Upload llava-v1.6-34b best check point

checkpoint-115/README.md

@@ -0,0 +1,202 @@
+---
+base_model: liuhaotian/llava-v1.6-34b
+library_name: peft
+---
+
+# 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 0.13.2

checkpoint-115/adapter_config.json

@@ -0,0 +1,34 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "liuhaotian/llava-v1.6-34b",
+  "bias": "none",
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "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": 8,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "gate_proj",
+    "o_proj",
+    "down_proj",
+    "k_proj",
+    "q_proj",
+    "up_proj",
+    "v_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "use_dora": false,
+  "use_rslora": false
+}

checkpoint-115/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8d9be90939d81f0c1f4d36995227114e9c12ac8f35dd4d8e5a34c74bef831e42
+size 125377072

checkpoint-115/global_step115/zero_pp_rank_0_mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:46c90fb3e46261dc582810d91b50bc25762ae35e7118476768539ff850970e83
+size 1053274

checkpoint-115/global_step115/zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ccac836bc3a4a2cc7bc0bc599f3afce15b4742d9e6ba51d8a8412c390b328d09
+size 364066797

checkpoint-115/global_step115/zero_pp_rank_1_mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:21f57bb03cae86feb7429e683a2e62c13540d5842b44981da8950ee072df4086
+size 1053274

checkpoint-115/global_step115/zero_pp_rank_1_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:863a9bdf590223856e2e7d61c7d53090b4684678849b5e9511557780760384b9
+size 364066797

checkpoint-115/global_step115/zero_pp_rank_2_mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:08e460641ebc4542861ee8be1e85a93efe4bb17faf3db95b08c3e65ea1806e69
+size 1053274

checkpoint-115/global_step115/zero_pp_rank_2_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:630242764e71ca747a013f697b05ca1d6a0529b1c32d4e9d95ce88202cf4a69c
+size 364066797

checkpoint-115/global_step115/zero_pp_rank_3_mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:06097032ffb7612fc38ee58012df396f2dcfcd622c56744e69ea5e5dcc93ff33
+size 1053274

checkpoint-115/global_step115/zero_pp_rank_3_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0c9cfc2e8af5b98d5fac044de7d9d24d57a323934e2406edb75502bb3f925754
+size 364066797

checkpoint-115/latest

@@ -0,0 +1 @@
+global_step115

checkpoint-115/rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:069bd959645cedd813117dc3cf2cf2b0a89083b6d3350b8829691245e10dde24
+size 14960

checkpoint-115/rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7cb4f9a0dea8611077615477ed649604ff1804e1407cd60ace3e921a95544fe2
+size 14960

checkpoint-115/rng_state_2.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:db9a85d43f5945cd0d97fe3501bd515353903f42c74096ea00db107bd6d194a8
+size 14960

checkpoint-115/rng_state_3.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2ecd48677ab11ff5286cb583a23fd6fb6d89adb5f28bf402849e24736516dd8d
+size 14960

checkpoint-115/special_tokens_map.json

@@ -0,0 +1,24 @@
+{
+  "bos_token": {
+    "content": "<|startoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<|im_end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "<unk>",
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

checkpoint-115/tokenizer.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:386c49cf943d71aa110361135338c50e38beeff0a66593480421f37b319e1a39
+size 1033105

checkpoint-115/tokenizer_config.json

@@ -0,0 +1,62 @@
+{
+  "add_bos_token": false,
+  "add_eos_token": false,
+  "add_prefix_space": true,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<|startoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "6": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "7": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<|startoftext|>",
+  "chat_template": "{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "legacy": true,
+  "model_max_length": 2048,
+  "pad_token": "<unk>",
+  "padding_side": "right",
+  "sp_model_kwargs": {},
+  "spaces_between_special_tokens": false,
+  "tokenizer_class": "LlamaTokenizer",
+  "trust_remote_code": false,
+  "unk_token": "<unk>",
+  "use_default_system_prompt": false,
+  "use_fast": true
+}

checkpoint-115/trainer_state.json

@@ -0,0 +1,1758 @@
+{
+  "best_metric": 0.7502281069755554,
+  "best_model_checkpoint": "./checkpoints/llava-v1.6-34b-chatml_direct-anyres/checkpoint-115",
+  "epoch": 3.59375,
+  "eval_steps": 1.0,
+  "global_step": 115,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.03125,
+      "grad_norm": 0.3345325833952841,
+      "learning_rate": 0.0,
+      "loss": 1.2677,
+      "step": 1
+    },
+    {
+      "epoch": 0.03125,
+      "eval_loss": 1.3042412996292114,
+      "eval_runtime": 195.8656,
+      "eval_samples_per_second": 1.021,
+      "eval_steps_per_second": 0.128,
+      "step": 1
+    },
+    {
+      "epoch": 0.0625,
+      "grad_norm": 0.35890614982474045,
+      "learning_rate": 8.613531161467863e-06,
+      "loss": 1.3505,
+      "step": 2
+    },
+    {
+      "epoch": 0.0625,
+      "eval_loss": 1.3042412996292114,
+      "eval_runtime": 190.4145,
+      "eval_samples_per_second": 1.05,
+      "eval_steps_per_second": 0.131,
+      "step": 2
+    },
+    {
+      "epoch": 0.09375,
+      "grad_norm": 0.3316701046946783,
+      "learning_rate": 1.3652123889719709e-05,
+      "loss": 1.2569,
+      "step": 3
+    },
+    {
+      "epoch": 0.09375,
+      "eval_loss": 1.2942205667495728,
+      "eval_runtime": 190.6475,
+      "eval_samples_per_second": 1.049,
+      "eval_steps_per_second": 0.131,
+      "step": 3
+    },
+    {
+      "epoch": 0.125,
+      "grad_norm": 0.32188096642751235,
+      "learning_rate": 1.7227062322935725e-05,
+      "loss": 1.2323,
+      "step": 4
+    },
+    {
+      "epoch": 0.125,
+      "eval_loss": 1.2789555788040161,
+      "eval_runtime": 189.7666,
+      "eval_samples_per_second": 1.054,
+      "eval_steps_per_second": 0.132,
+      "step": 4
+    },
+    {
+      "epoch": 0.15625,
+      "grad_norm": 0.3767527705004001,
+      "learning_rate": 2e-05,
+      "loss": 1.2785,
+      "step": 5
+    },
+    {
+      "epoch": 0.15625,
+      "eval_loss": 1.258152723312378,
+      "eval_runtime": 189.5935,
+      "eval_samples_per_second": 1.055,
+      "eval_steps_per_second": 0.132,
+      "step": 5
+    },
+    {
+      "epoch": 0.1875,
+      "grad_norm": 0.3287126070774628,
+      "learning_rate": 2e-05,
+      "loss": 1.2151,
+      "step": 6
+    },
+    {
+      "epoch": 0.1875,
+      "eval_loss": 1.2347216606140137,
+      "eval_runtime": 190.4111,
+      "eval_samples_per_second": 1.05,
+      "eval_steps_per_second": 0.131,
+      "step": 6
+    },
+    {
+      "epoch": 0.21875,
+      "grad_norm": 0.34451127286331007,
+      "learning_rate": 2e-05,
+      "loss": 1.2968,
+      "step": 7
+    },
+    {
+      "epoch": 0.21875,
+      "eval_loss": 1.210167646408081,
+      "eval_runtime": 190.9799,
+      "eval_samples_per_second": 1.047,
+      "eval_steps_per_second": 0.131,
+      "step": 7
+    },
+    {
+      "epoch": 0.25,
+      "grad_norm": 0.36105870692958336,
+      "learning_rate": 2e-05,
+      "loss": 1.2277,
+      "step": 8
+    },
+    {
+      "epoch": 0.25,
+      "eval_loss": 1.1862907409667969,
+      "eval_runtime": 190.685,
+      "eval_samples_per_second": 1.049,
+      "eval_steps_per_second": 0.131,
+      "step": 8
+    },
+    {
+      "epoch": 0.28125,
+      "grad_norm": 0.35460549637546845,
+      "learning_rate": 2e-05,
+      "loss": 1.2101,
+      "step": 9
+    },
+    {
+      "epoch": 0.28125,
+      "eval_loss": 1.1649302244186401,
+      "eval_runtime": 190.1569,
+      "eval_samples_per_second": 1.052,
+      "eval_steps_per_second": 0.131,
+      "step": 9
+    },
+    {
+      "epoch": 0.3125,
+      "grad_norm": 0.3134923556618721,
+      "learning_rate": 2e-05,
+      "loss": 1.1163,
+      "step": 10
+    },
+    {
+      "epoch": 0.3125,
+      "eval_loss": 1.144965410232544,
+      "eval_runtime": 190.0982,
+      "eval_samples_per_second": 1.052,
+      "eval_steps_per_second": 0.132,
+      "step": 10
+    },
+    {
+      "epoch": 0.34375,
+      "grad_norm": 0.3069481492118633,
+      "learning_rate": 2e-05,
+      "loss": 1.1483,
+      "step": 11
+    },
+    {
+      "epoch": 0.34375,
+      "eval_loss": 1.124668002128601,
+      "eval_runtime": 192.0572,
+      "eval_samples_per_second": 1.041,
+      "eval_steps_per_second": 0.13,
+      "step": 11
+    },
+    {
+      "epoch": 0.375,
+      "grad_norm": 0.2801324709168811,
+      "learning_rate": 2e-05,
+      "loss": 1.1172,
+      "step": 12
+    },
+    {
+      "epoch": 0.375,
+      "eval_loss": 1.1061824560165405,
+      "eval_runtime": 192.6406,
+      "eval_samples_per_second": 1.038,
+      "eval_steps_per_second": 0.13,
+      "step": 12
+    },
+    {
+      "epoch": 0.40625,
+      "grad_norm": 0.33156251919932406,
+      "learning_rate": 2e-05,
+      "loss": 1.1902,
+      "step": 13
+    },
+    {
+      "epoch": 0.40625,
+      "eval_loss": 1.0897018909454346,
+      "eval_runtime": 192.8064,
+      "eval_samples_per_second": 1.037,
+      "eval_steps_per_second": 0.13,
+      "step": 13
+    },
+    {
+      "epoch": 0.4375,
+      "grad_norm": 0.3307149375898363,
+      "learning_rate": 2e-05,
+      "loss": 1.1014,
+      "step": 14
+    },
+    {
+      "epoch": 0.4375,
+      "eval_loss": 1.075058937072754,
+      "eval_runtime": 192.1353,
+      "eval_samples_per_second": 1.041,
+      "eval_steps_per_second": 0.13,
+      "step": 14
+    },
+    {
+      "epoch": 0.46875,
+      "grad_norm": 0.31999611930431227,
+      "learning_rate": 2e-05,
+      "loss": 1.0847,
+      "step": 15
+    },
+    {
+      "epoch": 0.46875,
+      "eval_loss": 1.0613062381744385,
+      "eval_runtime": 192.0586,
+      "eval_samples_per_second": 1.041,
+      "eval_steps_per_second": 0.13,
+      "step": 15
+    },
+    {
+      "epoch": 0.5,
+      "grad_norm": 0.2494159223446848,
+      "learning_rate": 2e-05,
+      "loss": 1.0428,
+      "step": 16
+    },
+    {
+      "epoch": 0.5,
+      "eval_loss": 1.0484405755996704,
+      "eval_runtime": 192.52,
+      "eval_samples_per_second": 1.039,
+      "eval_steps_per_second": 0.13,
+      "step": 16
+    },
+    {
+      "epoch": 0.53125,
+      "grad_norm": 0.2899303168196212,
+      "learning_rate": 2e-05,
+      "loss": 1.122,
+      "step": 17
+    },
+    {
+      "epoch": 0.53125,
+      "eval_loss": 1.036120891571045,
+      "eval_runtime": 192.5716,
+      "eval_samples_per_second": 1.039,
+      "eval_steps_per_second": 0.13,
+      "step": 17
+    },
+    {
+      "epoch": 0.5625,
+      "grad_norm": 0.2995776829874209,
+      "learning_rate": 2e-05,
+      "loss": 1.0425,
+      "step": 18
+    },
+    {
+      "epoch": 0.5625,
+      "eval_loss": 1.0226774215698242,
+      "eval_runtime": 192.5256,
+      "eval_samples_per_second": 1.039,
+      "eval_steps_per_second": 0.13,
+      "step": 18
+    },
+    {
+      "epoch": 0.59375,
+      "grad_norm": 0.28709859243892955,
+      "learning_rate": 2e-05,
+      "loss": 1.0098,
+      "step": 19
+    },
+    {
+      "epoch": 0.59375,
+      "eval_loss": 1.0081837177276611,
+      "eval_runtime": 192.3486,
+      "eval_samples_per_second": 1.04,
+      "eval_steps_per_second": 0.13,
+      "step": 19
+    },
+    {
+      "epoch": 0.625,
+      "grad_norm": 0.27612474678791227,
+      "learning_rate": 2e-05,
+      "loss": 1.0563,
+      "step": 20
+    },
+    {
+      "epoch": 0.625,
+      "eval_loss": 0.994163990020752,
+      "eval_runtime": 191.9782,
+      "eval_samples_per_second": 1.042,
+      "eval_steps_per_second": 0.13,
+      "step": 20
+    },
+    {
+      "epoch": 0.65625,
+      "grad_norm": 0.24260720679126926,
+      "learning_rate": 2e-05,
+      "loss": 1.0355,
+      "step": 21
+    },
+    {
+      "epoch": 0.65625,
+      "eval_loss": 0.9819543361663818,
+      "eval_runtime": 191.9306,
+      "eval_samples_per_second": 1.042,
+      "eval_steps_per_second": 0.13,
+      "step": 21
+    },
+    {
+      "epoch": 0.6875,
+      "grad_norm": 0.25336536603884946,
+      "learning_rate": 2e-05,
+      "loss": 1.0525,
+      "step": 22
+    },
+    {
+      "epoch": 0.6875,
+      "eval_loss": 0.9709838032722473,
+      "eval_runtime": 192.9913,
+      "eval_samples_per_second": 1.036,
+      "eval_steps_per_second": 0.13,
+      "step": 22
+    },
+    {
+      "epoch": 0.71875,
+      "grad_norm": 0.24820839136364292,
+      "learning_rate": 2e-05,
+      "loss": 1.1392,
+      "step": 23
+    },
+    {
+      "epoch": 0.71875,
+      "eval_loss": 0.9616628885269165,
+      "eval_runtime": 192.6673,
+      "eval_samples_per_second": 1.038,
+      "eval_steps_per_second": 0.13,
+      "step": 23
+    },
+    {
+      "epoch": 0.75,
+      "grad_norm": 0.24589291203527217,
+      "learning_rate": 2e-05,
+      "loss": 1.058,
+      "step": 24
+    },
+    {
+      "epoch": 0.75,
+      "eval_loss": 0.9531083703041077,
+      "eval_runtime": 193.0994,
+      "eval_samples_per_second": 1.036,
+      "eval_steps_per_second": 0.129,
+      "step": 24
+    },
+    {
+      "epoch": 0.78125,
+      "grad_norm": 0.249532642718915,
+      "learning_rate": 2e-05,
+      "loss": 0.938,
+      "step": 25
+    },
+    {
+      "epoch": 0.78125,
+      "eval_loss": 0.9455437660217285,
+      "eval_runtime": 191.9941,
+      "eval_samples_per_second": 1.042,
+      "eval_steps_per_second": 0.13,
+      "step": 25
+    },
+    {
+      "epoch": 0.8125,
+      "grad_norm": 0.28034242585086017,
+      "learning_rate": 2e-05,
+      "loss": 0.9387,
+      "step": 26
+    },
+    {
+      "epoch": 0.8125,
+      "eval_loss": 0.93752121925354,
+      "eval_runtime": 195.4083,
+      "eval_samples_per_second": 1.023,
+      "eval_steps_per_second": 0.128,
+      "step": 26
+    },
+    {
+      "epoch": 0.84375,
+      "grad_norm": 0.2692565070546352,
+      "learning_rate": 2e-05,
+      "loss": 1.0474,
+      "step": 27
+    },
+    {
+      "epoch": 0.84375,
+      "eval_loss": 0.9300512075424194,
+      "eval_runtime": 195.3651,
+      "eval_samples_per_second": 1.024,
+      "eval_steps_per_second": 0.128,
+      "step": 27
+    },
+    {
+      "epoch": 0.875,
+      "grad_norm": 0.24705041646949316,
+      "learning_rate": 2e-05,
+      "loss": 0.9596,
+      "step": 28
+    },
+    {
+      "epoch": 0.875,
+      "eval_loss": 0.9226720929145813,
+      "eval_runtime": 195.4848,
+      "eval_samples_per_second": 1.023,
+      "eval_steps_per_second": 0.128,
+      "step": 28
+    },
+    {
+      "epoch": 0.90625,
+      "grad_norm": 0.24799871352606165,
+      "learning_rate": 2e-05,
+      "loss": 1.0172,
+      "step": 29
+    },
+    {
+      "epoch": 0.90625,
+      "eval_loss": 0.9159422516822815,
+      "eval_runtime": 196.157,
+      "eval_samples_per_second": 1.02,
+      "eval_steps_per_second": 0.127,
+      "step": 29
+    },
+    {
+      "epoch": 0.9375,
+      "grad_norm": 0.29755264040904106,
+      "learning_rate": 2e-05,
+      "loss": 0.9324,
+      "step": 30
+    },
+    {
+      "epoch": 0.9375,
+      "eval_loss": 0.9090733528137207,
+      "eval_runtime": 196.5295,
+      "eval_samples_per_second": 1.018,
+      "eval_steps_per_second": 0.127,
+      "step": 30
+    },
+    {
+      "epoch": 0.96875,
+      "grad_norm": 0.2629221961008751,
+      "learning_rate": 2e-05,
+      "loss": 0.9265,
+      "step": 31
+    },
+    {
+      "epoch": 0.96875,
+      "eval_loss": 0.9027940630912781,
+      "eval_runtime": 196.281,
+      "eval_samples_per_second": 1.019,
+      "eval_steps_per_second": 0.127,
+      "step": 31
+    },
+    {
+      "epoch": 1.0,
+      "grad_norm": 0.2901110704218056,
+      "learning_rate": 2e-05,
+      "loss": 0.9933,
+      "step": 32
+    },
+    {
+      "epoch": 1.0,
+      "eval_loss": 0.8970211148262024,
+      "eval_runtime": 190.2702,
+      "eval_samples_per_second": 1.051,
+      "eval_steps_per_second": 0.131,
+      "step": 32
+    },
+    {
+      "epoch": 1.03125,
+      "grad_norm": 0.27746608883483487,
+      "learning_rate": 2e-05,
+      "loss": 0.9339,
+      "step": 33
+    },
+    {
+      "epoch": 1.03125,
+      "eval_loss": 0.8916085958480835,
+      "eval_runtime": 189.4543,
+      "eval_samples_per_second": 1.056,
+      "eval_steps_per_second": 0.132,
+      "step": 33
+    },
+    {
+      "epoch": 1.0625,
+      "grad_norm": 0.26134437145600353,
+      "learning_rate": 2e-05,
+      "loss": 0.9438,
+      "step": 34
+    },
+    {
+      "epoch": 1.0625,
+      "eval_loss": 0.8867039680480957,
+      "eval_runtime": 189.6926,
+      "eval_samples_per_second": 1.054,
+      "eval_steps_per_second": 0.132,
+      "step": 34
+    },
+    {
+      "epoch": 1.09375,
+      "grad_norm": 0.252882507519195,
+      "learning_rate": 2e-05,
+      "loss": 0.8979,
+      "step": 35
+    },
+    {
+      "epoch": 1.09375,
+      "eval_loss": 0.8824067711830139,
+      "eval_runtime": 189.9217,
+      "eval_samples_per_second": 1.053,
+      "eval_steps_per_second": 0.132,
+      "step": 35
+    },
+    {
+      "epoch": 1.125,
+      "grad_norm": 0.25443025949474585,
+      "learning_rate": 2e-05,
+      "loss": 0.9411,
+      "step": 36
+    },
+    {
+      "epoch": 1.125,
+      "eval_loss": 0.8788293600082397,
+      "eval_runtime": 191.4083,
+      "eval_samples_per_second": 1.045,
+      "eval_steps_per_second": 0.131,
+      "step": 36
+    },
+    {
+      "epoch": 1.15625,
+      "grad_norm": 0.2559343621244427,
+      "learning_rate": 2e-05,
+      "loss": 0.9827,
+      "step": 37
+    },
+    {
+      "epoch": 1.15625,
+      "eval_loss": 0.8760793805122375,
+      "eval_runtime": 191.2732,
+      "eval_samples_per_second": 1.046,
+      "eval_steps_per_second": 0.131,
+      "step": 37
+    },
+    {
+      "epoch": 1.1875,
+      "grad_norm": 0.25403189851366254,
+      "learning_rate": 2e-05,
+      "loss": 0.8658,
+      "step": 38
+    },
+    {
+      "epoch": 1.1875,
+      "eval_loss": 0.8727380633354187,
+      "eval_runtime": 190.4281,
+      "eval_samples_per_second": 1.05,
+      "eval_steps_per_second": 0.131,
+      "step": 38
+    },
+    {
+      "epoch": 1.21875,
+      "grad_norm": 0.2493777578005398,
+      "learning_rate": 2e-05,
+      "loss": 1.0053,
+      "step": 39
+    },
+    {
+      "epoch": 1.21875,
+      "eval_loss": 0.869698703289032,
+      "eval_runtime": 190.3431,
+      "eval_samples_per_second": 1.051,
+      "eval_steps_per_second": 0.131,
+      "step": 39
+    },
+    {
+      "epoch": 1.25,
+      "grad_norm": 0.24823573574563138,
+      "learning_rate": 2e-05,
+      "loss": 0.8967,
+      "step": 40
+    },
+    {
+      "epoch": 1.25,
+      "eval_loss": 0.8664910793304443,
+      "eval_runtime": 189.9802,
+      "eval_samples_per_second": 1.053,
+      "eval_steps_per_second": 0.132,
+      "step": 40
+    },
+    {
+      "epoch": 1.28125,
+      "grad_norm": 0.25462243237743476,
+      "learning_rate": 2e-05,
+      "loss": 1.0064,
+      "step": 41
+    },
+    {
+      "epoch": 1.28125,
+      "eval_loss": 0.8638657927513123,
+      "eval_runtime": 195.3373,
+      "eval_samples_per_second": 1.024,
+      "eval_steps_per_second": 0.128,
+      "step": 41
+    },
+    {
+      "epoch": 1.3125,
+      "grad_norm": 0.2604089386111215,
+      "learning_rate": 2e-05,
+      "loss": 0.9898,
+      "step": 42
+    },
+    {
+      "epoch": 1.3125,
+      "eval_loss": 0.8607734441757202,
+      "eval_runtime": 195.219,
+      "eval_samples_per_second": 1.024,
+      "eval_steps_per_second": 0.128,
+      "step": 42
+    },
+    {
+      "epoch": 1.34375,
+      "grad_norm": 0.27139202440805793,
+      "learning_rate": 2e-05,
+      "loss": 1.0539,
+      "step": 43
+    },
+    {
+      "epoch": 1.34375,
+      "eval_loss": 0.8573687672615051,
+      "eval_runtime": 195.8828,
+      "eval_samples_per_second": 1.021,
+      "eval_steps_per_second": 0.128,
+      "step": 43
+    },
+    {
+      "epoch": 1.375,
+      "grad_norm": 0.27474433057157854,
+      "learning_rate": 2e-05,
+      "loss": 0.86,
+      "step": 44
+    },
+    {
+      "epoch": 1.375,
+      "eval_loss": 0.8537396192550659,
+      "eval_runtime": 194.9741,
+      "eval_samples_per_second": 1.026,
+      "eval_steps_per_second": 0.128,
+      "step": 44
+    },
+    {
+      "epoch": 1.40625,
+      "grad_norm": 0.2537208760747199,
+      "learning_rate": 2e-05,
+      "loss": 0.9562,
+      "step": 45
+    },
+    {
+      "epoch": 1.40625,
+      "eval_loss": 0.8497809767723083,
+      "eval_runtime": 194.9162,
+      "eval_samples_per_second": 1.026,
+      "eval_steps_per_second": 0.128,
+      "step": 45
+    },
+    {
+      "epoch": 1.4375,
+      "grad_norm": 0.27560461131090846,
+      "learning_rate": 2e-05,
+      "loss": 0.8767,
+      "step": 46
+    },
+    {
+      "epoch": 1.4375,
+      "eval_loss": 0.8458660244941711,
+      "eval_runtime": 195.1157,
+      "eval_samples_per_second": 1.025,
+      "eval_steps_per_second": 0.128,
+      "step": 46
+    },
+    {
+      "epoch": 1.46875,
+      "grad_norm": 0.2594536794112662,
+      "learning_rate": 2e-05,
+      "loss": 0.9256,
+      "step": 47
+    },
+    {
+      "epoch": 1.46875,
+      "eval_loss": 0.8429936766624451,
+      "eval_runtime": 195.048,
+      "eval_samples_per_second": 1.025,
+      "eval_steps_per_second": 0.128,
+      "step": 47
+    },
+    {
+      "epoch": 1.5,
+      "grad_norm": 0.28583207453838866,
+      "learning_rate": 2e-05,
+      "loss": 0.9858,
+      "step": 48
+    },
+    {
+      "epoch": 1.5,
+      "eval_loss": 0.84013831615448,
+      "eval_runtime": 194.3046,
+      "eval_samples_per_second": 1.029,
+      "eval_steps_per_second": 0.129,
+      "step": 48
+    },
+    {
+      "epoch": 1.53125,
+      "grad_norm": 0.28118976636788506,
+      "learning_rate": 2e-05,
+      "loss": 0.9158,
+      "step": 49
+    },
+    {
+      "epoch": 1.53125,
+      "eval_loss": 0.8369531035423279,
+      "eval_runtime": 194.521,
+      "eval_samples_per_second": 1.028,
+      "eval_steps_per_second": 0.129,
+      "step": 49
+    },
+    {
+      "epoch": 1.5625,
+      "grad_norm": 0.29276573776696546,
+      "learning_rate": 2e-05,
+      "loss": 0.8745,
+      "step": 50
+    },
+    {
+      "epoch": 1.5625,
+      "eval_loss": 0.8341982960700989,
+      "eval_runtime": 194.1114,
+      "eval_samples_per_second": 1.03,
+      "eval_steps_per_second": 0.129,
+      "step": 50
+    },
+    {
+      "epoch": 1.59375,
+      "grad_norm": 0.2860638141439372,
+      "learning_rate": 2e-05,
+      "loss": 0.854,
+      "step": 51
+    },
+    {
+      "epoch": 1.59375,
+      "eval_loss": 0.8317239284515381,
+      "eval_runtime": 198.1029,
+      "eval_samples_per_second": 1.01,
+      "eval_steps_per_second": 0.126,
+      "step": 51
+    },
+    {
+      "epoch": 1.625,
+      "grad_norm": 0.29960349722496704,
+      "learning_rate": 2e-05,
+      "loss": 0.8399,
+      "step": 52
+    },
+    {
+      "epoch": 1.625,
+      "eval_loss": 0.8290513753890991,
+      "eval_runtime": 198.0764,
+      "eval_samples_per_second": 1.01,
+      "eval_steps_per_second": 0.126,
+      "step": 52
+    },
+    {
+      "epoch": 1.65625,
+      "grad_norm": 0.2964234305808419,
+      "learning_rate": 2e-05,
+      "loss": 0.9694,
+      "step": 53
+    },
+    {
+      "epoch": 1.65625,
+      "eval_loss": 0.8267760276794434,
+      "eval_runtime": 197.8284,
+      "eval_samples_per_second": 1.011,
+      "eval_steps_per_second": 0.126,
+      "step": 53
+    },
+    {
+      "epoch": 1.6875,
+      "grad_norm": 0.26183932644077784,
+      "learning_rate": 2e-05,
+      "loss": 0.8153,
+      "step": 54
+    },
+    {
+      "epoch": 1.6875,
+      "eval_loss": 0.824044942855835,
+      "eval_runtime": 198.1694,
+      "eval_samples_per_second": 1.009,
+      "eval_steps_per_second": 0.126,
+      "step": 54
+    },
+    {
+      "epoch": 1.71875,
+      "grad_norm": 0.3067024314453144,
+      "learning_rate": 2e-05,
+      "loss": 0.883,
+      "step": 55
+    },
+    {
+      "epoch": 1.71875,
+      "eval_loss": 0.8216392397880554,
+      "eval_runtime": 198.0249,
+      "eval_samples_per_second": 1.01,
+      "eval_steps_per_second": 0.126,
+      "step": 55
+    },
+    {
+      "epoch": 1.75,
+      "grad_norm": 0.27888658705355013,
+      "learning_rate": 2e-05,
+      "loss": 0.8771,
+      "step": 56
+    },
+    {
+      "epoch": 1.75,
+      "eval_loss": 0.8194215297698975,
+      "eval_runtime": 195.4688,
+      "eval_samples_per_second": 1.023,
+      "eval_steps_per_second": 0.128,
+      "step": 56
+    },
+    {
+      "epoch": 1.78125,
+      "grad_norm": 0.32571765544245934,
+      "learning_rate": 2e-05,
+      "loss": 0.897,
+      "step": 57
+    },
+    {
+      "epoch": 1.78125,
+      "eval_loss": 0.8167170882225037,
+      "eval_runtime": 189.6243,
+      "eval_samples_per_second": 1.055,
+      "eval_steps_per_second": 0.132,
+      "step": 57
+    },
+    {
+      "epoch": 1.8125,
+      "grad_norm": 0.292216058855145,
+      "learning_rate": 2e-05,
+      "loss": 0.9277,
+      "step": 58
+    },
+    {
+      "epoch": 1.8125,
+      "eval_loss": 0.8145509958267212,
+      "eval_runtime": 190.2429,
+      "eval_samples_per_second": 1.051,
+      "eval_steps_per_second": 0.131,
+      "step": 58
+    },
+    {
+      "epoch": 1.84375,
+      "grad_norm": 0.29002612820437024,
+      "learning_rate": 2e-05,
+      "loss": 0.8971,
+      "step": 59
+    },
+    {
+      "epoch": 1.84375,
+      "eval_loss": 0.8122230768203735,
+      "eval_runtime": 189.9403,
+      "eval_samples_per_second": 1.053,
+      "eval_steps_per_second": 0.132,
+      "step": 59
+    },
+    {
+      "epoch": 1.875,
+      "grad_norm": 0.2926088029288858,
+      "learning_rate": 2e-05,
+      "loss": 0.9225,
+      "step": 60
+    },
+    {
+      "epoch": 1.875,
+      "eval_loss": 0.8100479245185852,
+      "eval_runtime": 190.2569,
+      "eval_samples_per_second": 1.051,
+      "eval_steps_per_second": 0.131,
+      "step": 60
+    },
+    {
+      "epoch": 1.90625,
+      "grad_norm": 0.30068993111077397,
+      "learning_rate": 2e-05,
+      "loss": 0.9134,
+      "step": 61
+    },
+    {
+      "epoch": 1.90625,
+      "eval_loss": 0.808087944984436,
+      "eval_runtime": 192.4896,
+      "eval_samples_per_second": 1.039,
+      "eval_steps_per_second": 0.13,
+      "step": 61
+    },
+    {
+      "epoch": 1.9375,
+      "grad_norm": 0.3157573686768343,
+      "learning_rate": 2e-05,
+      "loss": 0.8965,
+      "step": 62
+    },
+    {
+      "epoch": 1.9375,
+      "eval_loss": 0.8057371377944946,
+      "eval_runtime": 190.0158,
+      "eval_samples_per_second": 1.053,
+      "eval_steps_per_second": 0.132,
+      "step": 62
+    },
+    {
+      "epoch": 1.96875,
+      "grad_norm": 0.31215592754506605,
+      "learning_rate": 2e-05,
+      "loss": 0.7828,
+      "step": 63
+    },
+    {
+      "epoch": 1.96875,
+      "eval_loss": 0.8031384944915771,
+      "eval_runtime": 189.5204,
+      "eval_samples_per_second": 1.055,
+      "eval_steps_per_second": 0.132,
+      "step": 63
+    },
+    {
+      "epoch": 2.0,
+      "grad_norm": 0.29422828766227993,
+      "learning_rate": 2e-05,
+      "loss": 0.8196,
+      "step": 64
+    },
+    {
+      "epoch": 2.0,
+      "eval_loss": 0.8012601733207703,
+      "eval_runtime": 189.7041,
+      "eval_samples_per_second": 1.054,
+      "eval_steps_per_second": 0.132,
+      "step": 64
+    },
+    {
+      "epoch": 2.03125,
+      "grad_norm": 0.2885449518895793,
+      "learning_rate": 2e-05,
+      "loss": 0.9715,
+      "step": 65
+    },
+    {
+      "epoch": 2.03125,
+      "eval_loss": 0.8001161813735962,
+      "eval_runtime": 189.57,
+      "eval_samples_per_second": 1.055,
+      "eval_steps_per_second": 0.132,
+      "step": 65
+    },
+    {
+      "epoch": 2.0625,
+      "grad_norm": 0.30260184063348483,
+      "learning_rate": 2e-05,
+      "loss": 0.7912,
+      "step": 66
+    },
+    {
+      "epoch": 2.0625,
+      "eval_loss": 0.7989436388015747,
+      "eval_runtime": 193.0193,
+      "eval_samples_per_second": 1.036,
+      "eval_steps_per_second": 0.13,
+      "step": 66
+    },
+    {
+      "epoch": 2.09375,
+      "grad_norm": 0.32650294605024255,
+      "learning_rate": 2e-05,
+      "loss": 0.8176,
+      "step": 67
+    },
+    {
+      "epoch": 2.09375,
+      "eval_loss": 0.7972333431243896,
+      "eval_runtime": 193.2225,
+      "eval_samples_per_second": 1.035,
+      "eval_steps_per_second": 0.129,
+      "step": 67
+    },
+    {
+      "epoch": 2.125,
+      "grad_norm": 0.3382679480741134,
+      "learning_rate": 2e-05,
+      "loss": 0.8141,
+      "step": 68
+    },
+    {
+      "epoch": 2.125,
+      "eval_loss": 0.7950598001480103,
+      "eval_runtime": 193.2781,
+      "eval_samples_per_second": 1.035,
+      "eval_steps_per_second": 0.129,
+      "step": 68
+    },
+    {
+      "epoch": 2.15625,
+      "grad_norm": 0.3094090784935889,
+      "learning_rate": 2e-05,
+      "loss": 0.796,
+      "step": 69
+    },
+    {
+      "epoch": 2.15625,
+      "eval_loss": 0.7932476997375488,
+      "eval_runtime": 193.0282,
+      "eval_samples_per_second": 1.036,
+      "eval_steps_per_second": 0.13,
+      "step": 69
+    },
+    {
+      "epoch": 2.1875,
+      "grad_norm": 0.30209558834780514,
+      "learning_rate": 2e-05,
+      "loss": 0.927,
+      "step": 70
+    },
+    {
+      "epoch": 2.1875,
+      "eval_loss": 0.7922118902206421,
+      "eval_runtime": 192.8284,
+      "eval_samples_per_second": 1.037,
+      "eval_steps_per_second": 0.13,
+      "step": 70
+    },
+    {
+      "epoch": 2.21875,
+      "grad_norm": 0.35958652905266686,
+      "learning_rate": 2e-05,
+      "loss": 0.8224,
+      "step": 71
+    },
+    {
+      "epoch": 2.21875,
+      "eval_loss": 0.7909810543060303,
+      "eval_runtime": 202.3128,
+      "eval_samples_per_second": 0.989,
+      "eval_steps_per_second": 0.124,
+      "step": 71
+    },
+    {
+      "epoch": 2.25,
+      "grad_norm": 0.356338004067507,
+      "learning_rate": 2e-05,
+      "loss": 0.8376,
+      "step": 72
+    },
+    {
+      "epoch": 2.25,
+      "eval_loss": 0.7894656658172607,
+      "eval_runtime": 195.1481,
+      "eval_samples_per_second": 1.025,
+      "eval_steps_per_second": 0.128,
+      "step": 72
+    },
+    {
+      "epoch": 2.28125,
+      "grad_norm": 0.31886905989727465,
+      "learning_rate": 2e-05,
+      "loss": 0.8688,
+      "step": 73
+    },
+    {
+      "epoch": 2.28125,
+      "eval_loss": 0.7889463901519775,
+      "eval_runtime": 194.8418,
+      "eval_samples_per_second": 1.026,
+      "eval_steps_per_second": 0.128,
+      "step": 73
+    },
+    {
+      "epoch": 2.3125,
+      "grad_norm": 0.35606342918056466,
+      "learning_rate": 2e-05,
+      "loss": 0.835,
+      "step": 74
+    },
+    {
+      "epoch": 2.3125,
+      "eval_loss": 0.7875587344169617,
+      "eval_runtime": 194.7701,
+      "eval_samples_per_second": 1.027,
+      "eval_steps_per_second": 0.128,
+      "step": 74
+    },
+    {
+      "epoch": 2.34375,
+      "grad_norm": 0.3161858862696026,
+      "learning_rate": 2e-05,
+      "loss": 0.873,
+      "step": 75
+    },
+    {
+      "epoch": 2.34375,
+      "eval_loss": 0.7863460779190063,
+      "eval_runtime": 195.4811,
+      "eval_samples_per_second": 1.023,
+      "eval_steps_per_second": 0.128,
+      "step": 75
+    },
+    {
+      "epoch": 2.375,
+      "grad_norm": 0.35771781884741477,
+      "learning_rate": 2e-05,
+      "loss": 0.9021,
+      "step": 76
+    },
+    {
+      "epoch": 2.375,
+      "eval_loss": 0.7847577929496765,
+      "eval_runtime": 195.3724,
+      "eval_samples_per_second": 1.024,
+      "eval_steps_per_second": 0.128,
+      "step": 76
+    },
+    {
+      "epoch": 2.40625,
+      "grad_norm": 0.3549789155823785,
+      "learning_rate": 2e-05,
+      "loss": 0.9195,
+      "step": 77
+    },
+    {
+      "epoch": 2.40625,
+      "eval_loss": 0.783415675163269,
+      "eval_runtime": 190.226,
+      "eval_samples_per_second": 1.051,
+      "eval_steps_per_second": 0.131,
+      "step": 77
+    },
+    {
+      "epoch": 2.4375,
+      "grad_norm": 0.34734314309709374,
+      "learning_rate": 2e-05,
+      "loss": 0.8386,
+      "step": 78
+    },
+    {
+      "epoch": 2.4375,
+      "eval_loss": 0.7814657688140869,
+      "eval_runtime": 190.3177,
+      "eval_samples_per_second": 1.051,
+      "eval_steps_per_second": 0.131,
+      "step": 78
+    },
+    {
+      "epoch": 2.46875,
+      "grad_norm": 0.35540762574897183,
+      "learning_rate": 2e-05,
+      "loss": 0.851,
+      "step": 79
+    },
+    {
+      "epoch": 2.46875,
+      "eval_loss": 0.7798058390617371,
+      "eval_runtime": 190.7447,
+      "eval_samples_per_second": 1.049,
+      "eval_steps_per_second": 0.131,
+      "step": 79
+    },
+    {
+      "epoch": 2.5,
+      "grad_norm": 0.3844458514717174,
+      "learning_rate": 2e-05,
+      "loss": 0.767,
+      "step": 80
+    },
+    {
+      "epoch": 2.5,
+      "eval_loss": 0.7777827978134155,
+      "eval_runtime": 190.8548,
+      "eval_samples_per_second": 1.048,
+      "eval_steps_per_second": 0.131,
+      "step": 80
+    },
+    {
+      "epoch": 2.53125,
+      "grad_norm": 0.36232344175264375,
+      "learning_rate": 2e-05,
+      "loss": 0.8508,
+      "step": 81
+    },
+    {
+      "epoch": 2.53125,
+      "eval_loss": 0.7763205170631409,
+      "eval_runtime": 190.335,
+      "eval_samples_per_second": 1.051,
+      "eval_steps_per_second": 0.131,
+      "step": 81
+    },
+    {
+      "epoch": 2.5625,
+      "grad_norm": 0.36279843147857743,
+      "learning_rate": 2e-05,
+      "loss": 0.8331,
+      "step": 82
+    },
+    {
+      "epoch": 2.5625,
+      "eval_loss": 0.7757676839828491,
+      "eval_runtime": 190.9559,
+      "eval_samples_per_second": 1.047,
+      "eval_steps_per_second": 0.131,
+      "step": 82
+    },
+    {
+      "epoch": 2.59375,
+      "grad_norm": 0.395360566032837,
+      "learning_rate": 2e-05,
+      "loss": 0.847,
+      "step": 83
+    },
+    {
+      "epoch": 2.59375,
+      "eval_loss": 0.7743326425552368,
+      "eval_runtime": 190.6372,
+      "eval_samples_per_second": 1.049,
+      "eval_steps_per_second": 0.131,
+      "step": 83
+    },
+    {
+      "epoch": 2.625,
+      "grad_norm": 0.4268568783791123,
+      "learning_rate": 2e-05,
+      "loss": 0.869,
+      "step": 84
+    },
+    {
+      "epoch": 2.625,
+      "eval_loss": 0.772053062915802,
+      "eval_runtime": 190.2072,
+      "eval_samples_per_second": 1.051,
+      "eval_steps_per_second": 0.131,
+      "step": 84
+    },
+    {
+      "epoch": 2.65625,
+      "grad_norm": 0.3581495538253167,
+      "learning_rate": 2e-05,
+      "loss": 0.8591,
+      "step": 85
+    },
+    {
+      "epoch": 2.65625,
+      "eval_loss": 0.7711917757987976,
+      "eval_runtime": 190.4392,
+      "eval_samples_per_second": 1.05,
+      "eval_steps_per_second": 0.131,
+      "step": 85
+    },
+    {
+      "epoch": 2.6875,
+      "grad_norm": 0.3952841797586726,
+      "learning_rate": 2e-05,
+      "loss": 0.8167,
+      "step": 86
+    },
+    {
+      "epoch": 2.6875,
+      "eval_loss": 0.7714033722877502,
+      "eval_runtime": 193.8038,
+      "eval_samples_per_second": 1.032,
+      "eval_steps_per_second": 0.129,
+      "step": 86
+    },
+    {
+      "epoch": 2.71875,
+      "grad_norm": 0.41820009905687616,
+      "learning_rate": 2e-05,
+      "loss": 0.8165,
+      "step": 87
+    },
+    {
+      "epoch": 2.71875,
+      "eval_loss": 0.771486759185791,
+      "eval_runtime": 194.4791,
+      "eval_samples_per_second": 1.028,
+      "eval_steps_per_second": 0.129,
+      "step": 87
+    },
+    {
+      "epoch": 2.75,
+      "grad_norm": 0.3852566717747202,
+      "learning_rate": 2e-05,
+      "loss": 0.8459,
+      "step": 88
+    },
+    {
+      "epoch": 2.75,
+      "eval_loss": 0.7710732817649841,
+      "eval_runtime": 194.3404,
+      "eval_samples_per_second": 1.029,
+      "eval_steps_per_second": 0.129,
+      "step": 88
+    },
+    {
+      "epoch": 2.78125,
+      "grad_norm": 0.39909292055831935,
+      "learning_rate": 2e-05,
+      "loss": 0.8945,
+      "step": 89
+    },
+    {
+      "epoch": 2.78125,
+      "eval_loss": 0.7708308696746826,
+      "eval_runtime": 194.4483,
+      "eval_samples_per_second": 1.029,
+      "eval_steps_per_second": 0.129,
+      "step": 89
+    },
+    {
+      "epoch": 2.8125,
+      "grad_norm": 0.3916487629667217,
+      "learning_rate": 2e-05,
+      "loss": 0.8029,
+      "step": 90
+    },
+    {
+      "epoch": 2.8125,
+      "eval_loss": 0.7713395953178406,
+      "eval_runtime": 194.6045,
+      "eval_samples_per_second": 1.028,
+      "eval_steps_per_second": 0.128,
+      "step": 90
+    },
+    {
+      "epoch": 2.84375,
+      "grad_norm": 0.36969072235715195,
+      "learning_rate": 2e-05,
+      "loss": 0.7704,
+      "step": 91
+    },
+    {
+      "epoch": 2.84375,
+      "eval_loss": 0.7713618278503418,
+      "eval_runtime": 194.3895,
+      "eval_samples_per_second": 1.029,
+      "eval_steps_per_second": 0.129,
+      "step": 91
+    },
+    {
+      "epoch": 2.875,
+      "grad_norm": 0.3853248559868725,
+      "learning_rate": 2e-05,
+      "loss": 0.8247,
+      "step": 92
+    },
+    {
+      "epoch": 2.875,
+      "eval_loss": 0.7703633308410645,
+      "eval_runtime": 194.0457,
+      "eval_samples_per_second": 1.031,
+      "eval_steps_per_second": 0.129,
+      "step": 92
+    },
+    {
+      "epoch": 2.90625,
+      "grad_norm": 0.38111471762069055,
+      "learning_rate": 2e-05,
+      "loss": 0.855,
+      "step": 93
+    },
+    {
+      "epoch": 2.90625,
+      "eval_loss": 0.7690189480781555,
+      "eval_runtime": 194.3506,
+      "eval_samples_per_second": 1.029,
+      "eval_steps_per_second": 0.129,
+      "step": 93
+    },
+    {
+      "epoch": 2.9375,
+      "grad_norm": 0.3701270310997752,
+      "learning_rate": 2e-05,
+      "loss": 0.7518,
+      "step": 94
+    },
+    {
+      "epoch": 2.9375,
+      "eval_loss": 0.7675644159317017,
+      "eval_runtime": 194.3756,
+      "eval_samples_per_second": 1.029,
+      "eval_steps_per_second": 0.129,
+      "step": 94
+    },
+    {
+      "epoch": 2.96875,
+      "grad_norm": 0.40489524752286055,
+      "learning_rate": 2e-05,
+      "loss": 0.8559,
+      "step": 95
+    },
+    {
+      "epoch": 2.96875,
+      "eval_loss": 0.766002357006073,
+      "eval_runtime": 193.9472,
+      "eval_samples_per_second": 1.031,
+      "eval_steps_per_second": 0.129,
+      "step": 95
+    },
+    {
+      "epoch": 3.0,
+      "grad_norm": 0.39220887464051457,
+      "learning_rate": 2e-05,
+      "loss": 0.8629,
+      "step": 96
+    },
+    {
+      "epoch": 3.0,
+      "eval_loss": 0.7644355893135071,
+      "eval_runtime": 196.1686,
+      "eval_samples_per_second": 1.02,
+      "eval_steps_per_second": 0.127,
+      "step": 96
+    },
+    {
+      "epoch": 3.03125,
+      "grad_norm": 0.3644925708419195,
+      "learning_rate": 2e-05,
+      "loss": 0.7434,
+      "step": 97
+    },
+    {
+      "epoch": 3.03125,
+      "eval_loss": 0.7628399133682251,
+      "eval_runtime": 196.1515,
+      "eval_samples_per_second": 1.02,
+      "eval_steps_per_second": 0.127,
+      "step": 97
+    },
+    {
+      "epoch": 3.0625,
+      "grad_norm": 0.407089942317534,
+      "learning_rate": 2e-05,
+      "loss": 0.8038,
+      "step": 98
+    },
+    {
+      "epoch": 3.0625,
+      "eval_loss": 0.7609645128250122,
+      "eval_runtime": 196.8662,
+      "eval_samples_per_second": 1.016,
+      "eval_steps_per_second": 0.127,
+      "step": 98
+    },
+    {
+      "epoch": 3.09375,
+      "grad_norm": 0.38849177572880716,
+      "learning_rate": 2e-05,
+      "loss": 0.8106,
+      "step": 99
+    },
+    {
+      "epoch": 3.09375,
+      "eval_loss": 0.7598288059234619,
+      "eval_runtime": 196.1846,
+      "eval_samples_per_second": 1.019,
+      "eval_steps_per_second": 0.127,
+      "step": 99
+    },
+    {
+      "epoch": 3.125,
+      "grad_norm": 0.41885563528617265,
+      "learning_rate": 2e-05,
+      "loss": 0.808,
+      "step": 100
+    },
+    {
+      "epoch": 3.125,
+      "eval_loss": 0.7587143778800964,
+      "eval_runtime": 195.7296,
+      "eval_samples_per_second": 1.022,
+      "eval_steps_per_second": 0.128,
+      "step": 100
+    },
+    {
+      "epoch": 3.15625,
+      "grad_norm": 0.4003909227323588,
+      "learning_rate": 2e-05,
+      "loss": 0.791,
+      "step": 101
+    },
+    {
+      "epoch": 3.15625,
+      "eval_loss": 0.7578326463699341,
+      "eval_runtime": 195.2831,
+      "eval_samples_per_second": 1.024,
+      "eval_steps_per_second": 0.128,
+      "step": 101
+    },
+    {
+      "epoch": 3.1875,
+      "grad_norm": 0.4014550365826672,
+      "learning_rate": 2e-05,
+      "loss": 0.7402,
+      "step": 102
+    },
+    {
+      "epoch": 3.1875,
+      "eval_loss": 0.7573958039283752,
+      "eval_runtime": 189.5234,
+      "eval_samples_per_second": 1.055,
+      "eval_steps_per_second": 0.132,
+      "step": 102
+    },
+    {
+      "epoch": 3.21875,
+      "grad_norm": 0.4018554316691014,
+      "learning_rate": 2e-05,
+      "loss": 0.8165,
+      "step": 103
+    },
+    {
+      "epoch": 3.21875,
+      "eval_loss": 0.7571737766265869,
+      "eval_runtime": 190.0146,
+      "eval_samples_per_second": 1.053,
+      "eval_steps_per_second": 0.132,
+      "step": 103
+    },
+    {
+      "epoch": 3.25,
+      "grad_norm": 0.39691385018938347,
+      "learning_rate": 2e-05,
+      "loss": 0.7806,
+      "step": 104
+    },
+    {
+      "epoch": 3.25,
+      "eval_loss": 0.7581367492675781,
+      "eval_runtime": 190.1851,
+      "eval_samples_per_second": 1.052,
+      "eval_steps_per_second": 0.131,
+      "step": 104
+    },
+    {
+      "epoch": 3.28125,
+      "grad_norm": 0.390373263306042,
+      "learning_rate": 2e-05,
+      "loss": 0.7454,
+      "step": 105
+    },
+    {
+      "epoch": 3.28125,
+      "eval_loss": 0.7590533494949341,
+      "eval_runtime": 190.1255,
+      "eval_samples_per_second": 1.052,
+      "eval_steps_per_second": 0.131,
+      "step": 105
+    },
+    {
+      "epoch": 3.3125,
+      "grad_norm": 0.45093404603350434,
+      "learning_rate": 2e-05,
+      "loss": 0.8598,
+      "step": 106
+    },
+    {
+      "epoch": 3.3125,
+      "eval_loss": 0.7584137916564941,
+      "eval_runtime": 193.5956,
+      "eval_samples_per_second": 1.033,
+      "eval_steps_per_second": 0.129,
+      "step": 106
+    },
+    {
+      "epoch": 3.34375,
+      "grad_norm": 0.4112664411035318,
+      "learning_rate": 2e-05,
+      "loss": 0.8612,
+      "step": 107
+    },
+    {
+      "epoch": 3.34375,
+      "eval_loss": 0.757759690284729,
+      "eval_runtime": 191.7864,
+      "eval_samples_per_second": 1.043,
+      "eval_steps_per_second": 0.13,
+      "step": 107
+    },
+    {
+      "epoch": 3.375,
+      "grad_norm": 0.4158875890717671,
+      "learning_rate": 2e-05,
+      "loss": 0.7916,
+      "step": 108
+    },
+    {
+      "epoch": 3.375,
+      "eval_loss": 0.756908655166626,
+      "eval_runtime": 190.1833,
+      "eval_samples_per_second": 1.052,
+      "eval_steps_per_second": 0.131,
+      "step": 108
+    },
+    {
+      "epoch": 3.40625,
+      "grad_norm": 0.4234644828447959,
+      "learning_rate": 2e-05,
+      "loss": 0.798,
+      "step": 109
+    },
+    {
+      "epoch": 3.40625,
+      "eval_loss": 0.7559736371040344,
+      "eval_runtime": 190.0515,
+      "eval_samples_per_second": 1.052,
+      "eval_steps_per_second": 0.132,
+      "step": 109
+    },
+    {
+      "epoch": 3.4375,
+      "grad_norm": 0.480506693884699,
+      "learning_rate": 2e-05,
+      "loss": 0.7964,
+      "step": 110
+    },
+    {
+      "epoch": 3.4375,
+      "eval_loss": 0.7547717094421387,
+      "eval_runtime": 189.9579,
+      "eval_samples_per_second": 1.053,
+      "eval_steps_per_second": 0.132,
+      "step": 110
+    },
+    {
+      "epoch": 3.46875,
+      "grad_norm": 0.4540166631930203,
+      "learning_rate": 2e-05,
+      "loss": 0.8172,
+      "step": 111
+    },
+    {
+      "epoch": 3.46875,
+      "eval_loss": 0.7538467645645142,
+      "eval_runtime": 194.2136,
+      "eval_samples_per_second": 1.03,
+      "eval_steps_per_second": 0.129,
+      "step": 111
+    },
+    {
+      "epoch": 3.5,
+      "grad_norm": 0.46782505650509537,
+      "learning_rate": 2e-05,
+      "loss": 0.7826,
+      "step": 112
+    },
+    {
+      "epoch": 3.5,
+      "eval_loss": 0.7537503838539124,
+      "eval_runtime": 194.396,
+      "eval_samples_per_second": 1.029,
+      "eval_steps_per_second": 0.129,
+      "step": 112
+    },
+    {
+      "epoch": 3.53125,
+      "grad_norm": 0.5212465522615498,
+      "learning_rate": 2e-05,
+      "loss": 0.7715,
+      "step": 113
+    },
+    {
+      "epoch": 3.53125,
+      "eval_loss": 0.7527998089790344,
+      "eval_runtime": 193.1123,
+      "eval_samples_per_second": 1.036,
+      "eval_steps_per_second": 0.129,
+      "step": 113
+    },
+    {
+      "epoch": 3.5625,
+      "grad_norm": 0.4869709286188453,
+      "learning_rate": 2e-05,
+      "loss": 0.8007,
+      "step": 114
+    },
+    {
+      "epoch": 3.5625,
+      "eval_loss": 0.7510444521903992,
+      "eval_runtime": 193.3613,
+      "eval_samples_per_second": 1.034,
+      "eval_steps_per_second": 0.129,
+      "step": 114
+    },
+    {
+      "epoch": 3.59375,
+      "grad_norm": 0.45218808224844204,
+      "learning_rate": 2e-05,
+      "loss": 0.7183,
+      "step": 115
+    },
+    {
+      "epoch": 3.59375,
+      "eval_loss": 0.7502281069755554,
+      "eval_runtime": 192.9756,
+      "eval_samples_per_second": 1.036,
+      "eval_steps_per_second": 0.13,
+      "step": 115
+    }
+  ],
+  "logging_steps": 1.0,
+  "max_steps": 128,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 4,
+  "save_steps": 5,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 254953078521856.0,
+  "train_batch_size": 8,
+  "trial_name": null,
+  "trial_params": null
+}

checkpoint-115/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bad0ba8130e4e14127bba6f309b7c1a7c64fbb53a44b7e0d241978c150aa92ea
+size 8312

checkpoint-115/zero_to_fp32.py

@@ -0,0 +1,604 @@
+#!/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, exclude_frozen_parameters):
+    """
+    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,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+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,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _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,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _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, exclude_frozen_parameters=False):
+    """
+    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``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+
+    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, exclude_frozen_parameters)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None, exclude_frozen_parameters=False):
+    """
+    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``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag, exclude_frozen_parameters)
+    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("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    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,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)