Upload folder using huggingface_hub

.gitattributes

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

README.md

@@ -0,0 +1,144 @@
+---
+license: other
+library_name: peft
+tags:
+- generated_from_trainer
+base_model: google/gemma-2b-it
+model-index:
+- name: peft-gemma2b
+  results: []
+---
+
+
+[<img src="https://raw.githubusercontent.com/OpenAccess-AI-Collective/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/OpenAccess-AI-Collective/axolotl)
+<details><summary>See axolotl config</summary>
+
+axolotl version: `0.4.0`
+```yaml
+# use google/gemma-7b if you have access
+base_model: google/gemma-2b-it
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+
+load_in_8bit: false
+load_in_4bit: true
+strict: false
+
+# huggingface repo
+datasets:
+  - path: ./python-oasst/combined_chunk_2.jsonl
+    type: oasst
+val_set_size: 0.40
+output_dir: ./out3
+
+adapter: qlora
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+sequence_len: 4096
+sample_packing: true
+pad_to_sequence_len: true
+
+wandb_project: gemma-2b-it
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+
+gradient_accumulation_steps: 3
+micro_batch_size: 4
+num_epochs: 2
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: true
+group_by_length: false
+bf16: auto
+fp16:
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 4
+eval_table_size:
+eval_max_new_tokens: 256
+saves_per_epoch: 1
+debug:
+deepspeed: deepspeed_configs/zero1.json
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+
+```
+
+</details><br>
+
+# out3
+
+This model is a fine-tuned version of [google/gemma-2b-it](https://huggingface.co/google/gemma-2b-it) on the None dataset.
+It achieves the following results on the evaluation set:
+- Loss: 1.2430
+
+## Model description
+
+More information needed
+
+## Intended uses & limitations
+
+More information needed
+
+## Training and evaluation data
+
+More information needed
+
+## Training procedure
+
+### Training hyperparameters
+
+The following hyperparameters were used during training:
+- learning_rate: 0.0002
+- train_batch_size: 4
+- eval_batch_size: 4
+- seed: 42
+- distributed_type: multi-GPU
+- num_devices: 4
+- gradient_accumulation_steps: 3
+- total_train_batch_size: 48
+- total_eval_batch_size: 16
+- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
+- lr_scheduler_type: cosine
+- num_epochs: 2
+
+### Training results
+
+| Training Loss | Epoch | Step | Validation Loss |
+|:-------------:|:-----:|:----:|:---------------:|
+| 2.8926        | 0.02  | 1    | 2.7617          |
+| 1.4502        | 0.26  | 12   | 1.4564          |
+| 1.7617        | 0.52  | 24   | 1.3147          |
+| 1.2051        | 0.78  | 36   | 1.2781          |
+| 1.1353        | 1.01  | 48   | 1.2603          |
+| 1.1787        | 1.28  | 60   | 1.2498          |
+| 1.1416        | 1.54  | 72   | 1.2445          |
+| 1.1606        | 1.8   | 84   | 1.2430          |
+
+
+### Framework versions
+
+- PEFT 0.9.0
+- Transformers 4.38.0
+- Pytorch 2.0.1+cu117
+- Datasets 2.18.0
+- Tokenizers 0.15.0

adapter_config.json

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

adapter_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d7588baf2279a9decd1ae4ab0c5f6a03a9c05266fb6133696bade5771b733b19
+size 78486205

checkpoint-110/README.md

@@ -0,0 +1,202 @@
+---
+library_name: peft
+base_model: google/gemma-2b-it
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.9.0

checkpoint-110/adapter_config.json

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

checkpoint-110/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:808d55cfc0779724020543b0b6f2b69c829ed1a3d5efff889a6f890b8c1139be
+size 78480320

checkpoint-110/global_step110/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:747fb75ccb5a8b04441e183ba81925e338ded294ffbc1e5180c578f13a693e8a
+size 58886423

checkpoint-110/global_step110/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f22f475db28959ac7679e36700754b955b7f57f0fcf514207b849ed6030ee1ea
+size 58885463

checkpoint-110/global_step110/bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ea14779c099ee005ae1008ea80a4b0899c5154638318f569dc3d41ff21901e46
+size 58886487

checkpoint-110/global_step110/bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:939d4da58aa1f75a5c89110a76ec68e3e6badd2ef184d7382afa3dead722fbcb
+size 58885527

checkpoint-110/global_step110/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3da1e771f664c7850a57b8954bc62cf37a8d56bbb7db3bfbc65c349981318c43
+size 1159049457

checkpoint-110/latest

@@ -0,0 +1 @@
+global_step110

checkpoint-110/rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:24089072fa967ddda688610baf8c506a25f5771b731a9a05c0736c8a961273a7
+size 17655

checkpoint-110/rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:afe5f6f9755fd1245cedc10989cb0b4bb48aae05f1ccde8e9deb44e36aebc69b
+size 17655

checkpoint-110/rng_state_2.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d4e38510525d2854bf2b886708e15173236a90931480e673b0f69f8d620d8bf7
+size 17655

checkpoint-110/rng_state_3.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:53876e4b015161dc8aaad7580a291ce19694f401e74bf98e8ea9b34a5b8866a2
+size 17655

checkpoint-110/scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2fb173f58bc277f4540a71b9da714b6eab4897199fe96d39e8ced237b9480a43
+size 627

checkpoint-110/trainer_state.json

@@ -0,0 +1,855 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 1.9520958083832336,
+  "eval_steps": 14,
+  "global_step": 110,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.02,
+      "grad_norm": 0.866426735610868,
+      "learning_rate": 0.00019995921928281894,
+      "loss": 2.7031,
+      "step": 1
+    },
+    {
+      "epoch": 0.02,
+      "eval_loss": 2.6663479804992676,
+      "eval_runtime": 69.9146,
+      "eval_samples_per_second": 285.62,
+      "eval_steps_per_second": 17.865,
+      "step": 1
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 0.8776710816784347,
+      "learning_rate": 0.00019983691039261357,
+      "loss": 2.6299,
+      "step": 2
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 0.9796444288699809,
+      "learning_rate": 0.00019963317308626914,
+      "loss": 2.3027,
+      "step": 3
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 1.1160859766420685,
+      "learning_rate": 0.00019934817353485501,
+      "loss": 2.0674,
+      "step": 4
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 1.1551321141536093,
+      "learning_rate": 0.0001989821441880933,
+      "loss": 1.7861,
+      "step": 5
+    },
+    {
+      "epoch": 0.11,
+      "grad_norm": 0.890274648186441,
+      "learning_rate": 0.00019853538358476932,
+      "loss": 1.6172,
+      "step": 6
+    },
+    {
+      "epoch": 0.13,
+      "grad_norm": 0.7807325938639237,
+      "learning_rate": 0.00019800825610923934,
+      "loss": 1.6323,
+      "step": 7
+    },
+    {
+      "epoch": 0.14,
+      "grad_norm": 0.7513090074948452,
+      "learning_rate": 0.00019740119169423337,
+      "loss": 1.6172,
+      "step": 8
+    },
+    {
+      "epoch": 0.16,
+      "grad_norm": 0.6817700063779598,
+      "learning_rate": 0.00019671468547019573,
+      "loss": 1.5483,
+      "step": 9
+    },
+    {
+      "epoch": 0.18,
+      "grad_norm": 0.5485445966786001,
+      "learning_rate": 0.00019594929736144976,
+      "loss": 1.479,
+      "step": 10
+    },
+    {
+      "epoch": 0.2,
+      "grad_norm": 0.27041496308401886,
+      "learning_rate": 0.00019510565162951537,
+      "loss": 1.4292,
+      "step": 11
+    },
+    {
+      "epoch": 0.22,
+      "grad_norm": 0.24844480119928905,
+      "learning_rate": 0.00019418443636395248,
+      "loss": 1.4326,
+      "step": 12
+    },
+    {
+      "epoch": 0.23,
+      "grad_norm": 0.4516773777333339,
+      "learning_rate": 0.00019318640292114524,
+      "loss": 1.418,
+      "step": 13
+    },
+    {
+      "epoch": 0.25,
+      "grad_norm": 2.014260828449999,
+      "learning_rate": 0.000192112365311485,
+      "loss": 1.9883,
+      "step": 14
+    },
+    {
+      "epoch": 0.25,
+      "eval_loss": 1.441876769065857,
+      "eval_runtime": 71.2163,
+      "eval_samples_per_second": 280.399,
+      "eval_steps_per_second": 17.538,
+      "step": 14
+    },
+    {
+      "epoch": 0.27,
+      "grad_norm": 0.26474390032516393,
+      "learning_rate": 0.00019096319953545185,
+      "loss": 1.3955,
+      "step": 15
+    },
+    {
+      "epoch": 0.29,
+      "grad_norm": 0.22490458783426903,
+      "learning_rate": 0.00018973984286913584,
+      "loss": 1.376,
+      "step": 16
+    },
+    {
+      "epoch": 0.31,
+      "grad_norm": 0.33330700318335205,
+      "learning_rate": 0.00018844329309978145,
+      "loss": 1.3286,
+      "step": 17
+    },
+    {
+      "epoch": 0.32,
+      "grad_norm": 0.37492777208291783,
+      "learning_rate": 0.00018707460771197774,
+      "loss": 1.3359,
+      "step": 18
+    },
+    {
+      "epoch": 0.34,
+      "grad_norm": 0.29525135911036,
+      "learning_rate": 0.0001856349030251589,
+      "loss": 1.3198,
+      "step": 19
+    },
+    {
+      "epoch": 0.36,
+      "grad_norm": 0.22284214725380747,
+      "learning_rate": 0.00018412535328311814,
+      "loss": 1.3325,
+      "step": 20
+    },
+    {
+      "epoch": 0.38,
+      "grad_norm": 0.39903198098602605,
+      "learning_rate": 0.0001825471896962774,
+      "loss": 1.3252,
+      "step": 21
+    },
+    {
+      "epoch": 0.4,
+      "grad_norm": 0.4164131128200096,
+      "learning_rate": 0.00018090169943749476,
+      "loss": 1.2842,
+      "step": 22
+    },
+    {
+      "epoch": 0.41,
+      "grad_norm": 0.2610111064228617,
+      "learning_rate": 0.00017919022459222752,
+      "loss": 1.2979,
+      "step": 23
+    },
+    {
+      "epoch": 0.43,
+      "grad_norm": 0.2012045009785502,
+      "learning_rate": 0.00017741416106390826,
+      "loss": 1.2554,
+      "step": 24
+    },
+    {
+      "epoch": 0.45,
+      "grad_norm": 0.23300289011357986,
+      "learning_rate": 0.00017557495743542585,
+      "loss": 1.2461,
+      "step": 25
+    },
+    {
+      "epoch": 0.47,
+      "grad_norm": 0.23652936705914454,
+      "learning_rate": 0.0001736741137876405,
+      "loss": 1.271,
+      "step": 26
+    },
+    {
+      "epoch": 0.49,
+      "grad_norm": 0.1925895084011206,
+      "learning_rate": 0.00017171318047589637,
+      "loss": 1.2515,
+      "step": 27
+    },
+    {
+      "epoch": 0.5,
+      "grad_norm": 0.12096954328419277,
+      "learning_rate": 0.00016969375686552937,
+      "loss": 1.2461,
+      "step": 28
+    },
+    {
+      "epoch": 0.5,
+      "eval_loss": 1.3231805562973022,
+      "eval_runtime": 71.3812,
+      "eval_samples_per_second": 279.752,
+      "eval_steps_per_second": 17.498,
+      "step": 28
+    },
+    {
+      "epoch": 0.52,
+      "grad_norm": 0.12448721999954601,
+      "learning_rate": 0.00016761749002740193,
+      "loss": 1.2446,
+      "step": 29
+    },
+    {
+      "epoch": 0.54,
+      "grad_norm": 0.21969944256734386,
+      "learning_rate": 0.00016548607339452853,
+      "loss": 1.2783,
+      "step": 30
+    },
+    {
+      "epoch": 0.56,
+      "grad_norm": 0.2172107697789349,
+      "learning_rate": 0.00016330124538088705,
+      "loss": 1.2461,
+      "step": 31
+    },
+    {
+      "epoch": 0.57,
+      "grad_norm": 0.1547872862082663,
+      "learning_rate": 0.00016106478796354382,
+      "loss": 1.2241,
+      "step": 32
+    },
+    {
+      "epoch": 0.59,
+      "grad_norm": 0.08869884076741208,
+      "learning_rate": 0.00015877852522924732,
+      "loss": 1.1846,
+      "step": 33
+    },
+    {
+      "epoch": 0.61,
+      "grad_norm": 0.1538541877622166,
+      "learning_rate": 0.00015644432188667695,
+      "loss": 1.23,
+      "step": 34
+    },
+    {
+      "epoch": 0.63,
+      "grad_norm": 0.20399644873662537,
+      "learning_rate": 0.00015406408174555976,
+      "loss": 1.2114,
+      "step": 35
+    },
+    {
+      "epoch": 0.65,
+      "grad_norm": 0.19144417912447978,
+      "learning_rate": 0.0001516397461638962,
+      "loss": 1.2324,
+      "step": 36
+    },
+    {
+      "epoch": 0.66,
+      "grad_norm": 0.26912005474616185,
+      "learning_rate": 0.0001491732924645604,
+      "loss": 1.5522,
+      "step": 37
+    },
+    {
+      "epoch": 0.68,
+      "grad_norm": 0.09943230167494302,
+      "learning_rate": 0.00014666673232256738,
+      "loss": 1.1973,
+      "step": 38
+    },
+    {
+      "epoch": 0.7,
+      "grad_norm": 0.12111451370899122,
+      "learning_rate": 0.00014412211012432212,
+      "loss": 1.1685,
+      "step": 39
+    },
+    {
+      "epoch": 0.72,
+      "grad_norm": 0.13274421759806426,
+      "learning_rate": 0.00014154150130018866,
+      "loss": 1.1846,
+      "step": 40
+    },
+    {
+      "epoch": 0.74,
+      "grad_norm": 0.12218551159116417,
+      "learning_rate": 0.00013892701063173918,
+      "loss": 1.2173,
+      "step": 41
+    },
+    {
+      "epoch": 0.75,
+      "grad_norm": 0.09665309655921935,
+      "learning_rate": 0.0001362807705350641,
+      "loss": 1.2207,
+      "step": 42
+    },
+    {
+      "epoch": 0.75,
+      "eval_loss": 1.2895056009292603,
+      "eval_runtime": 71.3954,
+      "eval_samples_per_second": 279.696,
+      "eval_steps_per_second": 17.494,
+      "step": 42
+    },
+    {
+      "epoch": 0.77,
+      "grad_norm": 0.10941681743668577,
+      "learning_rate": 0.00013360493932154302,
+      "loss": 1.1943,
+      "step": 43
+    },
+    {
+      "epoch": 0.79,
+      "grad_norm": 0.13685075654053136,
+      "learning_rate": 0.00013090169943749476,
+      "loss": 1.1597,
+      "step": 44
+    },
+    {
+      "epoch": 0.81,
+      "grad_norm": 0.12907173631699106,
+      "learning_rate": 0.00012817325568414297,
+      "loss": 1.1851,
+      "step": 45
+    },
+    {
+      "epoch": 0.83,
+      "grad_norm": 0.1036179369816153,
+      "learning_rate": 0.00012542183341934872,
+      "loss": 1.1987,
+      "step": 46
+    },
+    {
+      "epoch": 0.84,
+      "grad_norm": 0.08667840513163542,
+      "learning_rate": 0.00012264967674257646,
+      "loss": 1.1724,
+      "step": 47
+    },
+    {
+      "epoch": 0.86,
+      "grad_norm": 0.1144689947654717,
+      "learning_rate": 0.00011985904666457455,
+      "loss": 1.1997,
+      "step": 48
+    },
+    {
+      "epoch": 0.88,
+      "grad_norm": 0.13847211768760592,
+      "learning_rate": 0.0001170522192632624,
+      "loss": 1.1602,
+      "step": 49
+    },
+    {
+      "epoch": 0.9,
+      "grad_norm": 0.11095253849183417,
+      "learning_rate": 0.00011423148382732853,
+      "loss": 1.1758,
+      "step": 50
+    },
+    {
+      "epoch": 0.92,
+      "grad_norm": 0.08573841203966975,
+      "learning_rate": 0.00011139914098905406,
+      "loss": 1.1846,
+      "step": 51
+    },
+    {
+      "epoch": 0.93,
+      "grad_norm": 0.0948235516667105,
+      "learning_rate": 0.00010855750084788398,
+      "loss": 1.1948,
+      "step": 52
+    },
+    {
+      "epoch": 0.95,
+      "grad_norm": 0.121781824636731,
+      "learning_rate": 0.00010570888108627681,
+      "loss": 1.1753,
+      "step": 53
+    },
+    {
+      "epoch": 0.97,
+      "grad_norm": 0.10546391970558401,
+      "learning_rate": 0.00010285560507936961,
+      "loss": 1.1699,
+      "step": 54
+    },
+    {
+      "epoch": 0.99,
+      "grad_norm": 0.34233328008898944,
+      "learning_rate": 0.0001,
+      "loss": 1.5063,
+      "step": 55
+    },
+    {
+      "epoch": 1.01,
+      "grad_norm": 0.11900909777508013,
+      "learning_rate": 9.71443949206304e-05,
+      "loss": 1.1479,
+      "step": 56
+    },
+    {
+      "epoch": 1.01,
+      "eval_loss": 1.2747071981430054,
+      "eval_runtime": 71.4286,
+      "eval_samples_per_second": 279.566,
+      "eval_steps_per_second": 17.486,
+      "step": 56
+    },
+    {
+      "epoch": 1.02,
+      "grad_norm": 0.11736589205080973,
+      "learning_rate": 9.42911189137232e-05,
+      "loss": 1.1411,
+      "step": 57
+    },
+    {
+      "epoch": 1.02,
+      "grad_norm": 0.12677764502316016,
+      "learning_rate": 9.144249915211605e-05,
+      "loss": 1.1514,
+      "step": 58
+    },
+    {
+      "epoch": 1.04,
+      "grad_norm": 0.11697640877801936,
+      "learning_rate": 8.860085901094595e-05,
+      "loss": 1.144,
+      "step": 59
+    },
+    {
+      "epoch": 1.05,
+      "grad_norm": 0.10851474914032534,
+      "learning_rate": 8.57685161726715e-05,
+      "loss": 1.1538,
+      "step": 60
+    },
+    {
+      "epoch": 1.07,
+      "grad_norm": 0.10591523712985791,
+      "learning_rate": 8.294778073673762e-05,
+      "loss": 1.1675,
+      "step": 61
+    },
+    {
+      "epoch": 1.09,
+      "grad_norm": 0.10109697997845449,
+      "learning_rate": 8.014095333542548e-05,
+      "loss": 1.1528,
+      "step": 62
+    },
+    {
+      "epoch": 1.11,
+      "grad_norm": 0.11774452860536459,
+      "learning_rate": 7.735032325742355e-05,
+      "loss": 1.144,
+      "step": 63
+    },
+    {
+      "epoch": 1.13,
+      "grad_norm": 0.11052397113271337,
+      "learning_rate": 7.457816658065134e-05,
+      "loss": 1.1533,
+      "step": 64
+    },
+    {
+      "epoch": 1.14,
+      "grad_norm": 0.09932538763675618,
+      "learning_rate": 7.182674431585704e-05,
+      "loss": 1.1562,
+      "step": 65
+    },
+    {
+      "epoch": 1.16,
+      "grad_norm": 0.100642669752315,
+      "learning_rate": 6.909830056250527e-05,
+      "loss": 1.1504,
+      "step": 66
+    },
+    {
+      "epoch": 1.18,
+      "grad_norm": 0.08717942179921322,
+      "learning_rate": 6.639506067845697e-05,
+      "loss": 1.1353,
+      "step": 67
+    },
+    {
+      "epoch": 1.2,
+      "grad_norm": 0.10484048248463712,
+      "learning_rate": 6.371922946493591e-05,
+      "loss": 1.1196,
+      "step": 68
+    },
+    {
+      "epoch": 1.22,
+      "grad_norm": 0.10322601363425112,
+      "learning_rate": 6.107298936826086e-05,
+      "loss": 1.1226,
+      "step": 69
+    },
+    {
+      "epoch": 1.23,
+      "grad_norm": 0.10689324440748724,
+      "learning_rate": 5.845849869981137e-05,
+      "loss": 1.1797,
+      "step": 70
+    },
+    {
+      "epoch": 1.23,
+      "eval_loss": 1.2690308094024658,
+      "eval_runtime": 71.4252,
+      "eval_samples_per_second": 279.579,
+      "eval_steps_per_second": 17.487,
+      "step": 70
+    },
+    {
+      "epoch": 1.25,
+      "grad_norm": 0.10301775684317682,
+      "learning_rate": 5.5877889875677845e-05,
+      "loss": 1.1265,
+      "step": 71
+    },
+    {
+      "epoch": 1.27,
+      "grad_norm": 0.08476164067920204,
+      "learning_rate": 5.333326767743263e-05,
+      "loss": 1.145,
+      "step": 72
+    },
+    {
+      "epoch": 1.29,
+      "grad_norm": 0.09070035744584001,
+      "learning_rate": 5.082670753543961e-05,
+      "loss": 1.1226,
+      "step": 73
+    },
+    {
+      "epoch": 1.31,
+      "grad_norm": 0.09969553773787451,
+      "learning_rate": 4.836025383610382e-05,
+      "loss": 1.1323,
+      "step": 74
+    },
+    {
+      "epoch": 1.32,
+      "grad_norm": 0.1100883382765975,
+      "learning_rate": 4.593591825444028e-05,
+      "loss": 1.1353,
+      "step": 75
+    },
+    {
+      "epoch": 1.34,
+      "grad_norm": 0.10084037635805906,
+      "learning_rate": 4.355567811332311e-05,
+      "loss": 1.1689,
+      "step": 76
+    },
+    {
+      "epoch": 1.36,
+      "grad_norm": 0.08865185209055175,
+      "learning_rate": 4.12214747707527e-05,
+      "loss": 1.1304,
+      "step": 77
+    },
+    {
+      "epoch": 1.38,
+      "grad_norm": 0.08317004226644682,
+      "learning_rate": 3.893521203645618e-05,
+      "loss": 1.0801,
+      "step": 78
+    },
+    {
+      "epoch": 1.4,
+      "grad_norm": 0.08437511567708315,
+      "learning_rate": 3.669875461911297e-05,
+      "loss": 1.1265,
+      "step": 79
+    },
+    {
+      "epoch": 1.41,
+      "grad_norm": 0.08737931667441079,
+      "learning_rate": 3.45139266054715e-05,
+      "loss": 1.1206,
+      "step": 80
+    },
+    {
+      "epoch": 1.43,
+      "grad_norm": 0.09676132223503787,
+      "learning_rate": 3.238250997259808e-05,
+      "loss": 1.0879,
+      "step": 81
+    },
+    {
+      "epoch": 1.45,
+      "grad_norm": 0.3064914875240821,
+      "learning_rate": 3.030624313447067e-05,
+      "loss": 1.4707,
+      "step": 82
+    },
+    {
+      "epoch": 1.47,
+      "grad_norm": 0.10108431705762476,
+      "learning_rate": 2.828681952410366e-05,
+      "loss": 1.1489,
+      "step": 83
+    },
+    {
+      "epoch": 1.49,
+      "grad_norm": 0.10356256932007075,
+      "learning_rate": 2.6325886212359498e-05,
+      "loss": 1.084,
+      "step": 84
+    },
+    {
+      "epoch": 1.49,
+      "eval_loss": 1.2660632133483887,
+      "eval_runtime": 71.4598,
+      "eval_samples_per_second": 279.444,
+      "eval_steps_per_second": 17.478,
+      "step": 84
+    },
+    {
+      "epoch": 1.5,
+      "grad_norm": 0.11000673888436871,
+      "learning_rate": 2.4425042564574184e-05,
+      "loss": 1.1494,
+      "step": 85
+    },
+    {
+      "epoch": 1.52,
+      "grad_norm": 0.09066127881300992,
+      "learning_rate": 2.2585838936091754e-05,
+      "loss": 1.1572,
+      "step": 86
+    },
+    {
+      "epoch": 1.54,
+      "grad_norm": 0.09757831324550774,
+      "learning_rate": 2.0809775407772503e-05,
+      "loss": 1.103,
+      "step": 87
+    },
+    {
+      "epoch": 1.56,
+      "grad_norm": 0.0934233089422467,
+      "learning_rate": 1.9098300562505266e-05,
+      "loss": 1.1001,
+      "step": 88
+    },
+    {
+      "epoch": 1.57,
+      "grad_norm": 0.09106355313835601,
+      "learning_rate": 1.74528103037226e-05,
+      "loss": 1.1406,
+      "step": 89
+    },
+    {
+      "epoch": 1.59,
+      "grad_norm": 0.09454088595436297,
+      "learning_rate": 1.587464671688187e-05,
+      "loss": 1.1387,
+      "step": 90
+    },
+    {
+      "epoch": 1.61,
+      "grad_norm": 0.09689276747664,
+      "learning_rate": 1.4365096974841108e-05,
+      "loss": 1.1318,
+      "step": 91
+    },
+    {
+      "epoch": 1.63,
+      "grad_norm": 0.08987470798352144,
+      "learning_rate": 1.2925392288022298e-05,
+      "loss": 1.1592,
+      "step": 92
+    },
+    {
+      "epoch": 1.65,
+      "grad_norm": 0.08771103153086109,
+      "learning_rate": 1.1556706900218572e-05,
+      "loss": 1.1582,
+      "step": 93
+    },
+    {
+      "epoch": 1.66,
+      "grad_norm": 0.08740633683699849,
+      "learning_rate": 1.026015713086418e-05,
+      "loss": 1.1787,
+      "step": 94
+    },
+    {
+      "epoch": 1.68,
+      "grad_norm": 0.08404626340301383,
+      "learning_rate": 9.036800464548157e-06,
+      "loss": 1.1362,
+      "step": 95
+    },
+    {
+      "epoch": 1.7,
+      "grad_norm": 0.08750778902246109,
+      "learning_rate": 7.887634688515e-06,
+      "loss": 1.1387,
+      "step": 96
+    },
+    {
+      "epoch": 1.72,
+      "grad_norm": 0.08581989813901233,
+      "learning_rate": 6.813597078854772e-06,
+      "loss": 1.1421,
+      "step": 97
+    },
+    {
+      "epoch": 1.74,
+      "grad_norm": 0.08777403515933545,
+      "learning_rate": 5.8155636360475385e-06,
+      "loss": 1.1035,
+      "step": 98
+    },
+    {
+      "epoch": 1.74,
+      "eval_loss": 1.2641561031341553,
+      "eval_runtime": 71.4805,
+      "eval_samples_per_second": 279.363,
+      "eval_steps_per_second": 17.473,
+      "step": 98
+    },
+    {
+      "epoch": 1.75,
+      "grad_norm": 0.08385195939000219,
+      "learning_rate": 4.8943483704846475e-06,
+      "loss": 1.1113,
+      "step": 99
+    },
+    {
+      "epoch": 1.77,
+      "grad_norm": 0.08908998416660925,
+      "learning_rate": 4.050702638550275e-06,
+      "loss": 1.1104,
+      "step": 100
+    },
+    {
+      "epoch": 1.79,
+      "grad_norm": 0.08249447233292928,
+      "learning_rate": 3.2853145298042953e-06,
+      "loss": 1.1396,
+      "step": 101
+    },
+    {
+      "epoch": 1.81,
+      "grad_norm": 0.08608973089588336,
+      "learning_rate": 2.5988083057666533e-06,
+      "loss": 1.1055,
+      "step": 102
+    },
+    {
+      "epoch": 1.83,
+      "grad_norm": 0.09951373726217344,
+      "learning_rate": 1.9917438907606556e-06,
+      "loss": 1.1221,
+      "step": 103
+    },
+    {
+      "epoch": 1.84,
+      "grad_norm": 0.08532100595350145,
+      "learning_rate": 1.4646164152307018e-06,
+      "loss": 1.0898,
+      "step": 104
+    },
+    {
+      "epoch": 1.86,
+      "grad_norm": 0.08282306531770366,
+      "learning_rate": 1.0178558119067315e-06,
+      "loss": 1.1118,
+      "step": 105
+    },
+    {
+      "epoch": 1.88,
+      "grad_norm": 0.08416795073468437,
+      "learning_rate": 6.518264651449779e-07,
+      "loss": 1.1133,
+      "step": 106
+    },
+    {
+      "epoch": 1.9,
+      "grad_norm": 0.08760805801842701,
+      "learning_rate": 3.6682691373086665e-07,
+      "loss": 1.123,
+      "step": 107
+    },
+    {
+      "epoch": 1.92,
+      "grad_norm": 0.08618073017190576,
+      "learning_rate": 1.630896073864352e-07,
+      "loss": 1.0981,
+      "step": 108
+    },
+    {
+      "epoch": 1.93,
+      "grad_norm": 0.0908365337693141,
+      "learning_rate": 4.078071718107701e-08,
+      "loss": 1.1328,
+      "step": 109
+    },
+    {
+      "epoch": 1.95,
+      "grad_norm": 0.09003969262620456,
+      "learning_rate": 0.0,
+      "loss": 1.1104,
+      "step": 110
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 110,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 2,
+  "save_steps": 55,
+  "total_flos": 2.6221185296708403e+17,
+  "train_batch_size": 4,
+  "trial_name": null,
+  "trial_params": null
+}

checkpoint-110/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:123d1fd699788f7d60d917b5964a4e0851773abfe7e74a3a42baed90eb8677c6
+size 6331

checkpoint-110/zero_to_fp32.py

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

checkpoint-46/README.md

@@ -0,0 +1,202 @@
+---
+library_name: peft
+base_model: google/gemma-2b-it
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.9.0

checkpoint-46/adapter_config.json

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

checkpoint-46/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c4f805062e6421d7c69a2ee547e3af8350f624451ce165140c7232181316db33
+size 78480320

checkpoint-46/global_step46/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9852910ac2adb1970801337fce7443a3bfdcaeabe07d654cff61bf2c432cbb3d
+size 58886423

checkpoint-46/global_step46/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c202b3d15fad6725b57469ab741165c4f095730417a3e390ce5fb29dda5a5c75
+size 58885463

checkpoint-46/global_step46/bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a7f5dc5545b3f194373b58996d0aed9e1376e5a54254941eb23eb2b0aaee3a8e
+size 58886487

checkpoint-46/global_step46/bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ec4ee84e8c134757373034201b648cb6e00fb8cc5f98385d4f02971bf99516e0
+size 58885527

checkpoint-46/global_step46/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:83d32f1971fddca0d11f12821e2366f4b0767156570446e5cc7bb231abc22851
+size 1159049457

checkpoint-46/latest

@@ -0,0 +1 @@
+global_step46

checkpoint-46/rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b165aaa144fad4f4fa49d6dd695add421ab6c181ffc890115a172d00266723ab
+size 17655

checkpoint-46/rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d430c8b2f065299cc8e2d4fdc64fdd296a0649dc1216084e10018c2c0f201921
+size 17655

checkpoint-46/rng_state_2.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1e513afae097b3d69141e7a925b55bd2f6657df9a7840cf02b10d4263576d744
+size 17655

checkpoint-46/rng_state_3.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4d013a562702b25cfd982378c7c0866be0e8fda59f98cb265cc32ba235b5bd6a
+size 17655

checkpoint-46/scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:761d7aba80bef53190f12e33a600d1e0cbb142bd0005601918622cc3b27ad6ff
+size 627

checkpoint-46/trainer_state.json

@@ -0,0 +1,375 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 1.0,
+  "eval_steps": 12,
+  "global_step": 46,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.02,
+      "grad_norm": 0.9959812474764307,
+      "learning_rate": 0.0001999417022366174,
+      "loss": 2.8926,
+      "step": 1
+    },
+    {
+      "epoch": 0.02,
+      "eval_loss": 2.7616989612579346,
+      "eval_runtime": 62.0214,
+      "eval_samples_per_second": 322.227,
+      "eval_steps_per_second": 20.154,
+      "step": 1
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 0.9553562375664082,
+      "learning_rate": 0.00019976687691905393,
+      "loss": 2.7178,
+      "step": 2
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 1.074690670043139,
+      "learning_rate": 0.00019947572788580947,
+      "loss": 2.3477,
+      "step": 3
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 1.2607308881446835,
+      "learning_rate": 0.00019906859460363307,
+      "loss": 2.0869,
+      "step": 4
+    },
+    {
+      "epoch": 0.11,
+      "grad_norm": 1.5197691743409516,
+      "learning_rate": 0.00019854595177171968,
+      "loss": 1.8477,
+      "step": 5
+    },
+    {
+      "epoch": 0.13,
+      "grad_norm": 0.7112299977659389,
+      "learning_rate": 0.00019790840876823232,
+      "loss": 1.7227,
+      "step": 6
+    },
+    {
+      "epoch": 0.15,
+      "grad_norm": 0.7680093267515216,
+      "learning_rate": 0.00019715670893979414,
+      "loss": 1.6553,
+      "step": 7
+    },
+    {
+      "epoch": 0.17,
+      "grad_norm": 0.7744797513985666,
+      "learning_rate": 0.00019629172873477995,
+      "loss": 1.6455,
+      "step": 8
+    },
+    {
+      "epoch": 0.2,
+      "grad_norm": 0.735518802506217,
+      "learning_rate": 0.00019531447668141608,
+      "loss": 1.5508,
+      "step": 9
+    },
+    {
+      "epoch": 0.22,
+      "grad_norm": 0.4535717241942791,
+      "learning_rate": 0.00019422609221188207,
+      "loss": 1.5039,
+      "step": 10
+    },
+    {
+      "epoch": 0.24,
+      "grad_norm": 0.29477652700703394,
+      "learning_rate": 0.0001930278443337833,
+      "loss": 1.502,
+      "step": 11
+    },
+    {
+      "epoch": 0.26,
+      "grad_norm": 0.24434475566463373,
+      "learning_rate": 0.00019172113015054532,
+      "loss": 1.4502,
+      "step": 12
+    },
+    {
+      "epoch": 0.26,
+      "eval_loss": 1.456370234489441,
+      "eval_runtime": 63.0041,
+      "eval_samples_per_second": 317.202,
+      "eval_steps_per_second": 19.84,
+      "step": 12
+    },
+    {
+      "epoch": 0.28,
+      "grad_norm": 0.45228370307484317,
+      "learning_rate": 0.00019030747323245327,
+      "loss": 1.4277,
+      "step": 13
+    },
+    {
+      "epoch": 0.3,
+      "grad_norm": 0.46523254227617566,
+      "learning_rate": 0.0001887885218402375,
+      "loss": 1.4097,
+      "step": 14
+    },
+    {
+      "epoch": 0.33,
+      "grad_norm": 0.2855872854921807,
+      "learning_rate": 0.00018716604700327514,
+      "loss": 1.3833,
+      "step": 15
+    },
+    {
+      "epoch": 0.35,
+      "grad_norm": 0.25447863105741503,
+      "learning_rate": 0.00018544194045464886,
+      "loss": 1.3652,
+      "step": 16
+    },
+    {
+      "epoch": 0.37,
+      "grad_norm": 0.3619444462250106,
+      "learning_rate": 0.0001836182124254711,
+      "loss": 1.3677,
+      "step": 17
+    },
+    {
+      "epoch": 0.39,
+      "grad_norm": 0.36635956863370217,
+      "learning_rate": 0.0001816969893010442,
+      "loss": 1.332,
+      "step": 18
+    },
+    {
+      "epoch": 0.41,
+      "grad_norm": 0.39509330975297474,
+      "learning_rate": 0.00017968051114159047,
+      "loss": 1.3086,
+      "step": 19
+    },
+    {
+      "epoch": 0.43,
+      "grad_norm": 0.3249913054000147,
+      "learning_rate": 0.000177571129070442,
+      "loss": 1.3071,
+      "step": 20
+    },
+    {
+      "epoch": 0.46,
+      "grad_norm": 0.30914838660163285,
+      "learning_rate": 0.00017537130253273613,
+      "loss": 1.2729,
+      "step": 21
+    },
+    {
+      "epoch": 0.48,
+      "grad_norm": 0.42289317853499514,
+      "learning_rate": 0.00017308359642781242,
+      "loss": 1.2773,
+      "step": 22
+    },
+    {
+      "epoch": 0.5,
+      "grad_norm": 0.35414226064550436,
+      "learning_rate": 0.00017071067811865476,
+      "loss": 1.2417,
+      "step": 23
+    },
+    {
+      "epoch": 0.52,
+      "grad_norm": 1.5532132927649234,
+      "learning_rate": 0.00016825531432186543,
+      "loss": 1.7617,
+      "step": 24
+    },
+    {
+      "epoch": 0.52,
+      "eval_loss": 1.314684271812439,
+      "eval_runtime": 63.1261,
+      "eval_samples_per_second": 316.588,
+      "eval_steps_per_second": 19.802,
+      "step": 24
+    },
+    {
+      "epoch": 0.54,
+      "grad_norm": 0.12961326680655122,
+      "learning_rate": 0.00016572036788179727,
+      "loss": 1.2183,
+      "step": 25
+    },
+    {
+      "epoch": 0.57,
+      "grad_norm": 0.1963943275104468,
+      "learning_rate": 0.00016310879443260528,
+      "loss": 1.2222,
+      "step": 26
+    },
+    {
+      "epoch": 0.59,
+      "grad_norm": 0.21010667932420385,
+      "learning_rate": 0.00016042363895210946,
+      "loss": 1.2163,
+      "step": 27
+    },
+    {
+      "epoch": 0.61,
+      "grad_norm": 0.2259721381332128,
+      "learning_rate": 0.00015766803221148673,
+      "loss": 1.2217,
+      "step": 28
+    },
+    {
+      "epoch": 0.63,
+      "grad_norm": 0.19095866752454876,
+      "learning_rate": 0.00015484518712493187,
+      "loss": 1.2212,
+      "step": 29
+    },
+    {
+      "epoch": 0.65,
+      "grad_norm": 0.11496636776764421,
+      "learning_rate": 0.00015195839500354335,
+      "loss": 1.2212,
+      "step": 30
+    },
+    {
+      "epoch": 0.67,
+      "grad_norm": 0.12075046364081166,
+      "learning_rate": 0.00014901102171780174,
+      "loss": 1.1616,
+      "step": 31
+    },
+    {
+      "epoch": 0.7,
+      "grad_norm": 0.21352779335608432,
+      "learning_rate": 0.00014600650377311522,
+      "loss": 1.2002,
+      "step": 32
+    },
+    {
+      "epoch": 0.72,
+      "grad_norm": 0.20188789347169253,
+      "learning_rate": 0.0001429483443030082,
+      "loss": 1.2236,
+      "step": 33
+    },
+    {
+      "epoch": 0.74,
+      "grad_norm": 0.1547188722220508,
+      "learning_rate": 0.00013984010898462416,
+      "loss": 1.1826,
+      "step": 34
+    },
+    {
+      "epoch": 0.76,
+      "grad_norm": 0.11634196980431331,
+      "learning_rate": 0.00013668542188130566,
+      "loss": 1.209,
+      "step": 35
+    },
+    {
+      "epoch": 0.78,
+      "grad_norm": 0.11144391112143442,
+      "learning_rate": 0.00013348796121709862,
+      "loss": 1.2051,
+      "step": 36
+    },
+    {
+      "epoch": 0.78,
+      "eval_loss": 1.2780728340148926,
+      "eval_runtime": 63.136,
+      "eval_samples_per_second": 316.539,
+      "eval_steps_per_second": 19.799,
+      "step": 36
+    },
+    {
+      "epoch": 0.8,
+      "grad_norm": 0.16076857117392804,
+      "learning_rate": 0.0001302514550881076,
+      "loss": 1.2046,
+      "step": 37
+    },
+    {
+      "epoch": 0.83,
+      "grad_norm": 0.15687324678725567,
+      "learning_rate": 0.00012697967711570242,
+      "loss": 1.21,
+      "step": 38
+    },
+    {
+      "epoch": 0.85,
+      "grad_norm": 0.4130016628721442,
+      "learning_rate": 0.00012367644204664468,
+      "loss": 1.5093,
+      "step": 39
+    },
+    {
+      "epoch": 0.87,
+      "grad_norm": 0.1343313559360696,
+      "learning_rate": 0.0001203456013052634,
+      "loss": 1.2051,
+      "step": 40
+    },
+    {
+      "epoch": 0.89,
+      "grad_norm": 0.11359323094812154,
+      "learning_rate": 0.00011699103850286669,
+      "loss": 1.2031,
+      "step": 41
+    },
+    {
+      "epoch": 0.91,
+      "grad_norm": 0.09939941549415923,
+      "learning_rate": 0.00011361666490962468,
+      "loss": 1.1929,
+      "step": 42
+    },
+    {
+      "epoch": 0.93,
+      "grad_norm": 0.09916599553279144,
+      "learning_rate": 0.00011022641489420342,
+      "loss": 1.1748,
+      "step": 43
+    },
+    {
+      "epoch": 0.96,
+      "grad_norm": 0.14209942605129547,
+      "learning_rate": 0.0001068242413364671,
+      "loss": 1.1665,
+      "step": 44
+    },
+    {
+      "epoch": 0.98,
+      "grad_norm": 0.12442345202001188,
+      "learning_rate": 0.00010341411101859679,
+      "loss": 1.1768,
+      "step": 45
+    },
+    {
+      "epoch": 1.0,
+      "grad_norm": 0.26217962366852043,
+      "learning_rate": 0.0001,
+      "loss": 1.5122,
+      "step": 46
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 92,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 2,
+  "save_steps": 46,
+  "total_flos": 1.096730008312873e+17,
+  "train_batch_size": 4,
+  "trial_name": null,
+  "trial_params": null
+}

checkpoint-46/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4c43810cdb8f146f2245bb960f4cc7afe1dd2bc3429988aca22dc36faf28059d
+size 6331

checkpoint-46/zero_to_fp32.py

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

checkpoint-55/README.md

@@ -0,0 +1,202 @@
+---
+library_name: peft
+base_model: google/gemma-2b-it
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.9.0

checkpoint-55/adapter_config.json

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

checkpoint-55/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6d1edc3d855436f9861a042b87d27ef93f67d417404e2545f1ad57eda3ca69ff
+size 78480320

checkpoint-55/global_step55/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2104a313d2624e7f4ac5e557b557aa19c29fe5d005f33e824b341a0e7eaa48e5
+size 58886423

checkpoint-55/global_step55/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cefbaff50a0cc2ed1ad8d1066051df27eabbba05cb291eaeeeb18e78082d200f
+size 58885463

checkpoint-55/global_step55/bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:07741327312405bbfdb835ca082556e9ec5dba02c874ac98c49b8475b8040f9d
+size 58886487

checkpoint-55/global_step55/bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c440b9c26a32e7e570f147fa8356307f748789400eaefc4b594c8c561c260c01
+size 58885527

checkpoint-55/global_step55/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:00511e35a1d4d0f7513c7ff2a4829d4c2204862cdc73180a03ad0d7b289ea754
+size 1159049457

checkpoint-55/latest

@@ -0,0 +1 @@
+global_step55

checkpoint-55/rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e8b2056fa67792ace30300b98f23b3ab3d644e32b902db24c352e0ab44404e78
+size 17655

checkpoint-55/rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a10faa18b876f22999c0fe0f8b3b63b8b85749c7015ceec8ad57f04b111a0272
+size 17655

checkpoint-55/rng_state_2.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:73b99b0efd4d198d3cc9d21ee39ad07baaefe8c2a4cc4854481ab7f40faa9d90
+size 17655