_per_second": 8.348,
+      "eval_steps_per_second": 0.526,
+      "step": 96
+    },
+    {
+      "epoch": 1.04,
+      "grad_norm": 0.2534616980189548,
+      "learning_rate": 0.00017237340381050703,
+      "loss": 0.9509,
+      "step": 97
+    },
+    {
+      "epoch": 1.05,
+      "grad_norm": 0.2354382873554396,
+      "learning_rate": 0.00017177347024911562,
+      "loss": 0.9611,
+      "step": 98
+    },
+    {
+      "epoch": 1.06,
+      "grad_norm": 0.2754259154521738,
+      "learning_rate": 0.00017116816083045602,
+      "loss": 0.9184,
+      "step": 99
+    },
+    {
+      "epoch": 1.07,
+      "grad_norm": 0.25868181129480755,
+      "learning_rate": 0.00017055752089240907,
+      "loss": 0.957,
+      "step": 100
+    },
+    {
+      "epoch": 1.08,
+      "grad_norm": 0.2383943586330267,
+      "learning_rate": 0.00016994159617211317,
+      "loss": 0.9638,
+      "step": 101
+    },
+    {
+      "epoch": 1.09,
+      "grad_norm": 0.2706420372628291,
+      "learning_rate": 0.0001693204328025389,
+      "loss": 0.9115,
+      "step": 102
+    },
+    {
+      "epoch": 1.1,
+      "grad_norm": 0.2751042656041904,
+      "learning_rate": 0.0001686940773090333,
+      "loss": 0.9277,
+      "step": 103
+    },
+    {
+      "epoch": 1.11,
+      "grad_norm": 0.27700872737428867,
+      "learning_rate": 0.00016806257660583534,
+      "loss": 0.9248,
+      "step": 104
+    },
+    {
+      "epoch": 1.12,
+      "grad_norm": 0.3350046312844708,
+      "learning_rate": 0.00016742597799256182,
+      "loss": 0.928,
+      "step": 105
+    },
+    {
+      "epoch": 1.13,
+      "grad_norm": 0.4055944986440079,
+      "learning_rate": 0.00016678432915066488,
+      "loss": 0.9074,
+      "step": 106
+    },
+    {
+      "epoch": 1.14,
+      "grad_norm": 0.2515177402600531,
+      "learning_rate": 0.00016613767813986044,
+      "loss": 0.9564,
+      "step": 107
+    },
+    {
+      "epoch": 1.15,
+      "grad_norm": 0.2571149695502646,
+      "learning_rate": 0.00016548607339452853,
+      "loss": 0.93,
+      "step": 108
+    },
+    {
+      "epoch": 1.16,
+      "grad_norm": 0.38608942941048996,
+      "learning_rate": 0.0001648295637200856,
+      "loss": 0.9281,
+      "step": 109
+    },
+    {
+      "epoch": 1.17,
+      "grad_norm": 0.31939838976976676,
+      "learning_rate": 0.000164168198289329,
+      "loss": 0.9914,
+      "step": 110
+    },
+    {
+      "epoch": 1.19,
+      "grad_norm": 0.30504937567650897,
+      "learning_rate": 0.00016350202663875386,
+      "loss": 0.9549,
+      "step": 111
+    },
+    {
+      "epoch": 1.2,
+      "grad_norm": 0.3320388344291162,
+      "learning_rate": 0.0001628310986648427,
+      "loss": 0.9086,
+      "step": 112
+    },
+    {
+      "epoch": 1.21,
+      "grad_norm": 0.27715569151296165,
+      "learning_rate": 0.0001621554646203284,
+      "loss": 0.8537,
+      "step": 113
+    },
+    {
+      "epoch": 1.22,
+      "grad_norm": 0.278787508566418,
+      "learning_rate": 0.0001614751751104301,
+      "loss": 0.9354,
+      "step": 114
+    },
+    {
+      "epoch": 1.23,
+      "grad_norm": 0.24483614460003267,
+      "learning_rate": 0.00016079028108906282,
+      "loss": 0.8996,
+      "step": 115
+    },
+    {
+      "epoch": 1.24,
+      "grad_norm": 0.37520609596400134,
+      "learning_rate": 0.0001601008338550211,
+      "loss": 0.9514,
+      "step": 116
+    },
+    {
+      "epoch": 1.25,
+      "grad_norm": 0.2565631505653599,
+      "learning_rate": 0.00015940688504813662,
+      "loss": 0.8984,
+      "step": 117
+    },
+    {
+      "epoch": 1.26,
+      "grad_norm": 0.26348552476529935,
+      "learning_rate": 0.00015870848664541044,
+      "loss": 0.8941,
+      "step": 118
+    },
+    {
+      "epoch": 1.27,
+      "grad_norm": 0.32431198985496534,
+      "learning_rate": 0.00015800569095711982,
+      "loss": 0.8876,
+      "step": 119
+    },
+    {
+      "epoch": 1.28,
+      "grad_norm": 0.29308039763069227,
+      "learning_rate": 0.00015729855062290022,
+      "loss": 0.9309,
+      "step": 120
+    },
+    {
+      "epoch": 1.28,
+      "eval_loss": 1.129751205444336,
+      "eval_runtime": 119.1497,
+      "eval_samples_per_second": 8.384,
+      "eval_steps_per_second": 0.529,
+      "step": 120
+    },
+    {
+      "epoch": 1.29,
+      "grad_norm": 0.2793291380060977,
+      "learning_rate": 0.0001565871186078025,
+      "loss": 0.9453,
+      "step": 121
+    },
+    {
+      "epoch": 1.3,
+      "grad_norm": 0.28873644301555734,
+      "learning_rate": 0.000155871448198326,
+      "loss": 0.9243,
+      "step": 122
+    },
+    {
+      "epoch": 1.31,
+      "grad_norm": 0.3086103724578039,
+      "learning_rate": 0.00015515159299842707,
+      "loss": 0.8877,
+      "step": 123
+    },
+    {
+      "epoch": 1.32,
+      "grad_norm": 0.30407892484693505,
+      "learning_rate": 0.00015442760692550443,
+      "loss": 0.9448,
+      "step": 124
+    },
+    {
+      "epoch": 1.33,
+      "grad_norm": 0.29771602861368474,
+      "learning_rate": 0.00015369954420636048,
+      "loss": 0.889,
+      "step": 125
+    },
+    {
+      "epoch": 1.35,
+      "grad_norm": 0.30480490158838136,
+      "learning_rate": 0.00015296745937313987,
+      "loss": 0.9405,
+      "step": 126
+    },
+    {
+      "epoch": 1.36,
+      "grad_norm": 0.2949192855418127,
+      "learning_rate": 0.00015223140725924495,
+      "loss": 0.9382,
+      "step": 127
+    },
+    {
+      "epoch": 1.37,
+      "grad_norm": 0.2813631863132807,
+      "learning_rate": 0.00015149144299522873,
+      "loss": 0.9526,
+      "step": 128
+    },
+    {
+      "epoch": 1.38,
+      "grad_norm": 0.28548924064070513,
+      "learning_rate": 0.00015074762200466556,
+      "loss": 0.9174,
+      "step": 129
+    },
+    {
+      "epoch": 1.39,
+      "grad_norm": 0.28137053449960464,
+      "learning_rate": 0.00015000000000000001,
+      "loss": 0.9244,
+      "step": 130
+    },
+    {
+      "epoch": 1.4,
+      "grad_norm": 0.2626750895717777,
+      "learning_rate": 0.00014924863297837378,
+      "loss": 0.9335,
+      "step": 131
+    },
+    {
+      "epoch": 1.41,
+      "grad_norm": 0.26686502371015536,
+      "learning_rate": 0.00014849357721743168,
+      "loss": 0.8948,
+      "step": 132
+    },
+    {
+      "epoch": 1.42,
+      "grad_norm": 0.3332273481179679,
+      "learning_rate": 0.00014773488927110633,
+      "loss": 0.9274,
+      "step": 133
+    },
+    {
+      "epoch": 1.43,
+      "grad_norm": 0.2528048763375234,
+      "learning_rate": 0.00014697262596538227,
+      "loss": 0.8731,
+      "step": 134
+    },
+    {
+      "epoch": 1.44,
+      "grad_norm": 0.27184211707488076,
+      "learning_rate": 0.00014620684439403962,
+      "loss": 0.9318,
+      "step": 135
+    },
+    {
+      "epoch": 1.45,
+      "grad_norm": 0.3051111137538683,
+      "learning_rate": 0.0001454376019143779,
+      "loss": 0.9447,
+      "step": 136
+    },
+    {
+      "epoch": 1.46,
+      "grad_norm": 0.28771401659835155,
+      "learning_rate": 0.00014466495614291977,
+      "loss": 0.9343,
+      "step": 137
+    },
+    {
+      "epoch": 1.47,
+      "grad_norm": 0.28995797921621524,
+      "learning_rate": 0.0001438889649510956,
+      "loss": 0.8978,
+      "step": 138
+    },
+    {
+      "epoch": 1.48,
+      "grad_norm": 0.2749930548874636,
+      "learning_rate": 0.00014310968646090883,
+      "loss": 0.924,
+      "step": 139
+    },
+    {
+      "epoch": 1.49,
+      "grad_norm": 0.3097189537380989,
+      "learning_rate": 0.0001423271790405828,
+      "loss": 0.9574,
+      "step": 140
+    },
+    {
+      "epoch": 1.51,
+      "grad_norm": 0.2449218990319832,
+      "learning_rate": 0.00014154150130018866,
+      "loss": 0.8475,
+      "step": 141
+    },
+    {
+      "epoch": 1.52,
+      "grad_norm": 0.24856388098419674,
+      "learning_rate": 0.0001407527120872557,
+      "loss": 0.9381,
+      "step": 142
+    },
+    {
+      "epoch": 1.53,
+      "grad_norm": 0.3169861882853132,
+      "learning_rate": 0.00013996087048236358,
+      "loss": 0.9141,
+      "step": 143
+    },
+    {
+      "epoch": 1.54,
+      "grad_norm": 0.30689184261103974,
+      "learning_rate": 0.00013916603579471705,
+      "loss": 0.9588,
+      "step": 144
+    },
+    {
+      "epoch": 1.54,
+      "eval_loss": 1.1242448091506958,
+      "eval_runtime": 119.0725,
+      "eval_samples_per_second": 8.39,
+      "eval_steps_per_second": 0.529,
+      "step": 144
+    },
+    {
+      "epoch": 1.55,
+      "grad_norm": 0.2961514212977567,
+      "learning_rate": 0.00013836826755770384,
+      "loss": 0.9371,
+      "step": 145
+    },
+    {
+      "epoch": 1.56,
+      "grad_norm": 0.30790856503439346,
+      "learning_rate": 0.00013756762552443553,
+      "loss": 0.9612,
+      "step": 146
+    },
+    {
+      "epoch": 1.57,
+      "grad_norm": 0.3517398492864053,
+      "learning_rate": 0.000136764169663272,
+      "loss": 0.9253,
+      "step": 147
+    },
+    {
+      "epoch": 1.58,
+      "grad_norm": 0.26375798832515857,
+      "learning_rate": 0.00013595796015332984,
+      "loss": 0.8977,
+      "step": 148
+    },
+    {
+      "epoch": 1.59,
+      "grad_norm": 0.274348892672977,
+      "learning_rate": 0.00013514905737997473,
+      "loss": 0.8817,
+      "step": 149
+    },
+    {
+      "epoch": 1.6,
+      "grad_norm": 0.35917564750751624,
+      "learning_rate": 0.00013433752193029886,
+      "loss": 0.886,
+      "step": 150
+    },
+    {
+      "epoch": 1.61,
+      "grad_norm": 0.38175124377914293,
+      "learning_rate": 0.00013352341458858265,
+      "loss": 0.8576,
+      "step": 151
+    },
+    {
+      "epoch": 1.62,
+      "grad_norm": 0.249633953215678,
+      "learning_rate": 0.00013270679633174218,
+      "loss": 1.0066,
+      "step": 152
+    },
+    {
+      "epoch": 1.63,
+      "grad_norm": 0.33494494430574784,
+      "learning_rate": 0.00013188772832476188,
+      "loss": 0.884,
+      "step": 153
+    },
+    {
+      "epoch": 1.64,
+      "grad_norm": 0.4176467296744032,
+      "learning_rate": 0.00013106627191611332,
+      "loss": 0.9041,
+      "step": 154
+    },
+    {
+      "epoch": 1.65,
+      "grad_norm": 0.27051479454532207,
+      "learning_rate": 0.00013024248863316012,
+      "loss": 0.8764,
+      "step": 155
+    },
+    {
+      "epoch": 1.67,
+      "grad_norm": 0.29302599029848847,
+      "learning_rate": 0.00012941644017754964,
+      "loss": 0.9786,
+      "step": 156
+    },
+    {
+      "epoch": 1.68,
+      "grad_norm": 0.3127378512248151,
+      "learning_rate": 0.00012858818842059145,
+      "loss": 0.9176,
+      "step": 157
+    },
+    {
+      "epoch": 1.69,
+      "grad_norm": 0.40647077063662906,
+      "learning_rate": 0.00012775779539862304,
+      "loss": 0.9387,
+      "step": 158
+    },
+    {
+      "epoch": 1.7,
+      "grad_norm": 0.29290601694481777,
+      "learning_rate": 0.00012692532330836346,
+      "loss": 0.9192,
+      "step": 159
+    },
+    {
+      "epoch": 1.71,
+      "grad_norm": 0.2819168741245354,
+      "learning_rate": 0.0001260908345022547,
+      "loss": 0.9253,
+      "step": 160
+    },
+    {
+      "epoch": 1.72,
+      "grad_norm": 0.3772714091394927,
+      "learning_rate": 0.00012525439148379128,
+      "loss": 0.9264,
+      "step": 161
+    },
+    {
+      "epoch": 1.73,
+      "grad_norm": 0.29399851067321503,
+      "learning_rate": 0.00012441605690283915,
+      "loss": 0.9357,
+      "step": 162
+    },
+    {
+      "epoch": 1.74,
+      "grad_norm": 0.2623180246832513,
+      "learning_rate": 0.00012357589355094275,
+      "loss": 0.8516,
+      "step": 163
+    },
+    {
+      "epoch": 1.75,
+      "grad_norm": 0.27796942024085824,
+      "learning_rate": 0.00012273396435662212,
+      "loss": 0.9328,
+      "step": 164
+    },
+    {
+      "epoch": 1.76,
+      "grad_norm": 0.3107670297529076,
+      "learning_rate": 0.0001218903323806595,
+      "loss": 0.8769,
+      "step": 165
+    },
+    {
+      "epoch": 1.77,
+      "grad_norm": 0.2865573350738354,
+      "learning_rate": 0.00012104506081137608,
+      "loss": 0.9015,
+      "step": 166
+    },
+    {
+      "epoch": 1.78,
+      "grad_norm": 0.30595087117636693,
+      "learning_rate": 0.00012019821295989912,
+      "loss": 0.94,
+      "step": 167
+    },
+    {
+      "epoch": 1.79,
+      "grad_norm": 0.32540365653257874,
+      "learning_rate": 0.00011934985225541998,
+      "loss": 0.8553,
+      "step": 168
+    },
+    {
+      "epoch": 1.79,
+      "eval_loss": 1.1259374618530273,
+      "eval_runtime": 119.4351,
+      "eval_samples_per_second": 8.364,
+      "eval_steps_per_second": 0.527,
+      "step": 168
+    },
+    {
+      "epoch": 1.8,
+      "grad_norm": 0.3058868303314457,
+      "learning_rate": 0.00011850004224044315,
+      "loss": 0.9074,
+      "step": 169
+    },
+    {
+      "epoch": 1.81,
+      "grad_norm": 0.33266760488242775,
+      "learning_rate": 0.0001176488465660271,
+      "loss": 0.8799,
+      "step": 170
+    },
+    {
+      "epoch": 1.83,
+      "grad_norm": 0.3101183375673487,
+      "learning_rate": 0.00011679632898701649,
+      "loss": 0.9004,
+      "step": 171
+    },
+    {
+      "epoch": 1.84,
+      "grad_norm": 0.31535579418195775,
+      "learning_rate": 0.00011594255335726724,
+      "loss": 0.9238,
+      "step": 172
+    },
+    {
+      "epoch": 1.85,
+      "grad_norm": 0.28341827112854334,
+      "learning_rate": 0.00011508758362486358,
+      "loss": 0.9138,
+      "step": 173
+    },
+    {
+      "epoch": 1.86,
+      "grad_norm": 0.25699888796695625,
+      "learning_rate": 0.00011423148382732853,
+      "loss": 0.9175,
+      "step": 174
+    },
+    {
+      "epoch": 1.87,
+      "grad_norm": 0.29504332662698246,
+      "learning_rate": 0.0001133743180868273,
+      "loss": 0.9023,
+      "step": 175
+    },
+    {
+      "epoch": 1.88,
+      "grad_norm": 0.2993175263873948,
+      "learning_rate": 0.0001125161506053646,
+      "loss": 0.8893,
+      "step": 176
+    },
+    {
+      "epoch": 1.89,
+      "grad_norm": 0.2762659379409218,
+      "learning_rate": 0.00011165704565997593,
+      "loss": 0.9071,
+      "step": 177
+    },
+    {
+      "epoch": 1.9,
+      "grad_norm": 0.23620994229530515,
+      "learning_rate": 0.00011079706759791311,
+      "loss": 0.8796,
+      "step": 178
+    },
+    {
+      "epoch": 1.91,
+      "grad_norm": 0.28317619721877,
+      "learning_rate": 0.00010993628083182467,
+      "loss": 0.8983,
+      "step": 179
+    },
+    {
+      "epoch": 1.92,
+      "grad_norm": 0.3252854551640304,
+      "learning_rate": 0.00010907474983493144,
+      "loss": 0.8947,
+      "step": 180
+    },
+    {
+      "epoch": 1.93,
+      "grad_norm": 0.2579136274422669,
+      "learning_rate": 0.00010821253913619726,
+      "loss": 0.8726,
+      "step": 181
+    },
+    {
+      "epoch": 1.94,
+      "grad_norm": 0.27201912720918364,
+      "learning_rate": 0.00010734971331549603,
+      "loss": 0.891,
+      "step": 182
+    },
+    {
+      "epoch": 1.95,
+      "grad_norm": 0.41257277193589503,
+      "learning_rate": 0.0001064863369987743,
+      "loss": 0.9188,
+      "step": 183
+    },
+    {
+      "epoch": 1.96,
+      "grad_norm": 0.264920112831242,
+      "learning_rate": 0.00010562247485321115,
+      "loss": 0.8761,
+      "step": 184
+    },
+    {
+      "epoch": 1.98,
+      "grad_norm": 0.28166441056422037,
+      "learning_rate": 0.00010475819158237425,
+      "loss": 0.8805,
+      "step": 185
+    },
+    {
+      "epoch": 1.99,
+      "grad_norm": 0.2818961139392159,
+      "learning_rate": 0.00010389355192137377,
+      "loss": 0.8934,
+      "step": 186
+    },
+    {
+      "epoch": 2.0,
+      "grad_norm": 0.27424787600345923,
+      "learning_rate": 0.00010302862063201367,
+      "loss": 0.9237,
+      "step": 187
+    },
+    {
+      "epoch": 2.01,
+      "grad_norm": 0.25570082666079225,
+      "learning_rate": 0.00010216346249794087,
+      "loss": 0.8656,
+      "step": 188
+    },
+    {
+      "epoch": 2.02,
+      "grad_norm": 0.2712359904481713,
+      "learning_rate": 0.0001012981423197931,
+      "loss": 0.7627,
+      "step": 189
+    },
+    {
+      "epoch": 2.03,
+      "grad_norm": 0.25054404547068676,
+      "learning_rate": 0.00010043272491034523,
+      "loss": 0.8142,
+      "step": 190
+    },
+    {
+      "epoch": 2.04,
+      "grad_norm": 0.28520868420260026,
+      "learning_rate": 9.956727508965481e-05,
+      "loss": 0.7953,
+      "step": 191
+    },
+    {
+      "epoch": 2.05,
+      "grad_norm": 0.29413880984694873,
+      "learning_rate": 9.870185768020693e-05,
+      "loss": 0.8231,
+      "step": 192
+    },
+    {
+      "epoch": 2.05,
+      "eval_loss": 1.144862413406372,
+      "eval_runtime": 119.3004,
+      "eval_samples_per_second": 8.374,
+      "eval_steps_per_second": 0.528,
+      "step": 192
+    },
+    {
+      "epoch": 2.06,
+      "grad_norm": 0.28378300985247035,
+      "learning_rate": 9.783653750205915e-05,
+      "loss": 0.7478,
+      "step": 193
+    },
+    {
+      "epoch": 2.07,
+      "grad_norm": 0.31792721348179676,
+      "learning_rate": 9.697137936798634e-05,
+      "loss": 0.7961,
+      "step": 194
+    },
+    {
+      "epoch": 2.08,
+      "grad_norm": 0.3291666436295964,
+      "learning_rate": 9.610644807862625e-05,
+      "loss": 0.7434,
+      "step": 195
+    },
+    {
+      "epoch": 2.09,
+      "grad_norm": 0.301579259001567,
+      "learning_rate": 9.524180841762577e-05,
+      "loss": 0.7779,
+      "step": 196
+    },
+    {
+      "epoch": 2.1,
+      "grad_norm": 0.30252161240414444,
+      "learning_rate": 9.437752514678887e-05,
+      "loss": 0.7689,
+      "step": 197
+    },
+    {
+      "epoch": 2.11,
+      "grad_norm": 0.3350657085129171,
+      "learning_rate": 9.35136630012257e-05,
+      "loss": 0.7574,
+      "step": 198
+    },
+    {
+      "epoch": 2.12,
+      "grad_norm": 0.3053109929956358,
+      "learning_rate": 9.265028668450402e-05,
+      "loss": 0.7729,
+      "step": 199
+    },
+    {
+      "epoch": 2.14,
+      "grad_norm": 0.30367223609567207,
+      "learning_rate": 9.178746086380275e-05,
+      "loss": 0.8111,
+      "step": 200
+    },
+    {
+      "epoch": 2.15,
+      "grad_norm": 0.3366440949136126,
+      "learning_rate": 9.092525016506858e-05,
+      "loss": 0.7986,
+      "step": 201
+    },
+    {
+      "epoch": 2.16,
+      "grad_norm": 0.3228036608413652,
+      "learning_rate": 9.006371916817534e-05,
+      "loss": 0.8382,
+      "step": 202
+    },
+    {
+      "epoch": 2.17,
+      "grad_norm": 0.2919040789403488,
+      "learning_rate": 8.920293240208694e-05,
+      "loss": 0.7696,
+      "step": 203
+    },
+    {
+      "epoch": 2.18,
+      "grad_norm": 0.30084198177583166,
+      "learning_rate": 8.83429543400241e-05,
+      "loss": 0.7671,
+      "step": 204
+    },
+    {
+      "epoch": 2.19,
+      "grad_norm": 0.33931609000743107,
+      "learning_rate": 8.748384939463543e-05,
+      "loss": 0.7553,
+      "step": 205
+    },
+    {
+      "epoch": 2.2,
+      "grad_norm": 0.30413284924824485,
+      "learning_rate": 8.662568191317273e-05,
+      "loss": 0.7324,
+      "step": 206
+    },
+    {
+      "epoch": 2.21,
+      "grad_norm": 0.3014038998090481,
+      "learning_rate": 8.57685161726715e-05,
+      "loss": 0.7567,
+      "step": 207
+    },
+    {
+      "epoch": 2.22,
+      "grad_norm": 0.3176466329519527,
+      "learning_rate": 8.491241637513644e-05,
+      "loss": 0.8222,
+      "step": 208
+    },
+    {
+      "epoch": 2.23,
+      "grad_norm": 0.29981213041628285,
+      "learning_rate": 8.405744664273278e-05,
+      "loss": 0.7077,
+      "step": 209
+    },
+    {
+      "epoch": 2.24,
+      "grad_norm": 0.2937916452228122,
+      "learning_rate": 8.320367101298351e-05,
+      "loss": 0.7231,
+      "step": 210
+    },
+    {
+      "epoch": 2.25,
+      "grad_norm": 0.32040684171320816,
+      "learning_rate": 8.235115343397295e-05,
+      "loss": 0.7556,
+      "step": 211
+    },
+    {
+      "epoch": 2.26,
+      "grad_norm": 0.31083028085316033,
+      "learning_rate": 8.149995775955686e-05,
+      "loss": 0.7514,
+      "step": 212
+    },
+    {
+      "epoch": 2.27,
+      "grad_norm": 0.3215465383581194,
+      "learning_rate": 8.065014774458003e-05,
+      "loss": 0.7933,
+      "step": 213
+    },
+    {
+      "epoch": 2.28,
+      "grad_norm": 0.3081200259196015,
+      "learning_rate": 7.980178704010089e-05,
+      "loss": 0.8062,
+      "step": 214
+    },
+    {
+      "epoch": 2.3,
+      "grad_norm": 0.3333248296288759,
+      "learning_rate": 7.895493918862396e-05,
+      "loss": 0.7784,
+      "step": 215
+    },
+    {
+      "epoch": 2.31,
+      "grad_norm": 0.3301326097292383,
+      "learning_rate": 7.810966761934053e-05,
+      "loss": 0.8154,
+      "step": 216
+    },
+    {
+      "epoch": 2.31,
+      "eval_loss": 1.1513652801513672,
+      "eval_runtime": 119.4371,
+      "eval_samples_per_second": 8.364,
+      "eval_steps_per_second": 0.527,
+      "step": 216
+    },
+    {
+      "epoch": 2.32,
+      "grad_norm": 0.3166760836422428,
+      "learning_rate": 7.726603564337791e-05,
+      "loss": 0.7486,
+      "step": 217
+    },
+    {
+      "epoch": 2.33,
+      "grad_norm": 0.31309757318131876,
+      "learning_rate": 7.642410644905726e-05,
+      "loss": 0.771,
+      "step": 218
+    },
+    {
+      "epoch": 2.34,
+      "grad_norm": 0.36968796131043985,
+      "learning_rate": 7.558394309716088e-05,
+      "loss": 0.8051,
+      "step": 219
+    },
+    {
+      "epoch": 2.35,
+      "grad_norm": 0.27537675917328025,
+      "learning_rate": 7.474560851620873e-05,
+      "loss": 0.7536,
+      "step": 220
+    },
+    {
+      "epoch": 2.36,
+      "grad_norm": 0.2878011945022053,
+      "learning_rate": 7.390916549774536e-05,
+      "loss": 0.8126,
+      "step": 221
+    },
+    {
+      "epoch": 2.37,
+      "grad_norm": 0.3172405217395398,
+      "learning_rate": 7.307467669163655e-05,
+      "loss": 0.8156,
+      "step": 222
+    },
+    {
+      "epoch": 2.38,
+      "grad_norm": 0.3183651086957915,
+      "learning_rate": 7.224220460137701e-05,
+      "loss": 0.7821,
+      "step": 223
+    },
+    {
+      "epoch": 2.39,
+      "grad_norm": 0.3318078467573977,
+      "learning_rate": 7.141181157940859e-05,
+      "loss": 0.7993,
+      "step": 224
+    },
+    {
+      "epoch": 2.4,
+      "grad_norm": 0.28446170407344085,
+      "learning_rate": 7.058355982245037e-05,
+      "loss": 0.7987,
+      "step": 225
+    },
+    {
+      "epoch": 2.41,
+      "grad_norm": 0.33568352702219995,
+      "learning_rate": 6.97575113668399e-05,
+      "loss": 0.773,
+      "step": 226
+    },
+    {
+      "epoch": 2.42,
+      "grad_norm": 0.30820575901544944,
+      "learning_rate": 6.893372808388675e-05,
+      "loss": 0.813,
+      "step": 227
+    },
+    {
+      "epoch": 2.43,
+      "grad_norm": 0.3121364386024255,
+      "learning_rate": 6.811227167523815e-05,
+      "loss": 0.7716,
+      "step": 228
+    },
+    {
+      "epoch": 2.44,
+      "grad_norm": 0.3211455560922844,
+      "learning_rate": 6.729320366825784e-05,
+      "loss": 0.7577,
+      "step": 229
+    },
+    {
+      "epoch": 2.46,
+      "grad_norm": 0.3315601260165869,
+      "learning_rate": 6.647658541141735e-05,
+      "loss": 0.779,
+      "step": 230
+    },
+    {
+      "epoch": 2.47,
+      "grad_norm": 0.35482236759964675,
+      "learning_rate": 6.566247806970119e-05,
+      "loss": 0.7936,
+      "step": 231
+    },
+    {
+      "epoch": 2.48,
+      "grad_norm": 0.3318703205331905,
+      "learning_rate": 6.485094262002529e-05,
+      "loss": 0.7721,
+      "step": 232
+    },
+    {
+      "epoch": 2.49,
+      "grad_norm": 0.313412585518615,
+      "learning_rate": 6.404203984667019e-05,
+      "loss": 0.7333,
+      "step": 233
+    },
+    {
+      "epoch": 2.5,
+      "grad_norm": 0.3389693444254627,
+      "learning_rate": 6.323583033672799e-05,
+      "loss": 0.6991,
+      "step": 234
+    },
+    {
+      "epoch": 2.51,
+      "grad_norm": 0.33056782619334757,
+      "learning_rate": 6.243237447556449e-05,
+      "loss": 0.7872,
+      "step": 235
+    },
+    {
+      "epoch": 2.52,
+      "grad_norm": 0.3064085209522584,
+      "learning_rate": 6.163173244229619e-05,
+      "loss": 0.7713,
+      "step": 236
+    },
+    {
+      "epoch": 2.53,
+      "grad_norm": 0.3109445125421656,
+      "learning_rate": 6.083396420528298e-05,
+      "loss": 0.8228,
+      "step": 237
+    },
+    {
+      "epoch": 2.54,
+      "grad_norm": 0.35767207742703394,
+      "learning_rate": 6.0039129517636435e-05,
+      "loss": 0.8167,
+      "step": 238
+    },
+    {
+      "epoch": 2.55,
+      "grad_norm": 0.32869196909020376,
+      "learning_rate": 5.924728791274432e-05,
+      "loss": 0.7893,
+      "step": 239
+    },
+    {
+      "epoch": 2.56,
+      "grad_norm": 0.31178216743238674,
+      "learning_rate": 5.845849869981137e-05,
+      "loss": 0.7354,
+      "step": 240
+    },
+    {
+      "epoch": 2.56,
+      "eval_loss": 1.1470853090286255,
+      "eval_runtime": 119.0749,
+      "eval_samples_per_second": 8.39,
+      "eval_steps_per_second": 0.529,
+      "step": 240
+    },
+    {
+      "epoch": 2.57,
+      "grad_norm": 0.3146586486940167,
+      "learning_rate": 5.7672820959417254e-05,
+      "loss": 0.785,
+      "step": 241
+    },
+    {
+      "epoch": 2.58,
+      "grad_norm": 0.3309473634570162,
+      "learning_rate": 5.68903135390912e-05,
+      "loss": 0.7007,
+      "step": 242
+    },
+    {
+      "epoch": 2.59,
+      "grad_norm": 0.2927704203363025,
+      "learning_rate": 5.611103504890444e-05,
+      "loss": 0.778,
+      "step": 243
+    },
+    {
+      "epoch": 2.6,
+      "grad_norm": 0.31346541530480915,
+      "learning_rate": 5.533504385708024e-05,
+      "loss": 0.7272,
+      "step": 244
+    },
+    {
+      "epoch": 2.62,
+      "grad_norm": 0.2996345434845278,
+      "learning_rate": 5.456239808562209e-05,
+      "loss": 0.8091,
+      "step": 245
+    },
+    {
+      "epoch": 2.63,
+      "grad_norm": 0.29407937930772826,
+      "learning_rate": 5.379315560596038e-05,
+      "loss": 0.7666,
+      "step": 246
+    },
+    {
+      "epoch": 2.64,
+      "grad_norm": 0.30530254935425627,
+      "learning_rate": 5.3027374034617785e-05,
+      "loss": 0.7982,
+      "step": 247
+    },
+    {
+      "epoch": 2.65,
+      "grad_norm": 0.3298149075133802,
+      "learning_rate": 5.226511072889371e-05,
+      "loss": 0.7962,
+      "step": 248
+    },
+    {
+      "epoch": 2.66,
+      "grad_norm": 0.33155001378615223,
+      "learning_rate": 5.1506422782568345e-05,
+      "loss": 0.8087,
+      "step": 249
+    },
+    {
+      "epoch": 2.67,
+      "grad_norm": 0.32891369446509405,
+      "learning_rate": 5.0751367021626215e-05,
+      "loss": 0.7702,
+      "step": 250
+    },
+    {
+      "epoch": 2.68,
+      "grad_norm": 0.3042328939887202,
+      "learning_rate": 5.000000000000002e-05,
+      "loss": 0.7924,
+      "step": 251
+    },
+    {
+      "epoch": 2.69,
+      "grad_norm": 0.3037799376581133,
+      "learning_rate": 4.9252377995334444e-05,
+      "loss": 0.7852,
+      "step": 252
+    },
+    {
+      "epoch": 2.7,
+      "grad_norm": 0.3435430445603929,
+      "learning_rate": 4.85085570047713e-05,
+      "loss": 0.7501,
+      "step": 253
+    },
+    {
+      "epoch": 2.71,
+      "grad_norm": 0.3072160193979946,
+      "learning_rate": 4.776859274075506e-05,
+      "loss": 0.7462,
+      "step": 254
+    },
+    {
+      "epoch": 2.72,
+      "grad_norm": 0.3223586439500028,
+      "learning_rate": 4.703254062686017e-05,
+      "loss": 0.775,
+      "step": 255
+    },
+    {
+      "epoch": 2.73,
+      "grad_norm": 0.3270406403084203,
+      "learning_rate": 4.630045579363957e-05,
+      "loss": 0.8306,
+      "step": 256
+    },
+    {
+      "epoch": 2.74,
+      "grad_norm": 0.3360192842512657,
+      "learning_rate": 4.557239307449561e-05,
+      "loss": 0.7697,
+      "step": 257
+    },
+    {
+      "epoch": 2.75,
+      "grad_norm": 0.34282816479900324,
+      "learning_rate": 4.484840700157295e-05,
+      "loss": 0.7654,
+      "step": 258
+    },
+    {
+      "epoch": 2.77,
+      "grad_norm": 0.30039142762313786,
+      "learning_rate": 4.412855180167406e-05,
+      "loss": 0.7703,
+      "step": 259
+    },
+    {
+      "epoch": 2.78,
+      "grad_norm": 0.34307884673711425,
+      "learning_rate": 4.3412881392197526e-05,
+      "loss": 0.7993,
+      "step": 260
+    },
+    {
+      "epoch": 2.79,
+      "grad_norm": 0.33685538845268104,
+      "learning_rate": 4.270144937709981e-05,
+      "loss": 0.7866,
+      "step": 261
+    },
+    {
+      "epoch": 2.8,
+      "grad_norm": 0.33166767859224683,
+      "learning_rate": 4.19943090428802e-05,
+      "loss": 0.8083,
+      "step": 262
+    },
+    {
+      "epoch": 2.81,
+      "grad_norm": 0.3086370003245581,
+      "learning_rate": 4.129151335458957e-05,
+      "loss": 0.7938,
+      "step": 263
+    },
+    {
+      "epoch": 2.82,
+      "grad_norm": 0.3715649674817313,
+      "learning_rate": 4.059311495186338e-05,
+      "loss": 0.7577,
+      "step": 264
+    },
+    {
+      "epoch": 2.82,
+      "eval_loss": 1.1478512287139893,
+      "eval_runtime": 119.1178,
+      "eval_samples_per_second": 8.387,
+      "eval_steps_per_second": 0.529,
+      "step": 264
+    },
+    {
+      "epoch": 2.83,
+      "grad_norm": 0.3298033298390841,
+      "learning_rate": 3.9899166144978904e-05,
+      "loss": 0.8296,
+      "step": 265
+    },
+    {
+      "epoch": 2.84,
+      "grad_norm": 0.3294808666769515,
+      "learning_rate": 3.920971891093718e-05,
+      "loss": 0.8206,
+      "step": 266
+    },
+    {
+      "epoch": 2.85,
+      "grad_norm": 0.3239672501165848,
+      "learning_rate": 3.852482488956992e-05,
+      "loss": 0.8116,
+      "step": 267
+    },
+    {
+      "epoch": 2.86,
+      "grad_norm": 0.3286742994048133,
+      "learning_rate": 3.784453537967161e-05,
+      "loss": 0.8096,
+      "step": 268
+    },
+    {
+      "epoch": 2.87,
+      "grad_norm": 0.31259050250842946,
+      "learning_rate": 3.7168901335157315e-05,
+      "loss": 0.7669,
+      "step": 269
+    },
+    {
+      "epoch": 2.88,
+      "grad_norm": 0.3308991711135206,
+      "learning_rate": 3.649797336124615e-05,
+      "loss": 0.8041,
+      "step": 270
+    },
+    {
+      "epoch": 2.89,
+      "grad_norm": 0.32757727002633424,
+      "learning_rate": 3.583180171067101e-05,
+      "loss": 0.7673,
+      "step": 271
+    },
+    {
+      "epoch": 2.9,
+      "grad_norm": 0.3342551756453125,
+      "learning_rate": 3.517043627991441e-05,
+      "loss": 0.8005,
+      "step": 272
+    },
+    {
+      "epoch": 2.91,
+      "grad_norm": 0.31643754309861705,
+      "learning_rate": 3.45139266054715e-05,
+      "loss": 0.787,
+      "step": 273
+    },
+    {
+      "epoch": 2.93,
+      "grad_norm": 0.3140452683879005,
+      "learning_rate": 3.3862321860139576e-05,
+      "loss": 0.7888,
+      "step": 274
+    },
+    {
+      "epoch": 2.94,
+      "grad_norm": 0.30706221155036223,
+      "learning_rate": 3.3215670849335155e-05,
+      "loss": 0.827,
+      "step": 275
+    },
+    {
+      "epoch": 2.95,
+      "grad_norm": 0.3185483102727301,
+      "learning_rate": 3.257402200743821e-05,
+      "loss": 0.7779,
+      "step": 276
+    },
+    {
+      "epoch": 2.96,
+      "grad_norm": 0.3032818796307545,
+      "learning_rate": 3.19374233941647e-05,
+      "loss": 0.7993,
+      "step": 277
+    },
+    {
+      "epoch": 2.97,
+      "grad_norm": 0.3057758504695884,
+      "learning_rate": 3.130592269096671e-05,
+      "loss": 0.768,
+      "step": 278
+    },
+    {
+      "epoch": 2.98,
+      "grad_norm": 0.3245404038219604,
+      "learning_rate": 3.0679567197461134e-05,
+      "loss": 0.7706,
+      "step": 279
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 372,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 4,
+  "save_steps": 93,
+  "total_flos": 5.168039211319689e+18,
+  "train_batch_size": 4,
+  "trial_name": null,
+  "trial_params": null
+}

checkpoint-279/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c56515a18cd914d4eee44c09952d3a756ea623b0b6e69e8dfaeb0dbc7b665f46
+size 6776

checkpoint-279/zero_to_fp32.py

@@ -0,0 +1,592 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example: python zero_to_fp32.py . pytorch_model.bin
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+
+    total_files = len(files)
+    state_dicts = []
+    for f in files:
+        state_dict = torch.load(f, map_location=device)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    if zero_stage <= 2:
+        fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    elif zero_stage == 3:
+        # if there is more than one param group, there will be multiple flattened tensors - one
+        # flattened tensor per group - for simplicity merge them into a single tensor
+        #
+        # XXX: could make the script more memory efficient for when there are multiple groups - it
+        # will require matching the sub-lists of param_shapes for each param group flattened tensor
+
+        fp32_flat_groups = [
+            torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
+        ]
+
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    for name, shape in param_shapes.items():
+
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # XXX: memory usage doubles here
+        state_dict[name] = torch.cat(
+            tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
+            0).narrow(0, 0, unpartitioned_numel).view(shape)
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``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``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    Note: this approach may not work if your application doesn't have sufficient free CPU memory and
+    you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
+        - ``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``
+    """
+
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+    print(f"Saving fp32 state dict to {output_file}")
+    torch.save(state_dict, output_file)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``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``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument(
+        "output_file",
+        type=str,
+        help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file, tag=args.tag)

checkpoint-336/README.md

@@ -0,0 +1,202 @@
+---
+library_name: peft
+base_model: google/gemma-2b
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- 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. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+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).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.9.0

checkpoint-336/adapter_config.json

@@ -0,0 +1,33 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "google/gemma-2b",
+  "bias": "none",
+  "fan_in_fan_out": null,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 16,
+  "lora_dropout": 0.05,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "up_proj",
+    "q_proj",
+    "v_proj",
+    "down_proj",
+    "gate_proj",
+    "k_proj",
+    "o_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "use_dora": false,
+  "use_rslora": false
+}

checkpoint-336/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5ce5c9479f7b4e2f4f1c71ed29d0ec95f79e1731de4be9d3f7759abe3043fcdc
+size 78480320

checkpoint-336/global_step336/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5f204930ec4f2a105b656f8596b32abc5228db4def6b1aa8c6f63fe8c492820e
+size 58886928

checkpoint-336/global_step336/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bd427c55f17c0510ec2ed53fe5e319eb0a2c4761d4083df28d11ba7aa84e5a15
+size 58885968

checkpoint-336/global_step336/bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6271e5b9edc1d160ad0326ac1a89d8d44ef09363904f40271525aff81aa3b01d
+size 58886992

checkpoint-336/global_step336/bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5abf7c61e69335a8e881c7220e7017eb5372fdf817a3b0d26486e4faab795701
+size 58886032

checkpoint-336/global_step336/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8dbda4b13cb1e71570782ac3ce184727dbacb34070d7b08deeb937890375555c
+size 1159049922

checkpoint-336/latest

@@ -0,0 +1 @@
+global_step336

checkpoint-336/rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:28b9cac536dcc2f0fcb0db1a7ed44d898a5e257f0e6a2dde4782893acb56ce7d
+size 15024

checkpoint-336/rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d3ee31ce56c4f2248ab7aaf5beaf8d895447d28644df750b83cc2177262498de
+size 15024

checkpoint-336/rng_state_2.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a615e7b3e06287a0e82a15b753b1c48c658347992fbb7d59ee5836d824655ebd
+size 15024

checkpoint-336/rng_state_3.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fe5c5388f4cf688aa51717160bed97071e825a07ba7d9a22897241c258de91d9
+size 15024

checkpoint-336/scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:033700c231840b794630147afe6dca04265ec61bb681c241b2e3012bcb9cc8a3
+size 1064

checkpoint-336/trainer_state.json

@@ -0,0 +1,2477 @@
+{
+  "best_metric": 1.203959345817566,
+  "best_model_checkpoint": "./out/checkpoint-112",
+  "epoch": 2.991097922848665,
+  "eval_steps": 28,
+  "global_step": 336,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.01,
+      "grad_norm": 4.313233023002325,
+      "learning_rate": 1.8181818181818182e-05,
+      "loss": 1.9528,
+      "step": 1
+    },
+    {
+      "epoch": 0.01,
+      "eval_loss": 2.1875686645507812,
+      "eval_runtime": 12.8608,
+      "eval_samples_per_second": 23.327,
+      "eval_steps_per_second": 2.955,
+      "step": 1
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 4.039172290955229,
+      "learning_rate": 3.6363636363636364e-05,
+      "loss": 1.8358,
+      "step": 2
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 4.504705512003857,
+      "learning_rate": 5.4545454545454546e-05,
+      "loss": 2.0207,
+      "step": 3
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 4.591862504847867,
+      "learning_rate": 7.272727272727273e-05,
+      "loss": 1.979,
+      "step": 4
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 3.812893581399005,
+      "learning_rate": 9.090909090909092e-05,
+      "loss": 1.8356,
+      "step": 5
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 0.42886752872747064,
+      "learning_rate": 0.00010909090909090909,
+      "loss": 1.6722,
+      "step": 6
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 0.22497294481851865,
+      "learning_rate": 0.00012727272727272728,
+      "loss": 1.6711,
+      "step": 7
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 0.20955259847301927,
+      "learning_rate": 0.00014545454545454546,
+      "loss": 1.8546,
+      "step": 8
+    },
+    {
+      "epoch": 0.08,
+      "grad_norm": 0.2200095325539683,
+      "learning_rate": 0.00016363636363636366,
+      "loss": 1.7538,
+      "step": 9
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 0.19187339879899318,
+      "learning_rate": 0.00018181818181818183,
+      "loss": 1.6137,
+      "step": 10
+    },
+    {
+      "epoch": 0.1,
+      "grad_norm": 0.2113395673717837,
+      "learning_rate": 0.0002,
+      "loss": 1.5225,
+      "step": 11
+    },
+    {
+      "epoch": 0.11,
+      "grad_norm": 0.17673768408382828,
+      "learning_rate": 0.00019999741592564903,
+      "loss": 1.5303,
+      "step": 12
+    },
+    {
+      "epoch": 0.12,
+      "grad_norm": 0.24120852820548402,
+      "learning_rate": 0.00019998966383614488,
+      "loss": 1.5089,
+      "step": 13
+    },
+    {
+      "epoch": 0.12,
+      "grad_norm": 0.3089489160535682,
+      "learning_rate": 0.00019997674413212708,
+      "loss": 1.4525,
+      "step": 14
+    },
+    {
+      "epoch": 0.13,
+      "grad_norm": 0.2656143410731927,
+      "learning_rate": 0.00019995865748130516,
+      "loss": 1.4648,
+      "step": 15
+    },
+    {
+      "epoch": 0.14,
+      "grad_norm": 3.769410316227205,
+      "learning_rate": 0.0001999354048184241,
+      "loss": 1.3439,
+      "step": 16
+    },
+    {
+      "epoch": 0.15,
+      "grad_norm": 0.32102180658823753,
+      "learning_rate": 0.00019990698734521613,
+      "loss": 1.4644,
+      "step": 17
+    },
+    {
+      "epoch": 0.16,
+      "grad_norm": 0.22094428128919438,
+      "learning_rate": 0.0001998734065303385,
+      "loss": 1.1927,
+      "step": 18
+    },
+    {
+      "epoch": 0.17,
+      "grad_norm": 0.22344487218098863,
+      "learning_rate": 0.00019983466410929764,
+      "loss": 1.2916,
+      "step": 19
+    },
+    {
+      "epoch": 0.18,
+      "grad_norm": 0.25036262498479456,
+      "learning_rate": 0.0001997907620843595,
+      "loss": 1.2982,
+      "step": 20
+    },
+    {
+      "epoch": 0.19,
+      "grad_norm": 0.22671119151539426,
+      "learning_rate": 0.00019974170272444604,
+      "loss": 1.2146,
+      "step": 21
+    },
+    {
+      "epoch": 0.2,
+      "grad_norm": 0.259249080403425,
+      "learning_rate": 0.00019968748856501788,
+      "loss": 1.2072,
+      "step": 22
+    },
+    {
+      "epoch": 0.2,
+      "grad_norm": 0.23538477651406017,
+      "learning_rate": 0.00019962812240794343,
+      "loss": 1.3281,
+      "step": 23
+    },
+    {
+      "epoch": 0.21,
+      "grad_norm": 0.2659115087625978,
+      "learning_rate": 0.000199563607321354,
+      "loss": 1.1396,
+      "step": 24
+    },
+    {
+      "epoch": 0.22,
+      "grad_norm": 0.23617264858854836,
+      "learning_rate": 0.0001994939466394851,
+      "loss": 1.1389,
+      "step": 25
+    },
+    {
+      "epoch": 0.23,
+      "grad_norm": 0.20514227454180176,
+      "learning_rate": 0.00019941914396250446,
+      "loss": 1.249,
+      "step": 26
+    },
+    {
+      "epoch": 0.24,
+      "grad_norm": 0.19660894225830144,
+      "learning_rate": 0.00019933920315632557,
+      "loss": 1.1776,
+      "step": 27
+    },
+    {
+      "epoch": 0.25,
+      "grad_norm": 0.2067663909729571,
+      "learning_rate": 0.00019925412835240826,
+      "loss": 1.1327,
+      "step": 28
+    },
+    {
+      "epoch": 0.25,
+      "eval_loss": 1.2991960048675537,
+      "eval_runtime": 13.153,
+      "eval_samples_per_second": 22.808,
+      "eval_steps_per_second": 2.889,
+      "step": 28
+    },
+    {
+      "epoch": 0.26,
+      "grad_norm": 0.1816588361901526,
+      "learning_rate": 0.0001991639239475448,
+      "loss": 1.1247,
+      "step": 29
+    },
+    {
+      "epoch": 0.27,
+      "grad_norm": 0.19626955153633807,
+      "learning_rate": 0.00019906859460363307,
+      "loss": 1.1212,
+      "step": 30
+    },
+    {
+      "epoch": 0.28,
+      "grad_norm": 0.21084275590405852,
+      "learning_rate": 0.00019896814524743528,
+      "loss": 0.9927,
+      "step": 31
+    },
+    {
+      "epoch": 0.28,
+      "grad_norm": 0.16560054949456768,
+      "learning_rate": 0.0001988625810703235,
+      "loss": 1.1249,
+      "step": 32
+    },
+    {
+      "epoch": 0.29,
+      "grad_norm": 0.14950879528294536,
+      "learning_rate": 0.0001987519075280114,
+      "loss": 1.1401,
+      "step": 33
+    },
+    {
+      "epoch": 0.3,
+      "grad_norm": 0.1777966882651237,
+      "learning_rate": 0.00019863613034027224,
+      "loss": 1.0769,
+      "step": 34
+    },
+    {
+      "epoch": 0.31,
+      "grad_norm": 0.1480537272052743,
+      "learning_rate": 0.00019851525549064323,
+      "loss": 1.0686,
+      "step": 35
+    },
+    {
+      "epoch": 0.32,
+      "grad_norm": 0.16911906750319078,
+      "learning_rate": 0.00019838928922611632,
+      "loss": 1.0253,
+      "step": 36
+    },
+    {
+      "epoch": 0.33,
+      "grad_norm": 0.15987682972555176,
+      "learning_rate": 0.00019825823805681543,
+      "loss": 1.0609,
+      "step": 37
+    },
+    {
+      "epoch": 0.34,
+      "grad_norm": 0.15757332939676763,
+      "learning_rate": 0.0001981221087556598,
+      "loss": 1.1086,
+      "step": 38
+    },
+    {
+      "epoch": 0.35,
+      "grad_norm": 0.13201845744757537,
+      "learning_rate": 0.00019798090835801418,
+      "loss": 1.073,
+      "step": 39
+    },
+    {
+      "epoch": 0.36,
+      "grad_norm": 0.12544508015984754,
+      "learning_rate": 0.00019783464416132506,
+      "loss": 1.0633,
+      "step": 40
+    },
+    {
+      "epoch": 0.36,
+      "grad_norm": 0.14645820383886451,
+      "learning_rate": 0.00019768332372474366,
+      "loss": 1.0653,
+      "step": 41
+    },
+    {
+      "epoch": 0.37,
+      "grad_norm": 0.14814101902137117,
+      "learning_rate": 0.00019752695486873517,
+      "loss": 1.0937,
+      "step": 42
+    },
+    {
+      "epoch": 0.38,
+      "grad_norm": 0.13888915595055443,
+      "learning_rate": 0.00019736554567467452,
+      "loss": 1.0462,
+      "step": 43
+    },
+    {
+      "epoch": 0.39,
+      "grad_norm": 0.13185349806639524,
+      "learning_rate": 0.00019719910448442893,
+      "loss": 1.2177,
+      "step": 44
+    },
+    {
+      "epoch": 0.4,
+      "grad_norm": 0.15271046712350847,
+      "learning_rate": 0.00019702763989992662,
+      "loss": 1.0237,
+      "step": 45
+    },
+    {
+      "epoch": 0.41,
+      "grad_norm": 0.17053588557430902,
+      "learning_rate": 0.00019685116078271223,
+      "loss": 1.0038,
+      "step": 46
+    },
+    {
+      "epoch": 0.42,
+      "grad_norm": 0.15641087356577812,
+      "learning_rate": 0.00019666967625348906,
+      "loss": 1.0886,
+      "step": 47
+    },
+    {
+      "epoch": 0.43,
+      "grad_norm": 0.1544028594191567,
+      "learning_rate": 0.00019648319569164736,
+      "loss": 1.1378,
+      "step": 48
+    },
+    {
+      "epoch": 0.44,
+      "grad_norm": 0.14794885994140625,
+      "learning_rate": 0.00019629172873477995,
+      "loss": 1.1495,
+      "step": 49
+    },
+    {
+      "epoch": 0.45,
+      "grad_norm": 0.1577684884028266,
+      "learning_rate": 0.0001960952852781838,
+      "loss": 1.0782,
+      "step": 50
+    },
+    {
+      "epoch": 0.45,
+      "grad_norm": 0.15961044045091288,
+      "learning_rate": 0.0001958938754743489,
+      "loss": 1.0107,
+      "step": 51
+    },
+    {
+      "epoch": 0.46,
+      "grad_norm": 0.14486696586022083,
+      "learning_rate": 0.0001956875097324334,
+      "loss": 1.0494,
+      "step": 52
+    },
+    {
+      "epoch": 0.47,
+      "grad_norm": 0.14250413725518896,
+      "learning_rate": 0.00019547619871772574,
+      "loss": 1.039,
+      "step": 53
+    },
+    {
+      "epoch": 0.48,
+      "grad_norm": 0.1196720279125328,
+      "learning_rate": 0.00019525995335109334,
+      "loss": 1.0966,
+      "step": 54
+    },
+    {
+      "epoch": 0.49,
+      "grad_norm": 0.14984795891635327,
+      "learning_rate": 0.0001950387848084183,
+      "loss": 1.0874,
+      "step": 55
+    },
+    {
+      "epoch": 0.5,
+      "grad_norm": 0.14891088442480416,
+      "learning_rate": 0.00019481270452001987,
+      "loss": 1.097,
+      "step": 56
+    },
+    {
+      "epoch": 0.5,
+      "eval_loss": 1.2264304161071777,
+      "eval_runtime": 13.2279,
+      "eval_samples_per_second": 22.679,
+      "eval_steps_per_second": 2.873,
+      "step": 56
+    },
+    {
+      "epoch": 0.51,
+      "grad_norm": 0.17814266552244534,
+      "learning_rate": 0.00019458172417006347,
+      "loss": 1.1372,
+      "step": 57
+    },
+    {
+      "epoch": 0.52,
+      "grad_norm": 0.16125636132578247,
+      "learning_rate": 0.00019434585569595708,
+      "loss": 1.0623,
+      "step": 58
+    },
+    {
+      "epoch": 0.53,
+      "grad_norm": 0.15203437202125702,
+      "learning_rate": 0.00019410511128773418,
+      "loss": 1.0399,
+      "step": 59
+    },
+    {
+      "epoch": 0.53,
+      "grad_norm": 0.1677461135605213,
+      "learning_rate": 0.0001938595033874238,
+      "loss": 1.0884,
+      "step": 60
+    },
+    {
+      "epoch": 0.54,
+      "grad_norm": 0.13564559875683407,
+      "learning_rate": 0.0001936090446884074,
+      "loss": 1.0176,
+      "step": 61
+    },
+    {
+      "epoch": 0.55,
+      "grad_norm": 0.1521886500642157,
+      "learning_rate": 0.00019335374813476302,
+      "loss": 1.0146,
+      "step": 62
+    },
+    {
+      "epoch": 0.56,
+      "grad_norm": 0.1410132122625916,
+      "learning_rate": 0.00019309362692059617,
+      "loss": 1.044,
+      "step": 63
+    },
+    {
+      "epoch": 0.57,
+      "grad_norm": 0.15237848179385577,
+      "learning_rate": 0.00019282869448935798,
+      "loss": 1.0354,
+      "step": 64
+    },
+    {
+      "epoch": 0.58,
+      "grad_norm": 0.13871660988504514,
+      "learning_rate": 0.00019255896453315052,
+      "loss": 1.0189,
+      "step": 65
+    },
+    {
+      "epoch": 0.59,
+      "grad_norm": 0.14863047478901453,
+      "learning_rate": 0.000192284450992019,
+      "loss": 1.0704,
+      "step": 66
+    },
+    {
+      "epoch": 0.6,
+      "grad_norm": 0.13794806124403974,
+      "learning_rate": 0.0001920051680532314,
+      "loss": 1.0996,
+      "step": 67
+    },
+    {
+      "epoch": 0.61,
+      "grad_norm": 0.13030507705779365,
+      "learning_rate": 0.00019172113015054532,
+      "loss": 1.0015,
+      "step": 68
+    },
+    {
+      "epoch": 0.61,
+      "grad_norm": 0.15092494718902358,
+      "learning_rate": 0.0001914323519634619,
+      "loss": 1.0822,
+      "step": 69
+    },
+    {
+      "epoch": 0.62,
+      "grad_norm": 0.1350212989006066,
+      "learning_rate": 0.00019113884841646736,
+      "loss": 1.0197,
+      "step": 70
+    },
+    {
+      "epoch": 0.63,
+      "grad_norm": 0.18991168066586347,
+      "learning_rate": 0.00019084063467826137,
+      "loss": 1.046,
+      "step": 71
+    },
+    {
+      "epoch": 0.64,
+      "grad_norm": 0.14884381774710187,
+      "learning_rate": 0.00019053772616097337,
+      "loss": 1.0346,
+      "step": 72
+    },
+    {
+      "epoch": 0.65,
+      "grad_norm": 0.15579311209945296,
+      "learning_rate": 0.000190230138519366,
+      "loss": 1.0505,
+      "step": 73
+    },
+    {
+      "epoch": 0.66,
+      "grad_norm": 0.16015337150592127,
+      "learning_rate": 0.000189917887650026,
+      "loss": 1.0504,
+      "step": 74
+    },
+    {
+      "epoch": 0.67,
+      "grad_norm": 0.1443969321518926,
+      "learning_rate": 0.00018960098969054255,
+      "loss": 1.0755,
+      "step": 75
+    },
+    {
+      "epoch": 0.68,
+      "grad_norm": 0.15722162227095848,
+      "learning_rate": 0.00018927946101867347,
+      "loss": 1.0541,
+      "step": 76
+    },
+    {
+      "epoch": 0.69,
+      "grad_norm": 0.17009697584926559,
+      "learning_rate": 0.0001889533182514986,
+      "loss": 1.0231,
+      "step": 77
+    },
+    {
+      "epoch": 0.69,
+      "grad_norm": 0.1256822726781221,
+      "learning_rate": 0.0001886225782445612,
+      "loss": 0.8814,
+      "step": 78
+    },
+    {
+      "epoch": 0.7,
+      "grad_norm": 0.14019958069756655,
+      "learning_rate": 0.00018828725809099655,
+      "loss": 1.0277,
+      "step": 79
+    },
+    {
+      "epoch": 0.71,
+      "grad_norm": 0.17159459150063183,
+      "learning_rate": 0.0001879473751206489,
+      "loss": 0.9495,
+      "step": 80
+    },
+    {
+      "epoch": 0.72,
+      "grad_norm": 0.146430011834186,
+      "learning_rate": 0.00018760294689917553,
+      "loss": 1.0598,
+      "step": 81
+    },
+    {
+      "epoch": 0.73,
+      "grad_norm": 0.16834256802992476,
+      "learning_rate": 0.00018725399122713912,
+      "loss": 1.0237,
+      "step": 82
+    },
+    {
+      "epoch": 0.74,
+      "grad_norm": 0.15663699267164208,
+      "learning_rate": 0.00018690052613908772,
+      "loss": 0.939,
+      "step": 83
+    },
+    {
+      "epoch": 0.75,
+      "grad_norm": 0.15655985150409854,
+      "learning_rate": 0.0001865425699026226,
+      "loss": 1.0302,
+      "step": 84
+    },
+    {
+      "epoch": 0.75,
+      "eval_loss": 1.2143030166625977,
+      "eval_runtime": 13.2387,
+      "eval_samples_per_second": 22.661,
+      "eval_steps_per_second": 2.87,
+      "step": 84
+    },
+    {
+      "epoch": 0.76,
+      "grad_norm": 0.15273470110260864,
+      "learning_rate": 0.00018618014101745442,
+      "loss": 1.0127,
+      "step": 85
+    },
+    {
+      "epoch": 0.77,
+      "grad_norm": 0.1723243680259614,
+      "learning_rate": 0.0001858132582144469,
+      "loss": 0.9306,
+      "step": 86
+    },
+    {
+      "epoch": 0.77,
+      "grad_norm": 0.14747098547446996,
+      "learning_rate": 0.00018544194045464886,
+      "loss": 1.0073,
+      "step": 87
+    },
+    {
+      "epoch": 0.78,
+      "grad_norm": 0.17208333285514918,
+      "learning_rate": 0.00018506620692831428,
+      "loss": 1.0328,
+      "step": 88
+    },
+    {
+      "epoch": 0.79,
+      "grad_norm": 0.14918051024971962,
+      "learning_rate": 0.0001846860770539105,
+      "loss": 1.0022,
+      "step": 89
+    },
+    {
+      "epoch": 0.8,
+      "grad_norm": 0.156315164090714,
+      "learning_rate": 0.00018430157047711474,
+      "loss": 1.0293,
+      "step": 90
+    },
+    {
+      "epoch": 0.81,
+      "grad_norm": 0.2013424548288477,
+      "learning_rate": 0.00018391270706979862,
+      "loss": 0.9395,
+      "step": 91
+    },
+    {
+      "epoch": 0.82,
+      "grad_norm": 0.17909726353002614,
+      "learning_rate": 0.00018351950692900126,
+      "loss": 0.9756,
+      "step": 92
+    },
+    {
+      "epoch": 0.83,
+      "grad_norm": 0.16939245158726288,
+      "learning_rate": 0.00018312199037589068,
+      "loss": 0.9576,
+      "step": 93
+    },
+    {
+      "epoch": 0.84,
+      "grad_norm": 0.14685720680893694,
+      "learning_rate": 0.00018272017795471345,
+      "loss": 1.0045,
+      "step": 94
+    },
+    {
+      "epoch": 0.85,
+      "grad_norm": 0.17464839085505987,
+      "learning_rate": 0.000182314090431733,
+      "loss": 0.9862,
+      "step": 95
+    },
+    {
+      "epoch": 0.85,
+      "grad_norm": 0.16060904136932572,
+      "learning_rate": 0.00018190374879415632,
+      "loss": 1.0022,
+      "step": 96
+    },
+    {
+      "epoch": 0.86,
+      "grad_norm": 0.18715193350083867,
+      "learning_rate": 0.00018148917424904953,
+      "loss": 1.042,
+      "step": 97
+    },
+    {
+      "epoch": 0.87,
+      "grad_norm": 0.1675573400576595,
+      "learning_rate": 0.0001810703882222415,
+      "loss": 1.0047,
+      "step": 98
+    },
+    {
+      "epoch": 0.88,
+      "grad_norm": 0.1871466286989249,
+      "learning_rate": 0.00018064741235721687,
+      "loss": 0.9834,
+      "step": 99
+    },
+    {
+      "epoch": 0.89,
+      "grad_norm": 0.17453934867565302,
+      "learning_rate": 0.00018022026851399737,
+      "loss": 0.9649,
+      "step": 100
+    },
+    {
+      "epoch": 0.9,
+      "grad_norm": 0.15960631507184767,
+      "learning_rate": 0.0001797889787680119,
+      "loss": 0.9673,
+      "step": 101
+    },
+    {
+      "epoch": 0.91,
+      "grad_norm": 0.17844936635366368,
+      "learning_rate": 0.00017935356540895597,
+      "loss": 1.0951,
+      "step": 102
+    },
+    {
+      "epoch": 0.92,
+      "grad_norm": 0.16733018789000254,
+      "learning_rate": 0.00017891405093963938,
+      "loss": 0.9954,
+      "step": 103
+    },
+    {
+      "epoch": 0.93,
+      "grad_norm": 0.17305556075296993,
+      "learning_rate": 0.00017847045807482345,
+      "loss": 0.892,
+      "step": 104
+    },
+    {
+      "epoch": 0.93,
+      "grad_norm": 0.17197614099805034,
+      "learning_rate": 0.00017802280974004716,
+      "loss": 1.0494,
+      "step": 105
+    },
+    {
+      "epoch": 0.94,
+      "grad_norm": 0.18063836817127235,
+      "learning_rate": 0.000177571129070442,
+      "loss": 1.0264,
+      "step": 106
+    },
+    {
+      "epoch": 0.95,
+      "grad_norm": 0.14597707005699143,
+      "learning_rate": 0.00017711543940953668,
+      "loss": 0.9532,
+      "step": 107
+    },
+    {
+      "epoch": 0.96,
+      "grad_norm": 0.1422048149465345,
+      "learning_rate": 0.00017665576430805053,
+      "loss": 0.97,
+      "step": 108
+    },
+    {
+      "epoch": 0.97,
+      "grad_norm": 0.18313914688655572,
+      "learning_rate": 0.0001761921275226763,
+      "loss": 0.9282,
+      "step": 109
+    },
+    {
+      "epoch": 0.98,
+      "grad_norm": 0.200679751171441,
+      "learning_rate": 0.00017572455301485249,
+      "loss": 1.0,
+      "step": 110
+    },
+    {
+      "epoch": 0.99,
+      "grad_norm": 0.17700985594898055,
+      "learning_rate": 0.00017525306494952498,
+      "loss": 1.0165,
+      "step": 111
+    },
+    {
+      "epoch": 1.0,
+      "grad_norm": 0.19925777202726191,
+      "learning_rate": 0.0001747776876938981,
+      "loss": 1.0346,
+      "step": 112
+    },
+    {
+      "epoch": 1.0,
+      "eval_loss": 1.203959345817566,
+      "eval_runtime": 13.2547,
+      "eval_samples_per_second": 22.634,
+      "eval_steps_per_second": 2.867,
+      "step": 112
+    },
+    {
+      "epoch": 1.01,
+      "grad_norm": 0.1606469603473709,
+      "learning_rate": 0.00017429844581617532,
+      "loss": 0.9832,
+      "step": 113
+    },
+    {
+      "epoch": 1.01,
+      "grad_norm": 0.16403912763780054,
+      "learning_rate": 0.00017381536408428948,
+      "loss": 0.9346,
+      "step": 114
+    },
+    {
+      "epoch": 1.02,
+      "grad_norm": 0.1936046893744468,
+      "learning_rate": 0.00017332846746462288,
+      "loss": 0.9382,
+      "step": 115
+    },
+    {
+      "epoch": 1.03,
+      "grad_norm": 0.14250769247239573,
+      "learning_rate": 0.0001728377811207168,
+      "loss": 0.8914,
+      "step": 116
+    },
+    {
+      "epoch": 1.04,
+      "grad_norm": 0.17889563599797687,
+      "learning_rate": 0.00017234333041197126,
+      "loss": 0.9736,
+      "step": 117
+    },
+    {
+      "epoch": 1.05,
+      "grad_norm": 0.20288960866045594,
+      "learning_rate": 0.00017184514089233405,
+      "loss": 0.8477,
+      "step": 118
+    },
+    {
+      "epoch": 1.06,
+      "grad_norm": 0.20926349930533472,
+      "learning_rate": 0.00017134323830898037,
+      "loss": 0.9933,
+      "step": 119
+    },
+    {
+      "epoch": 1.07,
+      "grad_norm": 0.21316934416499642,
+      "learning_rate": 0.00017083764860098205,
+      "loss": 0.9168,
+      "step": 120
+    },
+    {
+      "epoch": 1.08,
+      "grad_norm": 0.21654320387312692,
+      "learning_rate": 0.0001703283978979671,
+      "loss": 0.9584,
+      "step": 121
+    },
+    {
+      "epoch": 1.09,
+      "grad_norm": 0.23789742308175463,
+      "learning_rate": 0.00016981551251876904,
+      "loss": 1.0298,
+      "step": 122
+    },
+    {
+      "epoch": 1.09,
+      "grad_norm": 0.16433271793469648,
+      "learning_rate": 0.00016929901897006698,
+      "loss": 0.8833,
+      "step": 123
+    },
+    {
+      "epoch": 1.1,
+      "grad_norm": 0.16908727866207868,
+      "learning_rate": 0.0001687789439450156,
+      "loss": 1.0675,
+      "step": 124
+    },
+    {
+      "epoch": 1.11,
+      "grad_norm": 0.1670067931363302,
+      "learning_rate": 0.00016825531432186543,
+      "loss": 0.9515,
+      "step": 125
+    },
+    {
+      "epoch": 1.12,
+      "grad_norm": 0.17777465531550865,
+      "learning_rate": 0.00016772815716257412,
+      "loss": 0.8929,
+      "step": 126
+    },
+    {
+      "epoch": 1.13,
+      "grad_norm": 0.18442783204919333,
+      "learning_rate": 0.00016719749971140754,
+      "loss": 0.8388,
+      "step": 127
+    },
+    {
+      "epoch": 1.14,
+      "grad_norm": 0.19073362304284272,
+      "learning_rate": 0.0001666633693935319,
+      "loss": 0.9584,
+      "step": 128
+    },
+    {
+      "epoch": 1.15,
+      "grad_norm": 0.20189563405135308,
+      "learning_rate": 0.00016612579381359622,
+      "loss": 1.0264,
+      "step": 129
+    },
+    {
+      "epoch": 1.16,
+      "grad_norm": 0.1694138210313381,
+      "learning_rate": 0.00016558480075430594,
+      "loss": 0.9592,
+      "step": 130
+    },
+    {
+      "epoch": 1.17,
+      "grad_norm": 0.19195382946787184,
+      "learning_rate": 0.00016504041817498678,
+      "loss": 0.974,
+      "step": 131
+    },
+    {
+      "epoch": 1.18,
+      "grad_norm": 0.20684215619155688,
+      "learning_rate": 0.00016449267421013994,
+      "loss": 0.8499,
+      "step": 132
+    },
+    {
+      "epoch": 1.18,
+      "grad_norm": 0.22003490429847744,
+      "learning_rate": 0.00016394159716798807,
+      "loss": 0.9659,
+      "step": 133
+    },
+    {
+      "epoch": 1.19,
+      "grad_norm": 0.21977918206745437,
+      "learning_rate": 0.00016338721552901212,
+      "loss": 0.9213,
+      "step": 134
+    },
+    {
+      "epoch": 1.2,
+      "grad_norm": 0.2076993903333204,
+      "learning_rate": 0.0001628295579444796,
+      "loss": 0.8119,
+      "step": 135
+    },
+    {
+      "epoch": 1.21,
+      "grad_norm": 0.2001771499954729,
+      "learning_rate": 0.0001622686532349637,
+      "loss": 0.9183,
+      "step": 136
+    },
+    {
+      "epoch": 1.22,
+      "grad_norm": 0.18671550149366203,
+      "learning_rate": 0.00016170453038885394,
+      "loss": 0.8836,
+      "step": 137
+    },
+    {
+      "epoch": 1.23,
+      "grad_norm": 0.20867427207572573,
+      "learning_rate": 0.0001611372185608578,
+      "loss": 0.9964,
+      "step": 138
+    },
+    {
+      "epoch": 1.24,
+      "grad_norm": 0.20035138443113176,
+      "learning_rate": 0.0001605667470704942,
+      "loss": 0.9209,
+      "step": 139
+    },
+    {
+      "epoch": 1.25,
+      "grad_norm": 0.22696612020505577,
+      "learning_rate": 0.0001599931454005781,
+      "loss": 1.0162,
+      "step": 140
+    },
+    {
+      "epoch": 1.25,
+      "eval_loss": 1.2188584804534912,
+      "eval_runtime": 13.249,
+      "eval_samples_per_second": 22.643,
+      "eval_steps_per_second": 2.868,
+      "step": 140
+    },
+    {
+      "epoch": 1.26,
+      "grad_norm": 0.21554353495018647,
+      "learning_rate": 0.00015941644319569665,
+      "loss": 1.0487,
+      "step": 141
+    },
+    {
+      "epoch": 1.26,
+      "grad_norm": 0.22894492131909072,
+      "learning_rate": 0.00015883667026067745,
+      "loss": 0.9352,
+      "step": 142
+    },
+    {
+      "epoch": 1.27,
+      "grad_norm": 0.19145184577172686,
+      "learning_rate": 0.00015825385655904788,
+      "loss": 0.8878,
+      "step": 143
+    },
+    {
+      "epoch": 1.28,
+      "grad_norm": 0.22544664152936575,
+      "learning_rate": 0.00015766803221148673,
+      "loss": 1.0,
+      "step": 144
+    },
+    {
+      "epoch": 1.29,
+      "grad_norm": 0.26000661355557114,
+      "learning_rate": 0.00015707922749426737,
+      "loss": 0.9339,
+      "step": 145
+    },
+    {
+      "epoch": 1.3,
+      "grad_norm": 0.24433845134512236,
+      "learning_rate": 0.00015648747283769317,
+      "loss": 0.9474,
+      "step": 146
+    },
+    {
+      "epoch": 1.31,
+      "grad_norm": 0.21973931169609887,
+      "learning_rate": 0.00015589279882452476,
+      "loss": 0.9357,
+      "step": 147
+    },
+    {
+      "epoch": 1.32,
+      "grad_norm": 0.23929008733305812,
+      "learning_rate": 0.0001552952361883994,
+      "loss": 0.9985,
+      "step": 148
+    },
+    {
+      "epoch": 1.33,
+      "grad_norm": 0.23431856747573573,
+      "learning_rate": 0.00015469481581224272,
+      "loss": 0.8913,
+      "step": 149
+    },
+    {
+      "epoch": 1.34,
+      "grad_norm": 0.2233543327912565,
+      "learning_rate": 0.00015409156872667258,
+      "loss": 0.9877,
+      "step": 150
+    },
+    {
+      "epoch": 1.34,
+      "grad_norm": 0.21281207674183256,
+      "learning_rate": 0.0001534855261083954,
+      "loss": 0.9071,
+      "step": 151
+    },
+    {
+      "epoch": 1.35,
+      "grad_norm": 0.20314832700152685,
+      "learning_rate": 0.00015287671927859494,
+      "loss": 0.9373,
+      "step": 152
+    },
+    {
+      "epoch": 1.36,
+      "grad_norm": 0.19648565819019825,
+      "learning_rate": 0.00015226517970131343,
+      "loss": 0.9469,
+      "step": 153
+    },
+    {
+      "epoch": 1.37,
+      "grad_norm": 0.2262428264639853,
+      "learning_rate": 0.00015165093898182562,
+      "loss": 1.0066,
+      "step": 154
+    },
+    {
+      "epoch": 1.38,
+      "grad_norm": 0.22253433035020442,
+      "learning_rate": 0.00015103402886500525,
+      "loss": 0.8875,
+      "step": 155
+    },
+    {
+      "epoch": 1.39,
+      "grad_norm": 0.181161648904613,
+      "learning_rate": 0.00015041448123368455,
+      "loss": 0.9004,
+      "step": 156
+    },
+    {
+      "epoch": 1.4,
+      "grad_norm": 0.20968483802367816,
+      "learning_rate": 0.00014979232810700637,
+      "loss": 0.9133,
+      "step": 157
+    },
+    {
+      "epoch": 1.41,
+      "grad_norm": 0.20540509271288435,
+      "learning_rate": 0.0001491676016387694,
+      "loss": 0.8876,
+      "step": 158
+    },
+    {
+      "epoch": 1.42,
+      "grad_norm": 0.18762795731312454,
+      "learning_rate": 0.00014854033411576659,
+      "loss": 0.933,
+      "step": 159
+    },
+    {
+      "epoch": 1.42,
+      "grad_norm": 0.23223345997338857,
+      "learning_rate": 0.00014791055795611624,
+      "loss": 0.9182,
+      "step": 160
+    },
+    {
+      "epoch": 1.43,
+      "grad_norm": 0.21932384461027146,
+      "learning_rate": 0.00014727830570758678,
+      "loss": 0.9514,
+      "step": 161
+    },
+    {
+      "epoch": 1.44,
+      "grad_norm": 0.21819663730951108,
+      "learning_rate": 0.0001466436100459146,
+      "loss": 0.9162,
+      "step": 162
+    },
+    {
+      "epoch": 1.45,
+      "grad_norm": 0.2325813323476676,
+      "learning_rate": 0.00014600650377311522,
+      "loss": 0.9308,
+      "step": 163
+    },
+    {
+      "epoch": 1.46,
+      "grad_norm": 0.2568337182939043,
+      "learning_rate": 0.0001453670198157883,
+      "loss": 0.9995,
+      "step": 164
+    },
+    {
+      "epoch": 1.47,
+      "grad_norm": 0.22578454460723413,
+      "learning_rate": 0.00014472519122341566,
+      "loss": 0.9052,
+      "step": 165
+    },
+    {
+      "epoch": 1.48,
+      "grad_norm": 0.23564258958796755,
+      "learning_rate": 0.00014408105116665336,
+      "loss": 0.9714,
+      "step": 166
+    },
+    {
+      "epoch": 1.49,
+      "grad_norm": 0.24266133562839415,
+      "learning_rate": 0.00014343463293561734,
+      "loss": 0.9219,
+      "step": 167
+    },
+    {
+      "epoch": 1.5,
+      "grad_norm": 0.23472454708184465,
+      "learning_rate": 0.00014278596993816308,
+      "loss": 0.8762,
+      "step": 168
+    },
+    {
+      "epoch": 1.5,
+      "eval_loss": 1.2197421789169312,
+      "eval_runtime": 13.2616,
+      "eval_samples_per_second": 22.622,
+      "eval_steps_per_second": 2.865,
+      "step": 168
+    },
+    {
+      "epoch": 1.5,
+      "grad_norm": 0.23623633375452713,
+      "learning_rate": 0.00014213509569815884,
+      "loss": 0.8809,
+      "step": 169
+    },
+    {
+      "epoch": 1.51,
+      "grad_norm": 0.25344275204523486,
+      "learning_rate": 0.00014148204385375321,
+      "loss": 0.7972,
+      "step": 170
+    },
+    {
+      "epoch": 1.52,
+      "grad_norm": 0.23111396119549557,
+      "learning_rate": 0.0001408268481556366,
+      "loss": 0.8228,
+      "step": 171
+    },
+    {
+      "epoch": 1.53,
+      "grad_norm": 0.2510618369255398,
+      "learning_rate": 0.00014016954246529696,
+      "loss": 0.8849,
+      "step": 172
+    },
+    {
+      "epoch": 1.54,
+      "grad_norm": 0.2764366116622668,
+      "learning_rate": 0.0001395101607532698,
+      "loss": 0.8936,
+      "step": 173
+    },
+    {
+      "epoch": 1.55,
+      "grad_norm": 0.24325811719582827,
+      "learning_rate": 0.00013884873709738257,
+      "loss": 0.8602,
+      "step": 174
+    },
+    {
+      "epoch": 1.56,
+      "grad_norm": 0.213781513838486,
+      "learning_rate": 0.00013818530568099327,
+      "loss": 0.9492,
+      "step": 175
+    },
+    {
+      "epoch": 1.57,
+      "grad_norm": 0.2397396374239057,
+      "learning_rate": 0.00013751990079122412,
+      "loss": 1.0499,
+      "step": 176
+    },
+    {
+      "epoch": 1.58,
+      "grad_norm": 0.21579907170368723,
+      "learning_rate": 0.00013685255681718922,
+      "loss": 0.9438,
+      "step": 177
+    },
+    {
+      "epoch": 1.58,
+      "grad_norm": 0.2359312681928786,
+      "learning_rate": 0.0001361833082482175,
+      "loss": 0.9289,
+      "step": 178
+    },
+    {
+      "epoch": 1.59,
+      "grad_norm": 0.2618189093396496,
+      "learning_rate": 0.0001355121896720703,
+      "loss": 0.981,
+      "step": 179
+    },
+    {
+      "epoch": 1.6,
+      "grad_norm": 0.20876513773174135,
+      "learning_rate": 0.00013483923577315348,
+      "loss": 0.82,
+      "step": 180
+    },
+    {
+      "epoch": 1.61,
+      "grad_norm": 0.22162748553995645,
+      "learning_rate": 0.00013416448133072526,
+      "loss": 1.0131,
+      "step": 181
+    },
+    {
+      "epoch": 1.62,
+      "grad_norm": 0.20975549982451164,
+      "learning_rate": 0.00013348796121709862,
+      "loss": 0.8763,
+      "step": 182
+    },
+    {
+      "epoch": 1.63,
+      "grad_norm": 0.22840397707525473,
+      "learning_rate": 0.00013280971039583906,
+      "loss": 0.949,
+      "step": 183
+    },
+    {
+      "epoch": 1.64,
+      "grad_norm": 0.23384636230161737,
+      "learning_rate": 0.0001321297639199575,
+      "loss": 0.9567,
+      "step": 184
+    },
+    {
+      "epoch": 1.65,
+      "grad_norm": 0.22905979409902957,
+      "learning_rate": 0.000131448156930099,
+      "loss": 0.9153,
+      "step": 185
+    },
+    {
+      "epoch": 1.66,
+      "grad_norm": 0.27620894683694563,
+      "learning_rate": 0.0001307649246527263,
+      "loss": 0.8246,
+      "step": 186
+    },
+    {
+      "epoch": 1.66,
+      "grad_norm": 0.23004170633106227,
+      "learning_rate": 0.0001300801023982995,
+      "loss": 1.0181,
+      "step": 187
+    },
+    {
+      "epoch": 1.67,
+      "grad_norm": 0.2219849136264378,
+      "learning_rate": 0.00012939372555945112,
+      "loss": 0.9535,
+      "step": 188
+    },
+    {
+      "epoch": 1.68,
+      "grad_norm": 0.24458750452490116,
+      "learning_rate": 0.0001287058296091567,
+      "loss": 0.8968,
+      "step": 189
+    },
+    {
+      "epoch": 1.69,
+      "grad_norm": 0.2564337740159555,
+      "learning_rate": 0.00012801645009890195,
+      "loss": 0.7955,
+      "step": 190
+    },
+    {
+      "epoch": 1.7,
+      "grad_norm": 0.24100850371438767,
+      "learning_rate": 0.0001273256226568451,
+      "loss": 0.9235,
+      "step": 191
+    },
+    {
+      "epoch": 1.71,
+      "grad_norm": 0.24757089527873732,
+      "learning_rate": 0.00012663338298597563,
+      "loss": 1.007,
+      "step": 192
+    },
+    {
+      "epoch": 1.72,
+      "grad_norm": 0.24701038583742888,
+      "learning_rate": 0.00012593976686226904,
+      "loss": 0.9885,
+      "step": 193
+    },
+    {
+      "epoch": 1.73,
+      "grad_norm": 0.26373721125634964,
+      "learning_rate": 0.0001252448101328381,
+      "loss": 0.8785,
+      "step": 194
+    },
+    {
+      "epoch": 1.74,
+      "grad_norm": 0.2227761464470136,
+      "learning_rate": 0.00012454854871407994,
+      "loss": 0.8806,
+      "step": 195
+    },
+    {
+      "epoch": 1.74,
+      "grad_norm": 0.2283950634350429,
+      "learning_rate": 0.00012385101858982005,
+      "loss": 0.9053,
+      "step": 196
+    },
+    {
+      "epoch": 1.74,
+      "eval_loss": 1.2198154926300049,
+      "eval_runtime": 13.2208,
+      "eval_samples_per_second": 22.692,
+      "eval_steps_per_second": 2.874,
+      "step": 196
+    },
+    {
+      "epoch": 1.75,
+      "grad_norm": 0.23406423788354982,
+      "learning_rate": 0.00012315225580945252,
+      "loss": 0.9397,
+      "step": 197
+    },
+    {
+      "epoch": 1.76,
+      "grad_norm": 0.23807045727443327,
+      "learning_rate": 0.0001224522964860769,
+      "loss": 0.9712,
+      "step": 198
+    },
+    {
+      "epoch": 1.77,
+      "grad_norm": 0.2463614808838948,
+      "learning_rate": 0.00012175117679463187,
+      "loss": 0.8558,
+      "step": 199
+    },
+    {
+      "epoch": 1.78,
+      "grad_norm": 0.24737417059302014,
+      "learning_rate": 0.00012104893297002567,
+      "loss": 0.9723,
+      "step": 200
+    },
+    {
+      "epoch": 1.79,
+      "grad_norm": 0.243750688050595,
+      "learning_rate": 0.0001203456013052634,
+      "loss": 0.964,
+      "step": 201
+    },
+    {
+      "epoch": 1.8,
+      "grad_norm": 0.24572059557106538,
+      "learning_rate": 0.00011964121814957137,
+      "loss": 0.9109,
+      "step": 202
+    },
+    {
+      "epoch": 1.81,
+      "grad_norm": 0.24044117903962453,
+      "learning_rate": 0.00011893581990651848,
+      "loss": 1.0019,
+      "step": 203
+    },
+    {
+      "epoch": 1.82,
+      "grad_norm": 0.2737568489071465,
+      "learning_rate": 0.00011822944303213486,
+      "loss": 0.8893,
+      "step": 204
+    },
+    {
+      "epoch": 1.82,
+      "grad_norm": 0.24122455882790084,
+      "learning_rate": 0.00011752212403302784,
+      "loss": 0.9162,
+      "step": 205
+    },
+    {
+      "epoch": 1.83,
+      "grad_norm": 0.28991871401626856,
+      "learning_rate": 0.00011681389946449504,
+      "loss": 0.8555,
+      "step": 206
+    },
+    {
+      "epoch": 1.84,
+      "grad_norm": 0.23767408810646548,
+      "learning_rate": 0.00011610480592863531,
+      "loss": 0.9936,
+      "step": 207
+    },
+    {
+      "epoch": 1.85,
+      "grad_norm": 0.22614733706173062,
+      "learning_rate": 0.00011539488007245702,
+      "loss": 0.916,
+      "step": 208
+    },
+    {
+      "epoch": 1.86,
+      "grad_norm": 0.22471992425846515,
+      "learning_rate": 0.00011468415858598411,
+      "loss": 0.8872,
+      "step": 209
+    },
+    {
+      "epoch": 1.87,
+      "grad_norm": 0.22675717145909688,
+      "learning_rate": 0.00011397267820035986,
+      "loss": 0.8393,
+      "step": 210
+    },
+    {
+      "epoch": 1.88,
+      "grad_norm": 0.2727459336483823,
+      "learning_rate": 0.00011326047568594851,
+      "loss": 0.8265,
+      "step": 211
+    },
+    {
+      "epoch": 1.89,
+      "grad_norm": 0.25216778031670767,
+      "learning_rate": 0.00011254758785043515,
+      "loss": 0.9939,
+      "step": 212
+    },
+    {
+      "epoch": 1.9,
+      "grad_norm": 0.269147378424304,
+      "learning_rate": 0.0001118340515369232,
+      "loss": 0.9102,
+      "step": 213
+    },
+    {
+      "epoch": 1.91,
+      "grad_norm": 0.2216178370833471,
+      "learning_rate": 0.00011111990362203033,
+      "loss": 0.8778,
+      "step": 214
+    },
+    {
+      "epoch": 1.91,
+      "grad_norm": 0.2602474934716497,
+      "learning_rate": 0.00011040518101398276,
+      "loss": 0.9454,
+      "step": 215
+    },
+    {
+      "epoch": 1.92,
+      "grad_norm": 0.2658635078442998,
+      "learning_rate": 0.00010968992065070769,
+      "loss": 0.8098,
+      "step": 216
+    },
+    {
+      "epoch": 1.93,
+      "grad_norm": 0.20997905209488962,
+      "learning_rate": 0.00010897415949792427,
+      "loss": 0.9318,
+      "step": 217
+    },
+    {
+      "epoch": 1.94,
+      "grad_norm": 0.24752453752221557,
+      "learning_rate": 0.00010825793454723325,
+      "loss": 0.949,
+      "step": 218
+    },
+    {
+      "epoch": 1.95,
+      "grad_norm": 0.255579569750529,
+      "learning_rate": 0.0001075412828142051,
+      "loss": 0.915,
+      "step": 219
+    },
+    {
+      "epoch": 1.96,
+      "grad_norm": 0.23186981930561867,
+      "learning_rate": 0.0001068242413364671,
+      "loss": 0.9132,
+      "step": 220
+    },
+    {
+      "epoch": 1.97,
+      "grad_norm": 0.35685140391438824,
+      "learning_rate": 0.00010610684717178905,
+      "loss": 0.9398,
+      "step": 221
+    },
+    {
+      "epoch": 1.98,
+      "grad_norm": 0.27320389987223703,
+      "learning_rate": 0.00010538913739616816,
+      "loss": 0.857,
+      "step": 222
+    },
+    {
+      "epoch": 1.99,
+      "grad_norm": 0.2324276771141761,
+      "learning_rate": 0.00010467114910191289,
+      "loss": 0.8546,
+      "step": 223
+    },
+    {
+      "epoch": 1.99,
+      "grad_norm": 0.22820341349854167,
+      "learning_rate": 0.00010395291939572593,
+      "loss": 0.9301,
+      "step": 224
+    },
+    {
+      "epoch": 1.99,
+      "eval_loss": 1.2246263027191162,
+      "eval_runtime": 13.1981,
+      "eval_samples_per_second": 22.731,
+      "eval_steps_per_second": 2.879,
+      "step": 224
+    },
+    {
+      "epoch": 2.0,
+      "grad_norm": 0.2289800489154315,
+      "learning_rate": 0.00010323448539678653,
+      "loss": 0.9922,
+      "step": 225
+    },
+    {
+      "epoch": 2.01,
+      "grad_norm": 0.2673353778680862,
+      "learning_rate": 0.00010251588423483205,
+      "loss": 0.7779,
+      "step": 226
+    },
+    {
+      "epoch": 2.02,
+      "grad_norm": 0.2420933678952559,
+      "learning_rate": 0.0001017971530482392,
+      "loss": 0.8044,
+      "step": 227
+    },
+    {
+      "epoch": 2.03,
+      "grad_norm": 0.21799264660625498,
+      "learning_rate": 0.00010107832898210439,
+      "loss": 0.8773,
+      "step": 228
+    },
+    {
+      "epoch": 2.04,
+      "grad_norm": 0.21443255695871016,
+      "learning_rate": 0.00010035944918632429,
+      "loss": 0.9031,
+      "step": 229
+    },
+    {
+      "epoch": 2.05,
+      "grad_norm": 0.23983734165788242,
+      "learning_rate": 9.96405508136757e-05,
+      "loss": 0.9014,
+      "step": 230
+    },
+    {
+      "epoch": 2.06,
+      "grad_norm": 0.27915481475799336,
+      "learning_rate": 9.892167101789564e-05,
+      "loss": 0.8853,
+      "step": 231
+    },
+    {
+      "epoch": 2.07,
+      "grad_norm": 0.2688949371564916,
+      "learning_rate": 9.820284695176082e-05,
+      "loss": 0.8452,
+      "step": 232
+    },
+    {
+      "epoch": 2.07,
+      "grad_norm": 0.2623278518867105,
+      "learning_rate": 9.748411576516794e-05,
+      "loss": 0.8612,
+      "step": 233
+    },
+    {
+      "epoch": 2.08,
+      "grad_norm": 0.2710502639103885,
+      "learning_rate": 9.676551460321349e-05,
+      "loss": 0.8108,
+      "step": 234
+    },
+    {
+      "epoch": 2.09,
+      "grad_norm": 0.282572880285737,
+      "learning_rate": 9.60470806042741e-05,
+      "loss": 0.7866,
+      "step": 235
+    },
+    {
+      "epoch": 2.1,
+      "grad_norm": 0.2829396962922612,
+      "learning_rate": 9.532885089808713e-05,
+      "loss": 0.8557,
+      "step": 236
+    },
+    {
+      "epoch": 2.11,
+      "grad_norm": 0.2721172338857335,
+      "learning_rate": 9.461086260383187e-05,
+      "loss": 0.7933,
+      "step": 237
+    },
+    {
+      "epoch": 2.12,
+      "grad_norm": 0.29736638811364446,
+      "learning_rate": 9.389315282821097e-05,
+      "loss": 0.7674,
+      "step": 238
+    },
+    {
+      "epoch": 2.13,
+      "grad_norm": 0.28571679920981263,
+      "learning_rate": 9.317575866353292e-05,
+      "loss": 0.7442,
+      "step": 239
+    },
+    {
+      "epoch": 2.14,
+      "grad_norm": 0.264545167150173,
+      "learning_rate": 9.245871718579491e-05,
+      "loss": 0.8505,
+      "step": 240
+    },
+    {
+      "epoch": 2.15,
+      "grad_norm": 0.30691085134027757,
+      "learning_rate": 9.174206545276677e-05,
+      "loss": 0.7898,
+      "step": 241
+    },
+    {
+      "epoch": 2.15,
+      "grad_norm": 0.31375028121981235,
+      "learning_rate": 9.102584050207578e-05,
+      "loss": 0.7661,
+      "step": 242
+    },
+    {
+      "epoch": 2.16,
+      "grad_norm": 0.28421530221837016,
+      "learning_rate": 9.031007934929236e-05,
+      "loss": 0.8328,
+      "step": 243
+    },
+    {
+      "epoch": 2.17,
+      "grad_norm": 0.25601367811173414,
+      "learning_rate": 8.959481898601728e-05,
+      "loss": 0.8281,
+      "step": 244
+    },
+    {
+      "epoch": 2.18,
+      "grad_norm": 0.2983724947729522,
+      "learning_rate": 8.888009637796968e-05,
+      "loss": 0.8567,
+      "step": 245
+    },
+    {
+      "epoch": 2.19,
+      "grad_norm": 0.2545616786933236,
+      "learning_rate": 8.81659484630768e-05,
+      "loss": 0.9151,
+      "step": 246
+    },
+    {
+      "epoch": 2.2,
+      "grad_norm": 0.23873712362647942,
+      "learning_rate": 8.745241214956483e-05,
+      "loss": 0.8818,
+      "step": 247
+    },
+    {
+      "epoch": 2.21,
+      "grad_norm": 0.285331972404065,
+      "learning_rate": 8.673952431405148e-05,
+      "loss": 0.7983,
+      "step": 248
+    },
+    {
+      "epoch": 2.22,
+      "grad_norm": 0.23897707291689843,
+      "learning_rate": 8.602732179964017e-05,
+      "loss": 0.8758,
+      "step": 249
+    },
+    {
+      "epoch": 2.23,
+      "grad_norm": 0.2830966091447457,
+      "learning_rate": 8.531584141401591e-05,
+      "loss": 0.8714,
+      "step": 250
+    },
+    {
+      "epoch": 2.23,
+      "grad_norm": 0.28872599217076506,
+      "learning_rate": 8.4605119927543e-05,
+      "loss": 0.8387,
+      "step": 251
+    },
+    {
+      "epoch": 2.24,
+      "grad_norm": 0.2652236346400331,
+      "learning_rate": 8.38951940713647e-05,
+      "loss": 0.8232,
+      "step": 252
+    },
+    {
+      "epoch": 2.24,
+      "eval_loss": 1.2432794570922852,
+      "eval_runtime": 13.2405,
+      "eval_samples_per_second": 22.658,
+      "eval_steps_per_second": 2.87,
+      "step": 252
+    },
+    {
+      "epoch": 2.25,
+      "grad_norm": 0.299978013524394,
+      "learning_rate": 8.318610053550497e-05,
+      "loss": 0.7321,
+      "step": 253
+    },
+    {
+      "epoch": 2.26,
+      "grad_norm": 0.2740002835117391,
+      "learning_rate": 8.247787596697218e-05,
+      "loss": 0.7605,
+      "step": 254
+    },
+    {
+      "epoch": 2.27,
+      "grad_norm": 0.2848366030132808,
+      "learning_rate": 8.177055696786516e-05,
+      "loss": 0.8485,
+      "step": 255
+    },
+    {
+      "epoch": 2.28,
+      "grad_norm": 0.24847418856075218,
+      "learning_rate": 8.106418009348157e-05,
+      "loss": 0.7557,
+      "step": 256
+    },
+    {
+      "epoch": 2.29,
+      "grad_norm": 0.33515508602624905,
+      "learning_rate": 8.035878185042868e-05,
+      "loss": 0.8015,
+      "step": 257
+    },
+    {
+      "epoch": 2.3,
+      "grad_norm": 0.2905943721096322,
+      "learning_rate": 7.965439869473664e-05,
+      "loss": 0.8457,
+      "step": 258
+    },
+    {
+      "epoch": 2.31,
+      "grad_norm": 0.3140679719552616,
+      "learning_rate": 7.895106702997437e-05,
+      "loss": 0.8559,
+      "step": 259
+    },
+    {
+      "epoch": 2.31,
+      "grad_norm": 0.29745105018138573,
+      "learning_rate": 7.824882320536814e-05,
+      "loss": 0.7453,
+      "step": 260
+    },
+    {
+      "epoch": 2.32,
+      "grad_norm": 0.29818631731197365,
+      "learning_rate": 7.754770351392311e-05,
+      "loss": 0.8354,
+      "step": 261
+    },
+    {
+      "epoch": 2.33,
+      "grad_norm": 0.24721488944366407,
+      "learning_rate": 7.684774419054747e-05,
+      "loss": 0.7755,
+      "step": 262
+    },
+    {
+      "epoch": 2.34,
+      "grad_norm": 0.31210442779019465,
+      "learning_rate": 7.614898141017996e-05,
+      "loss": 0.7208,
+      "step": 263
+    },
+    {
+      "epoch": 2.35,
+      "grad_norm": 0.2873220240109992,
+      "learning_rate": 7.54514512859201e-05,
+      "loss": 0.7548,
+      "step": 264
+    },
+    {
+      "epoch": 2.36,
+      "grad_norm": 0.3006634171776217,
+      "learning_rate": 7.475518986716194e-05,
+      "loss": 0.7566,
+      "step": 265
+    },
+    {
+      "epoch": 2.37,
+      "grad_norm": 0.2799417613336026,
+      "learning_rate": 7.406023313773097e-05,
+      "loss": 0.727,
+      "step": 266
+    },
+    {
+      "epoch": 2.38,
+      "grad_norm": 0.2451761866231664,
+      "learning_rate": 7.336661701402439e-05,
+      "loss": 0.9641,
+      "step": 267
+    },
+    {
+      "epoch": 2.39,
+      "grad_norm": 0.305202611125298,
+      "learning_rate": 7.267437734315492e-05,
+      "loss": 0.7891,
+      "step": 268
+    },
+    {
+      "epoch": 2.39,
+      "grad_norm": 0.29107717848747816,
+      "learning_rate": 7.198354990109805e-05,
+      "loss": 0.9032,
+      "step": 269
+    },
+    {
+      "epoch": 2.4,
+      "grad_norm": 0.2688898665176787,
+      "learning_rate": 7.129417039084333e-05,
+      "loss": 0.8416,
+      "step": 270
+    },
+    {
+      "epoch": 2.41,
+      "grad_norm": 0.2814206029778395,
+      "learning_rate": 7.060627444054893e-05,
+      "loss": 0.8443,
+      "step": 271
+    },
+    {
+      "epoch": 2.42,
+      "grad_norm": 0.2862094867555512,
+      "learning_rate": 6.99198976017005e-05,
+      "loss": 0.8271,
+      "step": 272
+    },
+    {
+      "epoch": 2.43,
+      "grad_norm": 0.3214647340394826,
+      "learning_rate": 6.923507534727373e-05,
+      "loss": 0.7793,
+      "step": 273
+    },
+    {
+      "epoch": 2.44,
+      "grad_norm": 0.3033659714564417,
+      "learning_rate": 6.855184306990106e-05,
+      "loss": 0.7856,
+      "step": 274
+    },
+    {
+      "epoch": 2.45,
+      "grad_norm": 0.3024382342577774,
+      "learning_rate": 6.78702360800425e-05,
+      "loss": 0.8633,
+      "step": 275
+    },
+    {
+      "epoch": 2.46,
+      "grad_norm": 0.25803598196729505,
+      "learning_rate": 6.719028960416098e-05,
+      "loss": 0.8428,
+      "step": 276
+    },
+    {
+      "epoch": 2.47,
+      "grad_norm": 0.35469202971401803,
+      "learning_rate": 6.651203878290139e-05,
+      "loss": 0.8665,
+      "step": 277
+    },
+    {
+      "epoch": 2.47,
+      "grad_norm": 0.3122516837597691,
+      "learning_rate": 6.583551866927475e-05,
+      "loss": 0.8787,
+      "step": 278
+    },
+    {
+      "epoch": 2.48,
+      "grad_norm": 0.3305470786367901,
+      "learning_rate": 6.516076422684654e-05,
+      "loss": 0.8765,
+      "step": 279
+    },
+    {
+      "epoch": 2.49,
+      "grad_norm": 0.3324622666488467,
+      "learning_rate": 6.448781032792972e-05,
+      "loss": 0.8318,
+      "step": 280
+    },
+    {
+      "epoch": 2.49,
+      "eval_loss": 1.2546111345291138,
+      "eval_runtime": 13.2379,
+      "eval_samples_per_second": 22.662,
+      "eval_steps_per_second": 2.871,
+      "step": 280
+    },
+    {
+      "epoch": 2.5,
+      "grad_norm": 0.342341713579355,
+      "learning_rate": 6.381669175178248e-05,
+      "loss": 0.9517,
+      "step": 281
+    },
+    {
+      "epoch": 2.51,
+      "grad_norm": 0.33913458352374665,
+      "learning_rate": 6.31474431828108e-05,
+      "loss": 0.8564,
+      "step": 282
+    },
+    {
+      "epoch": 2.52,
+      "grad_norm": 0.30528689383480295,
+      "learning_rate": 6.248009920877592e-05,
+      "loss": 0.8199,
+      "step": 283
+    },
+    {
+      "epoch": 2.53,
+      "grad_norm": 0.29698648367254743,
+      "learning_rate": 6.181469431900672e-05,
+      "loss": 0.785,
+      "step": 284
+    },
+    {
+      "epoch": 2.54,
+      "grad_norm": 0.32239262939282626,
+      "learning_rate": 6.115126290261745e-05,
+      "loss": 0.7794,
+      "step": 285
+    },
+    {
+      "epoch": 2.55,
+      "grad_norm": 0.2694595905080167,
+      "learning_rate": 6.048983924673022e-05,
+      "loss": 0.8056,
+      "step": 286
+    },
+    {
+      "epoch": 2.55,
+      "grad_norm": 0.3045496751154443,
+      "learning_rate": 5.983045753470308e-05,
+      "loss": 0.8164,
+      "step": 287
+    },
+    {
+      "epoch": 2.56,
+      "grad_norm": 0.2927868214627918,
+      "learning_rate": 5.917315184436345e-05,
+      "loss": 0.8358,
+      "step": 288
+    },
+    {
+      "epoch": 2.57,
+      "grad_norm": 0.2931914055644858,
+      "learning_rate": 5.851795614624682e-05,
+      "loss": 0.8011,
+      "step": 289
+    },
+    {
+      "epoch": 2.58,
+      "grad_norm": 0.3158716819379082,
+      "learning_rate": 5.786490430184115e-05,
+      "loss": 0.8332,
+      "step": 290
+    },
+    {
+      "epoch": 2.59,
+      "grad_norm": 0.3482519147352008,
+      "learning_rate": 5.72140300618369e-05,
+      "loss": 0.7621,
+      "step": 291
+    },
+    {
+      "epoch": 2.6,
+      "grad_norm": 0.28652801822050894,
+      "learning_rate": 5.656536706438267e-05,
+      "loss": 0.77,
+      "step": 292
+    },
+    {
+      "epoch": 2.61,
+      "grad_norm": 0.29691290613407717,
+      "learning_rate": 5.591894883334667e-05,
+      "loss": 0.9394,
+      "step": 293
+    },
+    {
+      "epoch": 2.62,
+      "grad_norm": 0.26699581966985203,
+      "learning_rate": 5.5274808776584367e-05,
+      "loss": 0.7918,
+      "step": 294
+    },
+    {
+      "epoch": 2.63,
+      "grad_norm": 0.2926923719762685,
+      "learning_rate": 5.463298018421171e-05,
+      "loss": 0.8723,
+      "step": 295
+    },
+    {
+      "epoch": 2.64,
+      "grad_norm": 0.3403087263187063,
+      "learning_rate": 5.399349622688479e-05,
+      "loss": 0.8097,
+      "step": 296
+    },
+    {
+      "epoch": 2.64,
+      "grad_norm": 0.34261233464532476,
+      "learning_rate": 5.335638995408545e-05,
+      "loss": 0.9032,
+      "step": 297
+    },
+    {
+      "epoch": 2.65,
+      "grad_norm": 0.31315234759634086,
+      "learning_rate": 5.272169429241325e-05,
+      "loss": 0.82,
+      "step": 298
+    },
+    {
+      "epoch": 2.66,
+      "grad_norm": 0.3179759425444047,
+      "learning_rate": 5.208944204388377e-05,
+      "loss": 0.8864,
+      "step": 299
+    },
+    {
+      "epoch": 2.67,
+      "grad_norm": 0.3121296356843828,
+      "learning_rate": 5.145966588423341e-05,
+      "loss": 0.8258,
+      "step": 300
+    },
+    {
+      "epoch": 2.68,
+      "grad_norm": 0.268436849924173,
+      "learning_rate": 5.0832398361230596e-05,
+      "loss": 0.8906,
+      "step": 301
+    },
+    {
+      "epoch": 2.69,
+      "grad_norm": 0.2961161602467319,
+      "learning_rate": 5.020767189299369e-05,
+      "loss": 0.8828,
+      "step": 302
+    },
+    {
+      "epoch": 2.7,
+      "grad_norm": 0.27743957099992345,
+      "learning_rate": 4.9585518766315496e-05,
+      "loss": 0.8251,
+      "step": 303
+    },
+    {
+      "epoch": 2.71,
+      "grad_norm": 0.2949909861852426,
+      "learning_rate": 4.896597113499479e-05,
+      "loss": 0.7911,
+      "step": 304
+    },
+    {
+      "epoch": 2.72,
+      "grad_norm": 0.3161115451278363,
+      "learning_rate": 4.834906101817438e-05,
+      "loss": 0.8157,
+      "step": 305
+    },
+    {
+      "epoch": 2.72,
+      "grad_norm": 0.28720077046065867,
+      "learning_rate": 4.773482029868657e-05,
+      "loss": 0.82,
+      "step": 306
+    },
+    {
+      "epoch": 2.73,
+      "grad_norm": 0.4045319357608716,
+      "learning_rate": 4.712328072140505e-05,
+      "loss": 0.8414,
+      "step": 307
+    },
+    {
+      "epoch": 2.74,
+      "grad_norm": 0.3070232288390269,
+      "learning_rate": 4.651447389160458e-05,
+      "loss": 0.8427,
+      "step": 308
+    },
+    {
+      "epoch": 2.74,
+      "eval_loss": 1.2574400901794434,
+      "eval_runtime": 13.2473,
+      "eval_samples_per_second": 22.646,
+      "eval_steps_per_second": 2.869,
+      "step": 308
+    },
+    {
+      "epoch": 2.75,
+      "grad_norm": 0.3214782806968351,
+      "learning_rate": 4.5908431273327436e-05,
+      "loss": 0.8469,
+      "step": 309
+    },
+    {
+      "epoch": 2.76,
+      "grad_norm": 0.24241410698156174,
+      "learning_rate": 4.530518418775733e-05,
+      "loss": 0.8346,
+      "step": 310
+    },
+    {
+      "epoch": 2.77,
+      "grad_norm": 0.3303263594210879,
+      "learning_rate": 4.470476381160065e-05,
+      "loss": 0.8298,
+      "step": 311
+    },
+    {
+      "epoch": 2.78,
+      "grad_norm": 0.30711900849760865,
+      "learning_rate": 4.4107201175475275e-05,
+      "loss": 0.789,
+      "step": 312
+    },
+    {
+      "epoch": 2.79,
+      "grad_norm": 0.2954465859389713,
+      "learning_rate": 4.351252716230685e-05,
+      "loss": 0.8029,
+      "step": 313
+    },
+    {
+      "epoch": 2.8,
+      "grad_norm": 0.29925087091531116,
+      "learning_rate": 4.292077250573266e-05,
+      "loss": 0.8633,
+      "step": 314
+    },
+    {
+      "epoch": 2.8,
+      "grad_norm": 0.3177611223775825,
+      "learning_rate": 4.2331967788513295e-05,
+      "loss": 0.76,
+      "step": 315
+    },
+    {
+      "epoch": 2.81,
+      "grad_norm": 0.28642407848269513,
+      "learning_rate": 4.174614344095213e-05,
+      "loss": 0.823,
+      "step": 316
+    },
+    {
+      "epoch": 2.82,
+      "grad_norm": 0.3243224656005062,
+      "learning_rate": 4.116332973932256e-05,
+      "loss": 0.7831,
+      "step": 317
+    },
+    {
+      "epoch": 2.83,
+      "grad_norm": 0.34877334027822726,
+      "learning_rate": 4.058355680430337e-05,
+      "loss": 0.899,
+      "step": 318
+    },
+    {
+      "epoch": 2.84,
+      "grad_norm": 0.28640325479143114,
+      "learning_rate": 4.0006854599421926e-05,
+      "loss": 0.8292,
+      "step": 319
+    },
+    {
+      "epoch": 2.85,
+      "grad_norm": 0.3135316628014535,
+      "learning_rate": 3.943325292950579e-05,
+      "loss": 0.8731,
+      "step": 320
+    },
+    {
+      "epoch": 2.86,
+      "grad_norm": 0.2949970604257085,
+      "learning_rate": 3.886278143914219e-05,
+      "loss": 0.8402,
+      "step": 321
+    },
+    {
+      "epoch": 2.87,
+      "grad_norm": 0.30057896586780075,
+      "learning_rate": 3.829546961114607e-05,
+      "loss": 0.7713,
+      "step": 322
+    },
+    {
+      "epoch": 2.88,
+      "grad_norm": 0.3558574270285126,
+      "learning_rate": 3.773134676503629e-05,
+      "loss": 0.8435,
+      "step": 323
+    },
+    {
+      "epoch": 2.88,
+      "grad_norm": 0.29115288332943334,
+      "learning_rate": 3.7170442055520415e-05,
+      "loss": 0.9022,
+      "step": 324
+    },
+    {
+      "epoch": 2.89,
+      "grad_norm": 0.3192074718527619,
+      "learning_rate": 3.661278447098789e-05,
+      "loss": 0.7662,
+      "step": 325
+    },
+    {
+      "epoch": 2.9,
+      "grad_norm": 0.33335742888185405,
+      "learning_rate": 3.605840283201195e-05,
+      "loss": 0.8111,
+      "step": 326
+    },
+    {
+      "epoch": 2.91,
+      "grad_norm": 0.29748212071395186,
+      "learning_rate": 3.550732578986006e-05,
+      "loss": 0.7543,
+      "step": 327
+    },
+    {
+      "epoch": 2.92,
+      "grad_norm": 0.3680409192627914,
+      "learning_rate": 3.495958182501325e-05,
+      "loss": 0.8124,
+      "step": 328
+    },
+    {
+      "epoch": 2.93,
+      "grad_norm": 0.27807302364345643,
+      "learning_rate": 3.441519924569408e-05,
+      "loss": 0.7856,
+      "step": 329
+    },
+    {
+      "epoch": 2.94,
+      "grad_norm": 0.3050855823733691,
+      "learning_rate": 3.387420618640379e-05,
+      "loss": 0.8506,
+      "step": 330
+    },
+    {
+      "epoch": 2.95,
+      "grad_norm": 0.3322620263029238,
+      "learning_rate": 3.3336630606468134e-05,
+      "loss": 0.8771,
+      "step": 331
+    },
+    {
+      "epoch": 2.96,
+      "grad_norm": 0.3112008867427982,
+      "learning_rate": 3.280250028859248e-05,
+      "loss": 0.7785,
+      "step": 332
+    },
+    {
+      "epoch": 2.96,
+      "grad_norm": 0.2839548329095365,
+      "learning_rate": 3.227184283742591e-05,
+      "loss": 0.9153,
+      "step": 333
+    },
+    {
+      "epoch": 2.97,
+      "grad_norm": 0.34615397822650606,
+      "learning_rate": 3.174468567813461e-05,
+      "loss": 0.7753,
+      "step": 334
+    },
+    {
+      "epoch": 2.98,
+      "grad_norm": 0.34691866307772695,
+      "learning_rate": 3.122105605498442e-05,
+      "loss": 0.851,
+      "step": 335
+    },
+    {
+      "epoch": 2.99,
+      "grad_norm": 0.296369624391198,
+      "learning_rate": 3.070098102993302e-05,
+      "loss": 0.8572,
+      "step": 336
+    },
+    {
+      "epoch": 2.99,
+      "eval_loss": 1.2511259317398071,
+      "eval_runtime": 13.2202,
+      "eval_samples_per_second": 22.692,
+      "eval_steps_per_second": 2.874,
+      "step": 336
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 448,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 4,
+  "save_steps": 112,
+  "total_flos": 4.005448726012232e+17,
+  "train_batch_size": 2,
+  "trial_name": null,
+  "trial_params": null
+}

checkpoint-336/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d400c16f982c36b10268ff7e69e878c44d11f5fb692a61770a8e1efb50d4491c
+size 6776

checkpoint-336/zero_to_fp32.py

@@ -0,0 +1,592 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example: python zero_to_fp32.py . pytorch_model.bin
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+
+    total_files = len(files)
+    state_dicts = []
+    for f in files:
+        state_dict = torch.load(f, map_location=device)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    if zero_stage <= 2:
+        fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    elif zero_stage == 3:
+        # if there is more than one param group, there will be multiple flattened tensors - one
+        # flattened tensor per group - for simplicity merge them into a single tensor
+        #
+        # XXX: could make the script more memory efficient for when there are multiple groups - it
+        # will require matching the sub-lists of param_shapes for each param group flattened tensor
+
+        fp32_flat_groups = [
+            torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
+        ]
+
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    for name, shape in param_shapes.items():
+
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # XXX: memory usage doubles here
+        state_dict[name] = torch.cat(
+            tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
+            0).narrow(0, 0, unpartitioned_numel).view(shape)
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``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``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    Note: this approach may not work if your application doesn't have sufficient free CPU memory and
+    you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
+        - ``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``
+    """
+
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+    print(f"Saving fp32 state dict to {output_file}")
+    torch.save(state_dict, output_file)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``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``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument(
+        "output_file",
+        type=str,
+        help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file, tag=args.tag)

checkpoint-372/README.md

@@ -0,0 +1,202 @@
+---
+library_name: peft
+base_model: google/gemma-7b-it
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- 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. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+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).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.9.0

checkpoint-372/adapter_config.json

@@ -0,0 +1,33 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "google/gemma-7b-it",
+  "bias": "none",
+  "fan_in_fan_out": null,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 16,
+  "lora_dropout": 0.05,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "down_proj",
+    "o_proj",
+    "k_proj",
+    "q_proj",
+    "gate_proj",
+    "up_proj",
+    "v_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "use_dora": false,
+  "use_rslora": false
+}

checkpoint-372/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9c23b965687f7bf2e033e1e8051de69e24c99f3103c06606007e68485ebfabea
+size 200068904

checkpoint-372/global_step372/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:712292fc7e5a6d570c1376cc3be7e12dab2d34fb7ffe48281da38c8053603a39
+size 150126608

checkpoint-372/global_step372/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d3c6bbd461df3af80fb33496e1907ef542102bfa96434d50c174fe80c0dd98e4
+size 150126672

checkpoint-372/global_step372/bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:46fee0787345b6483d7a54f3ceeb3260a7a8bef008c22e24c18225027433ff01
+size 150126736

checkpoint-372/global_step372/bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0fbe9084b027f2164f5fa8039ea7d37a722d0e0f9f70b2a76fa605e462a2ad6e
+size 150126736

checkpoint-372/global_step372/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a38ea9a669e473fff57e6c134dd6703ddacc9f123121be165e81bcdcad09513b
+size 1896781286

checkpoint-372/latest

@@ -0,0 +1 @@
+global_step372

checkpoint-372/rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ebe22192845fac896cd970f52665ebcfd6b5796077804b55f0d8830fcfa32be
+size 15024

checkpoint-372/rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a5bbd2194b05d2155d794f7732bdab8deaa38ee92f4c49fa250d0c9f0fd5f532
+size 15024