Upload checkpoints

adapter_config.json

@@ -0,0 +1,29 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "openbmb/MiniCPM-2B-dpo-bf16",
+  "bias": "none",
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": "gaussian",
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 8,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "v_proj",
+    "q_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "use_dora": false,
+  "use_rslora": false
+}

adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bc12271f98dc5360edbd7b573143dee888c2dcd38b2bfc2fd39140291875ebb8
+size 5919456

checkpoints/checkpoint-1000/README.md

@@ -0,0 +1,202 @@
+---
+library_name: peft
+base_model: openbmb/MiniCPM-2B-dpo-bf16
+---
+
+# 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 0.10.0

checkpoints/checkpoint-1000/adapter_config.json

@@ -0,0 +1,29 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "openbmb/MiniCPM-2B-dpo-bf16",
+  "bias": "none",
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": "gaussian",
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 8,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "v_proj",
+    "q_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "use_dora": false,
+  "use_rslora": false
+}

checkpoints/checkpoint-1000/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9d0b71af557e40475bbd0ed5580b1163fc1e5813e6f18549c162fde489d38b7f
+size 5919456

checkpoints/checkpoint-1000/global_step1000/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:b4a4c216a05e414bd591cc3e065a79ed8d7d527e52eb808fce6d214c43d12a8a
+size 35393392

checkpoints/checkpoint-1000/global_step1000/zero_pp_rank_0_mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3394fcb77dc912f199fe9c2ebfb5a65dd81faa98b7bd60c9a238314ef859fcee
+size 183275

checkpoints/checkpoint-1000/latest

@@ -0,0 +1 @@
+global_step1000

checkpoints/checkpoint-1000/rng_state.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e8e0d9241eef1fbbb5da2f0e430c8a311fd2376d925c255c8ae5c0b1105086c8
+size 14244

checkpoints/checkpoint-1000/scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2fac7b4ec7da78a13fae6d5e24455e41750b089d8c7eca1f0a65f95de6385bb8
+size 1064

checkpoints/checkpoint-1000/special_tokens_map.json

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

checkpoints/checkpoint-1000/tokenizer_config.json

@@ -0,0 +1,42 @@
+{
+  "add_bos_token": true,
+  "add_eos_token": false,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<s>",
+  "chat_template": "{% for message in messages %}{% if message['role'] == 'user' %}{{'<用户>' + message['content'].strip() + '<AI>'}}{% else %}{{message['content'].strip()}}{% endif %}{% endfor %}",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "</s>",
+  "legacy": true,
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "</s>",
+  "sp_model_kwargs": {},
+  "spaces_between_special_tokens": false,
+  "tokenizer_class": "LlamaTokenizer",
+  "unk_token": "<unk>",
+  "use_default_system_prompt": false
+}

checkpoints/checkpoint-1000/trainer_state.json

@@ -0,0 +1,737 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.5350454788657036,
+  "eval_steps": 500,
+  "global_step": 1000,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.01,
+      "grad_norm": 7.3019086777975994,
+      "learning_rate": 5e-06,
+      "loss": 0.6939,
+      "step": 10
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 5.150989333266983,
+      "learning_rate": 1e-05,
+      "loss": 0.7167,
+      "step": 20
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 3.640558809610037,
+      "learning_rate": 1.5e-05,
+      "loss": 0.5683,
+      "step": 30
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 7.517999397731128,
+      "learning_rate": 2e-05,
+      "loss": 0.5472,
+      "step": 40
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 1.9687061679463425,
+      "learning_rate": 2.5e-05,
+      "loss": 0.4439,
+      "step": 50
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 3.643479206523606,
+      "learning_rate": 3e-05,
+      "loss": 0.2486,
+      "step": 60
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 2.2754773308695095,
+      "learning_rate": 3.5e-05,
+      "loss": 0.2217,
+      "step": 70
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 1.7144730049127388,
+      "learning_rate": 4e-05,
+      "loss": 0.169,
+      "step": 80
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 3.4702829704135114,
+      "learning_rate": 4.5e-05,
+      "loss": 0.1994,
+      "step": 90
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 1.3985127340985621,
+      "learning_rate": 5e-05,
+      "loss": 0.1612,
+      "step": 100
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 1.375992184386137,
+      "learning_rate": 4.982758620689655e-05,
+      "loss": 0.1576,
+      "step": 110
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 1.9528635753013313,
+      "learning_rate": 4.9655172413793107e-05,
+      "loss": 0.1393,
+      "step": 120
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 4.075169010198401,
+      "learning_rate": 4.9482758620689655e-05,
+      "loss": 0.1969,
+      "step": 130
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 2.0953991165751207,
+      "learning_rate": 4.931034482758621e-05,
+      "loss": 0.1294,
+      "step": 140
+    },
+    {
+      "epoch": 0.08,
+      "grad_norm": 1.942660591044849,
+      "learning_rate": 4.913793103448276e-05,
+      "loss": 0.1306,
+      "step": 150
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 3.1508904015728345,
+      "learning_rate": 4.896551724137931e-05,
+      "loss": 0.1526,
+      "step": 160
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 1.9862795165358471,
+      "learning_rate": 4.8793103448275864e-05,
+      "loss": 0.1186,
+      "step": 170
+    },
+    {
+      "epoch": 0.1,
+      "grad_norm": 2.633061833817991,
+      "learning_rate": 4.862068965517241e-05,
+      "loss": 0.1457,
+      "step": 180
+    },
+    {
+      "epoch": 0.1,
+      "grad_norm": 1.8017052368446178,
+      "learning_rate": 4.844827586206897e-05,
+      "loss": 0.1234,
+      "step": 190
+    },
+    {
+      "epoch": 0.11,
+      "grad_norm": 2.1560694100709803,
+      "learning_rate": 4.827586206896552e-05,
+      "loss": 0.1346,
+      "step": 200
+    },
+    {
+      "epoch": 0.11,
+      "grad_norm": 1.5737689267430703,
+      "learning_rate": 4.810344827586207e-05,
+      "loss": 0.116,
+      "step": 210
+    },
+    {
+      "epoch": 0.12,
+      "grad_norm": 1.957864677854788,
+      "learning_rate": 4.793103448275863e-05,
+      "loss": 0.1692,
+      "step": 220
+    },
+    {
+      "epoch": 0.12,
+      "grad_norm": 2.215039223521855,
+      "learning_rate": 4.7758620689655176e-05,
+      "loss": 0.1245,
+      "step": 230
+    },
+    {
+      "epoch": 0.13,
+      "grad_norm": 1.370517239734168,
+      "learning_rate": 4.7586206896551725e-05,
+      "loss": 0.1476,
+      "step": 240
+    },
+    {
+      "epoch": 0.13,
+      "grad_norm": 1.7341334563022532,
+      "learning_rate": 4.741379310344828e-05,
+      "loss": 0.1236,
+      "step": 250
+    },
+    {
+      "epoch": 0.14,
+      "grad_norm": 1.5994298113068974,
+      "learning_rate": 4.724137931034483e-05,
+      "loss": 0.1161,
+      "step": 260
+    },
+    {
+      "epoch": 0.14,
+      "grad_norm": 1.5317433190951963,
+      "learning_rate": 4.7068965517241385e-05,
+      "loss": 0.1035,
+      "step": 270
+    },
+    {
+      "epoch": 0.15,
+      "grad_norm": 2.191977732539556,
+      "learning_rate": 4.689655172413793e-05,
+      "loss": 0.1427,
+      "step": 280
+    },
+    {
+      "epoch": 0.16,
+      "grad_norm": 1.6038667570691656,
+      "learning_rate": 4.672413793103448e-05,
+      "loss": 0.1225,
+      "step": 290
+    },
+    {
+      "epoch": 0.16,
+      "grad_norm": 2.577572731831179,
+      "learning_rate": 4.655172413793104e-05,
+      "loss": 0.1399,
+      "step": 300
+    },
+    {
+      "epoch": 0.17,
+      "grad_norm": 1.6199241001441385,
+      "learning_rate": 4.6379310344827586e-05,
+      "loss": 0.1242,
+      "step": 310
+    },
+    {
+      "epoch": 0.17,
+      "grad_norm": 2.236577821186196,
+      "learning_rate": 4.6206896551724135e-05,
+      "loss": 0.1656,
+      "step": 320
+    },
+    {
+      "epoch": 0.18,
+      "grad_norm": 1.7294690605254757,
+      "learning_rate": 4.603448275862069e-05,
+      "loss": 0.1382,
+      "step": 330
+    },
+    {
+      "epoch": 0.18,
+      "grad_norm": 2.196527516378511,
+      "learning_rate": 4.586206896551724e-05,
+      "loss": 0.1257,
+      "step": 340
+    },
+    {
+      "epoch": 0.19,
+      "grad_norm": 2.1057444340221463,
+      "learning_rate": 4.5689655172413794e-05,
+      "loss": 0.1238,
+      "step": 350
+    },
+    {
+      "epoch": 0.19,
+      "grad_norm": 1.5409556870328274,
+      "learning_rate": 4.551724137931035e-05,
+      "loss": 0.1383,
+      "step": 360
+    },
+    {
+      "epoch": 0.2,
+      "grad_norm": 1.5204083616874053,
+      "learning_rate": 4.53448275862069e-05,
+      "loss": 0.1068,
+      "step": 370
+    },
+    {
+      "epoch": 0.2,
+      "grad_norm": 2.3557725298931746,
+      "learning_rate": 4.5172413793103454e-05,
+      "loss": 0.1071,
+      "step": 380
+    },
+    {
+      "epoch": 0.21,
+      "grad_norm": 3.2601538460418644,
+      "learning_rate": 4.5e-05,
+      "loss": 0.125,
+      "step": 390
+    },
+    {
+      "epoch": 0.21,
+      "grad_norm": 1.9031725385762286,
+      "learning_rate": 4.482758620689655e-05,
+      "loss": 0.0991,
+      "step": 400
+    },
+    {
+      "epoch": 0.22,
+      "grad_norm": 1.3946050262183123,
+      "learning_rate": 4.465517241379311e-05,
+      "loss": 0.1156,
+      "step": 410
+    },
+    {
+      "epoch": 0.22,
+      "grad_norm": 1.097644875106397,
+      "learning_rate": 4.4482758620689656e-05,
+      "loss": 0.1366,
+      "step": 420
+    },
+    {
+      "epoch": 0.23,
+      "grad_norm": 1.37846299019108,
+      "learning_rate": 4.431034482758621e-05,
+      "loss": 0.126,
+      "step": 430
+    },
+    {
+      "epoch": 0.24,
+      "grad_norm": 1.8340152889320331,
+      "learning_rate": 4.413793103448276e-05,
+      "loss": 0.1066,
+      "step": 440
+    },
+    {
+      "epoch": 0.24,
+      "grad_norm": 1.8304505611337867,
+      "learning_rate": 4.396551724137931e-05,
+      "loss": 0.0868,
+      "step": 450
+    },
+    {
+      "epoch": 0.25,
+      "grad_norm": 1.550196490898523,
+      "learning_rate": 4.3793103448275864e-05,
+      "loss": 0.1286,
+      "step": 460
+    },
+    {
+      "epoch": 0.25,
+      "grad_norm": 2.176112247796248,
+      "learning_rate": 4.362068965517241e-05,
+      "loss": 0.1206,
+      "step": 470
+    },
+    {
+      "epoch": 0.26,
+      "grad_norm": 1.6589263894091213,
+      "learning_rate": 4.344827586206897e-05,
+      "loss": 0.1008,
+      "step": 480
+    },
+    {
+      "epoch": 0.26,
+      "grad_norm": 1.8349611508902046,
+      "learning_rate": 4.327586206896552e-05,
+      "loss": 0.1198,
+      "step": 490
+    },
+    {
+      "epoch": 0.27,
+      "grad_norm": 2.1218964920724126,
+      "learning_rate": 4.3103448275862066e-05,
+      "loss": 0.1166,
+      "step": 500
+    },
+    {
+      "epoch": 0.27,
+      "eval_loss": 0.6078919172286987,
+      "eval_runtime": 116.8471,
+      "eval_samples_per_second": 11.288,
+      "eval_steps_per_second": 2.824,
+      "step": 500
+    },
+    {
+      "epoch": 0.27,
+      "grad_norm": 2.5775141311007856,
+      "learning_rate": 4.293103448275863e-05,
+      "loss": 0.1124,
+      "step": 510
+    },
+    {
+      "epoch": 0.28,
+      "grad_norm": 1.6019517017800202,
+      "learning_rate": 4.275862068965518e-05,
+      "loss": 0.1068,
+      "step": 520
+    },
+    {
+      "epoch": 0.28,
+      "grad_norm": 2.6901962755310205,
+      "learning_rate": 4.2586206896551725e-05,
+      "loss": 0.1286,
+      "step": 530
+    },
+    {
+      "epoch": 0.29,
+      "grad_norm": 3.9517995356721767,
+      "learning_rate": 4.241379310344828e-05,
+      "loss": 0.1149,
+      "step": 540
+    },
+    {
+      "epoch": 0.29,
+      "grad_norm": 2.0428896228074076,
+      "learning_rate": 4.224137931034483e-05,
+      "loss": 0.141,
+      "step": 550
+    },
+    {
+      "epoch": 0.3,
+      "grad_norm": 2.263258592133553,
+      "learning_rate": 4.2068965517241385e-05,
+      "loss": 0.0949,
+      "step": 560
+    },
+    {
+      "epoch": 0.3,
+      "grad_norm": 1.4823165953974604,
+      "learning_rate": 4.1896551724137934e-05,
+      "loss": 0.1365,
+      "step": 570
+    },
+    {
+      "epoch": 0.31,
+      "grad_norm": 1.4441377020989878,
+      "learning_rate": 4.172413793103448e-05,
+      "loss": 0.1015,
+      "step": 580
+    },
+    {
+      "epoch": 0.32,
+      "grad_norm": 1.4779059254436886,
+      "learning_rate": 4.155172413793104e-05,
+      "loss": 0.0988,
+      "step": 590
+    },
+    {
+      "epoch": 0.32,
+      "grad_norm": 1.7777823671018818,
+      "learning_rate": 4.1379310344827587e-05,
+      "loss": 0.1124,
+      "step": 600
+    },
+    {
+      "epoch": 0.33,
+      "grad_norm": 1.737579831138191,
+      "learning_rate": 4.120689655172414e-05,
+      "loss": 0.086,
+      "step": 610
+    },
+    {
+      "epoch": 0.33,
+      "grad_norm": 2.708453961232997,
+      "learning_rate": 4.103448275862069e-05,
+      "loss": 0.0933,
+      "step": 620
+    },
+    {
+      "epoch": 0.34,
+      "grad_norm": 1.8871805824236731,
+      "learning_rate": 4.086206896551724e-05,
+      "loss": 0.1407,
+      "step": 630
+    },
+    {
+      "epoch": 0.34,
+      "grad_norm": 1.7300112722427339,
+      "learning_rate": 4.0689655172413795e-05,
+      "loss": 0.1224,
+      "step": 640
+    },
+    {
+      "epoch": 0.35,
+      "grad_norm": 1.4631236252240614,
+      "learning_rate": 4.0517241379310344e-05,
+      "loss": 0.1014,
+      "step": 650
+    },
+    {
+      "epoch": 0.35,
+      "grad_norm": 1.2602431597419264,
+      "learning_rate": 4.03448275862069e-05,
+      "loss": 0.1583,
+      "step": 660
+    },
+    {
+      "epoch": 0.36,
+      "grad_norm": 1.2077937041919453,
+      "learning_rate": 4.0172413793103455e-05,
+      "loss": 0.1209,
+      "step": 670
+    },
+    {
+      "epoch": 0.36,
+      "grad_norm": 1.4386184566429954,
+      "learning_rate": 4e-05,
+      "loss": 0.1016,
+      "step": 680
+    },
+    {
+      "epoch": 0.37,
+      "grad_norm": 2.6160358835758584,
+      "learning_rate": 3.982758620689656e-05,
+      "loss": 0.1062,
+      "step": 690
+    },
+    {
+      "epoch": 0.37,
+      "grad_norm": 1.9278794640498955,
+      "learning_rate": 3.965517241379311e-05,
+      "loss": 0.1037,
+      "step": 700
+    },
+    {
+      "epoch": 0.38,
+      "grad_norm": 1.2872571900237024,
+      "learning_rate": 3.9482758620689656e-05,
+      "loss": 0.096,
+      "step": 710
+    },
+    {
+      "epoch": 0.39,
+      "grad_norm": 1.243554309347296,
+      "learning_rate": 3.931034482758621e-05,
+      "loss": 0.1084,
+      "step": 720
+    },
+    {
+      "epoch": 0.39,
+      "grad_norm": 1.5090589714253309,
+      "learning_rate": 3.913793103448276e-05,
+      "loss": 0.0877,
+      "step": 730
+    },
+    {
+      "epoch": 0.4,
+      "grad_norm": 2.1419550623025168,
+      "learning_rate": 3.896551724137931e-05,
+      "loss": 0.0994,
+      "step": 740
+    },
+    {
+      "epoch": 0.4,
+      "grad_norm": 1.7807417973632438,
+      "learning_rate": 3.8793103448275865e-05,
+      "loss": 0.1029,
+      "step": 750
+    },
+    {
+      "epoch": 0.41,
+      "grad_norm": 1.3342960263057682,
+      "learning_rate": 3.862068965517241e-05,
+      "loss": 0.1072,
+      "step": 760
+    },
+    {
+      "epoch": 0.41,
+      "grad_norm": 2.3865282340158136,
+      "learning_rate": 3.844827586206897e-05,
+      "loss": 0.1193,
+      "step": 770
+    },
+    {
+      "epoch": 0.42,
+      "grad_norm": 1.5428742248459941,
+      "learning_rate": 3.827586206896552e-05,
+      "loss": 0.1156,
+      "step": 780
+    },
+    {
+      "epoch": 0.42,
+      "grad_norm": 1.7660532115509044,
+      "learning_rate": 3.8103448275862066e-05,
+      "loss": 0.122,
+      "step": 790
+    },
+    {
+      "epoch": 0.43,
+      "grad_norm": 1.8149742752994733,
+      "learning_rate": 3.793103448275862e-05,
+      "loss": 0.1346,
+      "step": 800
+    },
+    {
+      "epoch": 0.43,
+      "grad_norm": 1.7456408876472995,
+      "learning_rate": 3.775862068965517e-05,
+      "loss": 0.1223,
+      "step": 810
+    },
+    {
+      "epoch": 0.44,
+      "grad_norm": 1.10163248244056,
+      "learning_rate": 3.7586206896551726e-05,
+      "loss": 0.1031,
+      "step": 820
+    },
+    {
+      "epoch": 0.44,
+      "grad_norm": 1.6441057737088702,
+      "learning_rate": 3.741379310344828e-05,
+      "loss": 0.1059,
+      "step": 830
+    },
+    {
+      "epoch": 0.45,
+      "grad_norm": 2.3999279790163484,
+      "learning_rate": 3.724137931034483e-05,
+      "loss": 0.1125,
+      "step": 840
+    },
+    {
+      "epoch": 0.45,
+      "grad_norm": 2.2081477934156903,
+      "learning_rate": 3.7068965517241385e-05,
+      "loss": 0.1266,
+      "step": 850
+    },
+    {
+      "epoch": 0.46,
+      "grad_norm": 2.366783771480017,
+      "learning_rate": 3.6896551724137934e-05,
+      "loss": 0.1127,
+      "step": 860
+    },
+    {
+      "epoch": 0.47,
+      "grad_norm": 1.3077873674136173,
+      "learning_rate": 3.672413793103448e-05,
+      "loss": 0.1095,
+      "step": 870
+    },
+    {
+      "epoch": 0.47,
+      "grad_norm": 1.8197812508114701,
+      "learning_rate": 3.655172413793104e-05,
+      "loss": 0.0932,
+      "step": 880
+    },
+    {
+      "epoch": 0.48,
+      "grad_norm": 1.0806192057981219,
+      "learning_rate": 3.637931034482759e-05,
+      "loss": 0.1166,
+      "step": 890
+    },
+    {
+      "epoch": 0.48,
+      "grad_norm": 1.520666439337001,
+      "learning_rate": 3.620689655172414e-05,
+      "loss": 0.0883,
+      "step": 900
+    },
+    {
+      "epoch": 0.49,
+      "grad_norm": 1.690002270629302,
+      "learning_rate": 3.603448275862069e-05,
+      "loss": 0.1199,
+      "step": 910
+    },
+    {
+      "epoch": 0.49,
+      "grad_norm": 1.4319374130118003,
+      "learning_rate": 3.586206896551724e-05,
+      "loss": 0.0991,
+      "step": 920
+    },
+    {
+      "epoch": 0.5,
+      "grad_norm": 1.0626084369653164,
+      "learning_rate": 3.5689655172413795e-05,
+      "loss": 0.0923,
+      "step": 930
+    },
+    {
+      "epoch": 0.5,
+      "grad_norm": 2.0848060597460902,
+      "learning_rate": 3.5517241379310344e-05,
+      "loss": 0.0979,
+      "step": 940
+    },
+    {
+      "epoch": 0.51,
+      "grad_norm": 1.4997189461483256,
+      "learning_rate": 3.53448275862069e-05,
+      "loss": 0.0949,
+      "step": 950
+    },
+    {
+      "epoch": 0.51,
+      "grad_norm": 1.7887817042743388,
+      "learning_rate": 3.517241379310345e-05,
+      "loss": 0.1135,
+      "step": 960
+    },
+    {
+      "epoch": 0.52,
+      "grad_norm": 3.242965692388458,
+      "learning_rate": 3.5e-05,
+      "loss": 0.1315,
+      "step": 970
+    },
+    {
+      "epoch": 0.52,
+      "grad_norm": 1.5034762176322083,
+      "learning_rate": 3.482758620689655e-05,
+      "loss": 0.1177,
+      "step": 980
+    },
+    {
+      "epoch": 0.53,
+      "grad_norm": 1.6679474444200848,
+      "learning_rate": 3.465517241379311e-05,
+      "loss": 0.1182,
+      "step": 990
+    },
+    {
+      "epoch": 0.54,
+      "grad_norm": 2.814574507251776,
+      "learning_rate": 3.4482758620689657e-05,
+      "loss": 0.0912,
+      "step": 1000
+    },
+    {
+      "epoch": 0.54,
+      "eval_loss": 0.6101276874542236,
+      "eval_runtime": 113.9995,
+      "eval_samples_per_second": 11.57,
+      "eval_steps_per_second": 2.895,
+      "step": 1000
+    }
+  ],
+  "logging_steps": 10,
+  "max_steps": 3000,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 2,
+  "save_steps": 500,
+  "total_flos": 28899016704000.0,
+  "train_batch_size": 4,
+  "trial_name": null,
+  "trial_params": null
+}

checkpoints/checkpoint-1000/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:22266252322e6977894297b61731ca9228c8fcfe4e8d788aef270218c40fec98
+size 6776

checkpoints/checkpoint-1000/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)

checkpoints/checkpoint-1500/README.md

@@ -0,0 +1,202 @@
+---
+library_name: peft
+base_model: openbmb/MiniCPM-2B-dpo-bf16
+---
+
+# 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 0.10.0

checkpoints/checkpoint-1500/adapter_config.json

@@ -0,0 +1,29 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "openbmb/MiniCPM-2B-dpo-bf16",
+  "bias": "none",
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": "gaussian",
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 8,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "v_proj",
+    "q_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "use_dora": false,
+  "use_rslora": false
+}

checkpoints/checkpoint-1500/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:de9e2856e08d5173f8e2af8a3753fd4b9d7ff846b54a5ff0c58f7c3f5766adbd
+size 5919456

checkpoints/checkpoint-1500/global_step1500/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:84bae28b1e6bf0e12a34d4a32f9f04cc7dacf4cb9fcc31079d3efe068b295712
+size 35393392

checkpoints/checkpoint-1500/global_step1500/zero_pp_rank_0_mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e17a037ee17783aebea4d60a2ca9224419f83386c47c8f19b0e070d416c5d008
+size 183275

checkpoints/checkpoint-1500/latest

@@ -0,0 +1 @@
+global_step1500

checkpoints/checkpoint-1500/rng_state.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9ff7bfe3bfc07bfdf7a752b2b29572987f4902cac1a5149315d94e9b25b3c90a
+size 14244

checkpoints/checkpoint-1500/scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bf4ee1a21cac7b0db9a6509d4738739668da3752503f4e308a0d5e0435a00c03
+size 1064

checkpoints/checkpoint-1500/special_tokens_map.json

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

checkpoints/checkpoint-1500/tokenizer_config.json

@@ -0,0 +1,42 @@
+{
+  "add_bos_token": true,
+  "add_eos_token": false,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<s>",
+  "chat_template": "{% for message in messages %}{% if message['role'] == 'user' %}{{'<用户>' + message['content'].strip() + '<AI>'}}{% else %}{{message['content'].strip()}}{% endif %}{% endfor %}",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "</s>",
+  "legacy": true,
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "</s>",
+  "sp_model_kwargs": {},
+  "spaces_between_special_tokens": false,
+  "tokenizer_class": "LlamaTokenizer",
+  "unk_token": "<unk>",
+  "use_default_system_prompt": false
+}

checkpoints/checkpoint-1500/trainer_state.json

@@ -0,0 +1,1095 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.8025682182985554,
+  "eval_steps": 500,
+  "global_step": 1500,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.01,
+      "grad_norm": 7.3019086777975994,
+      "learning_rate": 5e-06,
+      "loss": 0.6939,
+      "step": 10
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 5.150989333266983,
+      "learning_rate": 1e-05,
+      "loss": 0.7167,
+      "step": 20
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 3.640558809610037,
+      "learning_rate": 1.5e-05,
+      "loss": 0.5683,
+      "step": 30
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 7.517999397731128,
+      "learning_rate": 2e-05,
+      "loss": 0.5472,
+      "step": 40
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 1.9687061679463425,
+      "learning_rate": 2.5e-05,
+      "loss": 0.4439,
+      "step": 50
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 3.643479206523606,
+      "learning_rate": 3e-05,
+      "loss": 0.2486,
+      "step": 60
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 2.2754773308695095,
+      "learning_rate": 3.5e-05,
+      "loss": 0.2217,
+      "step": 70
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 1.7144730049127388,
+      "learning_rate": 4e-05,
+      "loss": 0.169,
+      "step": 80
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 3.4702829704135114,
+      "learning_rate": 4.5e-05,
+      "loss": 0.1994,
+      "step": 90
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 1.3985127340985621,
+      "learning_rate": 5e-05,
+      "loss": 0.1612,
+      "step": 100
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 1.375992184386137,
+      "learning_rate": 4.982758620689655e-05,
+      "loss": 0.1576,
+      "step": 110
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 1.9528635753013313,
+      "learning_rate": 4.9655172413793107e-05,
+      "loss": 0.1393,
+      "step": 120
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 4.075169010198401,
+      "learning_rate": 4.9482758620689655e-05,
+      "loss": 0.1969,
+      "step": 130
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 2.0953991165751207,
+      "learning_rate": 4.931034482758621e-05,
+      "loss": 0.1294,
+      "step": 140
+    },
+    {
+      "epoch": 0.08,
+      "grad_norm": 1.942660591044849,
+      "learning_rate": 4.913793103448276e-05,
+      "loss": 0.1306,
+      "step": 150
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 3.1508904015728345,
+      "learning_rate": 4.896551724137931e-05,
+      "loss": 0.1526,
+      "step": 160
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 1.9862795165358471,
+      "learning_rate": 4.8793103448275864e-05,
+      "loss": 0.1186,
+      "step": 170
+    },
+    {
+      "epoch": 0.1,
+      "grad_norm": 2.633061833817991,
+      "learning_rate": 4.862068965517241e-05,
+      "loss": 0.1457,
+      "step": 180
+    },
+    {
+      "epoch": 0.1,
+      "grad_norm": 1.8017052368446178,
+      "learning_rate": 4.844827586206897e-05,
+      "loss": 0.1234,
+      "step": 190
+    },
+    {
+      "epoch": 0.11,
+      "grad_norm": 2.1560694100709803,
+      "learning_rate": 4.827586206896552e-05,
+      "loss": 0.1346,
+      "step": 200
+    },
+    {
+      "epoch": 0.11,
+      "grad_norm": 1.5737689267430703,
+      "learning_rate": 4.810344827586207e-05,
+      "loss": 0.116,
+      "step": 210
+    },
+    {
+      "epoch": 0.12,
+      "grad_norm": 1.957864677854788,
+      "learning_rate": 4.793103448275863e-05,
+      "loss": 0.1692,
+      "step": 220
+    },
+    {
+      "epoch": 0.12,
+      "grad_norm": 2.215039223521855,
+      "learning_rate": 4.7758620689655176e-05,
+      "loss": 0.1245,
+      "step": 230
+    },
+    {
+      "epoch": 0.13,
+      "grad_norm": 1.370517239734168,
+      "learning_rate": 4.7586206896551725e-05,
+      "loss": 0.1476,
+      "step": 240
+    },
+    {
+      "epoch": 0.13,
+      "grad_norm": 1.7341334563022532,
+      "learning_rate": 4.741379310344828e-05,
+      "loss": 0.1236,
+      "step": 250
+    },
+    {
+      "epoch": 0.14,
+      "grad_norm": 1.5994298113068974,
+      "learning_rate": 4.724137931034483e-05,
+      "loss": 0.1161,
+      "step": 260
+    },
+    {
+      "epoch": 0.14,
+      "grad_norm": 1.5317433190951963,
+      "learning_rate": 4.7068965517241385e-05,
+      "loss": 0.1035,
+      "step": 270
+    },
+    {
+      "epoch": 0.15,
+      "grad_norm": 2.191977732539556,
+      "learning_rate": 4.689655172413793e-05,
+      "loss": 0.1427,
+      "step": 280
+    },
+    {
+      "epoch": 0.16,
+      "grad_norm": 1.6038667570691656,
+      "learning_rate": 4.672413793103448e-05,
+      "loss": 0.1225,
+      "step": 290
+    },
+    {
+      "epoch": 0.16,
+      "grad_norm": 2.577572731831179,
+      "learning_rate": 4.655172413793104e-05,
+      "loss": 0.1399,
+      "step": 300
+    },
+    {
+      "epoch": 0.17,
+      "grad_norm": 1.6199241001441385,
+      "learning_rate": 4.6379310344827586e-05,
+      "loss": 0.1242,
+      "step": 310
+    },
+    {
+      "epoch": 0.17,
+      "grad_norm": 2.236577821186196,
+      "learning_rate": 4.6206896551724135e-05,
+      "loss": 0.1656,
+      "step": 320
+    },
+    {
+      "epoch": 0.18,
+      "grad_norm": 1.7294690605254757,
+      "learning_rate": 4.603448275862069e-05,
+      "loss": 0.1382,
+      "step": 330
+    },
+    {
+      "epoch": 0.18,
+      "grad_norm": 2.196527516378511,
+      "learning_rate": 4.586206896551724e-05,
+      "loss": 0.1257,
+      "step": 340
+    },
+    {
+      "epoch": 0.19,
+      "grad_norm": 2.1057444340221463,
+      "learning_rate": 4.5689655172413794e-05,
+      "loss": 0.1238,
+      "step": 350
+    },
+    {
+      "epoch": 0.19,
+      "grad_norm": 1.5409556870328274,
+      "learning_rate": 4.551724137931035e-05,
+      "loss": 0.1383,
+      "step": 360
+    },
+    {
+      "epoch": 0.2,
+      "grad_norm": 1.5204083616874053,
+      "learning_rate": 4.53448275862069e-05,
+      "loss": 0.1068,
+      "step": 370
+    },
+    {
+      "epoch": 0.2,
+      "grad_norm": 2.3557725298931746,
+      "learning_rate": 4.5172413793103454e-05,
+      "loss": 0.1071,
+      "step": 380
+    },
+    {
+      "epoch": 0.21,
+      "grad_norm": 3.2601538460418644,
+      "learning_rate": 4.5e-05,
+      "loss": 0.125,
+      "step": 390
+    },
+    {
+      "epoch": 0.21,
+      "grad_norm": 1.9031725385762286,
+      "learning_rate": 4.482758620689655e-05,
+      "loss": 0.0991,
+      "step": 400
+    },
+    {
+      "epoch": 0.22,
+      "grad_norm": 1.3946050262183123,
+      "learning_rate": 4.465517241379311e-05,
+      "loss": 0.1156,
+      "step": 410
+    },
+    {
+      "epoch": 0.22,
+      "grad_norm": 1.097644875106397,
+      "learning_rate": 4.4482758620689656e-05,
+      "loss": 0.1366,
+      "step": 420
+    },
+    {
+      "epoch": 0.23,
+      "grad_norm": 1.37846299019108,
+      "learning_rate": 4.431034482758621e-05,
+      "loss": 0.126,
+      "step": 430
+    },
+    {
+      "epoch": 0.24,
+      "grad_norm": 1.8340152889320331,
+      "learning_rate": 4.413793103448276e-05,
+      "loss": 0.1066,
+      "step": 440
+    },
+    {
+      "epoch": 0.24,
+      "grad_norm": 1.8304505611337867,
+      "learning_rate": 4.396551724137931e-05,
+      "loss": 0.0868,
+      "step": 450
+    },
+    {
+      "epoch": 0.25,
+      "grad_norm": 1.550196490898523,
+      "learning_rate": 4.3793103448275864e-05,
+      "loss": 0.1286,
+      "step": 460
+    },
+    {
+      "epoch": 0.25,
+      "grad_norm": 2.176112247796248,
+      "learning_rate": 4.362068965517241e-05,
+      "loss": 0.1206,
+      "step": 470
+    },
+    {
+      "epoch": 0.26,
+      "grad_norm": 1.6589263894091213,
+      "learning_rate": 4.344827586206897e-05,
+      "loss": 0.1008,
+      "step": 480
+    },
+    {
+      "epoch": 0.26,
+      "grad_norm": 1.8349611508902046,
+      "learning_rate": 4.327586206896552e-05,
+      "loss": 0.1198,
+      "step": 490
+    },
+    {
+      "epoch": 0.27,
+      "grad_norm": 2.1218964920724126,
+      "learning_rate": 4.3103448275862066e-05,
+      "loss": 0.1166,
+      "step": 500
+    },
+    {
+      "epoch": 0.27,
+      "eval_loss": 0.6078919172286987,
+      "eval_runtime": 116.8471,
+      "eval_samples_per_second": 11.288,
+      "eval_steps_per_second": 2.824,
+      "step": 500
+    },
+    {
+      "epoch": 0.27,
+      "grad_norm": 2.5775141311007856,
+      "learning_rate": 4.293103448275863e-05,
+      "loss": 0.1124,
+      "step": 510
+    },
+    {
+      "epoch": 0.28,
+      "grad_norm": 1.6019517017800202,
+      "learning_rate": 4.275862068965518e-05,
+      "loss": 0.1068,
+      "step": 520
+    },
+    {
+      "epoch": 0.28,
+      "grad_norm": 2.6901962755310205,
+      "learning_rate": 4.2586206896551725e-05,
+      "loss": 0.1286,
+      "step": 530
+    },
+    {
+      "epoch": 0.29,
+      "grad_norm": 3.9517995356721767,
+      "learning_rate": 4.241379310344828e-05,
+      "loss": 0.1149,
+      "step": 540
+    },
+    {
+      "epoch": 0.29,
+      "grad_norm": 2.0428896228074076,
+      "learning_rate": 4.224137931034483e-05,
+      "loss": 0.141,
+      "step": 550
+    },
+    {
+      "epoch": 0.3,
+      "grad_norm": 2.263258592133553,
+      "learning_rate": 4.2068965517241385e-05,
+      "loss": 0.0949,
+      "step": 560
+    },
+    {
+      "epoch": 0.3,
+      "grad_norm": 1.4823165953974604,
+      "learning_rate": 4.1896551724137934e-05,
+      "loss": 0.1365,
+      "step": 570
+    },
+    {
+      "epoch": 0.31,
+      "grad_norm": 1.4441377020989878,
+      "learning_rate": 4.172413793103448e-05,
+      "loss": 0.1015,
+      "step": 580
+    },
+    {
+      "epoch": 0.32,
+      "grad_norm": 1.4779059254436886,
+      "learning_rate": 4.155172413793104e-05,
+      "loss": 0.0988,
+      "step": 590
+    },
+    {
+      "epoch": 0.32,
+      "grad_norm": 1.7777823671018818,
+      "learning_rate": 4.1379310344827587e-05,
+      "loss": 0.1124,
+      "step": 600
+    },
+    {
+      "epoch": 0.33,
+      "grad_norm": 1.737579831138191,
+      "learning_rate": 4.120689655172414e-05,
+      "loss": 0.086,
+      "step": 610
+    },
+    {
+      "epoch": 0.33,
+      "grad_norm": 2.708453961232997,
+      "learning_rate": 4.103448275862069e-05,
+      "loss": 0.0933,
+      "step": 620
+    },
+    {
+      "epoch": 0.34,
+      "grad_norm": 1.8871805824236731,
+      "learning_rate": 4.086206896551724e-05,
+      "loss": 0.1407,
+      "step": 630
+    },
+    {
+      "epoch": 0.34,
+      "grad_norm": 1.7300112722427339,
+      "learning_rate": 4.0689655172413795e-05,
+      "loss": 0.1224,
+      "step": 640
+    },
+    {
+      "epoch": 0.35,
+      "grad_norm": 1.4631236252240614,
+      "learning_rate": 4.0517241379310344e-05,
+      "loss": 0.1014,
+      "step": 650
+    },
+    {
+      "epoch": 0.35,
+      "grad_norm": 1.2602431597419264,
+      "learning_rate": 4.03448275862069e-05,
+      "loss": 0.1583,
+      "step": 660
+    },
+    {
+      "epoch": 0.36,
+      "grad_norm": 1.2077937041919453,
+      "learning_rate": 4.0172413793103455e-05,
+      "loss": 0.1209,
+      "step": 670
+    },
+    {
+      "epoch": 0.36,
+      "grad_norm": 1.4386184566429954,
+      "learning_rate": 4e-05,
+      "loss": 0.1016,
+      "step": 680
+    },
+    {
+      "epoch": 0.37,
+      "grad_norm": 2.6160358835758584,
+      "learning_rate": 3.982758620689656e-05,
+      "loss": 0.1062,
+      "step": 690
+    },
+    {
+      "epoch": 0.37,
+      "grad_norm": 1.9278794640498955,
+      "learning_rate": 3.965517241379311e-05,
+      "loss": 0.1037,
+      "step": 700
+    },
+    {
+      "epoch": 0.38,
+      "grad_norm": 1.2872571900237024,
+      "learning_rate": 3.9482758620689656e-05,
+      "loss": 0.096,
+      "step": 710
+    },
+    {
+      "epoch": 0.39,
+      "grad_norm": 1.243554309347296,
+      "learning_rate": 3.931034482758621e-05,
+      "loss": 0.1084,
+      "step": 720
+    },
+    {
+      "epoch": 0.39,
+      "grad_norm": 1.5090589714253309,
+      "learning_rate": 3.913793103448276e-05,
+      "loss": 0.0877,
+      "step": 730
+    },
+    {
+      "epoch": 0.4,
+      "grad_norm": 2.1419550623025168,
+      "learning_rate": 3.896551724137931e-05,
+      "loss": 0.0994,
+      "step": 740
+    },
+    {
+      "epoch": 0.4,
+      "grad_norm": 1.7807417973632438,
+      "learning_rate": 3.8793103448275865e-05,
+      "loss": 0.1029,
+      "step": 750
+    },
+    {
+      "epoch": 0.41,
+      "grad_norm": 1.3342960263057682,
+      "learning_rate": 3.862068965517241e-05,
+      "loss": 0.1072,
+      "step": 760
+    },
+    {
+      "epoch": 0.41,
+      "grad_norm": 2.3865282340158136,
+      "learning_rate": 3.844827586206897e-05,
+      "loss": 0.1193,
+      "step": 770
+    },
+    {
+      "epoch": 0.42,
+      "grad_norm": 1.5428742248459941,
+      "learning_rate": 3.827586206896552e-05,
+      "loss": 0.1156,
+      "step": 780
+    },
+    {
+      "epoch": 0.42,
+      "grad_norm": 1.7660532115509044,
+      "learning_rate": 3.8103448275862066e-05,
+      "loss": 0.122,
+      "step": 790
+    },
+    {
+      "epoch": 0.43,
+      "grad_norm": 1.8149742752994733,
+      "learning_rate": 3.793103448275862e-05,
+      "loss": 0.1346,
+      "step": 800
+    },
+    {
+      "epoch": 0.43,
+      "grad_norm": 1.7456408876472995,
+      "learning_rate": 3.775862068965517e-05,
+      "loss": 0.1223,
+      "step": 810
+    },
+    {
+      "epoch": 0.44,
+      "grad_norm": 1.10163248244056,
+      "learning_rate": 3.7586206896551726e-05,
+      "loss": 0.1031,
+      "step": 820
+    },
+    {
+      "epoch": 0.44,
+      "grad_norm": 1.6441057737088702,
+      "learning_rate": 3.741379310344828e-05,
+      "loss": 0.1059,
+      "step": 830
+    },
+    {
+      "epoch": 0.45,
+      "grad_norm": 2.3999279790163484,
+      "learning_rate": 3.724137931034483e-05,
+      "loss": 0.1125,
+      "step": 840
+    },
+    {
+      "epoch": 0.45,
+      "grad_norm": 2.2081477934156903,
+      "learning_rate": 3.7068965517241385e-05,
+      "loss": 0.1266,
+      "step": 850
+    },
+    {
+      "epoch": 0.46,
+      "grad_norm": 2.366783771480017,
+      "learning_rate": 3.6896551724137934e-05,
+      "loss": 0.1127,
+      "step": 860
+    },
+    {
+      "epoch": 0.47,
+      "grad_norm": 1.3077873674136173,
+      "learning_rate": 3.672413793103448e-05,
+      "loss": 0.1095,
+      "step": 870
+    },
+    {
+      "epoch": 0.47,
+      "grad_norm": 1.8197812508114701,
+      "learning_rate": 3.655172413793104e-05,
+      "loss": 0.0932,
+      "step": 880
+    },
+    {
+      "epoch": 0.48,
+      "grad_norm": 1.0806192057981219,
+      "learning_rate": 3.637931034482759e-05,
+      "loss": 0.1166,
+      "step": 890
+    },
+    {
+      "epoch": 0.48,
+      "grad_norm": 1.520666439337001,
+      "learning_rate": 3.620689655172414e-05,
+      "loss": 0.0883,
+      "step": 900
+    },
+    {
+      "epoch": 0.49,
+      "grad_norm": 1.690002270629302,
+      "learning_rate": 3.603448275862069e-05,
+      "loss": 0.1199,
+      "step": 910
+    },
+    {
+      "epoch": 0.49,
+      "grad_norm": 1.4319374130118003,
+      "learning_rate": 3.586206896551724e-05,
+      "loss": 0.0991,
+      "step": 920
+    },
+    {
+      "epoch": 0.5,
+      "grad_norm": 1.0626084369653164,
+      "learning_rate": 3.5689655172413795e-05,
+      "loss": 0.0923,
+      "step": 930
+    },
+    {
+      "epoch": 0.5,
+      "grad_norm": 2.0848060597460902,
+      "learning_rate": 3.5517241379310344e-05,
+      "loss": 0.0979,
+      "step": 940
+    },
+    {
+      "epoch": 0.51,
+      "grad_norm": 1.4997189461483256,
+      "learning_rate": 3.53448275862069e-05,
+      "loss": 0.0949,
+      "step": 950
+    },
+    {
+      "epoch": 0.51,
+      "grad_norm": 1.7887817042743388,
+      "learning_rate": 3.517241379310345e-05,
+      "loss": 0.1135,
+      "step": 960
+    },
+    {
+      "epoch": 0.52,
+      "grad_norm": 3.242965692388458,
+      "learning_rate": 3.5e-05,
+      "loss": 0.1315,
+      "step": 970
+    },
+    {
+      "epoch": 0.52,
+      "grad_norm": 1.5034762176322083,
+      "learning_rate": 3.482758620689655e-05,
+      "loss": 0.1177,
+      "step": 980
+    },
+    {
+      "epoch": 0.53,
+      "grad_norm": 1.6679474444200848,
+      "learning_rate": 3.465517241379311e-05,
+      "loss": 0.1182,
+      "step": 990
+    },
+    {
+      "epoch": 0.54,
+      "grad_norm": 2.814574507251776,
+      "learning_rate": 3.4482758620689657e-05,
+      "loss": 0.0912,
+      "step": 1000
+    },
+    {
+      "epoch": 0.54,
+      "eval_loss": 0.6101276874542236,
+      "eval_runtime": 113.9995,
+      "eval_samples_per_second": 11.57,
+      "eval_steps_per_second": 2.895,
+      "step": 1000
+    },
+    {
+      "epoch": 0.54,
+      "grad_norm": 2.1321319681580535,
+      "learning_rate": 3.431034482758621e-05,
+      "loss": 0.0983,
+      "step": 1010
+    },
+    {
+      "epoch": 0.55,
+      "grad_norm": 1.8915101367452352,
+      "learning_rate": 3.413793103448276e-05,
+      "loss": 0.1113,
+      "step": 1020
+    },
+    {
+      "epoch": 0.55,
+      "grad_norm": 1.4160985095468477,
+      "learning_rate": 3.3965517241379316e-05,
+      "loss": 0.1076,
+      "step": 1030
+    },
+    {
+      "epoch": 0.56,
+      "grad_norm": 1.8562788974878586,
+      "learning_rate": 3.3793103448275865e-05,
+      "loss": 0.1011,
+      "step": 1040
+    },
+    {
+      "epoch": 0.56,
+      "grad_norm": 1.3793192563691294,
+      "learning_rate": 3.3620689655172414e-05,
+      "loss": 0.0978,
+      "step": 1050
+    },
+    {
+      "epoch": 0.57,
+      "grad_norm": 1.4606563129628805,
+      "learning_rate": 3.344827586206897e-05,
+      "loss": 0.1069,
+      "step": 1060
+    },
+    {
+      "epoch": 0.57,
+      "grad_norm": 1.5680856211032999,
+      "learning_rate": 3.327586206896552e-05,
+      "loss": 0.0988,
+      "step": 1070
+    },
+    {
+      "epoch": 0.58,
+      "grad_norm": 1.5829345931951275,
+      "learning_rate": 3.310344827586207e-05,
+      "loss": 0.1256,
+      "step": 1080
+    },
+    {
+      "epoch": 0.58,
+      "grad_norm": 1.6200852939319585,
+      "learning_rate": 3.293103448275862e-05,
+      "loss": 0.097,
+      "step": 1090
+    },
+    {
+      "epoch": 0.59,
+      "grad_norm": 2.259656836213122,
+      "learning_rate": 3.275862068965517e-05,
+      "loss": 0.1137,
+      "step": 1100
+    },
+    {
+      "epoch": 0.59,
+      "grad_norm": 2.2483622341560645,
+      "learning_rate": 3.2586206896551726e-05,
+      "loss": 0.0999,
+      "step": 1110
+    },
+    {
+      "epoch": 0.6,
+      "grad_norm": 1.168198861956421,
+      "learning_rate": 3.2413793103448275e-05,
+      "loss": 0.1018,
+      "step": 1120
+    },
+    {
+      "epoch": 0.6,
+      "grad_norm": 2.4699824799031482,
+      "learning_rate": 3.2241379310344824e-05,
+      "loss": 0.1132,
+      "step": 1130
+    },
+    {
+      "epoch": 0.61,
+      "grad_norm": 1.2571654549549751,
+      "learning_rate": 3.206896551724138e-05,
+      "loss": 0.1054,
+      "step": 1140
+    },
+    {
+      "epoch": 0.62,
+      "grad_norm": 0.5559534032307631,
+      "learning_rate": 3.1896551724137935e-05,
+      "loss": 0.0789,
+      "step": 1150
+    },
+    {
+      "epoch": 0.62,
+      "grad_norm": 1.636369759504475,
+      "learning_rate": 3.172413793103448e-05,
+      "loss": 0.0902,
+      "step": 1160
+    },
+    {
+      "epoch": 0.63,
+      "grad_norm": 1.6137142935446496,
+      "learning_rate": 3.155172413793104e-05,
+      "loss": 0.1199,
+      "step": 1170
+    },
+    {
+      "epoch": 0.63,
+      "grad_norm": 1.7448003760796802,
+      "learning_rate": 3.137931034482759e-05,
+      "loss": 0.1295,
+      "step": 1180
+    },
+    {
+      "epoch": 0.64,
+      "grad_norm": 1.3261005358227276,
+      "learning_rate": 3.120689655172414e-05,
+      "loss": 0.1117,
+      "step": 1190
+    },
+    {
+      "epoch": 0.64,
+      "grad_norm": 1.7353127177901462,
+      "learning_rate": 3.103448275862069e-05,
+      "loss": 0.0951,
+      "step": 1200
+    },
+    {
+      "epoch": 0.65,
+      "grad_norm": 2.8569975913367074,
+      "learning_rate": 3.086206896551724e-05,
+      "loss": 0.112,
+      "step": 1210
+    },
+    {
+      "epoch": 0.65,
+      "grad_norm": 1.3481947218871082,
+      "learning_rate": 3.0689655172413796e-05,
+      "loss": 0.0876,
+      "step": 1220
+    },
+    {
+      "epoch": 0.66,
+      "grad_norm": 2.015933141613929,
+      "learning_rate": 3.0517241379310348e-05,
+      "loss": 0.0993,
+      "step": 1230
+    },
+    {
+      "epoch": 0.66,
+      "grad_norm": 1.0588164394448019,
+      "learning_rate": 3.0344827586206897e-05,
+      "loss": 0.1034,
+      "step": 1240
+    },
+    {
+      "epoch": 0.67,
+      "grad_norm": 1.3594986645993228,
+      "learning_rate": 3.017241379310345e-05,
+      "loss": 0.1004,
+      "step": 1250
+    },
+    {
+      "epoch": 0.67,
+      "grad_norm": 1.333098402625009,
+      "learning_rate": 3e-05,
+      "loss": 0.1259,
+      "step": 1260
+    },
+    {
+      "epoch": 0.68,
+      "grad_norm": 1.1324206075196583,
+      "learning_rate": 2.9827586206896553e-05,
+      "loss": 0.1155,
+      "step": 1270
+    },
+    {
+      "epoch": 0.68,
+      "grad_norm": 1.2270687927795876,
+      "learning_rate": 2.96551724137931e-05,
+      "loss": 0.092,
+      "step": 1280
+    },
+    {
+      "epoch": 0.69,
+      "grad_norm": 1.3085362507403875,
+      "learning_rate": 2.9482758620689654e-05,
+      "loss": 0.1064,
+      "step": 1290
+    },
+    {
+      "epoch": 0.7,
+      "grad_norm": 1.7135250277750762,
+      "learning_rate": 2.9310344827586206e-05,
+      "loss": 0.1132,
+      "step": 1300
+    },
+    {
+      "epoch": 0.7,
+      "grad_norm": 1.6121189707451158,
+      "learning_rate": 2.913793103448276e-05,
+      "loss": 0.1006,
+      "step": 1310
+    },
+    {
+      "epoch": 0.71,
+      "grad_norm": 1.3958680925504208,
+      "learning_rate": 2.8965517241379313e-05,
+      "loss": 0.0956,
+      "step": 1320
+    },
+    {
+      "epoch": 0.71,
+      "grad_norm": 1.546226150121884,
+      "learning_rate": 2.8793103448275865e-05,
+      "loss": 0.1142,
+      "step": 1330
+    },
+    {
+      "epoch": 0.72,
+      "grad_norm": 2.1837739995965415,
+      "learning_rate": 2.8620689655172417e-05,
+      "loss": 0.1127,
+      "step": 1340
+    },
+    {
+      "epoch": 0.72,
+      "grad_norm": 1.9402402206909504,
+      "learning_rate": 2.844827586206897e-05,
+      "loss": 0.0922,
+      "step": 1350
+    },
+    {
+      "epoch": 0.73,
+      "grad_norm": 1.7914401953164802,
+      "learning_rate": 2.8275862068965518e-05,
+      "loss": 0.1038,
+      "step": 1360
+    },
+    {
+      "epoch": 0.73,
+      "grad_norm": 1.505804090650568,
+      "learning_rate": 2.810344827586207e-05,
+      "loss": 0.1034,
+      "step": 1370
+    },
+    {
+      "epoch": 0.74,
+      "grad_norm": 1.9907350713586716,
+      "learning_rate": 2.7931034482758622e-05,
+      "loss": 0.103,
+      "step": 1380
+    },
+    {
+      "epoch": 0.74,
+      "grad_norm": 1.6948381773166858,
+      "learning_rate": 2.7758620689655175e-05,
+      "loss": 0.1091,
+      "step": 1390
+    },
+    {
+      "epoch": 0.75,
+      "grad_norm": 1.3995985437024723,
+      "learning_rate": 2.7586206896551727e-05,
+      "loss": 0.0852,
+      "step": 1400
+    },
+    {
+      "epoch": 0.75,
+      "grad_norm": 1.9347024029069393,
+      "learning_rate": 2.7413793103448275e-05,
+      "loss": 0.1393,
+      "step": 1410
+    },
+    {
+      "epoch": 0.76,
+      "grad_norm": 1.608776792445342,
+      "learning_rate": 2.7241379310344827e-05,
+      "loss": 0.0951,
+      "step": 1420
+    },
+    {
+      "epoch": 0.77,
+      "grad_norm": 1.6005483580619249,
+      "learning_rate": 2.706896551724138e-05,
+      "loss": 0.1037,
+      "step": 1430
+    },
+    {
+      "epoch": 0.77,
+      "grad_norm": 2.374208686020403,
+      "learning_rate": 2.689655172413793e-05,
+      "loss": 0.0926,
+      "step": 1440
+    },
+    {
+      "epoch": 0.78,
+      "grad_norm": 1.7554923995400171,
+      "learning_rate": 2.672413793103448e-05,
+      "loss": 0.1164,
+      "step": 1450
+    },
+    {
+      "epoch": 0.78,
+      "grad_norm": 1.2965114220197742,
+      "learning_rate": 2.6551724137931032e-05,
+      "loss": 0.1102,
+      "step": 1460
+    },
+    {
+      "epoch": 0.79,
+      "grad_norm": 1.8857303249108055,
+      "learning_rate": 2.637931034482759e-05,
+      "loss": 0.0868,
+      "step": 1470
+    },
+    {
+      "epoch": 0.79,
+      "grad_norm": 1.406207551120988,
+      "learning_rate": 2.620689655172414e-05,
+      "loss": 0.1179,
+      "step": 1480
+    },
+    {
+      "epoch": 0.8,
+      "grad_norm": 1.275728362064451,
+      "learning_rate": 2.6034482758620692e-05,
+      "loss": 0.1128,
+      "step": 1490
+    },
+    {
+      "epoch": 0.8,
+      "grad_norm": 1.7122434387720797,
+      "learning_rate": 2.5862068965517244e-05,
+      "loss": 0.1045,
+      "step": 1500
+    },
+    {
+      "epoch": 0.8,
+      "eval_loss": 0.6075221300125122,
+      "eval_runtime": 114.4297,
+      "eval_samples_per_second": 11.527,
+      "eval_steps_per_second": 2.884,
+      "step": 1500
+    }
+  ],
+  "logging_steps": 10,
+  "max_steps": 3000,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 2,
+  "save_steps": 500,
+  "total_flos": 43348525056000.0,
+  "train_batch_size": 4,
+  "trial_name": null,
+  "trial_params": null
+}

checkpoints/checkpoint-1500/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:22266252322e6977894297b61731ca9228c8fcfe4e8d788aef270218c40fec98
+size 6776

checkpoints/checkpoint-1500/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)

checkpoints/checkpoint-2000/README.md

@@ -0,0 +1,202 @@
+---
+library_name: peft
+base_model: openbmb/MiniCPM-2B-dpo-bf16
+---
+
+# 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 0.10.0

checkpoints/checkpoint-2000/adapter_config.json

@@ -0,0 +1,29 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "openbmb/MiniCPM-2B-dpo-bf16",
+  "bias": "none",
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": "gaussian",
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 8,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "v_proj",
+    "q_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "use_dora": false,
+  "use_rslora": false
+}

checkpoints/checkpoint-2000/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c1b8d76bd1aed8b1d8412eb6c56e84db43664bebd6ad688ad0950fd8209fc950
+size 5919456

checkpoints/checkpoint-2000/global_step2000/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:dda5a911ff898c12d55662a92778edfd52a8073eea19f9b26697459b3193b10a
+size 35393392

checkpoints/checkpoint-2000/global_step2000/zero_pp_rank_0_mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4e9a98f645b03701444c7e44c6c6a7f0356f803ea2720858c1e66a25b785255d
+size 183275

checkpoints/checkpoint-2000/latest

@@ -0,0 +1 @@
+global_step2000

checkpoints/checkpoint-2000/rng_state.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a818498d41465e958bf28c1ed45c89226bd9e0922529cc444009e8907e458b5c
+size 14244

checkpoints/checkpoint-2000/scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:047852d7b48f1f8dbf5d56d9f5244137e0e2504e7bca2d487065561fcb1ffc15
+size 1064

checkpoints/checkpoint-2000/special_tokens_map.json

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

checkpoints/checkpoint-2000/tokenizer_config.json

@@ -0,0 +1,42 @@
+{
+  "add_bos_token": true,
+  "add_eos_token": false,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<s>",
+  "chat_template": "{% for message in messages %}{% if message['role'] == 'user' %}{{'<用户>' + message['content'].strip() + '<AI>'}}{% else %}{{message['content'].strip()}}{% endif %}{% endfor %}",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "</s>",
+  "legacy": true,
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "</s>",
+  "sp_model_kwargs": {},
+  "spaces_between_special_tokens": false,
+  "tokenizer_class": "LlamaTokenizer",
+  "unk_token": "<unk>",
+  "use_default_system_prompt": false
+}

checkpoints/checkpoint-2000/trainer_state.json

@@ -0,0 +1,1453 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 1.070090957731407,
+  "eval_steps": 500,
+  "global_step": 2000,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.01,
+      "grad_norm": 7.3019086777975994,
+      "learning_rate": 5e-06,
+      "loss": 0.6939,
+      "step": 10
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 5.150989333266983,
+      "learning_rate": 1e-05,
+      "loss": 0.7167,
+      "step": 20
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 3.640558809610037,
+      "learning_rate": 1.5e-05,
+      "loss": 0.5683,
+      "step": 30
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 7.517999397731128,
+      "learning_rate": 2e-05,
+      "loss": 0.5472,
+      "step": 40
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 1.9687061679463425,
+      "learning_rate": 2.5e-05,
+      "loss": 0.4439,
+      "step": 50
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 3.643479206523606,
+      "learning_rate": 3e-05,
+      "loss": 0.2486,
+      "step": 60
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 2.2754773308695095,
+      "learning_rate": 3.5e-05,
+      "loss": 0.2217,
+      "step": 70
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 1.7144730049127388,
+      "learning_rate": 4e-05,
+      "loss": 0.169,
+      "step": 80
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 3.4702829704135114,
+      "learning_rate": 4.5e-05,
+      "loss": 0.1994,
+      "step": 90
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 1.3985127340985621,
+      "learning_rate": 5e-05,
+      "loss": 0.1612,
+      "step": 100
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 1.375992184386137,
+      "learning_rate": 4.982758620689655e-05,
+      "loss": 0.1576,
+      "step": 110
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 1.9528635753013313,
+      "learning_rate": 4.9655172413793107e-05,
+      "loss": 0.1393,
+      "step": 120
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 4.075169010198401,
+      "learning_rate": 4.9482758620689655e-05,
+      "loss": 0.1969,
+      "step": 130
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 2.0953991165751207,
+      "learning_rate": 4.931034482758621e-05,
+      "loss": 0.1294,
+      "step": 140
+    },
+    {
+      "epoch": 0.08,
+      "grad_norm": 1.942660591044849,
+      "learning_rate": 4.913793103448276e-05,
+      "loss": 0.1306,
+      "step": 150
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 3.1508904015728345,
+      "learning_rate": 4.896551724137931e-05,
+      "loss": 0.1526,
+      "step": 160
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 1.9862795165358471,
+      "learning_rate": 4.8793103448275864e-05,
+      "loss": 0.1186,
+      "step": 170
+    },
+    {
+      "epoch": 0.1,
+      "grad_norm": 2.633061833817991,
+      "learning_rate": 4.862068965517241e-05,
+      "loss": 0.1457,
+      "step": 180
+    },
+    {
+      "epoch": 0.1,
+      "grad_norm": 1.8017052368446178,
+      "learning_rate": 4.844827586206897e-05,
+      "loss": 0.1234,
+      "step": 190
+    },
+    {
+      "epoch": 0.11,
+      "grad_norm": 2.1560694100709803,
+      "learning_rate": 4.827586206896552e-05,
+      "loss": 0.1346,
+      "step": 200
+    },
+    {
+      "epoch": 0.11,
+      "grad_norm": 1.5737689267430703,
+      "learning_rate": 4.810344827586207e-05,
+      "loss": 0.116,
+      "step": 210
+    },
+    {
+      "epoch": 0.12,
+      "grad_norm": 1.957864677854788,
+      "learning_rate": 4.793103448275863e-05,
+      "loss": 0.1692,
+      "step": 220
+    },
+    {
+      "epoch": 0.12,
+      "grad_norm": 2.215039223521855,
+      "learning_rate": 4.7758620689655176e-05,
+      "loss": 0.1245,
+      "step": 230
+    },
+    {
+      "epoch": 0.13,
+      "grad_norm": 1.370517239734168,
+      "learning_rate": 4.7586206896551725e-05,
+      "loss": 0.1476,
+      "step": 240
+    },
+    {
+      "epoch": 0.13,
+      "grad_norm": 1.7341334563022532,
+      "learning_rate": 4.741379310344828e-05,
+      "loss": 0.1236,
+      "step": 250
+    },
+    {
+      "epoch": 0.14,
+      "grad_norm": 1.5994298113068974,
+      "learning_rate": 4.724137931034483e-05,
+      "loss": 0.1161,
+      "step": 260
+    },
+    {
+      "epoch": 0.14,
+      "grad_norm": 1.5317433190951963,
+      "learning_rate": 4.7068965517241385e-05,
+      "loss": 0.1035,
+      "step": 270
+    },
+    {
+      "epoch": 0.15,
+      "grad_norm": 2.191977732539556,
+      "learning_rate": 4.689655172413793e-05,
+      "loss": 0.1427,
+      "step": 280
+    },
+    {
+      "epoch": 0.16,
+      "grad_norm": 1.6038667570691656,
+      "learning_rate": 4.672413793103448e-05,
+      "loss": 0.1225,
+      "step": 290
+    },
+    {
+      "epoch": 0.16,
+      "grad_norm": 2.577572731831179,
+      "learning_rate": 4.655172413793104e-05,
+      "loss": 0.1399,
+      "step": 300
+    },
+    {
+      "epoch": 0.17,
+      "grad_norm": 1.6199241001441385,
+      "learning_rate": 4.6379310344827586e-05,
+      "loss": 0.1242,
+      "step": 310
+    },
+    {
+      "epoch": 0.17,
+      "grad_norm": 2.236577821186196,
+      "learning_rate": 4.6206896551724135e-05,
+      "loss": 0.1656,
+      "step": 320
+    },
+    {
+      "epoch": 0.18,
+      "grad_norm": 1.7294690605254757,
+      "learning_rate": 4.603448275862069e-05,
+      "loss": 0.1382,
+      "step": 330
+    },
+    {
+      "epoch": 0.18,
+      "grad_norm": 2.196527516378511,
+      "learning_rate": 4.586206896551724e-05,
+      "loss": 0.1257,
+      "step": 340
+    },
+    {
+      "epoch": 0.19,
+      "grad_norm": 2.1057444340221463,
+      "learning_rate": 4.5689655172413794e-05,
+      "loss": 0.1238,
+      "step": 350
+    },
+    {
+      "epoch": 0.19,
+      "grad_norm": 1.5409556870328274,
+      "learning_rate": 4.551724137931035e-05,
+      "loss": 0.1383,
+      "step": 360
+    },
+    {
+      "epoch": 0.2,
+      "grad_norm": 1.5204083616874053,
+      "learning_rate": 4.53448275862069e-05,
+      "loss": 0.1068,
+      "step": 370
+    },
+    {
+      "epoch": 0.2,
+      "grad_norm": 2.3557725298931746,
+      "learning_rate": 4.5172413793103454e-05,
+      "loss": 0.1071,
+      "step": 380
+    },
+    {
+      "epoch": 0.21,
+      "grad_norm": 3.2601538460418644,
+      "learning_rate": 4.5e-05,
+      "loss": 0.125,
+      "step": 390
+    },
+    {
+      "epoch": 0.21,
+      "grad_norm": 1.9031725385762286,
+      "learning_rate": 4.482758620689655e-05,
+      "loss": 0.0991,
+      "step": 400
+    },
+    {
+      "epoch": 0.22,
+      "grad_norm": 1.3946050262183123,
+      "learning_rate": 4.465517241379311e-05,
+      "loss": 0.1156,
+      "step": 410
+    },
+    {
+      "epoch": 0.22,
+      "grad_norm": 1.097644875106397,
+      "learning_rate": 4.4482758620689656e-05,
+      "loss": 0.1366,
+      "step": 420
+    },
+    {
+      "epoch": 0.23,
+      "grad_norm": 1.37846299019108,
+      "learning_rate": 4.431034482758621e-05,
+      "loss": 0.126,
+      "step": 430
+    },
+    {
+      "epoch": 0.24,
+      "grad_norm": 1.8340152889320331,
+      "learning_rate": 4.413793103448276e-05,
+      "loss": 0.1066,
+      "step": 440
+    },
+    {
+      "epoch": 0.24,
+      "grad_norm": 1.8304505611337867,
+      "learning_rate": 4.396551724137931e-05,
+      "loss": 0.0868,
+      "step": 450
+    },
+    {
+      "epoch": 0.25,
+      "grad_norm": 1.550196490898523,
+      "learning_rate": 4.3793103448275864e-05,
+      "loss": 0.1286,
+      "step": 460
+    },
+    {
+      "epoch": 0.25,
+      "grad_norm": 2.176112247796248,
+      "learning_rate": 4.362068965517241e-05,
+      "loss": 0.1206,
+      "step": 470
+    },
+    {
+      "epoch": 0.26,
+      "grad_norm": 1.6589263894091213,
+      "learning_rate": 4.344827586206897e-05,
+      "loss": 0.1008,
+      "step": 480
+    },
+    {
+      "epoch": 0.26,
+      "grad_norm": 1.8349611508902046,
+      "learning_rate": 4.327586206896552e-05,
+      "loss": 0.1198,
+      "step": 490
+    },
+    {
+      "epoch": 0.27,
+      "grad_norm": 2.1218964920724126,
+      "learning_rate": 4.3103448275862066e-05,
+      "loss": 0.1166,
+      "step": 500
+    },
+    {
+      "epoch": 0.27,
+      "eval_loss": 0.6078919172286987,
+      "eval_runtime": 116.8471,
+      "eval_samples_per_second": 11.288,
+      "eval_steps_per_second": 2.824,
+      "step": 500
+    },
+    {
+      "epoch": 0.27,
+      "grad_norm": 2.5775141311007856,
+      "learning_rate": 4.293103448275863e-05,
+      "loss": 0.1124,
+      "step": 510
+    },
+    {
+      "epoch": 0.28,
+      "grad_norm": 1.6019517017800202,
+      "learning_rate": 4.275862068965518e-05,
+      "loss": 0.1068,
+      "step": 520
+    },
+    {
+      "epoch": 0.28,
+      "grad_norm": 2.6901962755310205,
+      "learning_rate": 4.2586206896551725e-05,
+      "loss": 0.1286,
+      "step": 530
+    },
+    {
+      "epoch": 0.29,
+      "grad_norm": 3.9517995356721767,
+      "learning_rate": 4.241379310344828e-05,
+      "loss": 0.1149,
+      "step": 540
+    },
+    {
+      "epoch": 0.29,
+      "grad_norm": 2.0428896228074076,
+      "learning_rate": 4.224137931034483e-05,
+      "loss": 0.141,
+      "step": 550
+    },
+    {
+      "epoch": 0.3,
+      "grad_norm": 2.263258592133553,
+      "learning_rate": 4.2068965517241385e-05,
+      "loss": 0.0949,
+      "step": 560
+    },
+    {
+      "epoch": 0.3,
+      "grad_norm": 1.4823165953974604,
+      "learning_rate": 4.1896551724137934e-05,
+      "loss": 0.1365,
+      "step": 570
+    },
+    {
+      "epoch": 0.31,
+      "grad_norm": 1.4441377020989878,
+      "learning_rate": 4.172413793103448e-05,
+      "loss": 0.1015,
+      "step": 580
+    },
+    {
+      "epoch": 0.32,
+      "grad_norm": 1.4779059254436886,
+      "learning_rate": 4.155172413793104e-05,
+      "loss": 0.0988,
+      "step": 590
+    },
+    {
+      "epoch": 0.32,
+      "grad_norm": 1.7777823671018818,
+      "learning_rate": 4.1379310344827587e-05,
+      "loss": 0.1124,
+      "step": 600
+    },
+    {
+      "epoch": 0.33,
+      "grad_norm": 1.737579831138191,
+      "learning_rate": 4.120689655172414e-05,
+      "loss": 0.086,
+      "step": 610
+    },
+    {
+      "epoch": 0.33,
+      "grad_norm": 2.708453961232997,
+      "learning_rate": 4.103448275862069e-05,
+      "loss": 0.0933,
+      "step": 620
+    },
+    {
+      "epoch": 0.34,
+      "grad_norm": 1.8871805824236731,
+      "learning_rate": 4.086206896551724e-05,
+      "loss": 0.1407,
+      "step": 630
+    },
+    {
+      "epoch": 0.34,
+      "grad_norm": 1.7300112722427339,
+      "learning_rate": 4.0689655172413795e-05,
+      "loss": 0.1224,
+      "step": 640
+    },
+    {
+      "epoch": 0.35,
+      "grad_norm": 1.4631236252240614,
+      "learning_rate": 4.0517241379310344e-05,
+      "loss": 0.1014,
+      "step": 650
+    },
+    {
+      "epoch": 0.35,
+      "grad_norm": 1.2602431597419264,
+      "learning_rate": 4.03448275862069e-05,
+      "loss": 0.1583,
+      "step": 660
+    },
+    {
+      "epoch": 0.36,
+      "grad_norm": 1.2077937041919453,
+      "learning_rate": 4.0172413793103455e-05,
+      "loss": 0.1209,
+      "step": 670
+    },
+    {
+      "epoch": 0.36,
+      "grad_norm": 1.4386184566429954,
+      "learning_rate": 4e-05,
+      "loss": 0.1016,
+      "step": 680
+    },
+    {
+      "epoch": 0.37,
+      "grad_norm": 2.6160358835758584,
+      "learning_rate": 3.982758620689656e-05,
+      "loss": 0.1062,
+      "step": 690
+    },
+    {
+      "epoch": 0.37,
+      "grad_norm": 1.9278794640498955,
+      "learning_rate": 3.965517241379311e-05,
+      "loss": 0.1037,
+      "step": 700
+    },
+    {
+      "epoch": 0.38,
+      "grad_norm": 1.2872571900237024,
+      "learning_rate": 3.9482758620689656e-05,
+      "loss": 0.096,
+      "step": 710
+    },
+    {
+      "epoch": 0.39,
+      "grad_norm": 1.243554309347296,
+      "learning_rate": 3.931034482758621e-05,
+      "loss": 0.1084,
+      "step": 720
+    },
+    {
+      "epoch": 0.39,
+      "grad_norm": 1.5090589714253309,
+      "learning_rate": 3.913793103448276e-05,
+      "loss": 0.0877,
+      "step": 730
+    },
+    {
+      "epoch": 0.4,
+      "grad_norm": 2.1419550623025168,
+      "learning_rate": 3.896551724137931e-05,
+      "loss": 0.0994,
+      "step": 740
+    },
+    {
+      "epoch": 0.4,
+      "grad_norm": 1.7807417973632438,
+      "learning_rate": 3.8793103448275865e-05,
+      "loss": 0.1029,
+      "step": 750
+    },
+    {
+      "epoch": 0.41,
+      "grad_norm": 1.3342960263057682,
+      "learning_rate": 3.862068965517241e-05,
+      "loss": 0.1072,
+      "step": 760
+    },
+    {
+      "epoch": 0.41,
+      "grad_norm": 2.3865282340158136,
+      "learning_rate": 3.844827586206897e-05,
+      "loss": 0.1193,
+      "step": 770
+    },
+    {
+      "epoch": 0.42,
+      "grad_norm": 1.5428742248459941,
+      "learning_rate": 3.827586206896552e-05,
+      "loss": 0.1156,
+      "step": 780
+    },
+    {
+      "epoch": 0.42,
+      "grad_norm": 1.7660532115509044,
+      "learning_rate": 3.8103448275862066e-05,
+      "loss": 0.122,
+      "step": 790
+    },
+    {
+      "epoch": 0.43,
+      "grad_norm": 1.8149742752994733,
+      "learning_rate": 3.793103448275862e-05,
+      "loss": 0.1346,
+      "step": 800
+    },
+    {
+      "epoch": 0.43,
+      "grad_norm": 1.7456408876472995,
+      "learning_rate": 3.775862068965517e-05,
+      "loss": 0.1223,
+      "step": 810
+    },
+    {
+      "epoch": 0.44,
+      "grad_norm": 1.10163248244056,
+      "learning_rate": 3.7586206896551726e-05,
+      "loss": 0.1031,
+      "step": 820
+    },
+    {
+      "epoch": 0.44,
+      "grad_norm": 1.6441057737088702,
+      "learning_rate": 3.741379310344828e-05,
+      "loss": 0.1059,
+      "step": 830
+    },
+    {
+      "epoch": 0.45,
+      "grad_norm": 2.3999279790163484,
+      "learning_rate": 3.724137931034483e-05,
+      "loss": 0.1125,
+      "step": 840
+    },
+    {
+      "epoch": 0.45,
+      "grad_norm": 2.2081477934156903,
+      "learning_rate": 3.7068965517241385e-05,
+      "loss": 0.1266,
+      "step": 850
+    },
+    {
+      "epoch": 0.46,
+      "grad_norm": 2.366783771480017,
+      "learning_rate": 3.6896551724137934e-05,
+      "loss": 0.1127,
+      "step": 860
+    },
+    {
+      "epoch": 0.47,
+      "grad_norm": 1.3077873674136173,
+      "learning_rate": 3.672413793103448e-05,
+      "loss": 0.1095,
+      "step": 870
+    },
+    {
+      "epoch": 0.47,
+      "grad_norm": 1.8197812508114701,
+      "learning_rate": 3.655172413793104e-05,
+      "loss": 0.0932,
+      "step": 880
+    },
+    {
+      "epoch": 0.48,
+      "grad_norm": 1.0806192057981219,
+      "learning_rate": 3.637931034482759e-05,
+      "loss": 0.1166,
+      "step": 890
+    },
+    {
+      "epoch": 0.48,
+      "grad_norm": 1.520666439337001,
+      "learning_rate": 3.620689655172414e-05,
+      "loss": 0.0883,
+      "step": 900
+    },
+    {
+      "epoch": 0.49,
+      "grad_norm": 1.690002270629302,
+      "learning_rate": 3.603448275862069e-05,
+      "loss": 0.1199,
+      "step": 910
+    },
+    {
+      "epoch": 0.49,
+      "grad_norm": 1.4319374130118003,
+      "learning_rate": 3.586206896551724e-05,
+      "loss": 0.0991,
+      "step": 920
+    },
+    {
+      "epoch": 0.5,
+      "grad_norm": 1.0626084369653164,
+      "learning_rate": 3.5689655172413795e-05,
+      "loss": 0.0923,
+      "step": 930
+    },
+    {
+      "epoch": 0.5,
+      "grad_norm": 2.0848060597460902,
+      "learning_rate": 3.5517241379310344e-05,
+      "loss": 0.0979,
+      "step": 940
+    },
+    {
+      "epoch": 0.51,
+      "grad_norm": 1.4997189461483256,
+      "learning_rate": 3.53448275862069e-05,
+      "loss": 0.0949,
+      "step": 950
+    },
+    {
+      "epoch": 0.51,
+      "grad_norm": 1.7887817042743388,
+      "learning_rate": 3.517241379310345e-05,
+      "loss": 0.1135,
+      "step": 960
+    },
+    {
+      "epoch": 0.52,
+      "grad_norm": 3.242965692388458,
+      "learning_rate": 3.5e-05,
+      "loss": 0.1315,
+      "step": 970
+    },
+    {
+      "epoch": 0.52,
+      "grad_norm": 1.5034762176322083,
+      "learning_rate": 3.482758620689655e-05,
+      "loss": 0.1177,
+      "step": 980
+    },
+    {
+      "epoch": 0.53,
+      "grad_norm": 1.6679474444200848,
+      "learning_rate": 3.465517241379311e-05,
+      "loss": 0.1182,
+      "step": 990
+    },
+    {
+      "epoch": 0.54,
+      "grad_norm": 2.814574507251776,
+      "learning_rate": 3.4482758620689657e-05,
+      "loss": 0.0912,
+      "step": 1000
+    },
+    {
+      "epoch": 0.54,
+      "eval_loss": 0.6101276874542236,
+      "eval_runtime": 113.9995,
+      "eval_samples_per_second": 11.57,
+      "eval_steps_per_second": 2.895,
+      "step": 1000
+    },
+    {
+      "epoch": 0.54,
+      "grad_norm": 2.1321319681580535,
+      "learning_rate": 3.431034482758621e-05,
+      "loss": 0.0983,
+      "step": 1010
+    },
+    {
+      "epoch": 0.55,
+      "grad_norm": 1.8915101367452352,
+      "learning_rate": 3.413793103448276e-05,
+      "loss": 0.1113,
+      "step": 1020
+    },
+    {
+      "epoch": 0.55,
+      "grad_norm": 1.4160985095468477,
+      "learning_rate": 3.3965517241379316e-05,
+      "loss": 0.1076,
+      "step": 1030
+    },
+    {
+      "epoch": 0.56,
+      "grad_norm": 1.8562788974878586,
+      "learning_rate": 3.3793103448275865e-05,
+      "loss": 0.1011,
+      "step": 1040
+    },
+    {
+      "epoch": 0.56,
+      "grad_norm": 1.3793192563691294,
+      "learning_rate": 3.3620689655172414e-05,
+      "loss": 0.0978,
+      "step": 1050
+    },
+    {
+      "epoch": 0.57,
+      "grad_norm": 1.4606563129628805,
+      "learning_rate": 3.344827586206897e-05,
+      "loss": 0.1069,
+      "step": 1060
+    },
+    {
+      "epoch": 0.57,
+      "grad_norm": 1.5680856211032999,
+      "learning_rate": 3.327586206896552e-05,
+      "loss": 0.0988,
+      "step": 1070
+    },
+    {
+      "epoch": 0.58,
+      "grad_norm": 1.5829345931951275,
+      "learning_rate": 3.310344827586207e-05,
+      "loss": 0.1256,
+      "step": 1080
+    },
+    {
+      "epoch": 0.58,
+      "grad_norm": 1.6200852939319585,
+      "learning_rate": 3.293103448275862e-05,
+      "loss": 0.097,
+      "step": 1090
+    },
+    {
+      "epoch": 0.59,
+      "grad_norm": 2.259656836213122,
+      "learning_rate": 3.275862068965517e-05,
+      "loss": 0.1137,
+      "step": 1100
+    },
+    {
+      "epoch": 0.59,
+      "grad_norm": 2.2483622341560645,
+      "learning_rate": 3.2586206896551726e-05,
+      "loss": 0.0999,
+      "step": 1110
+    },
+    {
+      "epoch": 0.6,
+      "grad_norm": 1.168198861956421,
+      "learning_rate": 3.2413793103448275e-05,
+      "loss": 0.1018,
+      "step": 1120
+    },
+    {
+      "epoch": 0.6,
+      "grad_norm": 2.4699824799031482,
+      "learning_rate": 3.2241379310344824e-05,
+      "loss": 0.1132,
+      "step": 1130
+    },
+    {
+      "epoch": 0.61,
+      "grad_norm": 1.2571654549549751,
+      "learning_rate": 3.206896551724138e-05,
+      "loss": 0.1054,
+      "step": 1140
+    },
+    {
+      "epoch": 0.62,
+      "grad_norm": 0.5559534032307631,
+      "learning_rate": 3.1896551724137935e-05,
+      "loss": 0.0789,
+      "step": 1150
+    },
+    {
+      "epoch": 0.62,
+      "grad_norm": 1.636369759504475,
+      "learning_rate": 3.172413793103448e-05,
+      "loss": 0.0902,
+      "step": 1160
+    },
+    {
+      "epoch": 0.63,
+      "grad_norm": 1.6137142935446496,
+      "learning_rate": 3.155172413793104e-05,
+      "loss": 0.1199,
+      "step": 1170
+    },
+    {
+      "epoch": 0.63,
+      "grad_norm": 1.7448003760796802,
+      "learning_rate": 3.137931034482759e-05,
+      "loss": 0.1295,
+      "step": 1180
+    },
+    {
+      "epoch": 0.64,
+      "grad_norm": 1.3261005358227276,
+      "learning_rate": 3.120689655172414e-05,
+      "loss": 0.1117,
+      "step": 1190
+    },
+    {
+      "epoch": 0.64,
+      "grad_norm": 1.7353127177901462,
+      "learning_rate": 3.103448275862069e-05,
+      "loss": 0.0951,
+      "step": 1200
+    },
+    {
+      "epoch": 0.65,
+      "grad_norm": 2.8569975913367074,
+      "learning_rate": 3.086206896551724e-05,
+      "loss": 0.112,
+      "step": 1210
+    },
+    {
+      "epoch": 0.65,
+      "grad_norm": 1.3481947218871082,
+      "learning_rate": 3.0689655172413796e-05,
+      "loss": 0.0876,
+      "step": 1220
+    },
+    {
+      "epoch": 0.66,
+      "grad_norm": 2.015933141613929,
+      "learning_rate": 3.0517241379310348e-05,
+      "loss": 0.0993,
+      "step": 1230
+    },
+    {
+      "epoch": 0.66,
+      "grad_norm": 1.0588164394448019,
+      "learning_rate": 3.0344827586206897e-05,
+      "loss": 0.1034,
+      "step": 1240
+    },
+    {
+      "epoch": 0.67,
+      "grad_norm": 1.3594986645993228,
+      "learning_rate": 3.017241379310345e-05,
+      "loss": 0.1004,
+      "step": 1250
+    },
+    {
+      "epoch": 0.67,
+      "grad_norm": 1.333098402625009,
+      "learning_rate": 3e-05,
+      "loss": 0.1259,
+      "step": 1260
+    },
+    {
+      "epoch": 0.68,
+      "grad_norm": 1.1324206075196583,
+      "learning_rate": 2.9827586206896553e-05,
+      "loss": 0.1155,
+      "step": 1270
+    },
+    {
+      "epoch": 0.68,
+      "grad_norm": 1.2270687927795876,
+      "learning_rate": 2.96551724137931e-05,
+      "loss": 0.092,
+      "step": 1280
+    },
+    {
+      "epoch": 0.69,
+      "grad_norm": 1.3085362507403875,
+      "learning_rate": 2.9482758620689654e-05,
+      "loss": 0.1064,
+      "step": 1290
+    },
+    {
+      "epoch": 0.7,
+      "grad_norm": 1.7135250277750762,
+      "learning_rate": 2.9310344827586206e-05,
+      "loss": 0.1132,
+      "step": 1300
+    },
+    {
+      "epoch": 0.7,
+      "grad_norm": 1.6121189707451158,
+      "learning_rate": 2.913793103448276e-05,
+      "loss": 0.1006,
+      "step": 1310
+    },
+    {
+      "epoch": 0.71,
+      "grad_norm": 1.3958680925504208,
+      "learning_rate": 2.8965517241379313e-05,
+      "loss": 0.0956,
+      "step": 1320
+    },
+    {
+      "epoch": 0.71,
+      "grad_norm": 1.546226150121884,
+      "learning_rate": 2.8793103448275865e-05,
+      "loss": 0.1142,
+      "step": 1330
+    },
+    {
+      "epoch": 0.72,
+      "grad_norm": 2.1837739995965415,
+      "learning_rate": 2.8620689655172417e-05,
+      "loss": 0.1127,
+      "step": 1340
+    },
+    {
+      "epoch": 0.72,
+      "grad_norm": 1.9402402206909504,
+      "learning_rate": 2.844827586206897e-05,
+      "loss": 0.0922,
+      "step": 1350
+    },
+    {
+      "epoch": 0.73,
+      "grad_norm": 1.7914401953164802,
+      "learning_rate": 2.8275862068965518e-05,
+      "loss": 0.1038,
+      "step": 1360
+    },
+    {
+      "epoch": 0.73,
+      "grad_norm": 1.505804090650568,
+      "learning_rate": 2.810344827586207e-05,
+      "loss": 0.1034,
+      "step": 1370
+    },
+    {
+      "epoch": 0.74,
+      "grad_norm": 1.9907350713586716,
+      "learning_rate": 2.7931034482758622e-05,
+      "loss": 0.103,
+      "step": 1380
+    },
+    {
+      "epoch": 0.74,
+      "grad_norm": 1.6948381773166858,
+      "learning_rate": 2.7758620689655175e-05,
+      "loss": 0.1091,
+      "step": 1390
+    },
+    {
+      "epoch": 0.75,
+      "grad_norm": 1.3995985437024723,
+      "learning_rate": 2.7586206896551727e-05,
+      "loss": 0.0852,
+      "step": 1400
+    },
+    {
+      "epoch": 0.75,
+      "grad_norm": 1.9347024029069393,
+      "learning_rate": 2.7413793103448275e-05,
+      "loss": 0.1393,
+      "step": 1410
+    },
+    {
+      "epoch": 0.76,
+      "grad_norm": 1.608776792445342,
+      "learning_rate": 2.7241379310344827e-05,
+      "loss": 0.0951,
+      "step": 1420
+    },
+    {
+      "epoch": 0.77,
+      "grad_norm": 1.6005483580619249,
+      "learning_rate": 2.706896551724138e-05,
+      "loss": 0.1037,
+      "step": 1430
+    },
+    {
+      "epoch": 0.77,
+      "grad_norm": 2.374208686020403,
+      "learning_rate": 2.689655172413793e-05,
+      "loss": 0.0926,
+      "step": 1440
+    },
+    {
+      "epoch": 0.78,
+      "grad_norm": 1.7554923995400171,
+      "learning_rate": 2.672413793103448e-05,
+      "loss": 0.1164,
+      "step": 1450
+    },
+    {
+      "epoch": 0.78,
+      "grad_norm": 1.2965114220197742,
+      "learning_rate": 2.6551724137931032e-05,
+      "loss": 0.1102,
+      "step": 1460
+    },
+    {
+      "epoch": 0.79,
+      "grad_norm": 1.8857303249108055,
+      "learning_rate": 2.637931034482759e-05,
+      "loss": 0.0868,
+      "step": 1470
+    },
+    {
+      "epoch": 0.79,
+      "grad_norm": 1.406207551120988,
+      "learning_rate": 2.620689655172414e-05,
+      "loss": 0.1179,
+      "step": 1480
+    },
+    {
+      "epoch": 0.8,
+      "grad_norm": 1.275728362064451,
+      "learning_rate": 2.6034482758620692e-05,
+      "loss": 0.1128,
+      "step": 1490
+    },
+    {
+      "epoch": 0.8,
+      "grad_norm": 1.7122434387720797,
+      "learning_rate": 2.5862068965517244e-05,
+      "loss": 0.1045,
+      "step": 1500
+    },
+    {
+      "epoch": 0.8,
+      "eval_loss": 0.6075221300125122,
+      "eval_runtime": 114.4297,
+      "eval_samples_per_second": 11.527,
+      "eval_steps_per_second": 2.884,
+      "step": 1500
+    },
+    {
+      "epoch": 0.81,
+      "grad_norm": 1.3880467648229133,
+      "learning_rate": 2.5689655172413796e-05,
+      "loss": 0.1031,
+      "step": 1510
+    },
+    {
+      "epoch": 0.81,
+      "grad_norm": 2.3475935500456657,
+      "learning_rate": 2.551724137931035e-05,
+      "loss": 0.1064,
+      "step": 1520
+    },
+    {
+      "epoch": 0.82,
+      "grad_norm": 1.9746447047097486,
+      "learning_rate": 2.5344827586206897e-05,
+      "loss": 0.0995,
+      "step": 1530
+    },
+    {
+      "epoch": 0.82,
+      "grad_norm": 1.2244623226185645,
+      "learning_rate": 2.517241379310345e-05,
+      "loss": 0.1044,
+      "step": 1540
+    },
+    {
+      "epoch": 0.83,
+      "grad_norm": 2.2605012043265216,
+      "learning_rate": 2.5e-05,
+      "loss": 0.1017,
+      "step": 1550
+    },
+    {
+      "epoch": 0.83,
+      "grad_norm": 1.624612730097256,
+      "learning_rate": 2.4827586206896553e-05,
+      "loss": 0.0973,
+      "step": 1560
+    },
+    {
+      "epoch": 0.84,
+      "grad_norm": 1.3648662151461801,
+      "learning_rate": 2.4655172413793105e-05,
+      "loss": 0.0836,
+      "step": 1570
+    },
+    {
+      "epoch": 0.85,
+      "grad_norm": 1.4642386177378814,
+      "learning_rate": 2.4482758620689654e-05,
+      "loss": 0.1058,
+      "step": 1580
+    },
+    {
+      "epoch": 0.85,
+      "grad_norm": 1.3057087388796036,
+      "learning_rate": 2.4310344827586206e-05,
+      "loss": 0.1092,
+      "step": 1590
+    },
+    {
+      "epoch": 0.86,
+      "grad_norm": 2.092348689417081,
+      "learning_rate": 2.413793103448276e-05,
+      "loss": 0.0958,
+      "step": 1600
+    },
+    {
+      "epoch": 0.86,
+      "grad_norm": 1.3731115281087607,
+      "learning_rate": 2.3965517241379314e-05,
+      "loss": 0.0883,
+      "step": 1610
+    },
+    {
+      "epoch": 0.87,
+      "grad_norm": 1.3983722722394134,
+      "learning_rate": 2.3793103448275862e-05,
+      "loss": 0.0944,
+      "step": 1620
+    },
+    {
+      "epoch": 0.87,
+      "grad_norm": 2.4163442262002937,
+      "learning_rate": 2.3620689655172415e-05,
+      "loss": 0.1062,
+      "step": 1630
+    },
+    {
+      "epoch": 0.88,
+      "grad_norm": 1.5006079944233688,
+      "learning_rate": 2.3448275862068967e-05,
+      "loss": 0.1129,
+      "step": 1640
+    },
+    {
+      "epoch": 0.88,
+      "grad_norm": 1.54283270803711,
+      "learning_rate": 2.327586206896552e-05,
+      "loss": 0.0861,
+      "step": 1650
+    },
+    {
+      "epoch": 0.89,
+      "grad_norm": 2.269678144223195,
+      "learning_rate": 2.3103448275862067e-05,
+      "loss": 0.0978,
+      "step": 1660
+    },
+    {
+      "epoch": 0.89,
+      "grad_norm": 1.8105959776831768,
+      "learning_rate": 2.293103448275862e-05,
+      "loss": 0.1141,
+      "step": 1670
+    },
+    {
+      "epoch": 0.9,
+      "grad_norm": 1.9135579713863027,
+      "learning_rate": 2.2758620689655175e-05,
+      "loss": 0.0918,
+      "step": 1680
+    },
+    {
+      "epoch": 0.9,
+      "grad_norm": 1.6604450253018581,
+      "learning_rate": 2.2586206896551727e-05,
+      "loss": 0.096,
+      "step": 1690
+    },
+    {
+      "epoch": 0.91,
+      "grad_norm": 2.115987565053461,
+      "learning_rate": 2.2413793103448276e-05,
+      "loss": 0.1166,
+      "step": 1700
+    },
+    {
+      "epoch": 0.91,
+      "grad_norm": 1.4279927513160544,
+      "learning_rate": 2.2241379310344828e-05,
+      "loss": 0.1052,
+      "step": 1710
+    },
+    {
+      "epoch": 0.92,
+      "grad_norm": 1.185880441960968,
+      "learning_rate": 2.206896551724138e-05,
+      "loss": 0.1053,
+      "step": 1720
+    },
+    {
+      "epoch": 0.93,
+      "grad_norm": 1.6969029997828415,
+      "learning_rate": 2.1896551724137932e-05,
+      "loss": 0.106,
+      "step": 1730
+    },
+    {
+      "epoch": 0.93,
+      "grad_norm": 2.0330998697970286,
+      "learning_rate": 2.1724137931034484e-05,
+      "loss": 0.1073,
+      "step": 1740
+    },
+    {
+      "epoch": 0.94,
+      "grad_norm": 1.19027851417408,
+      "learning_rate": 2.1551724137931033e-05,
+      "loss": 0.091,
+      "step": 1750
+    },
+    {
+      "epoch": 0.94,
+      "grad_norm": 1.2470713090698218,
+      "learning_rate": 2.137931034482759e-05,
+      "loss": 0.0898,
+      "step": 1760
+    },
+    {
+      "epoch": 0.95,
+      "grad_norm": 2.235740059996042,
+      "learning_rate": 2.120689655172414e-05,
+      "loss": 0.1327,
+      "step": 1770
+    },
+    {
+      "epoch": 0.95,
+      "grad_norm": 1.5741742016710085,
+      "learning_rate": 2.1034482758620692e-05,
+      "loss": 0.1126,
+      "step": 1780
+    },
+    {
+      "epoch": 0.96,
+      "grad_norm": 0.9343547126371113,
+      "learning_rate": 2.086206896551724e-05,
+      "loss": 0.0819,
+      "step": 1790
+    },
+    {
+      "epoch": 0.96,
+      "grad_norm": 2.2764271447338937,
+      "learning_rate": 2.0689655172413793e-05,
+      "loss": 0.1204,
+      "step": 1800
+    },
+    {
+      "epoch": 0.97,
+      "grad_norm": 1.981384842073209,
+      "learning_rate": 2.0517241379310345e-05,
+      "loss": 0.1182,
+      "step": 1810
+    },
+    {
+      "epoch": 0.97,
+      "grad_norm": 1.044063198588911,
+      "learning_rate": 2.0344827586206897e-05,
+      "loss": 0.1005,
+      "step": 1820
+    },
+    {
+      "epoch": 0.98,
+      "grad_norm": 2.370183172473789,
+      "learning_rate": 2.017241379310345e-05,
+      "loss": 0.1174,
+      "step": 1830
+    },
+    {
+      "epoch": 0.98,
+      "grad_norm": 1.9052733799672823,
+      "learning_rate": 2e-05,
+      "loss": 0.1125,
+      "step": 1840
+    },
+    {
+      "epoch": 0.99,
+      "grad_norm": 1.628277530114902,
+      "learning_rate": 1.9827586206896554e-05,
+      "loss": 0.1015,
+      "step": 1850
+    },
+    {
+      "epoch": 1.0,
+      "grad_norm": 1.2522124245986863,
+      "learning_rate": 1.9655172413793106e-05,
+      "loss": 0.0924,
+      "step": 1860
+    },
+    {
+      "epoch": 1.0,
+      "grad_norm": 0.6911426489002421,
+      "learning_rate": 1.9482758620689655e-05,
+      "loss": 0.0965,
+      "step": 1870
+    },
+    {
+      "epoch": 1.01,
+      "grad_norm": 1.821890613342771,
+      "learning_rate": 1.9310344827586207e-05,
+      "loss": 0.081,
+      "step": 1880
+    },
+    {
+      "epoch": 1.01,
+      "grad_norm": 0.7643588179781782,
+      "learning_rate": 1.913793103448276e-05,
+      "loss": 0.0761,
+      "step": 1890
+    },
+    {
+      "epoch": 1.02,
+      "grad_norm": 1.1095002263403428,
+      "learning_rate": 1.896551724137931e-05,
+      "loss": 0.0871,
+      "step": 1900
+    },
+    {
+      "epoch": 1.02,
+      "grad_norm": 1.4639820667455608,
+      "learning_rate": 1.8793103448275863e-05,
+      "loss": 0.0805,
+      "step": 1910
+    },
+    {
+      "epoch": 1.03,
+      "grad_norm": 1.6214269161589794,
+      "learning_rate": 1.8620689655172415e-05,
+      "loss": 0.0902,
+      "step": 1920
+    },
+    {
+      "epoch": 1.03,
+      "grad_norm": 1.5979085316952373,
+      "learning_rate": 1.8448275862068967e-05,
+      "loss": 0.0967,
+      "step": 1930
+    },
+    {
+      "epoch": 1.04,
+      "grad_norm": 1.2001043976090235,
+      "learning_rate": 1.827586206896552e-05,
+      "loss": 0.069,
+      "step": 1940
+    },
+    {
+      "epoch": 1.04,
+      "grad_norm": 2.100190633629739,
+      "learning_rate": 1.810344827586207e-05,
+      "loss": 0.1024,
+      "step": 1950
+    },
+    {
+      "epoch": 1.05,
+      "grad_norm": 1.7393396532511867,
+      "learning_rate": 1.793103448275862e-05,
+      "loss": 0.0728,
+      "step": 1960
+    },
+    {
+      "epoch": 1.05,
+      "grad_norm": 1.873599965711283,
+      "learning_rate": 1.7758620689655172e-05,
+      "loss": 0.0735,
+      "step": 1970
+    },
+    {
+      "epoch": 1.06,
+      "grad_norm": 1.4460752726376342,
+      "learning_rate": 1.7586206896551724e-05,
+      "loss": 0.1,
+      "step": 1980
+    },
+    {
+      "epoch": 1.06,
+      "grad_norm": 0.8772715867399261,
+      "learning_rate": 1.7413793103448276e-05,
+      "loss": 0.0794,
+      "step": 1990
+    },
+    {
+      "epoch": 1.07,
+      "grad_norm": 1.398173054729605,
+      "learning_rate": 1.7241379310344828e-05,
+      "loss": 0.078,
+      "step": 2000
+    },
+    {
+      "epoch": 1.07,
+      "eval_loss": 0.6523420810699463,
+      "eval_runtime": 115.4048,
+      "eval_samples_per_second": 11.429,
+      "eval_steps_per_second": 2.859,
+      "step": 2000
+    }
+  ],
+  "logging_steps": 10,
+  "max_steps": 3000,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 2,
+  "save_steps": 500,
+  "total_flos": 57776359145472.0,
+  "train_batch_size": 4,
+  "trial_name": null,
+  "trial_params": null
+}

checkpoints/checkpoint-2000/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:22266252322e6977894297b61731ca9228c8fcfe4e8d788aef270218c40fec98
+size 6776

checkpoints/checkpoint-2000/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)

checkpoints/checkpoint-2500/README.md

@@ -0,0 +1,202 @@
+---
+library_name: peft
+base_model: openbmb/MiniCPM-2B-dpo-bf16
+---
+
+# 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 0.10.0

checkpoints/checkpoint-2500/adapter_config.json

@@ -0,0 +1,29 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "openbmb/MiniCPM-2B-dpo-bf16",
+  "bias": "none",
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": "gaussian",
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 8,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "v_proj",
+    "q_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "use_dora": false,
+  "use_rslora": false
+}

checkpoints/checkpoint-2500/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3930f22046333bd2b424020c5ba937de3e5f3239dc86cbd432b2b9f92c95a70a
+size 5919456

checkpoints/checkpoint-2500/global_step2500/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8e4efdd5c2b93b0c2c7671ca8a30ba480832bf7ecbcc638e1f13267ac9c26e5d
+size 35393392