Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,7 @@ 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
+checkpoint-1107/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+checkpoint-369/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+checkpoint-738/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,193 @@
+---
+library_name: peft
+tags:
+- generated_from_trainer
+base_model: mistralai_Mistral-Nemo-Instruct-2407
+model-index:
+- name: 12b-out
+  results: []
+---
+
+<!-- This model card has been generated automatically according to the information the Trainer had access to. You
+should probably proofread and complete it, then remove this comment. -->
+
+[<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/axolotl-ai-cloud/axolotl)
+<details><summary>See axolotl config</summary>
+
+axolotl version: `0.5.0`
+```yaml
+base_model: mistralai_Mistral-Nemo-Instruct-2407
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+
+plugins:
+  - axolotl.integrations.liger.LigerPlugin
+liger_rope: true
+liger_rms_norm: true
+liger_swiglu: true
+liger_fused_linear_cross_entropy: true
+
+load_in_8bit: false
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: NewEden/OpenCAI-ShareGPT
+    type: chat_template
+#    chat_template: mistralv3tekken
+    roles_to_train: ["gpt"]
+    field_messages: conversations
+    message_field_role: from
+    message_field_content: value
+    train_on_eos: turn
+  - path: NewEden/vanilla-backrooms-claude-sharegpt
+    type: chat_template
+#    chat_template: mistralv3tekken
+    roles_to_train: ["gpt"]
+    field_messages: conversations
+    message_field_role: from
+    message_field_content: value
+    train_on_eos: turn    
+  - path: anthracite-org/kalo_opus_misc_240827
+    type: chat_template
+#    chat_template: mistralv3tekken
+    roles_to_train: ["gpt"]
+    field_messages: conversations
+    message_field_role: from
+    message_field_content: value
+    train_on_eos: turn    
+  - path: anthracite-org/kalo_misc_part2
+    type: chat_template
+#    chat_template: mistralv3tekken
+    roles_to_train: ["gpt"]
+    field_messages: conversations
+    message_field_role: from
+    message_field_content: value
+    train_on_eos: turn
+  - path: NewEden/Roleplay-Logs-V2
+    type: chat_template
+#    chat_template: mistralv3tekken
+    roles_to_train: ["gpt"]
+    field_messages: conversations
+    message_field_role: from
+    message_field_content: value
+    train_on_eos: turn
+dataset_prepared_path: dataset_prepared
+val_set_size: 0.0
+output_dir: 12b-out
+
+sequence_len: 16384
+sample_packing: true
+pad_to_sequence_len: true
+
+adapter: lora
+lora_model_dir:
+lora_r: 128
+lora_alpha: 16
+lora_dropout: 0.05 
+  #lora_target_linear:
+  #lora_fan_in_fan_out: true
+peft_use_rslora: true
+lora_target_modules:
+  - gate_proj
+  - down_proj
+  - up_proj
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+
+wandb_project: 12b-control
+wandb_entity:
+wandb_watch:
+wandb_name: 12b-control
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 1
+num_epochs: 2
+optimizer: paged_adamw_8bit
+lr_scheduler: cosine
+learning_rate: 0.00001
+
+train_on_inputs: false
+group_by_length: false
+bf16: auto
+fp16:
+tf32: false
+
+gradient_checkpointing: unsloth
+  #gradient_checkpointing_kwargs:
+  #  use_reentrant: false
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_steps: 40
+evals_per_epoch: 
+eval_table_size:
+eval_max_new_tokens:
+saves_per_epoch: 2
+debug:
+deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
+weight_decay: 0.01
+fsdp:
+fsdp_config:
+special_tokens:
+  pad_token: <pad>
+
+
+```
+
+</details><br>
+
+# 12b-out
+
+This model was trained from scratch on the None dataset.
+
+## 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: 1e-05
+- train_batch_size: 1
+- eval_batch_size: 1
+- seed: 42
+- distributed_type: multi-GPU
+- num_devices: 4
+- gradient_accumulation_steps: 2
+- total_train_batch_size: 8
+- total_eval_batch_size: 4
+- optimizer: Use OptimizerNames.PAGED_ADAMW_8BIT with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
+- lr_scheduler_type: cosine
+- lr_scheduler_warmup_steps: 40
+- num_epochs: 2
+
+### Training results
+
+
+
+### Framework versions
+
+- PEFT 0.13.2
+- Transformers 4.46.1
+- Pytorch 2.3.1+cu121
+- Datasets 3.0.1
+- Tokenizers 0.20.3

adapter_config.json

@@ -0,0 +1,34 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "mistralai_Mistral-Nemo-Instruct-2407",
+  "bias": "none",
+  "fan_in_fan_out": null,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 16,
+  "lora_dropout": 0.05,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 128,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "v_proj",
+    "k_proj",
+    "down_proj",
+    "up_proj",
+    "o_proj",
+    "q_proj",
+    "gate_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "use_dora": false,
+  "use_rslora": true
+}

adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:912b74442aa9e9e1def6266ea27c4b417709435211cd92fc657859b8641c0104
+size 912337968

checkpoint-1107/README.md

@@ -0,0 +1,202 @@
+---
+base_model: mistralai_Mistral-Nemo-Instruct-2407
+library_name: peft
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.13.2

checkpoint-1107/adapter_config.json

@@ -0,0 +1,34 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "mistralai_Mistral-Nemo-Instruct-2407",
+  "bias": "none",
+  "fan_in_fan_out": null,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 16,
+  "lora_dropout": 0.05,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 128,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "v_proj",
+    "k_proj",
+    "down_proj",
+    "up_proj",
+    "o_proj",
+    "q_proj",
+    "gate_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "use_dora": false,
+  "use_rslora": true
+}

checkpoint-1107/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ae8e1b2856acfe3c8d8feca8956a11a6ba972c67694cd71544fffae4c302617f
+size 912337968

checkpoint-1107/global_step1107/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4034195a74bd46b115e5d5f294fbd25042ef871b9fa5762c9e6210d3d2b62d3c
+size 688705136

checkpoint-1107/global_step1107/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:c9f84e2f81127c931e01e4d2cebe4bfb63b2f76b1aa78070c572dd086d8b6d74
+size 688705136

checkpoint-1107/global_step1107/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:affbf02c93b03ebd0694d06e8d19765f12b5a7b79bcec058b347198929916251
+size 688705136

checkpoint-1107/global_step1107/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:a602a850a1cfdf4a410209a545cb10b74cf68d2d8334f62b7fe37a5de9dab395
+size 688705136

checkpoint-1107/global_step1107/zero_pp_rank_0_mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:08429cf699bfafd2ec35afc53817e4a80a75e90226d8743fa415672a6717a94f
+size 559966

checkpoint-1107/global_step1107/zero_pp_rank_1_mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0aafa27248abe8d62678c28fc595869d49d85eca9efe8c39b6da7264f1c0ff58
+size 559966

checkpoint-1107/global_step1107/zero_pp_rank_2_mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f3d8269e81285ea81a5e391ac9a6578570f8f9c9eb825d6bba4011b9126417b7
+size 559966

checkpoint-1107/global_step1107/zero_pp_rank_3_mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e25c46d38df422a82fd4d6b9a627e8267198b690fe714649cbe8a64ca33d7570
+size 559966

checkpoint-1107/latest

@@ -0,0 +1 @@
+global_step1107

checkpoint-1107/rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bdbde7b18f85248ea3e4f4d88e08c92f739aec2a596a8974df44a11d80cb0408
+size 14960

checkpoint-1107/rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:60979189c726d593c42335a265659f508a8e7d596c04618810386b3a1a375ff9
+size 14960

checkpoint-1107/rng_state_2.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:664c0c9f6f24169189fe9c8075991a85cc336b2fff8e71cc95e6b8a755dac3de
+size 14960

checkpoint-1107/rng_state_3.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f1efc02970f3fe747823398dd091a27d39b9c104365068ecbe58dfcc13a847cd
+size 14960

checkpoint-1107/scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:653c4f6368c97da8eb63b10e5349a30ffba2bb25ed49ba771de64d048a9419da
+size 1064

checkpoint-1107/special_tokens_map.json

@@ -0,0 +1,30 @@
+{
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<pad>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

checkpoint-1107/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b0240ce510f08e6c2041724e9043e33be9d251d1e4a4d94eb68cd47b954b61d2
+size 17078292

checkpoint-1107/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4bff29e7ff321b15e984196b58ce5d2fc8b550fddfbffd9146e631086009d7b1
+size 7992

checkpoint-1107/zero_to_fp32.py

@@ -0,0 +1,604 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example: python zero_to_fp32.py . pytorch_model.bin
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+
+    total_files = len(files)
+    state_dicts = []
+    for f in files:
+        state_dict = torch.load(f, map_location=device)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    if zero_stage <= 2:
+        fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    elif zero_stage == 3:
+        # if there is more than one param group, there will be multiple flattened tensors - one
+        # flattened tensor per group - for simplicity merge them into a single tensor
+        #
+        # XXX: could make the script more memory efficient for when there are multiple groups - it
+        # will require matching the sub-lists of param_shapes for each param group flattened tensor
+
+        fp32_flat_groups = [
+            torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
+        ]
+
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    for name, shape in param_shapes.items():
+
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # XXX: memory usage doubles here
+        state_dict[name] = torch.cat(
+            tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
+            0).narrow(0, 0, unpartitioned_numel).view(shape)
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None, exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+
+    Returns:
+        - pytorch ``state_dict``
+
+    Note: this approach may not work if your application doesn't have sufficient free CPU memory and
+    you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None, exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag, exclude_frozen_parameters)
+    print(f"Saving fp32 state dict to {output_file}")
+    torch.save(state_dict, output_file)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument(
+        "output_file",
+        type=str,
+        help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_file,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)

checkpoint-369/README.md

@@ -0,0 +1,202 @@
+---
+base_model: mistralai_Mistral-Nemo-Instruct-2407
+library_name: peft
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.13.2

checkpoint-369/adapter_config.json

@@ -0,0 +1,34 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "mistralai_Mistral-Nemo-Instruct-2407",
+  "bias": "none",
+  "fan_in_fan_out": null,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 16,
+  "lora_dropout": 0.05,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 128,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "v_proj",
+    "k_proj",
+    "down_proj",
+    "up_proj",
+    "o_proj",
+    "q_proj",
+    "gate_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "use_dora": false,
+  "use_rslora": true
+}

checkpoint-369/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b77d55a65d5a440d550746df1bb089e133a8fbfcb69beab2c6eac2020e1f0a6b
+size 912337968

checkpoint-369/global_step369/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:bf6aaa43db814b0db056fdd40ef81455cee2053a09a275f34eeb3268fb50ed8e
+size 688705136

checkpoint-369/global_step369/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:ad7cec1fabd4904e99c0bfb80bbf2918cc06956ba01019fb28b13ee2a459a8ac
+size 688705136

checkpoint-369/global_step369/bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:06735572849e4084f2483f8b1666ff621e07f816bca3a0fee06202cab332dd63
+size 688705136

checkpoint-369/global_step369/bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7aa816dc290f6b4b24e272eae6bb8b82a4bc6a73cf1960e0c26d634750590cc2
+size 688705136

checkpoint-369/global_step369/zero_pp_rank_0_mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6458a8381df22362adf1cb4d3050838367033aebb7ac0a389ac4a5fb9ca717d0
+size 559966

checkpoint-369/global_step369/zero_pp_rank_1_mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:87b3431f972816bfe24067f0a65e01f091ac7e3f68f7a33072477732aeb84519
+size 559966

checkpoint-369/global_step369/zero_pp_rank_2_mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:89c5d3704248213c78115ed270e39e01d4cc5b3afb7a751eab6dc0efaf8e8ee1
+size 559966

checkpoint-369/global_step369/zero_pp_rank_3_mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:85ccb68c73149ac50766ad701519fc0594324ae31dd127b9aa4c6eeafc903b88
+size 559966

checkpoint-369/latest

@@ -0,0 +1 @@
+global_step369

checkpoint-369/rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:193ab02b9ea2f38bccf8984657f0b1f6ceced083148df26f9961cbd53bd771b0
+size 14960

checkpoint-369/rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2ce80a2f8061d3aede17a14574a504bad06614b0a2bd8d3fdda770a7d8aa4e35
+size 14960

checkpoint-369/rng_state_2.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e2ccf94bd1219c5a51f81601c0db2d9a9da592801101d4ceaa1528d7202f671
+size 14960

checkpoint-369/rng_state_3.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:30255d17e25edfb1f161c2b27bde748d481bf4bfa14102c0b13a38e2177b78d4
+size 14960

checkpoint-369/scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a44a49a0ad33661f8ca98e6969165516cf2f862ed909c6c31a38267d265bc4b
+size 1064

checkpoint-369/special_tokens_map.json

@@ -0,0 +1,30 @@
+{
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<pad>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

checkpoint-369/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b0240ce510f08e6c2041724e9043e33be9d251d1e4a4d94eb68cd47b954b61d2
+size 17078292

checkpoint-369/trainer_state.json

@@ -0,0 +1,2616 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.4996614759647935,
+  "eval_steps": 500,
+  "global_step": 369,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.0013540961408259986,
+      "grad_norm": 2.4840032646087007,
+      "learning_rate": 2.5000000000000004e-07,
+      "loss": 3.1189,
+      "step": 1
+    },
+    {
+      "epoch": 0.002708192281651997,
+      "grad_norm": 2.92319290291185,
+      "learning_rate": 5.000000000000001e-07,
+      "loss": 2.6861,
+      "step": 2
+    },
+    {
+      "epoch": 0.004062288422477996,
+      "grad_norm": 2.8047579443092108,
+      "learning_rate": 7.5e-07,
+      "loss": 3.339,
+      "step": 3
+    },
+    {
+      "epoch": 0.005416384563303994,
+      "grad_norm": 2.4920981499635224,
+      "learning_rate": 1.0000000000000002e-06,
+      "loss": 2.4573,
+      "step": 4
+    },
+    {
+      "epoch": 0.006770480704129994,
+      "grad_norm": 2.6989232511239787,
+      "learning_rate": 1.25e-06,
+      "loss": 2.5324,
+      "step": 5
+    },
+    {
+      "epoch": 0.008124576844955992,
+      "grad_norm": 3.0185163841424987,
+      "learning_rate": 1.5e-06,
+      "loss": 2.7215,
+      "step": 6
+    },
+    {
+      "epoch": 0.009478672985781991,
+      "grad_norm": 2.0398408839964137,
+      "learning_rate": 1.75e-06,
+      "loss": 2.1644,
+      "step": 7
+    },
+    {
+      "epoch": 0.010832769126607989,
+      "grad_norm": 3.45002669711186,
+      "learning_rate": 2.0000000000000003e-06,
+      "loss": 2.9371,
+      "step": 8
+    },
+    {
+      "epoch": 0.012186865267433988,
+      "grad_norm": 2.2776378980264718,
+      "learning_rate": 2.25e-06,
+      "loss": 3.6517,
+      "step": 9
+    },
+    {
+      "epoch": 0.013540961408259987,
+      "grad_norm": 2.825903106798789,
+      "learning_rate": 2.5e-06,
+      "loss": 2.5649,
+      "step": 10
+    },
+    {
+      "epoch": 0.014895057549085985,
+      "grad_norm": 2.723314018937158,
+      "learning_rate": 2.7500000000000004e-06,
+      "loss": 3.0391,
+      "step": 11
+    },
+    {
+      "epoch": 0.016249153689911984,
+      "grad_norm": 3.172036134399068,
+      "learning_rate": 3e-06,
+      "loss": 2.8953,
+      "step": 12
+    },
+    {
+      "epoch": 0.017603249830737983,
+      "grad_norm": 2.838052809555009,
+      "learning_rate": 3.2500000000000002e-06,
+      "loss": 2.8184,
+      "step": 13
+    },
+    {
+      "epoch": 0.018957345971563982,
+      "grad_norm": 3.7070274580501756,
+      "learning_rate": 3.5e-06,
+      "loss": 2.9666,
+      "step": 14
+    },
+    {
+      "epoch": 0.020311442112389978,
+      "grad_norm": 2.2063413123307223,
+      "learning_rate": 3.7500000000000005e-06,
+      "loss": 2.8188,
+      "step": 15
+    },
+    {
+      "epoch": 0.021665538253215978,
+      "grad_norm": 2.5023444159347243,
+      "learning_rate": 4.000000000000001e-06,
+      "loss": 2.7621,
+      "step": 16
+    },
+    {
+      "epoch": 0.023019634394041977,
+      "grad_norm": 2.0716483514966266,
+      "learning_rate": 4.25e-06,
+      "loss": 2.5605,
+      "step": 17
+    },
+    {
+      "epoch": 0.024373730534867976,
+      "grad_norm": 1.9009417808791889,
+      "learning_rate": 4.5e-06,
+      "loss": 2.3993,
+      "step": 18
+    },
+    {
+      "epoch": 0.025727826675693975,
+      "grad_norm": 1.6281815273569142,
+      "learning_rate": 4.75e-06,
+      "loss": 2.8062,
+      "step": 19
+    },
+    {
+      "epoch": 0.027081922816519974,
+      "grad_norm": 2.7256934688885153,
+      "learning_rate": 5e-06,
+      "loss": 2.4145,
+      "step": 20
+    },
+    {
+      "epoch": 0.02843601895734597,
+      "grad_norm": 1.7492771266021097,
+      "learning_rate": 5.2500000000000006e-06,
+      "loss": 3.0841,
+      "step": 21
+    },
+    {
+      "epoch": 0.02979011509817197,
+      "grad_norm": 0.9056200631767789,
+      "learning_rate": 5.500000000000001e-06,
+      "loss": 2.7354,
+      "step": 22
+    },
+    {
+      "epoch": 0.03114421123899797,
+      "grad_norm": 0.9006011502765553,
+      "learning_rate": 5.75e-06,
+      "loss": 2.5312,
+      "step": 23
+    },
+    {
+      "epoch": 0.03249830737982397,
+      "grad_norm": 1.5068464694001362,
+      "learning_rate": 6e-06,
+      "loss": 2.4401,
+      "step": 24
+    },
+    {
+      "epoch": 0.033852403520649964,
+      "grad_norm": 1.1756450984633096,
+      "learning_rate": 6.25e-06,
+      "loss": 2.7984,
+      "step": 25
+    },
+    {
+      "epoch": 0.035206499661475966,
+      "grad_norm": 1.5609942359404068,
+      "learning_rate": 6.5000000000000004e-06,
+      "loss": 1.8799,
+      "step": 26
+    },
+    {
+      "epoch": 0.03656059580230196,
+      "grad_norm": 1.103542364498138,
+      "learning_rate": 6.750000000000001e-06,
+      "loss": 2.2657,
+      "step": 27
+    },
+    {
+      "epoch": 0.037914691943127965,
+      "grad_norm": 0.9971728919317084,
+      "learning_rate": 7e-06,
+      "loss": 2.4469,
+      "step": 28
+    },
+    {
+      "epoch": 0.03926878808395396,
+      "grad_norm": 1.2637045038992765,
+      "learning_rate": 7.25e-06,
+      "loss": 2.7085,
+      "step": 29
+    },
+    {
+      "epoch": 0.040622884224779957,
+      "grad_norm": 1.0087355495607542,
+      "learning_rate": 7.500000000000001e-06,
+      "loss": 2.647,
+      "step": 30
+    },
+    {
+      "epoch": 0.04197698036560596,
+      "grad_norm": 1.1422809581031197,
+      "learning_rate": 7.75e-06,
+      "loss": 2.666,
+      "step": 31
+    },
+    {
+      "epoch": 0.043331076506431955,
+      "grad_norm": 0.9656489652064195,
+      "learning_rate": 8.000000000000001e-06,
+      "loss": 2.674,
+      "step": 32
+    },
+    {
+      "epoch": 0.04468517264725796,
+      "grad_norm": 1.2694895548396936,
+      "learning_rate": 8.25e-06,
+      "loss": 2.6238,
+      "step": 33
+    },
+    {
+      "epoch": 0.046039268788083954,
+      "grad_norm": 0.9223153234403145,
+      "learning_rate": 8.5e-06,
+      "loss": 2.4833,
+      "step": 34
+    },
+    {
+      "epoch": 0.04739336492890995,
+      "grad_norm": 0.6904918931455449,
+      "learning_rate": 8.750000000000001e-06,
+      "loss": 2.3643,
+      "step": 35
+    },
+    {
+      "epoch": 0.04874746106973595,
+      "grad_norm": 1.3920707796397145,
+      "learning_rate": 9e-06,
+      "loss": 2.7138,
+      "step": 36
+    },
+    {
+      "epoch": 0.05010155721056195,
+      "grad_norm": 1.3111040526577917,
+      "learning_rate": 9.250000000000001e-06,
+      "loss": 3.0125,
+      "step": 37
+    },
+    {
+      "epoch": 0.05145565335138795,
+      "grad_norm": 1.280730826597217,
+      "learning_rate": 9.5e-06,
+      "loss": 3.0783,
+      "step": 38
+    },
+    {
+      "epoch": 0.052809749492213946,
+      "grad_norm": 1.0271755023288214,
+      "learning_rate": 9.75e-06,
+      "loss": 2.0996,
+      "step": 39
+    },
+    {
+      "epoch": 0.05416384563303995,
+      "grad_norm": 0.8222149166421281,
+      "learning_rate": 1e-05,
+      "loss": 2.5166,
+      "step": 40
+    },
+    {
+      "epoch": 0.055517941773865945,
+      "grad_norm": 0.9001955314444051,
+      "learning_rate": 9.999988034504136e-06,
+      "loss": 2.2754,
+      "step": 41
+    },
+    {
+      "epoch": 0.05687203791469194,
+      "grad_norm": 0.8230494015971713,
+      "learning_rate": 9.99995213807381e-06,
+      "loss": 2.2014,
+      "step": 42
+    },
+    {
+      "epoch": 0.05822613405551794,
+      "grad_norm": 0.8386219965539589,
+      "learning_rate": 9.99989231088083e-06,
+      "loss": 2.7913,
+      "step": 43
+    },
+    {
+      "epoch": 0.05958023019634394,
+      "grad_norm": 0.8301953727405478,
+      "learning_rate": 9.99980855321154e-06,
+      "loss": 2.8559,
+      "step": 44
+    },
+    {
+      "epoch": 0.06093432633716994,
+      "grad_norm": 1.435333986474164,
+      "learning_rate": 9.999700865466825e-06,
+      "loss": 2.7823,
+      "step": 45
+    },
+    {
+      "epoch": 0.06228842247799594,
+      "grad_norm": 0.7745604592769975,
+      "learning_rate": 9.999569248162095e-06,
+      "loss": 2.1645,
+      "step": 46
+    },
+    {
+      "epoch": 0.06364251861882193,
+      "grad_norm": 0.7872260023932528,
+      "learning_rate": 9.999413701927298e-06,
+      "loss": 2.6378,
+      "step": 47
+    },
+    {
+      "epoch": 0.06499661475964794,
+      "grad_norm": 0.9284676925561781,
+      "learning_rate": 9.999234227506912e-06,
+      "loss": 2.3144,
+      "step": 48
+    },
+    {
+      "epoch": 0.06635071090047394,
+      "grad_norm": 0.7702824584250887,
+      "learning_rate": 9.999030825759934e-06,
+      "loss": 2.5719,
+      "step": 49
+    },
+    {
+      "epoch": 0.06770480704129993,
+      "grad_norm": 1.7969417808741215,
+      "learning_rate": 9.998803497659885e-06,
+      "loss": 2.2117,
+      "step": 50
+    },
+    {
+      "epoch": 0.06905890318212593,
+      "grad_norm": 0.8303034868453852,
+      "learning_rate": 9.998552244294803e-06,
+      "loss": 2.4058,
+      "step": 51
+    },
+    {
+      "epoch": 0.07041299932295193,
+      "grad_norm": 1.0962548207238205,
+      "learning_rate": 9.998277066867236e-06,
+      "loss": 2.6942,
+      "step": 52
+    },
+    {
+      "epoch": 0.07176709546377792,
+      "grad_norm": 1.0119530000904415,
+      "learning_rate": 9.99797796669424e-06,
+      "loss": 2.6186,
+      "step": 53
+    },
+    {
+      "epoch": 0.07312119160460392,
+      "grad_norm": 0.8130591704344431,
+      "learning_rate": 9.997654945207368e-06,
+      "loss": 2.5114,
+      "step": 54
+    },
+    {
+      "epoch": 0.07447528774542993,
+      "grad_norm": 0.7598768980100439,
+      "learning_rate": 9.997308003952659e-06,
+      "loss": 2.5052,
+      "step": 55
+    },
+    {
+      "epoch": 0.07582938388625593,
+      "grad_norm": 0.7067903745494273,
+      "learning_rate": 9.99693714459065e-06,
+      "loss": 2.6569,
+      "step": 56
+    },
+    {
+      "epoch": 0.07718348002708192,
+      "grad_norm": 1.1966126433581616,
+      "learning_rate": 9.996542368896343e-06,
+      "loss": 2.1643,
+      "step": 57
+    },
+    {
+      "epoch": 0.07853757616790792,
+      "grad_norm": 0.6603354212066763,
+      "learning_rate": 9.996123678759214e-06,
+      "loss": 2.1989,
+      "step": 58
+    },
+    {
+      "epoch": 0.07989167230873392,
+      "grad_norm": 1.536917763640058,
+      "learning_rate": 9.995681076183197e-06,
+      "loss": 2.542,
+      "step": 59
+    },
+    {
+      "epoch": 0.08124576844955991,
+      "grad_norm": 1.076294402304339,
+      "learning_rate": 9.995214563286677e-06,
+      "loss": 2.6239,
+      "step": 60
+    },
+    {
+      "epoch": 0.08259986459038592,
+      "grad_norm": 0.8429191198665816,
+      "learning_rate": 9.994724142302474e-06,
+      "loss": 2.5167,
+      "step": 61
+    },
+    {
+      "epoch": 0.08395396073121192,
+      "grad_norm": 0.9566424524099907,
+      "learning_rate": 9.994209815577843e-06,
+      "loss": 2.4603,
+      "step": 62
+    },
+    {
+      "epoch": 0.08530805687203792,
+      "grad_norm": 0.770516812968522,
+      "learning_rate": 9.993671585574452e-06,
+      "loss": 2.615,
+      "step": 63
+    },
+    {
+      "epoch": 0.08666215301286391,
+      "grad_norm": 0.6919266782030121,
+      "learning_rate": 9.993109454868379e-06,
+      "loss": 2.3676,
+      "step": 64
+    },
+    {
+      "epoch": 0.08801624915368991,
+      "grad_norm": 0.75481755204904,
+      "learning_rate": 9.992523426150091e-06,
+      "loss": 2.4574,
+      "step": 65
+    },
+    {
+      "epoch": 0.08937034529451592,
+      "grad_norm": 0.7067208707696914,
+      "learning_rate": 9.991913502224438e-06,
+      "loss": 2.2171,
+      "step": 66
+    },
+    {
+      "epoch": 0.0907244414353419,
+      "grad_norm": 1.0878007201904665,
+      "learning_rate": 9.991279686010637e-06,
+      "loss": 3.1181,
+      "step": 67
+    },
+    {
+      "epoch": 0.09207853757616791,
+      "grad_norm": 0.7406127801462548,
+      "learning_rate": 9.990621980542258e-06,
+      "loss": 2.5117,
+      "step": 68
+    },
+    {
+      "epoch": 0.09343263371699391,
+      "grad_norm": 0.7363797966643979,
+      "learning_rate": 9.98994038896721e-06,
+      "loss": 2.769,
+      "step": 69
+    },
+    {
+      "epoch": 0.0947867298578199,
+      "grad_norm": 0.6997455244847285,
+      "learning_rate": 9.989234914547725e-06,
+      "loss": 2.3619,
+      "step": 70
+    },
+    {
+      "epoch": 0.0961408259986459,
+      "grad_norm": 0.6185449490741277,
+      "learning_rate": 9.988505560660344e-06,
+      "loss": 2.4934,
+      "step": 71
+    },
+    {
+      "epoch": 0.0974949221394719,
+      "grad_norm": 1.8182918826297632,
+      "learning_rate": 9.9877523307959e-06,
+      "loss": 2.7052,
+      "step": 72
+    },
+    {
+      "epoch": 0.0988490182802979,
+      "grad_norm": 0.9136867007851477,
+      "learning_rate": 9.9869752285595e-06,
+      "loss": 2.7127,
+      "step": 73
+    },
+    {
+      "epoch": 0.1002031144211239,
+      "grad_norm": 0.8177660946450126,
+      "learning_rate": 9.986174257670509e-06,
+      "loss": 2.5386,
+      "step": 74
+    },
+    {
+      "epoch": 0.1015572105619499,
+      "grad_norm": 0.7019364782921708,
+      "learning_rate": 9.985349421962532e-06,
+      "loss": 2.6846,
+      "step": 75
+    },
+    {
+      "epoch": 0.1029113067027759,
+      "grad_norm": 0.7776591538429096,
+      "learning_rate": 9.984500725383397e-06,
+      "loss": 2.7552,
+      "step": 76
+    },
+    {
+      "epoch": 0.10426540284360189,
+      "grad_norm": 0.6829070090456675,
+      "learning_rate": 9.983628171995136e-06,
+      "loss": 2.049,
+      "step": 77
+    },
+    {
+      "epoch": 0.10561949898442789,
+      "grad_norm": 0.8680053222596442,
+      "learning_rate": 9.98273176597396e-06,
+      "loss": 2.4778,
+      "step": 78
+    },
+    {
+      "epoch": 0.1069735951252539,
+      "grad_norm": 0.7964757819779202,
+      "learning_rate": 9.981811511610246e-06,
+      "loss": 2.609,
+      "step": 79
+    },
+    {
+      "epoch": 0.1083276912660799,
+      "grad_norm": 0.8603316916001541,
+      "learning_rate": 9.980867413308516e-06,
+      "loss": 2.6895,
+      "step": 80
+    },
+    {
+      "epoch": 0.10968178740690589,
+      "grad_norm": 0.9007993024274796,
+      "learning_rate": 9.97989947558741e-06,
+      "loss": 1.841,
+      "step": 81
+    },
+    {
+      "epoch": 0.11103588354773189,
+      "grad_norm": 1.022466421728221,
+      "learning_rate": 9.978907703079672e-06,
+      "loss": 2.368,
+      "step": 82
+    },
+    {
+      "epoch": 0.11238997968855789,
+      "grad_norm": 0.8845970984922622,
+      "learning_rate": 9.97789210053212e-06,
+      "loss": 2.6168,
+      "step": 83
+    },
+    {
+      "epoch": 0.11374407582938388,
+      "grad_norm": 0.7307749666567861,
+      "learning_rate": 9.976852672805625e-06,
+      "loss": 2.6611,
+      "step": 84
+    },
+    {
+      "epoch": 0.11509817197020988,
+      "grad_norm": 1.030662172928111,
+      "learning_rate": 9.975789424875104e-06,
+      "loss": 2.4778,
+      "step": 85
+    },
+    {
+      "epoch": 0.11645226811103589,
+      "grad_norm": 1.0269306202089545,
+      "learning_rate": 9.974702361829465e-06,
+      "loss": 2.0226,
+      "step": 86
+    },
+    {
+      "epoch": 0.11780636425186188,
+      "grad_norm": 0.9557403327101263,
+      "learning_rate": 9.97359148887161e-06,
+      "loss": 2.5957,
+      "step": 87
+    },
+    {
+      "epoch": 0.11916046039268788,
+      "grad_norm": 1.2342067337912177,
+      "learning_rate": 9.972456811318399e-06,
+      "loss": 2.3621,
+      "step": 88
+    },
+    {
+      "epoch": 0.12051455653351388,
+      "grad_norm": 0.9439454489946377,
+      "learning_rate": 9.97129833460062e-06,
+      "loss": 2.1986,
+      "step": 89
+    },
+    {
+      "epoch": 0.12186865267433988,
+      "grad_norm": 0.9410876234442836,
+      "learning_rate": 9.970116064262975e-06,
+      "loss": 2.7015,
+      "step": 90
+    },
+    {
+      "epoch": 0.12322274881516587,
+      "grad_norm": 0.7573572965312575,
+      "learning_rate": 9.968910005964043e-06,
+      "loss": 3.2786,
+      "step": 91
+    },
+    {
+      "epoch": 0.12457684495599188,
+      "grad_norm": 0.8749972433179365,
+      "learning_rate": 9.96768016547626e-06,
+      "loss": 2.5444,
+      "step": 92
+    },
+    {
+      "epoch": 0.12593094109681788,
+      "grad_norm": 0.9612381390622727,
+      "learning_rate": 9.966426548685883e-06,
+      "loss": 3.0084,
+      "step": 93
+    },
+    {
+      "epoch": 0.12728503723764387,
+      "grad_norm": 2.987117261751073,
+      "learning_rate": 9.965149161592973e-06,
+      "loss": 2.497,
+      "step": 94
+    },
+    {
+      "epoch": 0.12863913337846988,
+      "grad_norm": 0.9835961282773672,
+      "learning_rate": 9.96384801031136e-06,
+      "loss": 2.5352,
+      "step": 95
+    },
+    {
+      "epoch": 0.12999322951929587,
+      "grad_norm": 0.9193436131089541,
+      "learning_rate": 9.962523101068608e-06,
+      "loss": 2.8517,
+      "step": 96
+    },
+    {
+      "epoch": 0.13134732566012186,
+      "grad_norm": 0.7624957621675597,
+      "learning_rate": 9.961174440205998e-06,
+      "loss": 2.5317,
+      "step": 97
+    },
+    {
+      "epoch": 0.13270142180094788,
+      "grad_norm": 0.8220019285296615,
+      "learning_rate": 9.959802034178489e-06,
+      "loss": 2.4109,
+      "step": 98
+    },
+    {
+      "epoch": 0.13405551794177387,
+      "grad_norm": 0.7696635269910559,
+      "learning_rate": 9.958405889554687e-06,
+      "loss": 2.2455,
+      "step": 99
+    },
+    {
+      "epoch": 0.13540961408259986,
+      "grad_norm": 1.2821194483642973,
+      "learning_rate": 9.956986013016816e-06,
+      "loss": 2.568,
+      "step": 100
+    },
+    {
+      "epoch": 0.13676371022342587,
+      "grad_norm": 1.0396436901580886,
+      "learning_rate": 9.95554241136069e-06,
+      "loss": 2.1504,
+      "step": 101
+    },
+    {
+      "epoch": 0.13811780636425186,
+      "grad_norm": 1.222875873814822,
+      "learning_rate": 9.954075091495669e-06,
+      "loss": 1.8413,
+      "step": 102
+    },
+    {
+      "epoch": 0.13947190250507785,
+      "grad_norm": 0.7869762113593299,
+      "learning_rate": 9.95258406044464e-06,
+      "loss": 2.3467,
+      "step": 103
+    },
+    {
+      "epoch": 0.14082599864590387,
+      "grad_norm": 0.7464132847692124,
+      "learning_rate": 9.951069325343972e-06,
+      "loss": 2.3054,
+      "step": 104
+    },
+    {
+      "epoch": 0.14218009478672985,
+      "grad_norm": 0.7888132769649014,
+      "learning_rate": 9.949530893443489e-06,
+      "loss": 2.7196,
+      "step": 105
+    },
+    {
+      "epoch": 0.14353419092755584,
+      "grad_norm": 0.7956222737201969,
+      "learning_rate": 9.947968772106428e-06,
+      "loss": 2.4884,
+      "step": 106
+    },
+    {
+      "epoch": 0.14488828706838186,
+      "grad_norm": 0.7975440031710949,
+      "learning_rate": 9.946382968809413e-06,
+      "loss": 2.8331,
+      "step": 107
+    },
+    {
+      "epoch": 0.14624238320920785,
+      "grad_norm": 1.95160164876152,
+      "learning_rate": 9.944773491142416e-06,
+      "loss": 2.5936,
+      "step": 108
+    },
+    {
+      "epoch": 0.14759647935003387,
+      "grad_norm": 0.7707347796969224,
+      "learning_rate": 9.943140346808715e-06,
+      "loss": 1.9831,
+      "step": 109
+    },
+    {
+      "epoch": 0.14895057549085985,
+      "grad_norm": 1.0591140801414067,
+      "learning_rate": 9.94148354362486e-06,
+      "loss": 2.224,
+      "step": 110
+    },
+    {
+      "epoch": 0.15030467163168584,
+      "grad_norm": 1.0010907115642087,
+      "learning_rate": 9.939803089520641e-06,
+      "loss": 2.6739,
+      "step": 111
+    },
+    {
+      "epoch": 0.15165876777251186,
+      "grad_norm": 0.8318923981444863,
+      "learning_rate": 9.938098992539045e-06,
+      "loss": 2.2503,
+      "step": 112
+    },
+    {
+      "epoch": 0.15301286391333785,
+      "grad_norm": 0.7628830744549152,
+      "learning_rate": 9.93637126083622e-06,
+      "loss": 2.4004,
+      "step": 113
+    },
+    {
+      "epoch": 0.15436696005416384,
+      "grad_norm": 1.2732533307878235,
+      "learning_rate": 9.93461990268143e-06,
+      "loss": 2.3523,
+      "step": 114
+    },
+    {
+      "epoch": 0.15572105619498985,
+      "grad_norm": 0.8643222482800437,
+      "learning_rate": 9.932844926457024e-06,
+      "loss": 2.6932,
+      "step": 115
+    },
+    {
+      "epoch": 0.15707515233581584,
+      "grad_norm": 1.0982969820587143,
+      "learning_rate": 9.931046340658387e-06,
+      "loss": 2.6476,
+      "step": 116
+    },
+    {
+      "epoch": 0.15842924847664183,
+      "grad_norm": 0.8246096418384085,
+      "learning_rate": 9.929224153893913e-06,
+      "loss": 2.2256,
+      "step": 117
+    },
+    {
+      "epoch": 0.15978334461746785,
+      "grad_norm": 1.1291179847133526,
+      "learning_rate": 9.927378374884947e-06,
+      "loss": 2.521,
+      "step": 118
+    },
+    {
+      "epoch": 0.16113744075829384,
+      "grad_norm": 0.7541483793541535,
+      "learning_rate": 9.92550901246575e-06,
+      "loss": 1.8844,
+      "step": 119
+    },
+    {
+      "epoch": 0.16249153689911983,
+      "grad_norm": 0.8163459296091435,
+      "learning_rate": 9.923616075583465e-06,
+      "loss": 3.0096,
+      "step": 120
+    },
+    {
+      "epoch": 0.16384563303994584,
+      "grad_norm": 1.026618712540373,
+      "learning_rate": 9.921699573298063e-06,
+      "loss": 2.7169,
+      "step": 121
+    },
+    {
+      "epoch": 0.16519972918077183,
+      "grad_norm": 0.9163179029291131,
+      "learning_rate": 9.919759514782304e-06,
+      "loss": 2.2547,
+      "step": 122
+    },
+    {
+      "epoch": 0.16655382532159782,
+      "grad_norm": 0.637693899798208,
+      "learning_rate": 9.91779590932169e-06,
+      "loss": 2.1517,
+      "step": 123
+    },
+    {
+      "epoch": 0.16790792146242384,
+      "grad_norm": 0.9559710413913272,
+      "learning_rate": 9.91580876631443e-06,
+      "loss": 2.2203,
+      "step": 124
+    },
+    {
+      "epoch": 0.16926201760324983,
+      "grad_norm": 1.696116526101562,
+      "learning_rate": 9.913798095271385e-06,
+      "loss": 2.3273,
+      "step": 125
+    },
+    {
+      "epoch": 0.17061611374407584,
+      "grad_norm": 0.8410939323658712,
+      "learning_rate": 9.91176390581602e-06,
+      "loss": 2.5569,
+      "step": 126
+    },
+    {
+      "epoch": 0.17197020988490183,
+      "grad_norm": 0.9518930879821595,
+      "learning_rate": 9.909706207684377e-06,
+      "loss": 2.5826,
+      "step": 127
+    },
+    {
+      "epoch": 0.17332430602572782,
+      "grad_norm": 0.8464728478763351,
+      "learning_rate": 9.907625010724999e-06,
+      "loss": 2.5954,
+      "step": 128
+    },
+    {
+      "epoch": 0.17467840216655384,
+      "grad_norm": 1.7797343941367096,
+      "learning_rate": 9.905520324898914e-06,
+      "loss": 2.238,
+      "step": 129
+    },
+    {
+      "epoch": 0.17603249830737983,
+      "grad_norm": 1.200087573042979,
+      "learning_rate": 9.903392160279564e-06,
+      "loss": 2.7202,
+      "step": 130
+    },
+    {
+      "epoch": 0.17738659444820581,
+      "grad_norm": 0.8464685532106079,
+      "learning_rate": 9.901240527052764e-06,
+      "loss": 2.105,
+      "step": 131
+    },
+    {
+      "epoch": 0.17874069058903183,
+      "grad_norm": 0.940863564248954,
+      "learning_rate": 9.899065435516661e-06,
+      "loss": 2.5347,
+      "step": 132
+    },
+    {
+      "epoch": 0.18009478672985782,
+      "grad_norm": 0.7977576897280799,
+      "learning_rate": 9.896866896081675e-06,
+      "loss": 2.9814,
+      "step": 133
+    },
+    {
+      "epoch": 0.1814488828706838,
+      "grad_norm": 1.1663128930633815,
+      "learning_rate": 9.894644919270448e-06,
+      "loss": 2.5215,
+      "step": 134
+    },
+    {
+      "epoch": 0.18280297901150983,
+      "grad_norm": 1.728136703163258,
+      "learning_rate": 9.892399515717805e-06,
+      "loss": 3.053,
+      "step": 135
+    },
+    {
+      "epoch": 0.18415707515233581,
+      "grad_norm": 0.7851660984466424,
+      "learning_rate": 9.890130696170691e-06,
+      "loss": 1.9476,
+      "step": 136
+    },
+    {
+      "epoch": 0.1855111712931618,
+      "grad_norm": 0.7821494358810632,
+      "learning_rate": 9.887838471488126e-06,
+      "loss": 2.2599,
+      "step": 137
+    },
+    {
+      "epoch": 0.18686526743398782,
+      "grad_norm": 0.9644512402210179,
+      "learning_rate": 9.885522852641156e-06,
+      "loss": 2.352,
+      "step": 138
+    },
+    {
+      "epoch": 0.1882193635748138,
+      "grad_norm": 0.8127395523345673,
+      "learning_rate": 9.883183850712788e-06,
+      "loss": 2.1334,
+      "step": 139
+    },
+    {
+      "epoch": 0.1895734597156398,
+      "grad_norm": 1.041333455542229,
+      "learning_rate": 9.880821476897948e-06,
+      "loss": 1.7956,
+      "step": 140
+    },
+    {
+      "epoch": 0.1909275558564658,
+      "grad_norm": 0.905550712837899,
+      "learning_rate": 9.878435742503431e-06,
+      "loss": 2.4699,
+      "step": 141
+    },
+    {
+      "epoch": 0.1922816519972918,
+      "grad_norm": 2.4663654618066637,
+      "learning_rate": 9.87602665894783e-06,
+      "loss": 2.342,
+      "step": 142
+    },
+    {
+      "epoch": 0.19363574813811782,
+      "grad_norm": 0.937041815915097,
+      "learning_rate": 9.8735942377615e-06,
+      "loss": 2.2519,
+      "step": 143
+    },
+    {
+      "epoch": 0.1949898442789438,
+      "grad_norm": 1.0716246983522297,
+      "learning_rate": 9.871138490586489e-06,
+      "loss": 2.2515,
+      "step": 144
+    },
+    {
+      "epoch": 0.1963439404197698,
+      "grad_norm": 0.9279172063288832,
+      "learning_rate": 9.86865942917649e-06,
+      "loss": 2.1349,
+      "step": 145
+    },
+    {
+      "epoch": 0.1976980365605958,
+      "grad_norm": 1.091933592631405,
+      "learning_rate": 9.866157065396784e-06,
+      "loss": 2.465,
+      "step": 146
+    },
+    {
+      "epoch": 0.1990521327014218,
+      "grad_norm": 0.8345999787153139,
+      "learning_rate": 9.863631411224182e-06,
+      "loss": 2.3305,
+      "step": 147
+    },
+    {
+      "epoch": 0.2004062288422478,
+      "grad_norm": 0.8966768512424247,
+      "learning_rate": 9.861082478746962e-06,
+      "loss": 2.3482,
+      "step": 148
+    },
+    {
+      "epoch": 0.2017603249830738,
+      "grad_norm": 1.0000509829104935,
+      "learning_rate": 9.858510280164822e-06,
+      "loss": 2.7295,
+      "step": 149
+    },
+    {
+      "epoch": 0.2031144211238998,
+      "grad_norm": 0.826458573119617,
+      "learning_rate": 9.855914827788814e-06,
+      "loss": 2.4229,
+      "step": 150
+    },
+    {
+      "epoch": 0.20446851726472579,
+      "grad_norm": 1.0195204526871002,
+      "learning_rate": 9.853296134041292e-06,
+      "loss": 2.5522,
+      "step": 151
+    },
+    {
+      "epoch": 0.2058226134055518,
+      "grad_norm": 1.016538392862215,
+      "learning_rate": 9.850654211455837e-06,
+      "loss": 2.7309,
+      "step": 152
+    },
+    {
+      "epoch": 0.2071767095463778,
+      "grad_norm": 0.957564909378697,
+      "learning_rate": 9.84798907267722e-06,
+      "loss": 2.284,
+      "step": 153
+    },
+    {
+      "epoch": 0.20853080568720378,
+      "grad_norm": 0.7014654099006512,
+      "learning_rate": 9.84530073046132e-06,
+      "loss": 2.2114,
+      "step": 154
+    },
+    {
+      "epoch": 0.2098849018280298,
+      "grad_norm": 0.8362168323360238,
+      "learning_rate": 9.842589197675079e-06,
+      "loss": 2.5489,
+      "step": 155
+    },
+    {
+      "epoch": 0.21123899796885579,
+      "grad_norm": 0.88786168404071,
+      "learning_rate": 9.83985448729643e-06,
+      "loss": 2.3827,
+      "step": 156
+    },
+    {
+      "epoch": 0.21259309410968177,
+      "grad_norm": 0.9463226185764325,
+      "learning_rate": 9.837096612414238e-06,
+      "loss": 2.7418,
+      "step": 157
+    },
+    {
+      "epoch": 0.2139471902505078,
+      "grad_norm": 0.870063498618297,
+      "learning_rate": 9.83431558622824e-06,
+      "loss": 2.2803,
+      "step": 158
+    },
+    {
+      "epoch": 0.21530128639133378,
+      "grad_norm": 0.8764083614224288,
+      "learning_rate": 9.831511422048978e-06,
+      "loss": 2.1819,
+      "step": 159
+    },
+    {
+      "epoch": 0.2166553825321598,
+      "grad_norm": 0.9146169100705829,
+      "learning_rate": 9.828684133297738e-06,
+      "loss": 2.3676,
+      "step": 160
+    },
+    {
+      "epoch": 0.21800947867298578,
+      "grad_norm": 1.2142753558720913,
+      "learning_rate": 9.825833733506489e-06,
+      "loss": 2.402,
+      "step": 161
+    },
+    {
+      "epoch": 0.21936357481381177,
+      "grad_norm": 1.0480064346220106,
+      "learning_rate": 9.822960236317804e-06,
+      "loss": 2.5415,
+      "step": 162
+    },
+    {
+      "epoch": 0.2207176709546378,
+      "grad_norm": 0.7629917801668891,
+      "learning_rate": 9.820063655484813e-06,
+      "loss": 2.0297,
+      "step": 163
+    },
+    {
+      "epoch": 0.22207176709546378,
+      "grad_norm": 0.7990888394132036,
+      "learning_rate": 9.817144004871127e-06,
+      "loss": 2.8255,
+      "step": 164
+    },
+    {
+      "epoch": 0.22342586323628977,
+      "grad_norm": 0.7798600139592904,
+      "learning_rate": 9.814201298450773e-06,
+      "loss": 2.2763,
+      "step": 165
+    },
+    {
+      "epoch": 0.22477995937711578,
+      "grad_norm": 1.173741986019506,
+      "learning_rate": 9.811235550308127e-06,
+      "loss": 1.9561,
+      "step": 166
+    },
+    {
+      "epoch": 0.22613405551794177,
+      "grad_norm": 0.7570228405125989,
+      "learning_rate": 9.808246774637845e-06,
+      "loss": 2.8304,
+      "step": 167
+    },
+    {
+      "epoch": 0.22748815165876776,
+      "grad_norm": 1.2061240895038743,
+      "learning_rate": 9.805234985744804e-06,
+      "loss": 2.1653,
+      "step": 168
+    },
+    {
+      "epoch": 0.22884224779959378,
+      "grad_norm": 0.7756108337956075,
+      "learning_rate": 9.802200198044024e-06,
+      "loss": 2.5092,
+      "step": 169
+    },
+    {
+      "epoch": 0.23019634394041977,
+      "grad_norm": 1.1436553265186091,
+      "learning_rate": 9.799142426060595e-06,
+      "loss": 2.7414,
+      "step": 170
+    },
+    {
+      "epoch": 0.23155044008124576,
+      "grad_norm": 0.8234458954053101,
+      "learning_rate": 9.796061684429626e-06,
+      "loss": 2.2852,
+      "step": 171
+    },
+    {
+      "epoch": 0.23290453622207177,
+      "grad_norm": 1.257655370196978,
+      "learning_rate": 9.792957987896154e-06,
+      "loss": 2.6166,
+      "step": 172
+    },
+    {
+      "epoch": 0.23425863236289776,
+      "grad_norm": 1.0232152459073642,
+      "learning_rate": 9.78983135131509e-06,
+      "loss": 2.5478,
+      "step": 173
+    },
+    {
+      "epoch": 0.23561272850372375,
+      "grad_norm": 0.7837252990266905,
+      "learning_rate": 9.786681789651134e-06,
+      "loss": 2.3348,
+      "step": 174
+    },
+    {
+      "epoch": 0.23696682464454977,
+      "grad_norm": 1.227249953589518,
+      "learning_rate": 9.783509317978713e-06,
+      "loss": 2.2314,
+      "step": 175
+    },
+    {
+      "epoch": 0.23832092078537576,
+      "grad_norm": 1.0443451261742749,
+      "learning_rate": 9.780313951481904e-06,
+      "loss": 2.5914,
+      "step": 176
+    },
+    {
+      "epoch": 0.23967501692620177,
+      "grad_norm": 1.089642162382649,
+      "learning_rate": 9.77709570545437e-06,
+      "loss": 2.4722,
+      "step": 177
+    },
+    {
+      "epoch": 0.24102911306702776,
+      "grad_norm": 0.8415869612253108,
+      "learning_rate": 9.773854595299269e-06,
+      "loss": 2.259,
+      "step": 178
+    },
+    {
+      "epoch": 0.24238320920785375,
+      "grad_norm": 0.9995676515510443,
+      "learning_rate": 9.7705906365292e-06,
+      "loss": 2.4385,
+      "step": 179
+    },
+    {
+      "epoch": 0.24373730534867977,
+      "grad_norm": 0.8675225236988806,
+      "learning_rate": 9.767303844766118e-06,
+      "loss": 2.6306,
+      "step": 180
+    },
+    {
+      "epoch": 0.24509140148950576,
+      "grad_norm": 0.9866543693555193,
+      "learning_rate": 9.763994235741258e-06,
+      "loss": 2.721,
+      "step": 181
+    },
+    {
+      "epoch": 0.24644549763033174,
+      "grad_norm": 0.8263627006248617,
+      "learning_rate": 9.760661825295068e-06,
+      "loss": 2.4952,
+      "step": 182
+    },
+    {
+      "epoch": 0.24779959377115776,
+      "grad_norm": 0.9392502201652514,
+      "learning_rate": 9.757306629377122e-06,
+      "loss": 2.3426,
+      "step": 183
+    },
+    {
+      "epoch": 0.24915368991198375,
+      "grad_norm": 0.9008533634307534,
+      "learning_rate": 9.753928664046055e-06,
+      "loss": 2.237,
+      "step": 184
+    },
+    {
+      "epoch": 0.25050778605280977,
+      "grad_norm": 0.8810644743850296,
+      "learning_rate": 9.750527945469478e-06,
+      "loss": 2.1888,
+      "step": 185
+    },
+    {
+      "epoch": 0.25186188219363576,
+      "grad_norm": 1.0640477436832385,
+      "learning_rate": 9.747104489923907e-06,
+      "loss": 2.768,
+      "step": 186
+    },
+    {
+      "epoch": 0.25321597833446174,
+      "grad_norm": 1.1673458969812496,
+      "learning_rate": 9.743658313794676e-06,
+      "loss": 2.1024,
+      "step": 187
+    },
+    {
+      "epoch": 0.25457007447528773,
+      "grad_norm": 1.2908578737050618,
+      "learning_rate": 9.740189433575873e-06,
+      "loss": 2.7917,
+      "step": 188
+    },
+    {
+      "epoch": 0.2559241706161137,
+      "grad_norm": 0.8391729382821729,
+      "learning_rate": 9.736697865870241e-06,
+      "loss": 2.028,
+      "step": 189
+    },
+    {
+      "epoch": 0.25727826675693977,
+      "grad_norm": 0.8699715166345284,
+      "learning_rate": 9.733183627389117e-06,
+      "loss": 2.5548,
+      "step": 190
+    },
+    {
+      "epoch": 0.25863236289776576,
+      "grad_norm": 0.9116911866419519,
+      "learning_rate": 9.729646734952345e-06,
+      "loss": 2.4571,
+      "step": 191
+    },
+    {
+      "epoch": 0.25998645903859174,
+      "grad_norm": 1.299240651597945,
+      "learning_rate": 9.726087205488192e-06,
+      "loss": 2.5349,
+      "step": 192
+    },
+    {
+      "epoch": 0.26134055517941773,
+      "grad_norm": 0.743557436185181,
+      "learning_rate": 9.722505056033272e-06,
+      "loss": 2.2326,
+      "step": 193
+    },
+    {
+      "epoch": 0.2626946513202437,
+      "grad_norm": 0.8967055381021124,
+      "learning_rate": 9.718900303732465e-06,
+      "loss": 2.4696,
+      "step": 194
+    },
+    {
+      "epoch": 0.2640487474610697,
+      "grad_norm": 0.9220840880325898,
+      "learning_rate": 9.715272965838829e-06,
+      "loss": 2.9524,
+      "step": 195
+    },
+    {
+      "epoch": 0.26540284360189575,
+      "grad_norm": 0.9565379625707748,
+      "learning_rate": 9.711623059713522e-06,
+      "loss": 2.348,
+      "step": 196
+    },
+    {
+      "epoch": 0.26675693974272174,
+      "grad_norm": 0.8097073678685669,
+      "learning_rate": 9.70795060282572e-06,
+      "loss": 2.3975,
+      "step": 197
+    },
+    {
+      "epoch": 0.26811103588354773,
+      "grad_norm": 0.9063511706313765,
+      "learning_rate": 9.70425561275253e-06,
+      "loss": 1.8901,
+      "step": 198
+    },
+    {
+      "epoch": 0.2694651320243737,
+      "grad_norm": 1.077513556287683,
+      "learning_rate": 9.700538107178904e-06,
+      "loss": 2.6563,
+      "step": 199
+    },
+    {
+      "epoch": 0.2708192281651997,
+      "grad_norm": 0.9916331664645718,
+      "learning_rate": 9.696798103897567e-06,
+      "loss": 2.8377,
+      "step": 200
+    },
+    {
+      "epoch": 0.27217332430602575,
+      "grad_norm": 0.8205290104696272,
+      "learning_rate": 9.693035620808913e-06,
+      "loss": 2.0325,
+      "step": 201
+    },
+    {
+      "epoch": 0.27352742044685174,
+      "grad_norm": 1.1826064002287169,
+      "learning_rate": 9.689250675920932e-06,
+      "loss": 2.3692,
+      "step": 202
+    },
+    {
+      "epoch": 0.27488151658767773,
+      "grad_norm": 0.8204731687241611,
+      "learning_rate": 9.685443287349122e-06,
+      "loss": 2.1643,
+      "step": 203
+    },
+    {
+      "epoch": 0.2762356127285037,
+      "grad_norm": 0.8197753550893174,
+      "learning_rate": 9.6816134733164e-06,
+      "loss": 2.5373,
+      "step": 204
+    },
+    {
+      "epoch": 0.2775897088693297,
+      "grad_norm": 1.2751145677392224,
+      "learning_rate": 9.677761252153013e-06,
+      "loss": 2.9078,
+      "step": 205
+    },
+    {
+      "epoch": 0.2789438050101557,
+      "grad_norm": 1.1168584689300736,
+      "learning_rate": 9.67388664229646e-06,
+      "loss": 2.8702,
+      "step": 206
+    },
+    {
+      "epoch": 0.28029790115098174,
+      "grad_norm": 1.326117340861735,
+      "learning_rate": 9.669989662291391e-06,
+      "loss": 2.4586,
+      "step": 207
+    },
+    {
+      "epoch": 0.28165199729180773,
+      "grad_norm": 0.8654306997034001,
+      "learning_rate": 9.66607033078952e-06,
+      "loss": 2.2188,
+      "step": 208
+    },
+    {
+      "epoch": 0.2830060934326337,
+      "grad_norm": 0.9060281059984311,
+      "learning_rate": 9.662128666549552e-06,
+      "loss": 2.4796,
+      "step": 209
+    },
+    {
+      "epoch": 0.2843601895734597,
+      "grad_norm": 1.0187615360912865,
+      "learning_rate": 9.658164688437073e-06,
+      "loss": 2.3235,
+      "step": 210
+    },
+    {
+      "epoch": 0.2857142857142857,
+      "grad_norm": 0.9488215177960637,
+      "learning_rate": 9.654178415424463e-06,
+      "loss": 2.6938,
+      "step": 211
+    },
+    {
+      "epoch": 0.2870683818551117,
+      "grad_norm": 1.0938489378148304,
+      "learning_rate": 9.65016986659082e-06,
+      "loss": 2.0545,
+      "step": 212
+    },
+    {
+      "epoch": 0.28842247799593773,
+      "grad_norm": 2.7665169236406264,
+      "learning_rate": 9.646139061121852e-06,
+      "loss": 2.3696,
+      "step": 213
+    },
+    {
+      "epoch": 0.2897765741367637,
+      "grad_norm": 1.797717691337552,
+      "learning_rate": 9.642086018309798e-06,
+      "loss": 2.3102,
+      "step": 214
+    },
+    {
+      "epoch": 0.2911306702775897,
+      "grad_norm": 0.8164527845654876,
+      "learning_rate": 9.638010757553317e-06,
+      "loss": 2.4098,
+      "step": 215
+    },
+    {
+      "epoch": 0.2924847664184157,
+      "grad_norm": 1.2899587042696066,
+      "learning_rate": 9.63391329835742e-06,
+      "loss": 2.2318,
+      "step": 216
+    },
+    {
+      "epoch": 0.2938388625592417,
+      "grad_norm": 1.074459385295116,
+      "learning_rate": 9.62979366033336e-06,
+      "loss": 2.139,
+      "step": 217
+    },
+    {
+      "epoch": 0.29519295870006773,
+      "grad_norm": 1.1189933959983496,
+      "learning_rate": 9.625651863198538e-06,
+      "loss": 2.4854,
+      "step": 218
+    },
+    {
+      "epoch": 0.2965470548408937,
+      "grad_norm": 0.8519091467054853,
+      "learning_rate": 9.62148792677642e-06,
+      "loss": 2.3087,
+      "step": 219
+    },
+    {
+      "epoch": 0.2979011509817197,
+      "grad_norm": 0.8015738365902023,
+      "learning_rate": 9.617301870996432e-06,
+      "loss": 2.1197,
+      "step": 220
+    },
+    {
+      "epoch": 0.2992552471225457,
+      "grad_norm": 0.8518091208180533,
+      "learning_rate": 9.613093715893863e-06,
+      "loss": 2.7451,
+      "step": 221
+    },
+    {
+      "epoch": 0.3006093432633717,
+      "grad_norm": 0.8996345060862967,
+      "learning_rate": 9.608863481609784e-06,
+      "loss": 2.3018,
+      "step": 222
+    },
+    {
+      "epoch": 0.3019634394041977,
+      "grad_norm": 0.8891028358657084,
+      "learning_rate": 9.604611188390932e-06,
+      "loss": 2.6009,
+      "step": 223
+    },
+    {
+      "epoch": 0.3033175355450237,
+      "grad_norm": 0.7802164858157057,
+      "learning_rate": 9.600336856589622e-06,
+      "loss": 2.5882,
+      "step": 224
+    },
+    {
+      "epoch": 0.3046716316858497,
+      "grad_norm": 0.7990982765184432,
+      "learning_rate": 9.596040506663662e-06,
+      "loss": 2.3743,
+      "step": 225
+    },
+    {
+      "epoch": 0.3060257278266757,
+      "grad_norm": 1.022524178945592,
+      "learning_rate": 9.591722159176229e-06,
+      "loss": 2.8266,
+      "step": 226
+    },
+    {
+      "epoch": 0.3073798239675017,
+      "grad_norm": 0.9867120584703192,
+      "learning_rate": 9.587381834795792e-06,
+      "loss": 2.3584,
+      "step": 227
+    },
+    {
+      "epoch": 0.3087339201083277,
+      "grad_norm": 1.4753323206729818,
+      "learning_rate": 9.583019554296004e-06,
+      "loss": 2.0803,
+      "step": 228
+    },
+    {
+      "epoch": 0.31008801624915366,
+      "grad_norm": 1.789632341284706,
+      "learning_rate": 9.578635338555607e-06,
+      "loss": 2.1242,
+      "step": 229
+    },
+    {
+      "epoch": 0.3114421123899797,
+      "grad_norm": 0.8811955080251023,
+      "learning_rate": 9.574229208558322e-06,
+      "loss": 2.0273,
+      "step": 230
+    },
+    {
+      "epoch": 0.3127962085308057,
+      "grad_norm": 1.1507808612384305,
+      "learning_rate": 9.569801185392768e-06,
+      "loss": 1.7925,
+      "step": 231
+    },
+    {
+      "epoch": 0.3141503046716317,
+      "grad_norm": 0.9196147703328809,
+      "learning_rate": 9.565351290252339e-06,
+      "loss": 2.6128,
+      "step": 232
+    },
+    {
+      "epoch": 0.3155044008124577,
+      "grad_norm": 0.961361586988772,
+      "learning_rate": 9.560879544435114e-06,
+      "loss": 2.4787,
+      "step": 233
+    },
+    {
+      "epoch": 0.31685849695328366,
+      "grad_norm": 0.8613093158718477,
+      "learning_rate": 9.556385969343756e-06,
+      "loss": 2.158,
+      "step": 234
+    },
+    {
+      "epoch": 0.3182125930941097,
+      "grad_norm": 1.4391522005243542,
+      "learning_rate": 9.551870586485408e-06,
+      "loss": 2.4246,
+      "step": 235
+    },
+    {
+      "epoch": 0.3195666892349357,
+      "grad_norm": 1.2006343600943894,
+      "learning_rate": 9.547333417471589e-06,
+      "loss": 2.3493,
+      "step": 236
+    },
+    {
+      "epoch": 0.3209207853757617,
+      "grad_norm": 1.0080927833403484,
+      "learning_rate": 9.542774484018087e-06,
+      "loss": 2.8215,
+      "step": 237
+    },
+    {
+      "epoch": 0.3222748815165877,
+      "grad_norm": 0.8450008133514967,
+      "learning_rate": 9.538193807944864e-06,
+      "loss": 2.4001,
+      "step": 238
+    },
+    {
+      "epoch": 0.32362897765741366,
+      "grad_norm": 0.7976914162195807,
+      "learning_rate": 9.533591411175942e-06,
+      "loss": 2.1287,
+      "step": 239
+    },
+    {
+      "epoch": 0.32498307379823965,
+      "grad_norm": 0.9190646647609063,
+      "learning_rate": 9.528967315739308e-06,
+      "loss": 2.3067,
+      "step": 240
+    },
+    {
+      "epoch": 0.3263371699390657,
+      "grad_norm": 1.0380510321795802,
+      "learning_rate": 9.524321543766797e-06,
+      "loss": 2.7597,
+      "step": 241
+    },
+    {
+      "epoch": 0.3276912660798917,
+      "grad_norm": 0.8826085637159452,
+      "learning_rate": 9.519654117493996e-06,
+      "loss": 2.4152,
+      "step": 242
+    },
+    {
+      "epoch": 0.3290453622207177,
+      "grad_norm": 0.8084663287024824,
+      "learning_rate": 9.514965059260133e-06,
+      "loss": 2.1707,
+      "step": 243
+    },
+    {
+      "epoch": 0.33039945836154366,
+      "grad_norm": 2.295312303080644,
+      "learning_rate": 9.510254391507971e-06,
+      "loss": 2.4209,
+      "step": 244
+    },
+    {
+      "epoch": 0.33175355450236965,
+      "grad_norm": 1.025364157662343,
+      "learning_rate": 9.505522136783703e-06,
+      "loss": 2.3565,
+      "step": 245
+    },
+    {
+      "epoch": 0.33310765064319564,
+      "grad_norm": 1.0666622437173356,
+      "learning_rate": 9.500768317736832e-06,
+      "loss": 2.2188,
+      "step": 246
+    },
+    {
+      "epoch": 0.3344617467840217,
+      "grad_norm": 1.1850709868668183,
+      "learning_rate": 9.495992957120083e-06,
+      "loss": 2.3777,
+      "step": 247
+    },
+    {
+      "epoch": 0.3358158429248477,
+      "grad_norm": 0.873142498936687,
+      "learning_rate": 9.49119607778928e-06,
+      "loss": 1.8418,
+      "step": 248
+    },
+    {
+      "epoch": 0.33716993906567366,
+      "grad_norm": 0.8446038012686401,
+      "learning_rate": 9.486377702703237e-06,
+      "loss": 2.0972,
+      "step": 249
+    },
+    {
+      "epoch": 0.33852403520649965,
+      "grad_norm": 0.8356250880759191,
+      "learning_rate": 9.481537854923654e-06,
+      "loss": 2.1706,
+      "step": 250
+    },
+    {
+      "epoch": 0.33987813134732564,
+      "grad_norm": 1.7976623194509354,
+      "learning_rate": 9.476676557615002e-06,
+      "loss": 2.4971,
+      "step": 251
+    },
+    {
+      "epoch": 0.3412322274881517,
+      "grad_norm": 0.8019936660785039,
+      "learning_rate": 9.471793834044416e-06,
+      "loss": 2.6548,
+      "step": 252
+    },
+    {
+      "epoch": 0.3425863236289777,
+      "grad_norm": 0.8674056692586112,
+      "learning_rate": 9.466889707581575e-06,
+      "loss": 2.42,
+      "step": 253
+    },
+    {
+      "epoch": 0.34394041976980366,
+      "grad_norm": 0.9089186351830855,
+      "learning_rate": 9.461964201698604e-06,
+      "loss": 2.4349,
+      "step": 254
+    },
+    {
+      "epoch": 0.34529451591062965,
+      "grad_norm": 1.0982880519520153,
+      "learning_rate": 9.457017339969952e-06,
+      "loss": 2.6149,
+      "step": 255
+    },
+    {
+      "epoch": 0.34664861205145564,
+      "grad_norm": 0.8124889170645071,
+      "learning_rate": 9.452049146072278e-06,
+      "loss": 2.0518,
+      "step": 256
+    },
+    {
+      "epoch": 0.34800270819228163,
+      "grad_norm": 1.0328011672628001,
+      "learning_rate": 9.447059643784345e-06,
+      "loss": 1.8293,
+      "step": 257
+    },
+    {
+      "epoch": 0.3493568043331077,
+      "grad_norm": 0.8959245575010989,
+      "learning_rate": 9.442048856986899e-06,
+      "loss": 2.2905,
+      "step": 258
+    },
+    {
+      "epoch": 0.35071090047393366,
+      "grad_norm": 0.8715537744503126,
+      "learning_rate": 9.437016809662562e-06,
+      "loss": 2.5979,
+      "step": 259
+    },
+    {
+      "epoch": 0.35206499661475965,
+      "grad_norm": 0.8301936180619232,
+      "learning_rate": 9.431963525895709e-06,
+      "loss": 2.1647,
+      "step": 260
+    },
+    {
+      "epoch": 0.35341909275558564,
+      "grad_norm": 1.4578867670552098,
+      "learning_rate": 9.426889029872358e-06,
+      "loss": 2.3302,
+      "step": 261
+    },
+    {
+      "epoch": 0.35477318889641163,
+      "grad_norm": 1.218953962641191,
+      "learning_rate": 9.421793345880055e-06,
+      "loss": 2.1792,
+      "step": 262
+    },
+    {
+      "epoch": 0.3561272850372376,
+      "grad_norm": 1.054421263121385,
+      "learning_rate": 9.416676498307753e-06,
+      "loss": 2.4324,
+      "step": 263
+    },
+    {
+      "epoch": 0.35748138117806366,
+      "grad_norm": 1.0196674429556223,
+      "learning_rate": 9.4115385116457e-06,
+      "loss": 1.9509,
+      "step": 264
+    },
+    {
+      "epoch": 0.35883547731888965,
+      "grad_norm": 0.9596518686839914,
+      "learning_rate": 9.406379410485318e-06,
+      "loss": 2.1244,
+      "step": 265
+    },
+    {
+      "epoch": 0.36018957345971564,
+      "grad_norm": 0.8223441202319559,
+      "learning_rate": 9.401199219519088e-06,
+      "loss": 2.2646,
+      "step": 266
+    },
+    {
+      "epoch": 0.36154366960054163,
+      "grad_norm": 1.0821402171143544,
+      "learning_rate": 9.395997963540435e-06,
+      "loss": 2.2111,
+      "step": 267
+    },
+    {
+      "epoch": 0.3628977657413676,
+      "grad_norm": 0.8651589636687357,
+      "learning_rate": 9.390775667443602e-06,
+      "loss": 1.9122,
+      "step": 268
+    },
+    {
+      "epoch": 0.36425186188219366,
+      "grad_norm": 0.8620843604100115,
+      "learning_rate": 9.385532356223527e-06,
+      "loss": 2.342,
+      "step": 269
+    },
+    {
+      "epoch": 0.36560595802301965,
+      "grad_norm": 0.9397786133798538,
+      "learning_rate": 9.380268054975745e-06,
+      "loss": 2.779,
+      "step": 270
+    },
+    {
+      "epoch": 0.36696005416384564,
+      "grad_norm": 1.1073908774978374,
+      "learning_rate": 9.374982788896241e-06,
+      "loss": 2.6926,
+      "step": 271
+    },
+    {
+      "epoch": 0.36831415030467163,
+      "grad_norm": 1.04556124980846,
+      "learning_rate": 9.36967658328135e-06,
+      "loss": 2.6823,
+      "step": 272
+    },
+    {
+      "epoch": 0.3696682464454976,
+      "grad_norm": 5.754521524944833,
+      "learning_rate": 9.364349463527625e-06,
+      "loss": 2.132,
+      "step": 273
+    },
+    {
+      "epoch": 0.3710223425863236,
+      "grad_norm": 1.0093248053929325,
+      "learning_rate": 9.359001455131713e-06,
+      "loss": 2.5921,
+      "step": 274
+    },
+    {
+      "epoch": 0.37237643872714965,
+      "grad_norm": 1.0149649889598005,
+      "learning_rate": 9.353632583690249e-06,
+      "loss": 2.0404,
+      "step": 275
+    },
+    {
+      "epoch": 0.37373053486797564,
+      "grad_norm": 0.7933286967641568,
+      "learning_rate": 9.34824287489971e-06,
+      "loss": 2.349,
+      "step": 276
+    },
+    {
+      "epoch": 0.37508463100880163,
+      "grad_norm": 1.4731819170091527,
+      "learning_rate": 9.342832354556319e-06,
+      "loss": 2.1622,
+      "step": 277
+    },
+    {
+      "epoch": 0.3764387271496276,
+      "grad_norm": 0.8780445261970546,
+      "learning_rate": 9.337401048555892e-06,
+      "loss": 2.2959,
+      "step": 278
+    },
+    {
+      "epoch": 0.3777928232904536,
+      "grad_norm": 0.9934701059856305,
+      "learning_rate": 9.331948982893741e-06,
+      "loss": 2.2491,
+      "step": 279
+    },
+    {
+      "epoch": 0.3791469194312796,
+      "grad_norm": 1.1682396086558167,
+      "learning_rate": 9.326476183664535e-06,
+      "loss": 2.4071,
+      "step": 280
+    },
+    {
+      "epoch": 0.38050101557210564,
+      "grad_norm": 1.1443828908579956,
+      "learning_rate": 9.320982677062173e-06,
+      "loss": 2.4467,
+      "step": 281
+    },
+    {
+      "epoch": 0.3818551117129316,
+      "grad_norm": 0.9041072793605435,
+      "learning_rate": 9.315468489379668e-06,
+      "loss": 2.2458,
+      "step": 282
+    },
+    {
+      "epoch": 0.3832092078537576,
+      "grad_norm": 0.9075478380778889,
+      "learning_rate": 9.309933647009018e-06,
+      "loss": 1.8672,
+      "step": 283
+    },
+    {
+      "epoch": 0.3845633039945836,
+      "grad_norm": 0.9178385682224962,
+      "learning_rate": 9.304378176441076e-06,
+      "loss": 2.3703,
+      "step": 284
+    },
+    {
+      "epoch": 0.3859174001354096,
+      "grad_norm": 2.075191172215327,
+      "learning_rate": 9.298802104265423e-06,
+      "loss": 2.533,
+      "step": 285
+    },
+    {
+      "epoch": 0.38727149627623564,
+      "grad_norm": 0.924446368315556,
+      "learning_rate": 9.29320545717025e-06,
+      "loss": 2.0906,
+      "step": 286
+    },
+    {
+      "epoch": 0.3886255924170616,
+      "grad_norm": 0.8276545894642513,
+      "learning_rate": 9.287588261942216e-06,
+      "loss": 2.2143,
+      "step": 287
+    },
+    {
+      "epoch": 0.3899796885578876,
+      "grad_norm": 1.03799747305039,
+      "learning_rate": 9.281950545466336e-06,
+      "loss": 2.3721,
+      "step": 288
+    },
+    {
+      "epoch": 0.3913337846987136,
+      "grad_norm": 1.0317596432913245,
+      "learning_rate": 9.276292334725838e-06,
+      "loss": 1.9487,
+      "step": 289
+    },
+    {
+      "epoch": 0.3926878808395396,
+      "grad_norm": 0.8956343315407433,
+      "learning_rate": 9.27061365680204e-06,
+      "loss": 2.2824,
+      "step": 290
+    },
+    {
+      "epoch": 0.3940419769803656,
+      "grad_norm": 0.9035145142320139,
+      "learning_rate": 9.264914538874221e-06,
+      "loss": 2.3565,
+      "step": 291
+    },
+    {
+      "epoch": 0.3953960731211916,
+      "grad_norm": 1.1869989906886842,
+      "learning_rate": 9.25919500821949e-06,
+      "loss": 2.5587,
+      "step": 292
+    },
+    {
+      "epoch": 0.3967501692620176,
+      "grad_norm": 0.8987379863372337,
+      "learning_rate": 9.253455092212656e-06,
+      "loss": 2.2668,
+      "step": 293
+    },
+    {
+      "epoch": 0.3981042654028436,
+      "grad_norm": 0.8756739379853996,
+      "learning_rate": 9.247694818326092e-06,
+      "loss": 2.123,
+      "step": 294
+    },
+    {
+      "epoch": 0.3994583615436696,
+      "grad_norm": 0.9540331255503948,
+      "learning_rate": 9.241914214129618e-06,
+      "loss": 2.6226,
+      "step": 295
+    },
+    {
+      "epoch": 0.4008124576844956,
+      "grad_norm": 1.102045569540978,
+      "learning_rate": 9.236113307290345e-06,
+      "loss": 2.5496,
+      "step": 296
+    },
+    {
+      "epoch": 0.40216655382532157,
+      "grad_norm": 0.9598540264304664,
+      "learning_rate": 9.230292125572569e-06,
+      "loss": 2.3411,
+      "step": 297
+    },
+    {
+      "epoch": 0.4035206499661476,
+      "grad_norm": 1.1343281097624651,
+      "learning_rate": 9.224450696837617e-06,
+      "loss": 2.5745,
+      "step": 298
+    },
+    {
+      "epoch": 0.4048747461069736,
+      "grad_norm": 1.1066468488776415,
+      "learning_rate": 9.218589049043726e-06,
+      "loss": 2.5049,
+      "step": 299
+    },
+    {
+      "epoch": 0.4062288422477996,
+      "grad_norm": 0.9154064589009782,
+      "learning_rate": 9.212707210245908e-06,
+      "loss": 2.0704,
+      "step": 300
+    },
+    {
+      "epoch": 0.4075829383886256,
+      "grad_norm": 1.1554701199935409,
+      "learning_rate": 9.206805208595805e-06,
+      "loss": 2.1283,
+      "step": 301
+    },
+    {
+      "epoch": 0.40893703452945157,
+      "grad_norm": 2.566380710482718,
+      "learning_rate": 9.200883072341573e-06,
+      "loss": 2.3335,
+      "step": 302
+    },
+    {
+      "epoch": 0.4102911306702776,
+      "grad_norm": 1.0638420692642232,
+      "learning_rate": 9.194940829827728e-06,
+      "loss": 2.0887,
+      "step": 303
+    },
+    {
+      "epoch": 0.4116452268111036,
+      "grad_norm": 1.026901985546702,
+      "learning_rate": 9.188978509495022e-06,
+      "loss": 2.358,
+      "step": 304
+    },
+    {
+      "epoch": 0.4129993229519296,
+      "grad_norm": 1.0271668374927614,
+      "learning_rate": 9.1829961398803e-06,
+      "loss": 2.1089,
+      "step": 305
+    },
+    {
+      "epoch": 0.4143534190927556,
+      "grad_norm": 0.9951227785408013,
+      "learning_rate": 9.176993749616374e-06,
+      "loss": 2.3089,
+      "step": 306
+    },
+    {
+      "epoch": 0.41570751523358157,
+      "grad_norm": 0.9365004671579732,
+      "learning_rate": 9.17097136743187e-06,
+      "loss": 2.2569,
+      "step": 307
+    },
+    {
+      "epoch": 0.41706161137440756,
+      "grad_norm": 0.9319443863400425,
+      "learning_rate": 9.164929022151106e-06,
+      "loss": 2.3883,
+      "step": 308
+    },
+    {
+      "epoch": 0.4184157075152336,
+      "grad_norm": 0.8611859165777872,
+      "learning_rate": 9.158866742693947e-06,
+      "loss": 2.8113,
+      "step": 309
+    },
+    {
+      "epoch": 0.4197698036560596,
+      "grad_norm": 0.9865030843425563,
+      "learning_rate": 9.15278455807566e-06,
+      "loss": 2.3777,
+      "step": 310
+    },
+    {
+      "epoch": 0.4211238997968856,
+      "grad_norm": 0.9592735653322616,
+      "learning_rate": 9.146682497406792e-06,
+      "loss": 2.2102,
+      "step": 311
+    },
+    {
+      "epoch": 0.42247799593771157,
+      "grad_norm": 1.1176241826216586,
+      "learning_rate": 9.140560589893012e-06,
+      "loss": 2.5398,
+      "step": 312
+    },
+    {
+      "epoch": 0.42383209207853756,
+      "grad_norm": 0.8322094005243622,
+      "learning_rate": 9.134418864834986e-06,
+      "loss": 2.0246,
+      "step": 313
+    },
+    {
+      "epoch": 0.42518618821936355,
+      "grad_norm": 1.0231281197044073,
+      "learning_rate": 9.128257351628224e-06,
+      "loss": 2.0589,
+      "step": 314
+    },
+    {
+      "epoch": 0.4265402843601896,
+      "grad_norm": 0.800090080992752,
+      "learning_rate": 9.122076079762958e-06,
+      "loss": 1.9548,
+      "step": 315
+    },
+    {
+      "epoch": 0.4278943805010156,
+      "grad_norm": 1.1046831999194016,
+      "learning_rate": 9.115875078823975e-06,
+      "loss": 2.1922,
+      "step": 316
+    },
+    {
+      "epoch": 0.42924847664184157,
+      "grad_norm": 1.1057141700908926,
+      "learning_rate": 9.109654378490495e-06,
+      "loss": 2.3618,
+      "step": 317
+    },
+    {
+      "epoch": 0.43060257278266756,
+      "grad_norm": 0.9676545844336863,
+      "learning_rate": 9.103414008536029e-06,
+      "loss": 2.4645,
+      "step": 318
+    },
+    {
+      "epoch": 0.43195666892349355,
+      "grad_norm": 0.865022307271061,
+      "learning_rate": 9.09715399882822e-06,
+      "loss": 2.2775,
+      "step": 319
+    },
+    {
+      "epoch": 0.4333107650643196,
+      "grad_norm": 0.8835769231615183,
+      "learning_rate": 9.09087437932872e-06,
+      "loss": 2.5524,
+      "step": 320
+    },
+    {
+      "epoch": 0.4346648612051456,
+      "grad_norm": 0.9432892345765396,
+      "learning_rate": 9.084575180093031e-06,
+      "loss": 2.5083,
+      "step": 321
+    },
+    {
+      "epoch": 0.43601895734597157,
+      "grad_norm": 0.931907462171622,
+      "learning_rate": 9.07825643127037e-06,
+      "loss": 2.4523,
+      "step": 322
+    },
+    {
+      "epoch": 0.43737305348679756,
+      "grad_norm": 0.9405696755624474,
+      "learning_rate": 9.071918163103524e-06,
+      "loss": 2.4703,
+      "step": 323
+    },
+    {
+      "epoch": 0.43872714962762355,
+      "grad_norm": 0.962612723073031,
+      "learning_rate": 9.065560405928699e-06,
+      "loss": 2.847,
+      "step": 324
+    },
+    {
+      "epoch": 0.44008124576844954,
+      "grad_norm": 1.0577696194404815,
+      "learning_rate": 9.059183190175386e-06,
+      "loss": 2.4337,
+      "step": 325
+    },
+    {
+      "epoch": 0.4414353419092756,
+      "grad_norm": 1.4201981580037544,
+      "learning_rate": 9.0527865463662e-06,
+      "loss": 2.4028,
+      "step": 326
+    },
+    {
+      "epoch": 0.44278943805010157,
+      "grad_norm": 0.8701075062948284,
+      "learning_rate": 9.046370505116748e-06,
+      "loss": 2.2132,
+      "step": 327
+    },
+    {
+      "epoch": 0.44414353419092756,
+      "grad_norm": 1.1362568277712843,
+      "learning_rate": 9.039935097135479e-06,
+      "loss": 2.4098,
+      "step": 328
+    },
+    {
+      "epoch": 0.44549763033175355,
+      "grad_norm": 1.0852398132134922,
+      "learning_rate": 9.03348035322353e-06,
+      "loss": 2.136,
+      "step": 329
+    },
+    {
+      "epoch": 0.44685172647257954,
+      "grad_norm": 0.9806916801248307,
+      "learning_rate": 9.027006304274584e-06,
+      "loss": 2.7784,
+      "step": 330
+    },
+    {
+      "epoch": 0.4482058226134055,
+      "grad_norm": 0.9341321077617596,
+      "learning_rate": 9.020512981274726e-06,
+      "loss": 1.8668,
+      "step": 331
+    },
+    {
+      "epoch": 0.44955991875423157,
+      "grad_norm": 1.1941574187846893,
+      "learning_rate": 9.014000415302286e-06,
+      "loss": 2.7557,
+      "step": 332
+    },
+    {
+      "epoch": 0.45091401489505756,
+      "grad_norm": 1.0291782053412863,
+      "learning_rate": 9.007468637527697e-06,
+      "loss": 2.5349,
+      "step": 333
+    },
+    {
+      "epoch": 0.45226811103588355,
+      "grad_norm": 0.9440659662613932,
+      "learning_rate": 9.000917679213344e-06,
+      "loss": 2.2004,
+      "step": 334
+    },
+    {
+      "epoch": 0.45362220717670954,
+      "grad_norm": 0.883811084632338,
+      "learning_rate": 8.99434757171341e-06,
+      "loss": 1.8828,
+      "step": 335
+    },
+    {
+      "epoch": 0.4549763033175355,
+      "grad_norm": 0.8690517488922695,
+      "learning_rate": 8.987758346473739e-06,
+      "loss": 2.5318,
+      "step": 336
+    },
+    {
+      "epoch": 0.45633039945836157,
+      "grad_norm": 0.9009315334756164,
+      "learning_rate": 8.981150035031663e-06,
+      "loss": 2.1954,
+      "step": 337
+    },
+    {
+      "epoch": 0.45768449559918756,
+      "grad_norm": 0.8858212672786216,
+      "learning_rate": 8.974522669015872e-06,
+      "loss": 2.2137,
+      "step": 338
+    },
+    {
+      "epoch": 0.45903859174001355,
+      "grad_norm": 0.8328355239198729,
+      "learning_rate": 8.96787628014626e-06,
+      "loss": 2.3718,
+      "step": 339
+    },
+    {
+      "epoch": 0.46039268788083954,
+      "grad_norm": 0.999826123063525,
+      "learning_rate": 8.961210900233757e-06,
+      "loss": 2.6404,
+      "step": 340
+    },
+    {
+      "epoch": 0.4617467840216655,
+      "grad_norm": 0.8192874415252067,
+      "learning_rate": 8.954526561180195e-06,
+      "loss": 2.2227,
+      "step": 341
+    },
+    {
+      "epoch": 0.4631008801624915,
+      "grad_norm": 0.8800622522324091,
+      "learning_rate": 8.947823294978147e-06,
+      "loss": 2.3034,
+      "step": 342
+    },
+    {
+      "epoch": 0.46445497630331756,
+      "grad_norm": 0.9239289219678168,
+      "learning_rate": 8.941101133710772e-06,
+      "loss": 2.3034,
+      "step": 343
+    },
+    {
+      "epoch": 0.46580907244414355,
+      "grad_norm": 1.0400589249817949,
+      "learning_rate": 8.934360109551671e-06,
+      "loss": 2.355,
+      "step": 344
+    },
+    {
+      "epoch": 0.46716316858496953,
+      "grad_norm": 0.9418406311279763,
+      "learning_rate": 8.927600254764721e-06,
+      "loss": 2.1249,
+      "step": 345
+    },
+    {
+      "epoch": 0.4685172647257955,
+      "grad_norm": 1.286887601031447,
+      "learning_rate": 8.920821601703927e-06,
+      "loss": 2.3553,
+      "step": 346
+    },
+    {
+      "epoch": 0.4698713608666215,
+      "grad_norm": 0.9461254286423676,
+      "learning_rate": 8.914024182813268e-06,
+      "loss": 2.882,
+      "step": 347
+    },
+    {
+      "epoch": 0.4712254570074475,
+      "grad_norm": 0.9709976554539759,
+      "learning_rate": 8.907208030626538e-06,
+      "loss": 2.3406,
+      "step": 348
+    },
+    {
+      "epoch": 0.47257955314827355,
+      "grad_norm": 1.0768202606705715,
+      "learning_rate": 8.900373177767196e-06,
+      "loss": 2.1405,
+      "step": 349
+    },
+    {
+      "epoch": 0.47393364928909953,
+      "grad_norm": 1.0960778664068742,
+      "learning_rate": 8.8935196569482e-06,
+      "loss": 2.1959,
+      "step": 350
+    },
+    {
+      "epoch": 0.4752877454299255,
+      "grad_norm": 0.9024940161441628,
+      "learning_rate": 8.886647500971862e-06,
+      "loss": 1.9828,
+      "step": 351
+    },
+    {
+      "epoch": 0.4766418415707515,
+      "grad_norm": 0.8351151637681756,
+      "learning_rate": 8.879756742729683e-06,
+      "loss": 2.4813,
+      "step": 352
+    },
+    {
+      "epoch": 0.4779959377115775,
+      "grad_norm": 1.0637419955818261,
+      "learning_rate": 8.8728474152022e-06,
+      "loss": 2.7144,
+      "step": 353
+    },
+    {
+      "epoch": 0.47935003385240355,
+      "grad_norm": 0.9971260587925014,
+      "learning_rate": 8.865919551458823e-06,
+      "loss": 2.4754,
+      "step": 354
+    },
+    {
+      "epoch": 0.48070412999322953,
+      "grad_norm": 0.9455478449985035,
+      "learning_rate": 8.858973184657682e-06,
+      "loss": 2.0941,
+      "step": 355
+    },
+    {
+      "epoch": 0.4820582261340555,
+      "grad_norm": 1.0025865640257856,
+      "learning_rate": 8.852008348045468e-06,
+      "loss": 2.1712,
+      "step": 356
+    },
+    {
+      "epoch": 0.4834123222748815,
+      "grad_norm": 1.0847129459559228,
+      "learning_rate": 8.84502507495727e-06,
+      "loss": 2.2485,
+      "step": 357
+    },
+    {
+      "epoch": 0.4847664184157075,
+      "grad_norm": 0.9727542785581984,
+      "learning_rate": 8.838023398816417e-06,
+      "loss": 2.2295,
+      "step": 358
+    },
+    {
+      "epoch": 0.4861205145565335,
+      "grad_norm": 1.0006473633765713,
+      "learning_rate": 8.831003353134321e-06,
+      "loss": 2.4642,
+      "step": 359
+    },
+    {
+      "epoch": 0.48747461069735953,
+      "grad_norm": 0.9316788259717722,
+      "learning_rate": 8.823964971510313e-06,
+      "loss": 2.7107,
+      "step": 360
+    },
+    {
+      "epoch": 0.4888287068381855,
+      "grad_norm": 0.9186715038840205,
+      "learning_rate": 8.816908287631483e-06,
+      "loss": 2.374,
+      "step": 361
+    },
+    {
+      "epoch": 0.4901828029790115,
+      "grad_norm": 1.145278113777804,
+      "learning_rate": 8.809833335272517e-06,
+      "loss": 2.616,
+      "step": 362
+    },
+    {
+      "epoch": 0.4915368991198375,
+      "grad_norm": 1.0240670602436293,
+      "learning_rate": 8.802740148295546e-06,
+      "loss": 2.4254,
+      "step": 363
+    },
+    {
+      "epoch": 0.4928909952606635,
+      "grad_norm": 1.167577693850126,
+      "learning_rate": 8.795628760649965e-06,
+      "loss": 2.3852,
+      "step": 364
+    },
+    {
+      "epoch": 0.4942450914014895,
+      "grad_norm": 1.0151735496069052,
+      "learning_rate": 8.78849920637229e-06,
+      "loss": 2.262,
+      "step": 365
+    },
+    {
+      "epoch": 0.4955991875423155,
+      "grad_norm": 20.48704343249556,
+      "learning_rate": 8.781351519585978e-06,
+      "loss": 2.8307,
+      "step": 366
+    },
+    {
+      "epoch": 0.4969532836831415,
+      "grad_norm": 1.0123187878318054,
+      "learning_rate": 8.774185734501277e-06,
+      "loss": 1.9856,
+      "step": 367
+    },
+    {
+      "epoch": 0.4983073798239675,
+      "grad_norm": 1.1847404703515332,
+      "learning_rate": 8.767001885415055e-06,
+      "loss": 2.7989,
+      "step": 368
+    },
+    {
+      "epoch": 0.4996614759647935,
+      "grad_norm": 1.197682332560284,
+      "learning_rate": 8.759800006710641e-06,
+      "loss": 2.3791,
+      "step": 369
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 1476,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 2,
+  "save_steps": 369,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 9.74929356914688e+16,
+  "train_batch_size": 1,
+  "trial_name": null,
+  "trial_params": null
+}

checkpoint-369/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4bff29e7ff321b15e984196b58ce5d2fc8b550fddfbffd9146e631086009d7b1
+size 7992

checkpoint-369/zero_to_fp32.py

@@ -0,0 +1,604 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example: python zero_to_fp32.py . pytorch_model.bin
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+
+    total_files = len(files)
+    state_dicts = []
+    for f in files:
+        state_dict = torch.load(f, map_location=device)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    if zero_stage <= 2:
+        fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    elif zero_stage == 3:
+        # if there is more than one param group, there will be multiple flattened tensors - one
+        # flattened tensor per group - for simplicity merge them into a single tensor
+        #
+        # XXX: could make the script more memory efficient for when there are multiple groups - it
+        # will require matching the sub-lists of param_shapes for each param group flattened tensor
+
+        fp32_flat_groups = [
+            torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
+        ]
+
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    for name, shape in param_shapes.items():
+
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # XXX: memory usage doubles here
+        state_dict[name] = torch.cat(
+            tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
+            0).narrow(0, 0, unpartitioned_numel).view(shape)
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None, exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+
+    Returns:
+        - pytorch ``state_dict``
+
+    Note: this approach may not work if your application doesn't have sufficient free CPU memory and
+    you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None, exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag, exclude_frozen_parameters)
+    print(f"Saving fp32 state dict to {output_file}")
+    torch.save(state_dict, output_file)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument(
+        "output_file",
+        type=str,
+        help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_file,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)