diff --git a/checkpoint-100/README.md b/checkpoint-100/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..e7d4f706aeefc6bb2aa84e963870f345a14c4af9
--- /dev/null
+++ b/checkpoint-100/README.md
@@ -0,0 +1,202 @@
+---
+library_name: peft
+base_model: ../ckpts/Meta-Llama-3-8B-Instruct
+---
+
+# 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 0.11.1
\ No newline at end of file
diff --git a/checkpoint-100/adapter_config.json b/checkpoint-100/adapter_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..ad16c073ed8946d227350e8fe4e99d689a8ede54
--- /dev/null
+++ b/checkpoint-100/adapter_config.json
@@ -0,0 +1,35 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "../ckpts/Meta-Llama-3-8B-Instruct",
+  "bias": "none",
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 16,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "down_proj",
+    "gate_proj",
+    "q_proj",
+    "lm_head",
+    "o_proj",
+    "v_proj",
+    "up_proj",
+    "k_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "use_dora": false,
+  "use_rslora": false
+}
\ No newline at end of file
diff --git a/checkpoint-100/adapter_model.safetensors b/checkpoint-100/adapter_model.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..1a2473ee58d5209b28e8a00f8e330e31ce259574
--- /dev/null
+++ b/checkpoint-100/adapter_model.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:baca06cc42b3510c3c3bf848f73f79b72b1baea5a4f963941e134e37b3ab777d
+size 1138856856
diff --git a/checkpoint-100/global_step100/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt b/checkpoint-100/global_step100/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..b5c76794e67be718271ed6e69c8658a4f1f9bdb1
--- /dev/null
+++ b/checkpoint-100/global_step100/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:808ab758e2efff532198a64b20a7f9c499ac3216290ea2598de18360f006e24c
+size 528781328
diff --git a/checkpoint-100/global_step100/mp_rank_00_model_states.pt b/checkpoint-100/global_step100/mp_rank_00_model_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..bad4b0b42fdf8faa0e3313ccca8475b6cba6a7be
--- /dev/null
+++ b/checkpoint-100/global_step100/mp_rank_00_model_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cba7f894e1d634e3bd3acbb30c43a7817dcd9bbc30549de0142cf5b4970bdd25
+size 199905337
diff --git a/checkpoint-100/latest b/checkpoint-100/latest
new file mode 100644
index 0000000000000000000000000000000000000000..744ae7dbad571b6f37ec6c7066549494261bb59e
--- /dev/null
+++ b/checkpoint-100/latest
@@ -0,0 +1 @@
+global_step100
\ No newline at end of file
diff --git a/checkpoint-100/rng_state.pth b/checkpoint-100/rng_state.pth
new file mode 100644
index 0000000000000000000000000000000000000000..7bf5862c14e4daf3b373f7fc1969da8616b4035b
--- /dev/null
+++ b/checkpoint-100/rng_state.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1a2d9af3fcfc54434e74034cabf1cd700dd64663d32e5f7d97b79b71b2284409
+size 14244
diff --git a/checkpoint-100/trainer_state.json b/checkpoint-100/trainer_state.json
new file mode 100644
index 0000000000000000000000000000000000000000..06634638d4c911f70bea701ef4ca483d8d376221
--- /dev/null
+++ b/checkpoint-100/trainer_state.json
@@ -0,0 +1,733 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.8533333333333334,
+  "eval_steps": 500,
+  "global_step": 100,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.008533333333333334,
+      "grad_norm": 100.21848203113535,
+      "learning_rate": 0.0,
+      "loss": 7.1962,
+      "step": 1
+    },
+    {
+      "epoch": 0.017066666666666667,
+      "grad_norm": 96.43006188910957,
+      "learning_rate": 3.0102999566398115e-06,
+      "loss": 6.9414,
+      "step": 2
+    },
+    {
+      "epoch": 0.0256,
+      "grad_norm": 97.35803466618715,
+      "learning_rate": 4.771212547196624e-06,
+      "loss": 7.0102,
+      "step": 3
+    },
+    {
+      "epoch": 0.034133333333333335,
+      "grad_norm": 95.14837816372646,
+      "learning_rate": 6.020599913279623e-06,
+      "loss": 6.5295,
+      "step": 4
+    },
+    {
+      "epoch": 0.042666666666666665,
+      "grad_norm": 91.76544275692784,
+      "learning_rate": 6.989700043360187e-06,
+      "loss": 6.4806,
+      "step": 5
+    },
+    {
+      "epoch": 0.0512,
+      "grad_norm": 84.4494318688335,
+      "learning_rate": 7.781512503836437e-06,
+      "loss": 6.4194,
+      "step": 6
+    },
+    {
+      "epoch": 0.05973333333333333,
+      "grad_norm": 71.37977490595638,
+      "learning_rate": 8.450980400142568e-06,
+      "loss": 5.4953,
+      "step": 7
+    },
+    {
+      "epoch": 0.06826666666666667,
+      "grad_norm": 49.31153456754566,
+      "learning_rate": 9.030899869919434e-06,
+      "loss": 5.4123,
+      "step": 8
+    },
+    {
+      "epoch": 0.0768,
+      "grad_norm": 20.37296364560341,
+      "learning_rate": 9.542425094393249e-06,
+      "loss": 5.2334,
+      "step": 9
+    },
+    {
+      "epoch": 0.08533333333333333,
+      "grad_norm": 7.968467079076881,
+      "learning_rate": 9.999999999999999e-06,
+      "loss": 5.0282,
+      "step": 10
+    },
+    {
+      "epoch": 0.09386666666666667,
+      "grad_norm": 3.559446532055649,
+      "learning_rate": 1.041392685158225e-05,
+      "loss": 4.612,
+      "step": 11
+    },
+    {
+      "epoch": 0.1024,
+      "grad_norm": 3.5528846947995674,
+      "learning_rate": 1.0791812460476248e-05,
+      "loss": 4.9475,
+      "step": 12
+    },
+    {
+      "epoch": 0.11093333333333333,
+      "grad_norm": 3.541968897471334,
+      "learning_rate": 1.1139433523068365e-05,
+      "loss": 4.2777,
+      "step": 13
+    },
+    {
+      "epoch": 0.11946666666666667,
+      "grad_norm": 3.54718070036198,
+      "learning_rate": 1.1461280356782378e-05,
+      "loss": 4.3507,
+      "step": 14
+    },
+    {
+      "epoch": 0.128,
+      "grad_norm": 3.8632334830606747,
+      "learning_rate": 1.1760912590556813e-05,
+      "loss": 4.5364,
+      "step": 15
+    },
+    {
+      "epoch": 0.13653333333333334,
+      "grad_norm": 3.6637424744054004,
+      "learning_rate": 1.2041199826559246e-05,
+      "loss": 3.9672,
+      "step": 16
+    },
+    {
+      "epoch": 0.14506666666666668,
+      "grad_norm": 3.919802904818311,
+      "learning_rate": 1.230448921378274e-05,
+      "loss": 4.0618,
+      "step": 17
+    },
+    {
+      "epoch": 0.1536,
+      "grad_norm": 4.71904950738746,
+      "learning_rate": 1.2552725051033058e-05,
+      "loss": 4.6656,
+      "step": 18
+    },
+    {
+      "epoch": 0.16213333333333332,
+      "grad_norm": 4.6656317698690835,
+      "learning_rate": 1.2787536009528288e-05,
+      "loss": 4.1131,
+      "step": 19
+    },
+    {
+      "epoch": 0.17066666666666666,
+      "grad_norm": 5.145138692367417,
+      "learning_rate": 1.301029995663981e-05,
+      "loss": 4.0989,
+      "step": 20
+    },
+    {
+      "epoch": 0.1792,
+      "grad_norm": 5.923538014759818,
+      "learning_rate": 1.3222192947339192e-05,
+      "loss": 4.4991,
+      "step": 21
+    },
+    {
+      "epoch": 0.18773333333333334,
+      "grad_norm": 5.941056962941364,
+      "learning_rate": 1.3424226808222062e-05,
+      "loss": 4.0836,
+      "step": 22
+    },
+    {
+      "epoch": 0.19626666666666667,
+      "grad_norm": 6.171026012117947,
+      "learning_rate": 1.3617278360175927e-05,
+      "loss": 3.6861,
+      "step": 23
+    },
+    {
+      "epoch": 0.2048,
+      "grad_norm": 7.130542138930838,
+      "learning_rate": 1.380211241711606e-05,
+      "loss": 4.0958,
+      "step": 24
+    },
+    {
+      "epoch": 0.21333333333333335,
+      "grad_norm": 7.328837606110418,
+      "learning_rate": 1.3979400086720374e-05,
+      "loss": 3.9524,
+      "step": 25
+    },
+    {
+      "epoch": 0.22186666666666666,
+      "grad_norm": 6.923489005711429,
+      "learning_rate": 1.4149733479708178e-05,
+      "loss": 3.6062,
+      "step": 26
+    },
+    {
+      "epoch": 0.2304,
+      "grad_norm": 6.409498926059221,
+      "learning_rate": 1.4313637641589872e-05,
+      "loss": 3.2034,
+      "step": 27
+    },
+    {
+      "epoch": 0.23893333333333333,
+      "grad_norm": 5.781628405584682,
+      "learning_rate": 1.4471580313422191e-05,
+      "loss": 2.8158,
+      "step": 28
+    },
+    {
+      "epoch": 0.24746666666666667,
+      "grad_norm": 6.2927590068243315,
+      "learning_rate": 1.4623979978989559e-05,
+      "loss": 2.9803,
+      "step": 29
+    },
+    {
+      "epoch": 0.256,
+      "grad_norm": 6.103844678630006,
+      "learning_rate": 1.4771212547196623e-05,
+      "loss": 2.847,
+      "step": 30
+    },
+    {
+      "epoch": 0.26453333333333334,
+      "grad_norm": 7.656341925867144,
+      "learning_rate": 1.4913616938342726e-05,
+      "loss": 3.0907,
+      "step": 31
+    },
+    {
+      "epoch": 0.2730666666666667,
+      "grad_norm": 6.324242877501844,
+      "learning_rate": 1.5051499783199059e-05,
+      "loss": 2.3467,
+      "step": 32
+    },
+    {
+      "epoch": 0.2816,
+      "grad_norm": 7.606313101162739,
+      "learning_rate": 1.5185139398778874e-05,
+      "loss": 2.5292,
+      "step": 33
+    },
+    {
+      "epoch": 0.29013333333333335,
+      "grad_norm": 8.553792493849265,
+      "learning_rate": 1.531478917042255e-05,
+      "loss": 2.4547,
+      "step": 34
+    },
+    {
+      "epoch": 0.2986666666666667,
+      "grad_norm": 8.483368703272543,
+      "learning_rate": 1.5440680443502753e-05,
+      "loss": 2.1956,
+      "step": 35
+    },
+    {
+      "epoch": 0.3072,
+      "grad_norm": 8.364739352838077,
+      "learning_rate": 1.5563025007672873e-05,
+      "loss": 1.8552,
+      "step": 36
+    },
+    {
+      "epoch": 0.3157333333333333,
+      "grad_norm": 9.37663682000104,
+      "learning_rate": 1.5682017240669948e-05,
+      "loss": 1.9228,
+      "step": 37
+    },
+    {
+      "epoch": 0.32426666666666665,
+      "grad_norm": 9.823047193440066,
+      "learning_rate": 1.57978359661681e-05,
+      "loss": 1.7033,
+      "step": 38
+    },
+    {
+      "epoch": 0.3328,
+      "grad_norm": 9.692618512955894,
+      "learning_rate": 1.591064607026499e-05,
+      "loss": 1.3768,
+      "step": 39
+    },
+    {
+      "epoch": 0.3413333333333333,
+      "grad_norm": 9.08889021911031,
+      "learning_rate": 1.6020599913279622e-05,
+      "loss": 1.3015,
+      "step": 40
+    },
+    {
+      "epoch": 0.34986666666666666,
+      "grad_norm": 8.081534221516058,
+      "learning_rate": 1.6127838567197353e-05,
+      "loss": 0.9228,
+      "step": 41
+    },
+    {
+      "epoch": 0.3584,
+      "grad_norm": 6.238638048950311,
+      "learning_rate": 1.6232492903979005e-05,
+      "loss": 0.7267,
+      "step": 42
+    },
+    {
+      "epoch": 0.36693333333333333,
+      "grad_norm": 3.4058036861773604,
+      "learning_rate": 1.6334684555795865e-05,
+      "loss": 0.5875,
+      "step": 43
+    },
+    {
+      "epoch": 0.37546666666666667,
+      "grad_norm": 2.079163829467713,
+      "learning_rate": 1.6434526764861872e-05,
+      "loss": 0.6355,
+      "step": 44
+    },
+    {
+      "epoch": 0.384,
+      "grad_norm": 1.5597487833024746,
+      "learning_rate": 1.6532125137753435e-05,
+      "loss": 0.5106,
+      "step": 45
+    },
+    {
+      "epoch": 0.39253333333333335,
+      "grad_norm": 2.491689602375256,
+      "learning_rate": 1.662757831681574e-05,
+      "loss": 0.6454,
+      "step": 46
+    },
+    {
+      "epoch": 0.4010666666666667,
+      "grad_norm": 2.010880438195854,
+      "learning_rate": 1.672097857935717e-05,
+      "loss": 0.4757,
+      "step": 47
+    },
+    {
+      "epoch": 0.4096,
+      "grad_norm": 1.9452805114322096,
+      "learning_rate": 1.681241237375587e-05,
+      "loss": 0.4133,
+      "step": 48
+    },
+    {
+      "epoch": 0.41813333333333336,
+      "grad_norm": 1.7620848552306103,
+      "learning_rate": 1.6901960800285137e-05,
+      "loss": 0.4004,
+      "step": 49
+    },
+    {
+      "epoch": 0.4266666666666667,
+      "grad_norm": 1.278224489774809,
+      "learning_rate": 1.6989700043360187e-05,
+      "loss": 0.3523,
+      "step": 50
+    },
+    {
+      "epoch": 0.4352,
+      "grad_norm": 1.6151354758303231,
+      "learning_rate": 1.7075701760979363e-05,
+      "loss": 0.4317,
+      "step": 51
+    },
+    {
+      "epoch": 0.4437333333333333,
+      "grad_norm": 1.3451396055695035,
+      "learning_rate": 1.716003343634799e-05,
+      "loss": 0.3474,
+      "step": 52
+    },
+    {
+      "epoch": 0.45226666666666665,
+      "grad_norm": 1.6814977782362666,
+      "learning_rate": 1.724275869600789e-05,
+      "loss": 0.3706,
+      "step": 53
+    },
+    {
+      "epoch": 0.4608,
+      "grad_norm": 1.1682442432688667,
+      "learning_rate": 1.7323937598229687e-05,
+      "loss": 0.3488,
+      "step": 54
+    },
+    {
+      "epoch": 0.4693333333333333,
+      "grad_norm": 0.8839814540462471,
+      "learning_rate": 1.7403626894942437e-05,
+      "loss": 0.293,
+      "step": 55
+    },
+    {
+      "epoch": 0.47786666666666666,
+      "grad_norm": 0.7974304806295485,
+      "learning_rate": 1.7481880270062003e-05,
+      "loss": 0.2717,
+      "step": 56
+    },
+    {
+      "epoch": 0.4864,
+      "grad_norm": 0.6232203657585239,
+      "learning_rate": 1.7558748556724913e-05,
+      "loss": 0.1741,
+      "step": 57
+    },
+    {
+      "epoch": 0.49493333333333334,
+      "grad_norm": 0.6850688604563008,
+      "learning_rate": 1.763427993562937e-05,
+      "loss": 0.228,
+      "step": 58
+    },
+    {
+      "epoch": 0.5034666666666666,
+      "grad_norm": 0.5923826384300431,
+      "learning_rate": 1.7708520116421443e-05,
+      "loss": 0.2131,
+      "step": 59
+    },
+    {
+      "epoch": 0.512,
+      "grad_norm": 0.489754430485032,
+      "learning_rate": 1.7781512503836432e-05,
+      "loss": 0.165,
+      "step": 60
+    },
+    {
+      "epoch": 0.5205333333333333,
+      "grad_norm": 0.5280940052395061,
+      "learning_rate": 1.7853298350107667e-05,
+      "loss": 0.1658,
+      "step": 61
+    },
+    {
+      "epoch": 0.5290666666666667,
+      "grad_norm": 0.4750905992036739,
+      "learning_rate": 1.7923916894982537e-05,
+      "loss": 0.1438,
+      "step": 62
+    },
+    {
+      "epoch": 0.5376,
+      "grad_norm": 0.4162998392722401,
+      "learning_rate": 1.7993405494535815e-05,
+      "loss": 0.1555,
+      "step": 63
+    },
+    {
+      "epoch": 0.5461333333333334,
+      "grad_norm": 0.26863266246370443,
+      "learning_rate": 1.806179973983887e-05,
+      "loss": 0.1323,
+      "step": 64
+    },
+    {
+      "epoch": 0.5546666666666666,
+      "grad_norm": 0.26534221125601215,
+      "learning_rate": 1.8129133566428553e-05,
+      "loss": 0.1671,
+      "step": 65
+    },
+    {
+      "epoch": 0.5632,
+      "grad_norm": 0.2548662962257576,
+      "learning_rate": 1.8195439355418686e-05,
+      "loss": 0.1308,
+      "step": 66
+    },
+    {
+      "epoch": 0.5717333333333333,
+      "grad_norm": 0.18045394638578796,
+      "learning_rate": 1.8260748027008263e-05,
+      "loss": 0.1262,
+      "step": 67
+    },
+    {
+      "epoch": 0.5802666666666667,
+      "grad_norm": 0.17070388073612064,
+      "learning_rate": 1.8325089127062364e-05,
+      "loss": 0.1192,
+      "step": 68
+    },
+    {
+      "epoch": 0.5888,
+      "grad_norm": 0.1531381679244776,
+      "learning_rate": 1.8388490907372553e-05,
+      "loss": 0.1274,
+      "step": 69
+    },
+    {
+      "epoch": 0.5973333333333334,
+      "grad_norm": 0.18196434993687946,
+      "learning_rate": 1.8450980400142568e-05,
+      "loss": 0.1375,
+      "step": 70
+    },
+    {
+      "epoch": 0.6058666666666667,
+      "grad_norm": 0.15324416972951205,
+      "learning_rate": 1.8512583487190752e-05,
+      "loss": 0.1599,
+      "step": 71
+    },
+    {
+      "epoch": 0.6144,
+      "grad_norm": 0.10884462064503801,
+      "learning_rate": 1.857332496431268e-05,
+      "loss": 0.1041,
+      "step": 72
+    },
+    {
+      "epoch": 0.6229333333333333,
+      "grad_norm": 0.12915133528192668,
+      "learning_rate": 1.8633228601204555e-05,
+      "loss": 0.1406,
+      "step": 73
+    },
+    {
+      "epoch": 0.6314666666666666,
+      "grad_norm": 0.12553425699952878,
+      "learning_rate": 1.8692317197309763e-05,
+      "loss": 0.1256,
+      "step": 74
+    },
+    {
+      "epoch": 0.64,
+      "grad_norm": 0.11976960918968543,
+      "learning_rate": 1.8750612633916997e-05,
+      "loss": 0.1144,
+      "step": 75
+    },
+    {
+      "epoch": 0.6485333333333333,
+      "grad_norm": 0.115805998298789,
+      "learning_rate": 1.8808135922807914e-05,
+      "loss": 0.1528,
+      "step": 76
+    },
+    {
+      "epoch": 0.6570666666666667,
+      "grad_norm": 0.10325948496697443,
+      "learning_rate": 1.8864907251724818e-05,
+      "loss": 0.1044,
+      "step": 77
+    },
+    {
+      "epoch": 0.6656,
+      "grad_norm": 0.09595064346541006,
+      "learning_rate": 1.8920946026904802e-05,
+      "loss": 0.1534,
+      "step": 78
+    },
+    {
+      "epoch": 0.6741333333333334,
+      "grad_norm": 0.08796742845240496,
+      "learning_rate": 1.8976270912904414e-05,
+      "loss": 0.1155,
+      "step": 79
+    },
+    {
+      "epoch": 0.6826666666666666,
+      "grad_norm": 0.08218991738379527,
+      "learning_rate": 1.9030899869919434e-05,
+      "loss": 0.1311,
+      "step": 80
+    },
+    {
+      "epoch": 0.6912,
+      "grad_norm": 0.08290815261109215,
+      "learning_rate": 1.9084850188786497e-05,
+      "loss": 0.11,
+      "step": 81
+    },
+    {
+      "epoch": 0.6997333333333333,
+      "grad_norm": 0.0794132180176064,
+      "learning_rate": 1.9138138523837165e-05,
+      "loss": 0.1135,
+      "step": 82
+    },
+    {
+      "epoch": 0.7082666666666667,
+      "grad_norm": 0.06934410705255296,
+      "learning_rate": 1.919078092376074e-05,
+      "loss": 0.109,
+      "step": 83
+    },
+    {
+      "epoch": 0.7168,
+      "grad_norm": 0.09000563031870593,
+      "learning_rate": 1.9242792860618813e-05,
+      "loss": 0.12,
+      "step": 84
+    },
+    {
+      "epoch": 0.7253333333333334,
+      "grad_norm": 0.1134042277879818,
+      "learning_rate": 1.929418925714293e-05,
+      "loss": 0.1223,
+      "step": 85
+    },
+    {
+      "epoch": 0.7338666666666667,
+      "grad_norm": 0.09118764690233076,
+      "learning_rate": 1.9344984512435673e-05,
+      "loss": 0.1459,
+      "step": 86
+    },
+    {
+      "epoch": 0.7424,
+      "grad_norm": 0.07873016754353963,
+      "learning_rate": 1.9395192526186183e-05,
+      "loss": 0.1422,
+      "step": 87
+    },
+    {
+      "epoch": 0.7509333333333333,
+      "grad_norm": 0.1796495874463076,
+      "learning_rate": 1.9444826721501687e-05,
+      "loss": 0.1291,
+      "step": 88
+    },
+    {
+      "epoch": 0.7594666666666666,
+      "grad_norm": 0.0679589944174269,
+      "learning_rate": 1.9493900066449125e-05,
+      "loss": 0.108,
+      "step": 89
+    },
+    {
+      "epoch": 0.768,
+      "grad_norm": 0.08174688574235538,
+      "learning_rate": 1.9542425094393246e-05,
+      "loss": 0.1081,
+      "step": 90
+    },
+    {
+      "epoch": 0.7765333333333333,
+      "grad_norm": 0.057137370501406756,
+      "learning_rate": 1.9590413923210934e-05,
+      "loss": 0.0934,
+      "step": 91
+    },
+    {
+      "epoch": 0.7850666666666667,
+      "grad_norm": 0.06578111924908255,
+      "learning_rate": 1.9637878273455555e-05,
+      "loss": 0.1085,
+      "step": 92
+    },
+    {
+      "epoch": 0.7936,
+      "grad_norm": 0.08945990540906254,
+      "learning_rate": 1.968482948553935e-05,
+      "loss": 0.1747,
+      "step": 93
+    },
+    {
+      "epoch": 0.8021333333333334,
+      "grad_norm": 0.06183863311044229,
+      "learning_rate": 1.9731278535996986e-05,
+      "loss": 0.1136,
+      "step": 94
+    },
+    {
+      "epoch": 0.8106666666666666,
+      "grad_norm": 0.05777899602544702,
+      "learning_rate": 1.9777236052888476e-05,
+      "loss": 0.0984,
+      "step": 95
+    },
+    {
+      "epoch": 0.8192,
+      "grad_norm": 0.08130851607693534,
+      "learning_rate": 1.9822712330395683e-05,
+      "loss": 0.187,
+      "step": 96
+    },
+    {
+      "epoch": 0.8277333333333333,
+      "grad_norm": 0.06426546202002927,
+      "learning_rate": 1.986771734266245e-05,
+      "loss": 0.1296,
+      "step": 97
+    },
+    {
+      "epoch": 0.8362666666666667,
+      "grad_norm": 0.069692313707994,
+      "learning_rate": 1.991226075692495e-05,
+      "loss": 0.1404,
+      "step": 98
+    },
+    {
+      "epoch": 0.8448,
+      "grad_norm": 0.05494542266886729,
+      "learning_rate": 1.9956351945975496e-05,
+      "loss": 0.116,
+      "step": 99
+    },
+    {
+      "epoch": 0.8533333333333334,
+      "grad_norm": 0.07571686966840627,
+      "learning_rate": 1.9999999999999998e-05,
+      "loss": 0.1539,
+      "step": 100
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 351,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 3,
+  "save_steps": 20,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 1.4183197916633498e+18,
+  "train_batch_size": 16,
+  "trial_name": null,
+  "trial_params": null
+}
diff --git a/checkpoint-100/training_args.bin b/checkpoint-100/training_args.bin
new file mode 100644
index 0000000000000000000000000000000000000000..8ac842d29fc6ad3e132cb7208dda5b61dd1079ee
--- /dev/null
+++ b/checkpoint-100/training_args.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aa4edad1da5557fcd52a6da980443588016cead6f0444a3562cfa68029c66a04
+size 6840
diff --git a/checkpoint-100/zero_to_fp32.py b/checkpoint-100/zero_to_fp32.py
new file mode 100644
index 0000000000000000000000000000000000000000..49b846633d6eb1e836e34681e44033581f4edb7b
--- /dev/null
+++ b/checkpoint-100/zero_to_fp32.py
@@ -0,0 +1,592 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example: python zero_to_fp32.py . pytorch_model.bin
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+
+    total_files = len(files)
+    state_dicts = []
+    for f in files:
+        state_dict = torch.load(f, map_location=device)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    if zero_stage <= 2:
+        fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    elif zero_stage == 3:
+        # if there is more than one param group, there will be multiple flattened tensors - one
+        # flattened tensor per group - for simplicity merge them into a single tensor
+        #
+        # XXX: could make the script more memory efficient for when there are multiple groups - it
+        # will require matching the sub-lists of param_shapes for each param group flattened tensor
+
+        fp32_flat_groups = [
+            torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
+        ]
+
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    for name, shape in param_shapes.items():
+
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # XXX: memory usage doubles here
+        state_dict[name] = torch.cat(
+            tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
+            0).narrow(0, 0, unpartitioned_numel).view(shape)
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    Note: this approach may not work if your application doesn't have sufficient free CPU memory and
+    you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+    """
+
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+    print(f"Saving fp32 state dict to {output_file}")
+    torch.save(state_dict, output_file)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument(
+        "output_file",
+        type=str,
+        help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file, tag=args.tag)
diff --git a/checkpoint-120/README.md b/checkpoint-120/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..e7d4f706aeefc6bb2aa84e963870f345a14c4af9
--- /dev/null
+++ b/checkpoint-120/README.md
@@ -0,0 +1,202 @@
+---
+library_name: peft
+base_model: ../ckpts/Meta-Llama-3-8B-Instruct
+---
+
+# 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 0.11.1
\ No newline at end of file
diff --git a/checkpoint-120/adapter_config.json b/checkpoint-120/adapter_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..ad16c073ed8946d227350e8fe4e99d689a8ede54
--- /dev/null
+++ b/checkpoint-120/adapter_config.json
@@ -0,0 +1,35 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "../ckpts/Meta-Llama-3-8B-Instruct",
+  "bias": "none",
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 16,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "down_proj",
+    "gate_proj",
+    "q_proj",
+    "lm_head",
+    "o_proj",
+    "v_proj",
+    "up_proj",
+    "k_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "use_dora": false,
+  "use_rslora": false
+}
\ No newline at end of file
diff --git a/checkpoint-120/adapter_model.safetensors b/checkpoint-120/adapter_model.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..9493442001d73e61f62fce764c9b35b520637662
--- /dev/null
+++ b/checkpoint-120/adapter_model.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:33e6839258702615da06d8aa30c9fd67b2657cce43937f62cf3927f85b0a8362
+size 1138856856
diff --git a/checkpoint-120/global_step120/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt b/checkpoint-120/global_step120/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..4283e37a935e0b67841c3e3e74b279dbc27fbb4e
--- /dev/null
+++ b/checkpoint-120/global_step120/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:11b436ebd244ff8f8485f00d23b9c863c009715ce5d24e64a356edf305f17b6e
+size 528781328
diff --git a/checkpoint-120/global_step120/mp_rank_00_model_states.pt b/checkpoint-120/global_step120/mp_rank_00_model_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..d29971694e7a6b321cc8175a2ddc958f767579f3
--- /dev/null
+++ b/checkpoint-120/global_step120/mp_rank_00_model_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2c17e1cbdbbc34d79755357fd83899cf1a03c70d31c45b9c5104634809569283
+size 199905337
diff --git a/checkpoint-120/latest b/checkpoint-120/latest
new file mode 100644
index 0000000000000000000000000000000000000000..0e13e0563ec45a863d519305a1251d3e72b9e3e4
--- /dev/null
+++ b/checkpoint-120/latest
@@ -0,0 +1 @@
+global_step120
\ No newline at end of file
diff --git a/checkpoint-120/rng_state.pth b/checkpoint-120/rng_state.pth
new file mode 100644
index 0000000000000000000000000000000000000000..07287c645140a43d7c9c2d0aa6ee90520966e222
--- /dev/null
+++ b/checkpoint-120/rng_state.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9ae17b05257a556f53e7168228629d1bb3ae5e5c9000a8d7dd9d2cc833ef3948
+size 14244
diff --git a/checkpoint-120/trainer_state.json b/checkpoint-120/trainer_state.json
new file mode 100644
index 0000000000000000000000000000000000000000..8fa28c6e36ab8566052930f668692e15e2519237
--- /dev/null
+++ b/checkpoint-120/trainer_state.json
@@ -0,0 +1,873 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 1.024,
+  "eval_steps": 500,
+  "global_step": 120,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.008533333333333334,
+      "grad_norm": 100.21848203113535,
+      "learning_rate": 0.0,
+      "loss": 7.1962,
+      "step": 1
+    },
+    {
+      "epoch": 0.017066666666666667,
+      "grad_norm": 96.43006188910957,
+      "learning_rate": 3.0102999566398115e-06,
+      "loss": 6.9414,
+      "step": 2
+    },
+    {
+      "epoch": 0.0256,
+      "grad_norm": 97.35803466618715,
+      "learning_rate": 4.771212547196624e-06,
+      "loss": 7.0102,
+      "step": 3
+    },
+    {
+      "epoch": 0.034133333333333335,
+      "grad_norm": 95.14837816372646,
+      "learning_rate": 6.020599913279623e-06,
+      "loss": 6.5295,
+      "step": 4
+    },
+    {
+      "epoch": 0.042666666666666665,
+      "grad_norm": 91.76544275692784,
+      "learning_rate": 6.989700043360187e-06,
+      "loss": 6.4806,
+      "step": 5
+    },
+    {
+      "epoch": 0.0512,
+      "grad_norm": 84.4494318688335,
+      "learning_rate": 7.781512503836437e-06,
+      "loss": 6.4194,
+      "step": 6
+    },
+    {
+      "epoch": 0.05973333333333333,
+      "grad_norm": 71.37977490595638,
+      "learning_rate": 8.450980400142568e-06,
+      "loss": 5.4953,
+      "step": 7
+    },
+    {
+      "epoch": 0.06826666666666667,
+      "grad_norm": 49.31153456754566,
+      "learning_rate": 9.030899869919434e-06,
+      "loss": 5.4123,
+      "step": 8
+    },
+    {
+      "epoch": 0.0768,
+      "grad_norm": 20.37296364560341,
+      "learning_rate": 9.542425094393249e-06,
+      "loss": 5.2334,
+      "step": 9
+    },
+    {
+      "epoch": 0.08533333333333333,
+      "grad_norm": 7.968467079076881,
+      "learning_rate": 9.999999999999999e-06,
+      "loss": 5.0282,
+      "step": 10
+    },
+    {
+      "epoch": 0.09386666666666667,
+      "grad_norm": 3.559446532055649,
+      "learning_rate": 1.041392685158225e-05,
+      "loss": 4.612,
+      "step": 11
+    },
+    {
+      "epoch": 0.1024,
+      "grad_norm": 3.5528846947995674,
+      "learning_rate": 1.0791812460476248e-05,
+      "loss": 4.9475,
+      "step": 12
+    },
+    {
+      "epoch": 0.11093333333333333,
+      "grad_norm": 3.541968897471334,
+      "learning_rate": 1.1139433523068365e-05,
+      "loss": 4.2777,
+      "step": 13
+    },
+    {
+      "epoch": 0.11946666666666667,
+      "grad_norm": 3.54718070036198,
+      "learning_rate": 1.1461280356782378e-05,
+      "loss": 4.3507,
+      "step": 14
+    },
+    {
+      "epoch": 0.128,
+      "grad_norm": 3.8632334830606747,
+      "learning_rate": 1.1760912590556813e-05,
+      "loss": 4.5364,
+      "step": 15
+    },
+    {
+      "epoch": 0.13653333333333334,
+      "grad_norm": 3.6637424744054004,
+      "learning_rate": 1.2041199826559246e-05,
+      "loss": 3.9672,
+      "step": 16
+    },
+    {
+      "epoch": 0.14506666666666668,
+      "grad_norm": 3.919802904818311,
+      "learning_rate": 1.230448921378274e-05,
+      "loss": 4.0618,
+      "step": 17
+    },
+    {
+      "epoch": 0.1536,
+      "grad_norm": 4.71904950738746,
+      "learning_rate": 1.2552725051033058e-05,
+      "loss": 4.6656,
+      "step": 18
+    },
+    {
+      "epoch": 0.16213333333333332,
+      "grad_norm": 4.6656317698690835,
+      "learning_rate": 1.2787536009528288e-05,
+      "loss": 4.1131,
+      "step": 19
+    },
+    {
+      "epoch": 0.17066666666666666,
+      "grad_norm": 5.145138692367417,
+      "learning_rate": 1.301029995663981e-05,
+      "loss": 4.0989,
+      "step": 20
+    },
+    {
+      "epoch": 0.1792,
+      "grad_norm": 5.923538014759818,
+      "learning_rate": 1.3222192947339192e-05,
+      "loss": 4.4991,
+      "step": 21
+    },
+    {
+      "epoch": 0.18773333333333334,
+      "grad_norm": 5.941056962941364,
+      "learning_rate": 1.3424226808222062e-05,
+      "loss": 4.0836,
+      "step": 22
+    },
+    {
+      "epoch": 0.19626666666666667,
+      "grad_norm": 6.171026012117947,
+      "learning_rate": 1.3617278360175927e-05,
+      "loss": 3.6861,
+      "step": 23
+    },
+    {
+      "epoch": 0.2048,
+      "grad_norm": 7.130542138930838,
+      "learning_rate": 1.380211241711606e-05,
+      "loss": 4.0958,
+      "step": 24
+    },
+    {
+      "epoch": 0.21333333333333335,
+      "grad_norm": 7.328837606110418,
+      "learning_rate": 1.3979400086720374e-05,
+      "loss": 3.9524,
+      "step": 25
+    },
+    {
+      "epoch": 0.22186666666666666,
+      "grad_norm": 6.923489005711429,
+      "learning_rate": 1.4149733479708178e-05,
+      "loss": 3.6062,
+      "step": 26
+    },
+    {
+      "epoch": 0.2304,
+      "grad_norm": 6.409498926059221,
+      "learning_rate": 1.4313637641589872e-05,
+      "loss": 3.2034,
+      "step": 27
+    },
+    {
+      "epoch": 0.23893333333333333,
+      "grad_norm": 5.781628405584682,
+      "learning_rate": 1.4471580313422191e-05,
+      "loss": 2.8158,
+      "step": 28
+    },
+    {
+      "epoch": 0.24746666666666667,
+      "grad_norm": 6.2927590068243315,
+      "learning_rate": 1.4623979978989559e-05,
+      "loss": 2.9803,
+      "step": 29
+    },
+    {
+      "epoch": 0.256,
+      "grad_norm": 6.103844678630006,
+      "learning_rate": 1.4771212547196623e-05,
+      "loss": 2.847,
+      "step": 30
+    },
+    {
+      "epoch": 0.26453333333333334,
+      "grad_norm": 7.656341925867144,
+      "learning_rate": 1.4913616938342726e-05,
+      "loss": 3.0907,
+      "step": 31
+    },
+    {
+      "epoch": 0.2730666666666667,
+      "grad_norm": 6.324242877501844,
+      "learning_rate": 1.5051499783199059e-05,
+      "loss": 2.3467,
+      "step": 32
+    },
+    {
+      "epoch": 0.2816,
+      "grad_norm": 7.606313101162739,
+      "learning_rate": 1.5185139398778874e-05,
+      "loss": 2.5292,
+      "step": 33
+    },
+    {
+      "epoch": 0.29013333333333335,
+      "grad_norm": 8.553792493849265,
+      "learning_rate": 1.531478917042255e-05,
+      "loss": 2.4547,
+      "step": 34
+    },
+    {
+      "epoch": 0.2986666666666667,
+      "grad_norm": 8.483368703272543,
+      "learning_rate": 1.5440680443502753e-05,
+      "loss": 2.1956,
+      "step": 35
+    },
+    {
+      "epoch": 0.3072,
+      "grad_norm": 8.364739352838077,
+      "learning_rate": 1.5563025007672873e-05,
+      "loss": 1.8552,
+      "step": 36
+    },
+    {
+      "epoch": 0.3157333333333333,
+      "grad_norm": 9.37663682000104,
+      "learning_rate": 1.5682017240669948e-05,
+      "loss": 1.9228,
+      "step": 37
+    },
+    {
+      "epoch": 0.32426666666666665,
+      "grad_norm": 9.823047193440066,
+      "learning_rate": 1.57978359661681e-05,
+      "loss": 1.7033,
+      "step": 38
+    },
+    {
+      "epoch": 0.3328,
+      "grad_norm": 9.692618512955894,
+      "learning_rate": 1.591064607026499e-05,
+      "loss": 1.3768,
+      "step": 39
+    },
+    {
+      "epoch": 0.3413333333333333,
+      "grad_norm": 9.08889021911031,
+      "learning_rate": 1.6020599913279622e-05,
+      "loss": 1.3015,
+      "step": 40
+    },
+    {
+      "epoch": 0.34986666666666666,
+      "grad_norm": 8.081534221516058,
+      "learning_rate": 1.6127838567197353e-05,
+      "loss": 0.9228,
+      "step": 41
+    },
+    {
+      "epoch": 0.3584,
+      "grad_norm": 6.238638048950311,
+      "learning_rate": 1.6232492903979005e-05,
+      "loss": 0.7267,
+      "step": 42
+    },
+    {
+      "epoch": 0.36693333333333333,
+      "grad_norm": 3.4058036861773604,
+      "learning_rate": 1.6334684555795865e-05,
+      "loss": 0.5875,
+      "step": 43
+    },
+    {
+      "epoch": 0.37546666666666667,
+      "grad_norm": 2.079163829467713,
+      "learning_rate": 1.6434526764861872e-05,
+      "loss": 0.6355,
+      "step": 44
+    },
+    {
+      "epoch": 0.384,
+      "grad_norm": 1.5597487833024746,
+      "learning_rate": 1.6532125137753435e-05,
+      "loss": 0.5106,
+      "step": 45
+    },
+    {
+      "epoch": 0.39253333333333335,
+      "grad_norm": 2.491689602375256,
+      "learning_rate": 1.662757831681574e-05,
+      "loss": 0.6454,
+      "step": 46
+    },
+    {
+      "epoch": 0.4010666666666667,
+      "grad_norm": 2.010880438195854,
+      "learning_rate": 1.672097857935717e-05,
+      "loss": 0.4757,
+      "step": 47
+    },
+    {
+      "epoch": 0.4096,
+      "grad_norm": 1.9452805114322096,
+      "learning_rate": 1.681241237375587e-05,
+      "loss": 0.4133,
+      "step": 48
+    },
+    {
+      "epoch": 0.41813333333333336,
+      "grad_norm": 1.7620848552306103,
+      "learning_rate": 1.6901960800285137e-05,
+      "loss": 0.4004,
+      "step": 49
+    },
+    {
+      "epoch": 0.4266666666666667,
+      "grad_norm": 1.278224489774809,
+      "learning_rate": 1.6989700043360187e-05,
+      "loss": 0.3523,
+      "step": 50
+    },
+    {
+      "epoch": 0.4352,
+      "grad_norm": 1.6151354758303231,
+      "learning_rate": 1.7075701760979363e-05,
+      "loss": 0.4317,
+      "step": 51
+    },
+    {
+      "epoch": 0.4437333333333333,
+      "grad_norm": 1.3451396055695035,
+      "learning_rate": 1.716003343634799e-05,
+      "loss": 0.3474,
+      "step": 52
+    },
+    {
+      "epoch": 0.45226666666666665,
+      "grad_norm": 1.6814977782362666,
+      "learning_rate": 1.724275869600789e-05,
+      "loss": 0.3706,
+      "step": 53
+    },
+    {
+      "epoch": 0.4608,
+      "grad_norm": 1.1682442432688667,
+      "learning_rate": 1.7323937598229687e-05,
+      "loss": 0.3488,
+      "step": 54
+    },
+    {
+      "epoch": 0.4693333333333333,
+      "grad_norm": 0.8839814540462471,
+      "learning_rate": 1.7403626894942437e-05,
+      "loss": 0.293,
+      "step": 55
+    },
+    {
+      "epoch": 0.47786666666666666,
+      "grad_norm": 0.7974304806295485,
+      "learning_rate": 1.7481880270062003e-05,
+      "loss": 0.2717,
+      "step": 56
+    },
+    {
+      "epoch": 0.4864,
+      "grad_norm": 0.6232203657585239,
+      "learning_rate": 1.7558748556724913e-05,
+      "loss": 0.1741,
+      "step": 57
+    },
+    {
+      "epoch": 0.49493333333333334,
+      "grad_norm": 0.6850688604563008,
+      "learning_rate": 1.763427993562937e-05,
+      "loss": 0.228,
+      "step": 58
+    },
+    {
+      "epoch": 0.5034666666666666,
+      "grad_norm": 0.5923826384300431,
+      "learning_rate": 1.7708520116421443e-05,
+      "loss": 0.2131,
+      "step": 59
+    },
+    {
+      "epoch": 0.512,
+      "grad_norm": 0.489754430485032,
+      "learning_rate": 1.7781512503836432e-05,
+      "loss": 0.165,
+      "step": 60
+    },
+    {
+      "epoch": 0.5205333333333333,
+      "grad_norm": 0.5280940052395061,
+      "learning_rate": 1.7853298350107667e-05,
+      "loss": 0.1658,
+      "step": 61
+    },
+    {
+      "epoch": 0.5290666666666667,
+      "grad_norm": 0.4750905992036739,
+      "learning_rate": 1.7923916894982537e-05,
+      "loss": 0.1438,
+      "step": 62
+    },
+    {
+      "epoch": 0.5376,
+      "grad_norm": 0.4162998392722401,
+      "learning_rate": 1.7993405494535815e-05,
+      "loss": 0.1555,
+      "step": 63
+    },
+    {
+      "epoch": 0.5461333333333334,
+      "grad_norm": 0.26863266246370443,
+      "learning_rate": 1.806179973983887e-05,
+      "loss": 0.1323,
+      "step": 64
+    },
+    {
+      "epoch": 0.5546666666666666,
+      "grad_norm": 0.26534221125601215,
+      "learning_rate": 1.8129133566428553e-05,
+      "loss": 0.1671,
+      "step": 65
+    },
+    {
+      "epoch": 0.5632,
+      "grad_norm": 0.2548662962257576,
+      "learning_rate": 1.8195439355418686e-05,
+      "loss": 0.1308,
+      "step": 66
+    },
+    {
+      "epoch": 0.5717333333333333,
+      "grad_norm": 0.18045394638578796,
+      "learning_rate": 1.8260748027008263e-05,
+      "loss": 0.1262,
+      "step": 67
+    },
+    {
+      "epoch": 0.5802666666666667,
+      "grad_norm": 0.17070388073612064,
+      "learning_rate": 1.8325089127062364e-05,
+      "loss": 0.1192,
+      "step": 68
+    },
+    {
+      "epoch": 0.5888,
+      "grad_norm": 0.1531381679244776,
+      "learning_rate": 1.8388490907372553e-05,
+      "loss": 0.1274,
+      "step": 69
+    },
+    {
+      "epoch": 0.5973333333333334,
+      "grad_norm": 0.18196434993687946,
+      "learning_rate": 1.8450980400142568e-05,
+      "loss": 0.1375,
+      "step": 70
+    },
+    {
+      "epoch": 0.6058666666666667,
+      "grad_norm": 0.15324416972951205,
+      "learning_rate": 1.8512583487190752e-05,
+      "loss": 0.1599,
+      "step": 71
+    },
+    {
+      "epoch": 0.6144,
+      "grad_norm": 0.10884462064503801,
+      "learning_rate": 1.857332496431268e-05,
+      "loss": 0.1041,
+      "step": 72
+    },
+    {
+      "epoch": 0.6229333333333333,
+      "grad_norm": 0.12915133528192668,
+      "learning_rate": 1.8633228601204555e-05,
+      "loss": 0.1406,
+      "step": 73
+    },
+    {
+      "epoch": 0.6314666666666666,
+      "grad_norm": 0.12553425699952878,
+      "learning_rate": 1.8692317197309763e-05,
+      "loss": 0.1256,
+      "step": 74
+    },
+    {
+      "epoch": 0.64,
+      "grad_norm": 0.11976960918968543,
+      "learning_rate": 1.8750612633916997e-05,
+      "loss": 0.1144,
+      "step": 75
+    },
+    {
+      "epoch": 0.6485333333333333,
+      "grad_norm": 0.115805998298789,
+      "learning_rate": 1.8808135922807914e-05,
+      "loss": 0.1528,
+      "step": 76
+    },
+    {
+      "epoch": 0.6570666666666667,
+      "grad_norm": 0.10325948496697443,
+      "learning_rate": 1.8864907251724818e-05,
+      "loss": 0.1044,
+      "step": 77
+    },
+    {
+      "epoch": 0.6656,
+      "grad_norm": 0.09595064346541006,
+      "learning_rate": 1.8920946026904802e-05,
+      "loss": 0.1534,
+      "step": 78
+    },
+    {
+      "epoch": 0.6741333333333334,
+      "grad_norm": 0.08796742845240496,
+      "learning_rate": 1.8976270912904414e-05,
+      "loss": 0.1155,
+      "step": 79
+    },
+    {
+      "epoch": 0.6826666666666666,
+      "grad_norm": 0.08218991738379527,
+      "learning_rate": 1.9030899869919434e-05,
+      "loss": 0.1311,
+      "step": 80
+    },
+    {
+      "epoch": 0.6912,
+      "grad_norm": 0.08290815261109215,
+      "learning_rate": 1.9084850188786497e-05,
+      "loss": 0.11,
+      "step": 81
+    },
+    {
+      "epoch": 0.6997333333333333,
+      "grad_norm": 0.0794132180176064,
+      "learning_rate": 1.9138138523837165e-05,
+      "loss": 0.1135,
+      "step": 82
+    },
+    {
+      "epoch": 0.7082666666666667,
+      "grad_norm": 0.06934410705255296,
+      "learning_rate": 1.919078092376074e-05,
+      "loss": 0.109,
+      "step": 83
+    },
+    {
+      "epoch": 0.7168,
+      "grad_norm": 0.09000563031870593,
+      "learning_rate": 1.9242792860618813e-05,
+      "loss": 0.12,
+      "step": 84
+    },
+    {
+      "epoch": 0.7253333333333334,
+      "grad_norm": 0.1134042277879818,
+      "learning_rate": 1.929418925714293e-05,
+      "loss": 0.1223,
+      "step": 85
+    },
+    {
+      "epoch": 0.7338666666666667,
+      "grad_norm": 0.09118764690233076,
+      "learning_rate": 1.9344984512435673e-05,
+      "loss": 0.1459,
+      "step": 86
+    },
+    {
+      "epoch": 0.7424,
+      "grad_norm": 0.07873016754353963,
+      "learning_rate": 1.9395192526186183e-05,
+      "loss": 0.1422,
+      "step": 87
+    },
+    {
+      "epoch": 0.7509333333333333,
+      "grad_norm": 0.1796495874463076,
+      "learning_rate": 1.9444826721501687e-05,
+      "loss": 0.1291,
+      "step": 88
+    },
+    {
+      "epoch": 0.7594666666666666,
+      "grad_norm": 0.0679589944174269,
+      "learning_rate": 1.9493900066449125e-05,
+      "loss": 0.108,
+      "step": 89
+    },
+    {
+      "epoch": 0.768,
+      "grad_norm": 0.08174688574235538,
+      "learning_rate": 1.9542425094393246e-05,
+      "loss": 0.1081,
+      "step": 90
+    },
+    {
+      "epoch": 0.7765333333333333,
+      "grad_norm": 0.057137370501406756,
+      "learning_rate": 1.9590413923210934e-05,
+      "loss": 0.0934,
+      "step": 91
+    },
+    {
+      "epoch": 0.7850666666666667,
+      "grad_norm": 0.06578111924908255,
+      "learning_rate": 1.9637878273455555e-05,
+      "loss": 0.1085,
+      "step": 92
+    },
+    {
+      "epoch": 0.7936,
+      "grad_norm": 0.08945990540906254,
+      "learning_rate": 1.968482948553935e-05,
+      "loss": 0.1747,
+      "step": 93
+    },
+    {
+      "epoch": 0.8021333333333334,
+      "grad_norm": 0.06183863311044229,
+      "learning_rate": 1.9731278535996986e-05,
+      "loss": 0.1136,
+      "step": 94
+    },
+    {
+      "epoch": 0.8106666666666666,
+      "grad_norm": 0.05777899602544702,
+      "learning_rate": 1.9777236052888476e-05,
+      "loss": 0.0984,
+      "step": 95
+    },
+    {
+      "epoch": 0.8192,
+      "grad_norm": 0.08130851607693534,
+      "learning_rate": 1.9822712330395683e-05,
+      "loss": 0.187,
+      "step": 96
+    },
+    {
+      "epoch": 0.8277333333333333,
+      "grad_norm": 0.06426546202002927,
+      "learning_rate": 1.986771734266245e-05,
+      "loss": 0.1296,
+      "step": 97
+    },
+    {
+      "epoch": 0.8362666666666667,
+      "grad_norm": 0.069692313707994,
+      "learning_rate": 1.991226075692495e-05,
+      "loss": 0.1404,
+      "step": 98
+    },
+    {
+      "epoch": 0.8448,
+      "grad_norm": 0.05494542266886729,
+      "learning_rate": 1.9956351945975496e-05,
+      "loss": 0.116,
+      "step": 99
+    },
+    {
+      "epoch": 0.8533333333333334,
+      "grad_norm": 0.07571686966840627,
+      "learning_rate": 1.9999999999999998e-05,
+      "loss": 0.1539,
+      "step": 100
+    },
+    {
+      "epoch": 0.8618666666666667,
+      "grad_norm": 0.054351059117603705,
+      "learning_rate": 2e-05,
+      "loss": 0.1037,
+      "step": 101
+    },
+    {
+      "epoch": 0.8704,
+      "grad_norm": 0.06531899611551092,
+      "learning_rate": 2e-05,
+      "loss": 0.0827,
+      "step": 102
+    },
+    {
+      "epoch": 0.8789333333333333,
+      "grad_norm": 0.06131678646504652,
+      "learning_rate": 2e-05,
+      "loss": 0.1266,
+      "step": 103
+    },
+    {
+      "epoch": 0.8874666666666666,
+      "grad_norm": 0.06850220540661824,
+      "learning_rate": 2e-05,
+      "loss": 0.1456,
+      "step": 104
+    },
+    {
+      "epoch": 0.896,
+      "grad_norm": 0.05806908951252483,
+      "learning_rate": 2e-05,
+      "loss": 0.0954,
+      "step": 105
+    },
+    {
+      "epoch": 0.9045333333333333,
+      "grad_norm": 0.06503642452033717,
+      "learning_rate": 2e-05,
+      "loss": 0.1417,
+      "step": 106
+    },
+    {
+      "epoch": 0.9130666666666667,
+      "grad_norm": 0.050486271853277066,
+      "learning_rate": 2e-05,
+      "loss": 0.0959,
+      "step": 107
+    },
+    {
+      "epoch": 0.9216,
+      "grad_norm": 0.07746063813802379,
+      "learning_rate": 2e-05,
+      "loss": 0.1256,
+      "step": 108
+    },
+    {
+      "epoch": 0.9301333333333334,
+      "grad_norm": 0.051231172380840004,
+      "learning_rate": 2e-05,
+      "loss": 0.1116,
+      "step": 109
+    },
+    {
+      "epoch": 0.9386666666666666,
+      "grad_norm": 0.056296443557859455,
+      "learning_rate": 2e-05,
+      "loss": 0.1056,
+      "step": 110
+    },
+    {
+      "epoch": 0.9472,
+      "grad_norm": 0.05058663240713958,
+      "learning_rate": 2e-05,
+      "loss": 0.0971,
+      "step": 111
+    },
+    {
+      "epoch": 0.9557333333333333,
+      "grad_norm": 0.05532886570130611,
+      "learning_rate": 2e-05,
+      "loss": 0.1086,
+      "step": 112
+    },
+    {
+      "epoch": 0.9642666666666667,
+      "grad_norm": 0.05327811326654907,
+      "learning_rate": 2e-05,
+      "loss": 0.0989,
+      "step": 113
+    },
+    {
+      "epoch": 0.9728,
+      "grad_norm": 0.05663279364147864,
+      "learning_rate": 2e-05,
+      "loss": 0.0958,
+      "step": 114
+    },
+    {
+      "epoch": 0.9813333333333333,
+      "grad_norm": 0.04930904541225805,
+      "learning_rate": 2e-05,
+      "loss": 0.0887,
+      "step": 115
+    },
+    {
+      "epoch": 0.9898666666666667,
+      "grad_norm": 0.06096947951115022,
+      "learning_rate": 2e-05,
+      "loss": 0.106,
+      "step": 116
+    },
+    {
+      "epoch": 0.9984,
+      "grad_norm": 0.050092322361182495,
+      "learning_rate": 2e-05,
+      "loss": 0.0931,
+      "step": 117
+    },
+    {
+      "epoch": 1.0069333333333332,
+      "grad_norm": 0.04980408443758999,
+      "learning_rate": 2e-05,
+      "loss": 0.0955,
+      "step": 118
+    },
+    {
+      "epoch": 1.0154666666666667,
+      "grad_norm": 0.051183082721834305,
+      "learning_rate": 2e-05,
+      "loss": 0.1049,
+      "step": 119
+    },
+    {
+      "epoch": 1.024,
+      "grad_norm": 0.04332220265802814,
+      "learning_rate": 2e-05,
+      "loss": 0.0983,
+      "step": 120
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 351,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 3,
+  "save_steps": 20,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 1.7047845829953454e+18,
+  "train_batch_size": 16,
+  "trial_name": null,
+  "trial_params": null
+}
diff --git a/checkpoint-120/training_args.bin b/checkpoint-120/training_args.bin
new file mode 100644
index 0000000000000000000000000000000000000000..8ac842d29fc6ad3e132cb7208dda5b61dd1079ee
--- /dev/null
+++ b/checkpoint-120/training_args.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aa4edad1da5557fcd52a6da980443588016cead6f0444a3562cfa68029c66a04
+size 6840
diff --git a/checkpoint-120/zero_to_fp32.py b/checkpoint-120/zero_to_fp32.py
new file mode 100644
index 0000000000000000000000000000000000000000..49b846633d6eb1e836e34681e44033581f4edb7b
--- /dev/null
+++ b/checkpoint-120/zero_to_fp32.py
@@ -0,0 +1,592 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example: python zero_to_fp32.py . pytorch_model.bin
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+
+    total_files = len(files)
+    state_dicts = []
+    for f in files:
+        state_dict = torch.load(f, map_location=device)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    if zero_stage <= 2:
+        fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    elif zero_stage == 3:
+        # if there is more than one param group, there will be multiple flattened tensors - one
+        # flattened tensor per group - for simplicity merge them into a single tensor
+        #
+        # XXX: could make the script more memory efficient for when there are multiple groups - it
+        # will require matching the sub-lists of param_shapes for each param group flattened tensor
+
+        fp32_flat_groups = [
+            torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
+        ]
+
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    for name, shape in param_shapes.items():
+
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # XXX: memory usage doubles here
+        state_dict[name] = torch.cat(
+            tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
+            0).narrow(0, 0, unpartitioned_numel).view(shape)
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    Note: this approach may not work if your application doesn't have sufficient free CPU memory and
+    you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+    """
+
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+    print(f"Saving fp32 state dict to {output_file}")
+    torch.save(state_dict, output_file)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument(
+        "output_file",
+        type=str,
+        help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file, tag=args.tag)
diff --git a/checkpoint-140/README.md b/checkpoint-140/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..e7d4f706aeefc6bb2aa84e963870f345a14c4af9
--- /dev/null
+++ b/checkpoint-140/README.md
@@ -0,0 +1,202 @@
+---
+library_name: peft
+base_model: ../ckpts/Meta-Llama-3-8B-Instruct
+---
+
+# 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 0.11.1
\ No newline at end of file
diff --git a/checkpoint-140/adapter_config.json b/checkpoint-140/adapter_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..ad16c073ed8946d227350e8fe4e99d689a8ede54
--- /dev/null
+++ b/checkpoint-140/adapter_config.json
@@ -0,0 +1,35 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "../ckpts/Meta-Llama-3-8B-Instruct",
+  "bias": "none",
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 16,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "down_proj",
+    "gate_proj",
+    "q_proj",
+    "lm_head",
+    "o_proj",
+    "v_proj",
+    "up_proj",
+    "k_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "use_dora": false,
+  "use_rslora": false
+}
\ No newline at end of file
diff --git a/checkpoint-140/adapter_model.safetensors b/checkpoint-140/adapter_model.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..f90ce0829cbffa6e09b32434bd4cff5e73374710
--- /dev/null
+++ b/checkpoint-140/adapter_model.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b12f85af7122bcce5f6997bec8d9a8a845cf9ab8d2c4f241d37562051e34c216
+size 1138856856
diff --git a/checkpoint-140/global_step140/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt b/checkpoint-140/global_step140/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..9bf4b0bdf91ea891ccefd945807690103d866a32
--- /dev/null
+++ b/checkpoint-140/global_step140/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8993927bbec291df7a8a041058bc3762b88d9ca9fcfbdccdabc686717c9c5f5d
+size 528781328
diff --git a/checkpoint-140/global_step140/mp_rank_00_model_states.pt b/checkpoint-140/global_step140/mp_rank_00_model_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..ab1a0d5e2b82ed338fe9cc9f94bd2806ec8ca2cb
--- /dev/null
+++ b/checkpoint-140/global_step140/mp_rank_00_model_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a4dca3099b75369d573830bab987eb6f984e58440ba145f2f47df24a9212b7c4
+size 199905337
diff --git a/checkpoint-140/latest b/checkpoint-140/latest
new file mode 100644
index 0000000000000000000000000000000000000000..fd2b9aef86529798137c2868d556e873a23c785c
--- /dev/null
+++ b/checkpoint-140/latest
@@ -0,0 +1 @@
+global_step140
\ No newline at end of file
diff --git a/checkpoint-140/rng_state.pth b/checkpoint-140/rng_state.pth
new file mode 100644
index 0000000000000000000000000000000000000000..180e748a0f61603af87ab992e933142ff04a6b93
--- /dev/null
+++ b/checkpoint-140/rng_state.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cb5b9188bbc87bb2ff7dba4844a90c5c236962fc66f3da23e5af9ee0eda34a1b
+size 14244
diff --git a/checkpoint-140/trainer_state.json b/checkpoint-140/trainer_state.json
new file mode 100644
index 0000000000000000000000000000000000000000..d3c8bcf3579b682d1042be3748a2d489e81fb7ed
--- /dev/null
+++ b/checkpoint-140/trainer_state.json
@@ -0,0 +1,1013 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 1.1946666666666665,
+  "eval_steps": 500,
+  "global_step": 140,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.008533333333333334,
+      "grad_norm": 100.21848203113535,
+      "learning_rate": 0.0,
+      "loss": 7.1962,
+      "step": 1
+    },
+    {
+      "epoch": 0.017066666666666667,
+      "grad_norm": 96.43006188910957,
+      "learning_rate": 3.0102999566398115e-06,
+      "loss": 6.9414,
+      "step": 2
+    },
+    {
+      "epoch": 0.0256,
+      "grad_norm": 97.35803466618715,
+      "learning_rate": 4.771212547196624e-06,
+      "loss": 7.0102,
+      "step": 3
+    },
+    {
+      "epoch": 0.034133333333333335,
+      "grad_norm": 95.14837816372646,
+      "learning_rate": 6.020599913279623e-06,
+      "loss": 6.5295,
+      "step": 4
+    },
+    {
+      "epoch": 0.042666666666666665,
+      "grad_norm": 91.76544275692784,
+      "learning_rate": 6.989700043360187e-06,
+      "loss": 6.4806,
+      "step": 5
+    },
+    {
+      "epoch": 0.0512,
+      "grad_norm": 84.4494318688335,
+      "learning_rate": 7.781512503836437e-06,
+      "loss": 6.4194,
+      "step": 6
+    },
+    {
+      "epoch": 0.05973333333333333,
+      "grad_norm": 71.37977490595638,
+      "learning_rate": 8.450980400142568e-06,
+      "loss": 5.4953,
+      "step": 7
+    },
+    {
+      "epoch": 0.06826666666666667,
+      "grad_norm": 49.31153456754566,
+      "learning_rate": 9.030899869919434e-06,
+      "loss": 5.4123,
+      "step": 8
+    },
+    {
+      "epoch": 0.0768,
+      "grad_norm": 20.37296364560341,
+      "learning_rate": 9.542425094393249e-06,
+      "loss": 5.2334,
+      "step": 9
+    },
+    {
+      "epoch": 0.08533333333333333,
+      "grad_norm": 7.968467079076881,
+      "learning_rate": 9.999999999999999e-06,
+      "loss": 5.0282,
+      "step": 10
+    },
+    {
+      "epoch": 0.09386666666666667,
+      "grad_norm": 3.559446532055649,
+      "learning_rate": 1.041392685158225e-05,
+      "loss": 4.612,
+      "step": 11
+    },
+    {
+      "epoch": 0.1024,
+      "grad_norm": 3.5528846947995674,
+      "learning_rate": 1.0791812460476248e-05,
+      "loss": 4.9475,
+      "step": 12
+    },
+    {
+      "epoch": 0.11093333333333333,
+      "grad_norm": 3.541968897471334,
+      "learning_rate": 1.1139433523068365e-05,
+      "loss": 4.2777,
+      "step": 13
+    },
+    {
+      "epoch": 0.11946666666666667,
+      "grad_norm": 3.54718070036198,
+      "learning_rate": 1.1461280356782378e-05,
+      "loss": 4.3507,
+      "step": 14
+    },
+    {
+      "epoch": 0.128,
+      "grad_norm": 3.8632334830606747,
+      "learning_rate": 1.1760912590556813e-05,
+      "loss": 4.5364,
+      "step": 15
+    },
+    {
+      "epoch": 0.13653333333333334,
+      "grad_norm": 3.6637424744054004,
+      "learning_rate": 1.2041199826559246e-05,
+      "loss": 3.9672,
+      "step": 16
+    },
+    {
+      "epoch": 0.14506666666666668,
+      "grad_norm": 3.919802904818311,
+      "learning_rate": 1.230448921378274e-05,
+      "loss": 4.0618,
+      "step": 17
+    },
+    {
+      "epoch": 0.1536,
+      "grad_norm": 4.71904950738746,
+      "learning_rate": 1.2552725051033058e-05,
+      "loss": 4.6656,
+      "step": 18
+    },
+    {
+      "epoch": 0.16213333333333332,
+      "grad_norm": 4.6656317698690835,
+      "learning_rate": 1.2787536009528288e-05,
+      "loss": 4.1131,
+      "step": 19
+    },
+    {
+      "epoch": 0.17066666666666666,
+      "grad_norm": 5.145138692367417,
+      "learning_rate": 1.301029995663981e-05,
+      "loss": 4.0989,
+      "step": 20
+    },
+    {
+      "epoch": 0.1792,
+      "grad_norm": 5.923538014759818,
+      "learning_rate": 1.3222192947339192e-05,
+      "loss": 4.4991,
+      "step": 21
+    },
+    {
+      "epoch": 0.18773333333333334,
+      "grad_norm": 5.941056962941364,
+      "learning_rate": 1.3424226808222062e-05,
+      "loss": 4.0836,
+      "step": 22
+    },
+    {
+      "epoch": 0.19626666666666667,
+      "grad_norm": 6.171026012117947,
+      "learning_rate": 1.3617278360175927e-05,
+      "loss": 3.6861,
+      "step": 23
+    },
+    {
+      "epoch": 0.2048,
+      "grad_norm": 7.130542138930838,
+      "learning_rate": 1.380211241711606e-05,
+      "loss": 4.0958,
+      "step": 24
+    },
+    {
+      "epoch": 0.21333333333333335,
+      "grad_norm": 7.328837606110418,
+      "learning_rate": 1.3979400086720374e-05,
+      "loss": 3.9524,
+      "step": 25
+    },
+    {
+      "epoch": 0.22186666666666666,
+      "grad_norm": 6.923489005711429,
+      "learning_rate": 1.4149733479708178e-05,
+      "loss": 3.6062,
+      "step": 26
+    },
+    {
+      "epoch": 0.2304,
+      "grad_norm": 6.409498926059221,
+      "learning_rate": 1.4313637641589872e-05,
+      "loss": 3.2034,
+      "step": 27
+    },
+    {
+      "epoch": 0.23893333333333333,
+      "grad_norm": 5.781628405584682,
+      "learning_rate": 1.4471580313422191e-05,
+      "loss": 2.8158,
+      "step": 28
+    },
+    {
+      "epoch": 0.24746666666666667,
+      "grad_norm": 6.2927590068243315,
+      "learning_rate": 1.4623979978989559e-05,
+      "loss": 2.9803,
+      "step": 29
+    },
+    {
+      "epoch": 0.256,
+      "grad_norm": 6.103844678630006,
+      "learning_rate": 1.4771212547196623e-05,
+      "loss": 2.847,
+      "step": 30
+    },
+    {
+      "epoch": 0.26453333333333334,
+      "grad_norm": 7.656341925867144,
+      "learning_rate": 1.4913616938342726e-05,
+      "loss": 3.0907,
+      "step": 31
+    },
+    {
+      "epoch": 0.2730666666666667,
+      "grad_norm": 6.324242877501844,
+      "learning_rate": 1.5051499783199059e-05,
+      "loss": 2.3467,
+      "step": 32
+    },
+    {
+      "epoch": 0.2816,
+      "grad_norm": 7.606313101162739,
+      "learning_rate": 1.5185139398778874e-05,
+      "loss": 2.5292,
+      "step": 33
+    },
+    {
+      "epoch": 0.29013333333333335,
+      "grad_norm": 8.553792493849265,
+      "learning_rate": 1.531478917042255e-05,
+      "loss": 2.4547,
+      "step": 34
+    },
+    {
+      "epoch": 0.2986666666666667,
+      "grad_norm": 8.483368703272543,
+      "learning_rate": 1.5440680443502753e-05,
+      "loss": 2.1956,
+      "step": 35
+    },
+    {
+      "epoch": 0.3072,
+      "grad_norm": 8.364739352838077,
+      "learning_rate": 1.5563025007672873e-05,
+      "loss": 1.8552,
+      "step": 36
+    },
+    {
+      "epoch": 0.3157333333333333,
+      "grad_norm": 9.37663682000104,
+      "learning_rate": 1.5682017240669948e-05,
+      "loss": 1.9228,
+      "step": 37
+    },
+    {
+      "epoch": 0.32426666666666665,
+      "grad_norm": 9.823047193440066,
+      "learning_rate": 1.57978359661681e-05,
+      "loss": 1.7033,
+      "step": 38
+    },
+    {
+      "epoch": 0.3328,
+      "grad_norm": 9.692618512955894,
+      "learning_rate": 1.591064607026499e-05,
+      "loss": 1.3768,
+      "step": 39
+    },
+    {
+      "epoch": 0.3413333333333333,
+      "grad_norm": 9.08889021911031,
+      "learning_rate": 1.6020599913279622e-05,
+      "loss": 1.3015,
+      "step": 40
+    },
+    {
+      "epoch": 0.34986666666666666,
+      "grad_norm": 8.081534221516058,
+      "learning_rate": 1.6127838567197353e-05,
+      "loss": 0.9228,
+      "step": 41
+    },
+    {
+      "epoch": 0.3584,
+      "grad_norm": 6.238638048950311,
+      "learning_rate": 1.6232492903979005e-05,
+      "loss": 0.7267,
+      "step": 42
+    },
+    {
+      "epoch": 0.36693333333333333,
+      "grad_norm": 3.4058036861773604,
+      "learning_rate": 1.6334684555795865e-05,
+      "loss": 0.5875,
+      "step": 43
+    },
+    {
+      "epoch": 0.37546666666666667,
+      "grad_norm": 2.079163829467713,
+      "learning_rate": 1.6434526764861872e-05,
+      "loss": 0.6355,
+      "step": 44
+    },
+    {
+      "epoch": 0.384,
+      "grad_norm": 1.5597487833024746,
+      "learning_rate": 1.6532125137753435e-05,
+      "loss": 0.5106,
+      "step": 45
+    },
+    {
+      "epoch": 0.39253333333333335,
+      "grad_norm": 2.491689602375256,
+      "learning_rate": 1.662757831681574e-05,
+      "loss": 0.6454,
+      "step": 46
+    },
+    {
+      "epoch": 0.4010666666666667,
+      "grad_norm": 2.010880438195854,
+      "learning_rate": 1.672097857935717e-05,
+      "loss": 0.4757,
+      "step": 47
+    },
+    {
+      "epoch": 0.4096,
+      "grad_norm": 1.9452805114322096,
+      "learning_rate": 1.681241237375587e-05,
+      "loss": 0.4133,
+      "step": 48
+    },
+    {
+      "epoch": 0.41813333333333336,
+      "grad_norm": 1.7620848552306103,
+      "learning_rate": 1.6901960800285137e-05,
+      "loss": 0.4004,
+      "step": 49
+    },
+    {
+      "epoch": 0.4266666666666667,
+      "grad_norm": 1.278224489774809,
+      "learning_rate": 1.6989700043360187e-05,
+      "loss": 0.3523,
+      "step": 50
+    },
+    {
+      "epoch": 0.4352,
+      "grad_norm": 1.6151354758303231,
+      "learning_rate": 1.7075701760979363e-05,
+      "loss": 0.4317,
+      "step": 51
+    },
+    {
+      "epoch": 0.4437333333333333,
+      "grad_norm": 1.3451396055695035,
+      "learning_rate": 1.716003343634799e-05,
+      "loss": 0.3474,
+      "step": 52
+    },
+    {
+      "epoch": 0.45226666666666665,
+      "grad_norm": 1.6814977782362666,
+      "learning_rate": 1.724275869600789e-05,
+      "loss": 0.3706,
+      "step": 53
+    },
+    {
+      "epoch": 0.4608,
+      "grad_norm": 1.1682442432688667,
+      "learning_rate": 1.7323937598229687e-05,
+      "loss": 0.3488,
+      "step": 54
+    },
+    {
+      "epoch": 0.4693333333333333,
+      "grad_norm": 0.8839814540462471,
+      "learning_rate": 1.7403626894942437e-05,
+      "loss": 0.293,
+      "step": 55
+    },
+    {
+      "epoch": 0.47786666666666666,
+      "grad_norm": 0.7974304806295485,
+      "learning_rate": 1.7481880270062003e-05,
+      "loss": 0.2717,
+      "step": 56
+    },
+    {
+      "epoch": 0.4864,
+      "grad_norm": 0.6232203657585239,
+      "learning_rate": 1.7558748556724913e-05,
+      "loss": 0.1741,
+      "step": 57
+    },
+    {
+      "epoch": 0.49493333333333334,
+      "grad_norm": 0.6850688604563008,
+      "learning_rate": 1.763427993562937e-05,
+      "loss": 0.228,
+      "step": 58
+    },
+    {
+      "epoch": 0.5034666666666666,
+      "grad_norm": 0.5923826384300431,
+      "learning_rate": 1.7708520116421443e-05,
+      "loss": 0.2131,
+      "step": 59
+    },
+    {
+      "epoch": 0.512,
+      "grad_norm": 0.489754430485032,
+      "learning_rate": 1.7781512503836432e-05,
+      "loss": 0.165,
+      "step": 60
+    },
+    {
+      "epoch": 0.5205333333333333,
+      "grad_norm": 0.5280940052395061,
+      "learning_rate": 1.7853298350107667e-05,
+      "loss": 0.1658,
+      "step": 61
+    },
+    {
+      "epoch": 0.5290666666666667,
+      "grad_norm": 0.4750905992036739,
+      "learning_rate": 1.7923916894982537e-05,
+      "loss": 0.1438,
+      "step": 62
+    },
+    {
+      "epoch": 0.5376,
+      "grad_norm": 0.4162998392722401,
+      "learning_rate": 1.7993405494535815e-05,
+      "loss": 0.1555,
+      "step": 63
+    },
+    {
+      "epoch": 0.5461333333333334,
+      "grad_norm": 0.26863266246370443,
+      "learning_rate": 1.806179973983887e-05,
+      "loss": 0.1323,
+      "step": 64
+    },
+    {
+      "epoch": 0.5546666666666666,
+      "grad_norm": 0.26534221125601215,
+      "learning_rate": 1.8129133566428553e-05,
+      "loss": 0.1671,
+      "step": 65
+    },
+    {
+      "epoch": 0.5632,
+      "grad_norm": 0.2548662962257576,
+      "learning_rate": 1.8195439355418686e-05,
+      "loss": 0.1308,
+      "step": 66
+    },
+    {
+      "epoch": 0.5717333333333333,
+      "grad_norm": 0.18045394638578796,
+      "learning_rate": 1.8260748027008263e-05,
+      "loss": 0.1262,
+      "step": 67
+    },
+    {
+      "epoch": 0.5802666666666667,
+      "grad_norm": 0.17070388073612064,
+      "learning_rate": 1.8325089127062364e-05,
+      "loss": 0.1192,
+      "step": 68
+    },
+    {
+      "epoch": 0.5888,
+      "grad_norm": 0.1531381679244776,
+      "learning_rate": 1.8388490907372553e-05,
+      "loss": 0.1274,
+      "step": 69
+    },
+    {
+      "epoch": 0.5973333333333334,
+      "grad_norm": 0.18196434993687946,
+      "learning_rate": 1.8450980400142568e-05,
+      "loss": 0.1375,
+      "step": 70
+    },
+    {
+      "epoch": 0.6058666666666667,
+      "grad_norm": 0.15324416972951205,
+      "learning_rate": 1.8512583487190752e-05,
+      "loss": 0.1599,
+      "step": 71
+    },
+    {
+      "epoch": 0.6144,
+      "grad_norm": 0.10884462064503801,
+      "learning_rate": 1.857332496431268e-05,
+      "loss": 0.1041,
+      "step": 72
+    },
+    {
+      "epoch": 0.6229333333333333,
+      "grad_norm": 0.12915133528192668,
+      "learning_rate": 1.8633228601204555e-05,
+      "loss": 0.1406,
+      "step": 73
+    },
+    {
+      "epoch": 0.6314666666666666,
+      "grad_norm": 0.12553425699952878,
+      "learning_rate": 1.8692317197309763e-05,
+      "loss": 0.1256,
+      "step": 74
+    },
+    {
+      "epoch": 0.64,
+      "grad_norm": 0.11976960918968543,
+      "learning_rate": 1.8750612633916997e-05,
+      "loss": 0.1144,
+      "step": 75
+    },
+    {
+      "epoch": 0.6485333333333333,
+      "grad_norm": 0.115805998298789,
+      "learning_rate": 1.8808135922807914e-05,
+      "loss": 0.1528,
+      "step": 76
+    },
+    {
+      "epoch": 0.6570666666666667,
+      "grad_norm": 0.10325948496697443,
+      "learning_rate": 1.8864907251724818e-05,
+      "loss": 0.1044,
+      "step": 77
+    },
+    {
+      "epoch": 0.6656,
+      "grad_norm": 0.09595064346541006,
+      "learning_rate": 1.8920946026904802e-05,
+      "loss": 0.1534,
+      "step": 78
+    },
+    {
+      "epoch": 0.6741333333333334,
+      "grad_norm": 0.08796742845240496,
+      "learning_rate": 1.8976270912904414e-05,
+      "loss": 0.1155,
+      "step": 79
+    },
+    {
+      "epoch": 0.6826666666666666,
+      "grad_norm": 0.08218991738379527,
+      "learning_rate": 1.9030899869919434e-05,
+      "loss": 0.1311,
+      "step": 80
+    },
+    {
+      "epoch": 0.6912,
+      "grad_norm": 0.08290815261109215,
+      "learning_rate": 1.9084850188786497e-05,
+      "loss": 0.11,
+      "step": 81
+    },
+    {
+      "epoch": 0.6997333333333333,
+      "grad_norm": 0.0794132180176064,
+      "learning_rate": 1.9138138523837165e-05,
+      "loss": 0.1135,
+      "step": 82
+    },
+    {
+      "epoch": 0.7082666666666667,
+      "grad_norm": 0.06934410705255296,
+      "learning_rate": 1.919078092376074e-05,
+      "loss": 0.109,
+      "step": 83
+    },
+    {
+      "epoch": 0.7168,
+      "grad_norm": 0.09000563031870593,
+      "learning_rate": 1.9242792860618813e-05,
+      "loss": 0.12,
+      "step": 84
+    },
+    {
+      "epoch": 0.7253333333333334,
+      "grad_norm": 0.1134042277879818,
+      "learning_rate": 1.929418925714293e-05,
+      "loss": 0.1223,
+      "step": 85
+    },
+    {
+      "epoch": 0.7338666666666667,
+      "grad_norm": 0.09118764690233076,
+      "learning_rate": 1.9344984512435673e-05,
+      "loss": 0.1459,
+      "step": 86
+    },
+    {
+      "epoch": 0.7424,
+      "grad_norm": 0.07873016754353963,
+      "learning_rate": 1.9395192526186183e-05,
+      "loss": 0.1422,
+      "step": 87
+    },
+    {
+      "epoch": 0.7509333333333333,
+      "grad_norm": 0.1796495874463076,
+      "learning_rate": 1.9444826721501687e-05,
+      "loss": 0.1291,
+      "step": 88
+    },
+    {
+      "epoch": 0.7594666666666666,
+      "grad_norm": 0.0679589944174269,
+      "learning_rate": 1.9493900066449125e-05,
+      "loss": 0.108,
+      "step": 89
+    },
+    {
+      "epoch": 0.768,
+      "grad_norm": 0.08174688574235538,
+      "learning_rate": 1.9542425094393246e-05,
+      "loss": 0.1081,
+      "step": 90
+    },
+    {
+      "epoch": 0.7765333333333333,
+      "grad_norm": 0.057137370501406756,
+      "learning_rate": 1.9590413923210934e-05,
+      "loss": 0.0934,
+      "step": 91
+    },
+    {
+      "epoch": 0.7850666666666667,
+      "grad_norm": 0.06578111924908255,
+      "learning_rate": 1.9637878273455555e-05,
+      "loss": 0.1085,
+      "step": 92
+    },
+    {
+      "epoch": 0.7936,
+      "grad_norm": 0.08945990540906254,
+      "learning_rate": 1.968482948553935e-05,
+      "loss": 0.1747,
+      "step": 93
+    },
+    {
+      "epoch": 0.8021333333333334,
+      "grad_norm": 0.06183863311044229,
+      "learning_rate": 1.9731278535996986e-05,
+      "loss": 0.1136,
+      "step": 94
+    },
+    {
+      "epoch": 0.8106666666666666,
+      "grad_norm": 0.05777899602544702,
+      "learning_rate": 1.9777236052888476e-05,
+      "loss": 0.0984,
+      "step": 95
+    },
+    {
+      "epoch": 0.8192,
+      "grad_norm": 0.08130851607693534,
+      "learning_rate": 1.9822712330395683e-05,
+      "loss": 0.187,
+      "step": 96
+    },
+    {
+      "epoch": 0.8277333333333333,
+      "grad_norm": 0.06426546202002927,
+      "learning_rate": 1.986771734266245e-05,
+      "loss": 0.1296,
+      "step": 97
+    },
+    {
+      "epoch": 0.8362666666666667,
+      "grad_norm": 0.069692313707994,
+      "learning_rate": 1.991226075692495e-05,
+      "loss": 0.1404,
+      "step": 98
+    },
+    {
+      "epoch": 0.8448,
+      "grad_norm": 0.05494542266886729,
+      "learning_rate": 1.9956351945975496e-05,
+      "loss": 0.116,
+      "step": 99
+    },
+    {
+      "epoch": 0.8533333333333334,
+      "grad_norm": 0.07571686966840627,
+      "learning_rate": 1.9999999999999998e-05,
+      "loss": 0.1539,
+      "step": 100
+    },
+    {
+      "epoch": 0.8618666666666667,
+      "grad_norm": 0.054351059117603705,
+      "learning_rate": 2e-05,
+      "loss": 0.1037,
+      "step": 101
+    },
+    {
+      "epoch": 0.8704,
+      "grad_norm": 0.06531899611551092,
+      "learning_rate": 2e-05,
+      "loss": 0.0827,
+      "step": 102
+    },
+    {
+      "epoch": 0.8789333333333333,
+      "grad_norm": 0.06131678646504652,
+      "learning_rate": 2e-05,
+      "loss": 0.1266,
+      "step": 103
+    },
+    {
+      "epoch": 0.8874666666666666,
+      "grad_norm": 0.06850220540661824,
+      "learning_rate": 2e-05,
+      "loss": 0.1456,
+      "step": 104
+    },
+    {
+      "epoch": 0.896,
+      "grad_norm": 0.05806908951252483,
+      "learning_rate": 2e-05,
+      "loss": 0.0954,
+      "step": 105
+    },
+    {
+      "epoch": 0.9045333333333333,
+      "grad_norm": 0.06503642452033717,
+      "learning_rate": 2e-05,
+      "loss": 0.1417,
+      "step": 106
+    },
+    {
+      "epoch": 0.9130666666666667,
+      "grad_norm": 0.050486271853277066,
+      "learning_rate": 2e-05,
+      "loss": 0.0959,
+      "step": 107
+    },
+    {
+      "epoch": 0.9216,
+      "grad_norm": 0.07746063813802379,
+      "learning_rate": 2e-05,
+      "loss": 0.1256,
+      "step": 108
+    },
+    {
+      "epoch": 0.9301333333333334,
+      "grad_norm": 0.051231172380840004,
+      "learning_rate": 2e-05,
+      "loss": 0.1116,
+      "step": 109
+    },
+    {
+      "epoch": 0.9386666666666666,
+      "grad_norm": 0.056296443557859455,
+      "learning_rate": 2e-05,
+      "loss": 0.1056,
+      "step": 110
+    },
+    {
+      "epoch": 0.9472,
+      "grad_norm": 0.05058663240713958,
+      "learning_rate": 2e-05,
+      "loss": 0.0971,
+      "step": 111
+    },
+    {
+      "epoch": 0.9557333333333333,
+      "grad_norm": 0.05532886570130611,
+      "learning_rate": 2e-05,
+      "loss": 0.1086,
+      "step": 112
+    },
+    {
+      "epoch": 0.9642666666666667,
+      "grad_norm": 0.05327811326654907,
+      "learning_rate": 2e-05,
+      "loss": 0.0989,
+      "step": 113
+    },
+    {
+      "epoch": 0.9728,
+      "grad_norm": 0.05663279364147864,
+      "learning_rate": 2e-05,
+      "loss": 0.0958,
+      "step": 114
+    },
+    {
+      "epoch": 0.9813333333333333,
+      "grad_norm": 0.04930904541225805,
+      "learning_rate": 2e-05,
+      "loss": 0.0887,
+      "step": 115
+    },
+    {
+      "epoch": 0.9898666666666667,
+      "grad_norm": 0.06096947951115022,
+      "learning_rate": 2e-05,
+      "loss": 0.106,
+      "step": 116
+    },
+    {
+      "epoch": 0.9984,
+      "grad_norm": 0.050092322361182495,
+      "learning_rate": 2e-05,
+      "loss": 0.0931,
+      "step": 117
+    },
+    {
+      "epoch": 1.0069333333333332,
+      "grad_norm": 0.04980408443758999,
+      "learning_rate": 2e-05,
+      "loss": 0.0955,
+      "step": 118
+    },
+    {
+      "epoch": 1.0154666666666667,
+      "grad_norm": 0.051183082721834305,
+      "learning_rate": 2e-05,
+      "loss": 0.1049,
+      "step": 119
+    },
+    {
+      "epoch": 1.024,
+      "grad_norm": 0.04332220265802814,
+      "learning_rate": 2e-05,
+      "loss": 0.0983,
+      "step": 120
+    },
+    {
+      "epoch": 1.0325333333333333,
+      "grad_norm": 0.07211731499677299,
+      "learning_rate": 2e-05,
+      "loss": 0.1386,
+      "step": 121
+    },
+    {
+      "epoch": 1.0410666666666666,
+      "grad_norm": 0.06550870223740553,
+      "learning_rate": 2e-05,
+      "loss": 0.1334,
+      "step": 122
+    },
+    {
+      "epoch": 1.0496,
+      "grad_norm": 0.05331839690767287,
+      "learning_rate": 2e-05,
+      "loss": 0.1014,
+      "step": 123
+    },
+    {
+      "epoch": 1.0581333333333334,
+      "grad_norm": 0.05227685628767905,
+      "learning_rate": 2e-05,
+      "loss": 0.1098,
+      "step": 124
+    },
+    {
+      "epoch": 1.0666666666666667,
+      "grad_norm": 0.07641693882491171,
+      "learning_rate": 2e-05,
+      "loss": 0.1127,
+      "step": 125
+    },
+    {
+      "epoch": 1.0752,
+      "grad_norm": 0.052835367770791786,
+      "learning_rate": 2e-05,
+      "loss": 0.1231,
+      "step": 126
+    },
+    {
+      "epoch": 1.0837333333333334,
+      "grad_norm": 0.07520329755025788,
+      "learning_rate": 2e-05,
+      "loss": 0.085,
+      "step": 127
+    },
+    {
+      "epoch": 1.0922666666666667,
+      "grad_norm": 0.07670066152157425,
+      "learning_rate": 2e-05,
+      "loss": 0.1071,
+      "step": 128
+    },
+    {
+      "epoch": 1.1008,
+      "grad_norm": 0.052832906560645154,
+      "learning_rate": 2e-05,
+      "loss": 0.1093,
+      "step": 129
+    },
+    {
+      "epoch": 1.1093333333333333,
+      "grad_norm": 0.06573889037311398,
+      "learning_rate": 2e-05,
+      "loss": 0.1193,
+      "step": 130
+    },
+    {
+      "epoch": 1.1178666666666666,
+      "grad_norm": 0.05175471296566334,
+      "learning_rate": 2e-05,
+      "loss": 0.1184,
+      "step": 131
+    },
+    {
+      "epoch": 1.1264,
+      "grad_norm": 0.05912231419793496,
+      "learning_rate": 2e-05,
+      "loss": 0.1154,
+      "step": 132
+    },
+    {
+      "epoch": 1.1349333333333333,
+      "grad_norm": 0.04899140475981105,
+      "learning_rate": 2e-05,
+      "loss": 0.0957,
+      "step": 133
+    },
+    {
+      "epoch": 1.1434666666666666,
+      "grad_norm": 0.05939916939142137,
+      "learning_rate": 2e-05,
+      "loss": 0.0979,
+      "step": 134
+    },
+    {
+      "epoch": 1.152,
+      "grad_norm": 0.0516819217599706,
+      "learning_rate": 2e-05,
+      "loss": 0.0834,
+      "step": 135
+    },
+    {
+      "epoch": 1.1605333333333334,
+      "grad_norm": 0.05456440346454737,
+      "learning_rate": 2e-05,
+      "loss": 0.1183,
+      "step": 136
+    },
+    {
+      "epoch": 1.1690666666666667,
+      "grad_norm": 0.059906464476343235,
+      "learning_rate": 2e-05,
+      "loss": 0.1048,
+      "step": 137
+    },
+    {
+      "epoch": 1.1776,
+      "grad_norm": 0.0720112680204319,
+      "learning_rate": 2e-05,
+      "loss": 0.1168,
+      "step": 138
+    },
+    {
+      "epoch": 1.1861333333333333,
+      "grad_norm": 0.04940202805828527,
+      "learning_rate": 2e-05,
+      "loss": 0.0948,
+      "step": 139
+    },
+    {
+      "epoch": 1.1946666666666665,
+      "grad_norm": 0.060088609545130046,
+      "learning_rate": 2e-05,
+      "loss": 0.0952,
+      "step": 140
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 351,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 3,
+  "save_steps": 20,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 1.9872207486907843e+18,
+  "train_batch_size": 16,
+  "trial_name": null,
+  "trial_params": null
+}
diff --git a/checkpoint-140/training_args.bin b/checkpoint-140/training_args.bin
new file mode 100644
index 0000000000000000000000000000000000000000..8ac842d29fc6ad3e132cb7208dda5b61dd1079ee
--- /dev/null
+++ b/checkpoint-140/training_args.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aa4edad1da5557fcd52a6da980443588016cead6f0444a3562cfa68029c66a04
+size 6840
diff --git a/checkpoint-140/zero_to_fp32.py b/checkpoint-140/zero_to_fp32.py
new file mode 100644
index 0000000000000000000000000000000000000000..49b846633d6eb1e836e34681e44033581f4edb7b
--- /dev/null
+++ b/checkpoint-140/zero_to_fp32.py
@@ -0,0 +1,592 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example: python zero_to_fp32.py . pytorch_model.bin
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+
+    total_files = len(files)
+    state_dicts = []
+    for f in files:
+        state_dict = torch.load(f, map_location=device)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    if zero_stage <= 2:
+        fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    elif zero_stage == 3:
+        # if there is more than one param group, there will be multiple flattened tensors - one
+        # flattened tensor per group - for simplicity merge them into a single tensor
+        #
+        # XXX: could make the script more memory efficient for when there are multiple groups - it
+        # will require matching the sub-lists of param_shapes for each param group flattened tensor
+
+        fp32_flat_groups = [
+            torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
+        ]
+
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    for name, shape in param_shapes.items():
+
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # XXX: memory usage doubles here
+        state_dict[name] = torch.cat(
+            tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
+            0).narrow(0, 0, unpartitioned_numel).view(shape)
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    Note: this approach may not work if your application doesn't have sufficient free CPU memory and
+    you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+    """
+
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+    print(f"Saving fp32 state dict to {output_file}")
+    torch.save(state_dict, output_file)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument(
+        "output_file",
+        type=str,
+        help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file, tag=args.tag)
diff --git a/checkpoint-160/README.md b/checkpoint-160/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..e7d4f706aeefc6bb2aa84e963870f345a14c4af9
--- /dev/null
+++ b/checkpoint-160/README.md
@@ -0,0 +1,202 @@
+---
+library_name: peft
+base_model: ../ckpts/Meta-Llama-3-8B-Instruct
+---
+
+# 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 0.11.1
\ No newline at end of file
diff --git a/checkpoint-160/adapter_config.json b/checkpoint-160/adapter_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..ad16c073ed8946d227350e8fe4e99d689a8ede54
--- /dev/null
+++ b/checkpoint-160/adapter_config.json
@@ -0,0 +1,35 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "../ckpts/Meta-Llama-3-8B-Instruct",
+  "bias": "none",
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 16,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "down_proj",
+    "gate_proj",
+    "q_proj",
+    "lm_head",
+    "o_proj",
+    "v_proj",
+    "up_proj",
+    "k_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "use_dora": false,
+  "use_rslora": false
+}
\ No newline at end of file
diff --git a/checkpoint-160/adapter_model.safetensors b/checkpoint-160/adapter_model.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..813aaf9317a1f1800e56036b6f520e6aa25eafda
--- /dev/null
+++ b/checkpoint-160/adapter_model.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6b6b5a540a8357ce98d65b484037f08c6e0295863490398dfd6671ec06a2916b
+size 1138856856
diff --git a/checkpoint-160/global_step160/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt b/checkpoint-160/global_step160/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..6617b79eaaaa719a52a57a9e3d16913739827008
--- /dev/null
+++ b/checkpoint-160/global_step160/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:f1a136bd5515088baf044edf425835b86e73f43dc41097181483574d0a3ec311
+size 528781328
diff --git a/checkpoint-160/global_step160/mp_rank_00_model_states.pt b/checkpoint-160/global_step160/mp_rank_00_model_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..f5d86ace886f802632302817ee6f555ea016c2f8
--- /dev/null
+++ b/checkpoint-160/global_step160/mp_rank_00_model_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e64b47cf0673bd0721e3b682a0e4fd592936923b3a1fc6966a1236c9b9f1a4ec
+size 199905337
diff --git a/checkpoint-160/latest b/checkpoint-160/latest
new file mode 100644
index 0000000000000000000000000000000000000000..3df30ded267d950ff3ca04cffb9660be12079ca6
--- /dev/null
+++ b/checkpoint-160/latest
@@ -0,0 +1 @@
+global_step160
\ No newline at end of file
diff --git a/checkpoint-160/rng_state.pth b/checkpoint-160/rng_state.pth
new file mode 100644
index 0000000000000000000000000000000000000000..809cc01625a9fe2af3abe387e6eeb078e560a412
--- /dev/null
+++ b/checkpoint-160/rng_state.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cefbcec9b3679fcdb9b96cdb9ed85fb37b575748c126094c451cc80787e18e73
+size 14244
diff --git a/checkpoint-160/trainer_state.json b/checkpoint-160/trainer_state.json
new file mode 100644
index 0000000000000000000000000000000000000000..8bbf024f29df623a4b6c5971930774e7d6804c77
--- /dev/null
+++ b/checkpoint-160/trainer_state.json
@@ -0,0 +1,1153 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 1.3653333333333333,
+  "eval_steps": 500,
+  "global_step": 160,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.008533333333333334,
+      "grad_norm": 100.21848203113535,
+      "learning_rate": 0.0,
+      "loss": 7.1962,
+      "step": 1
+    },
+    {
+      "epoch": 0.017066666666666667,
+      "grad_norm": 96.43006188910957,
+      "learning_rate": 3.0102999566398115e-06,
+      "loss": 6.9414,
+      "step": 2
+    },
+    {
+      "epoch": 0.0256,
+      "grad_norm": 97.35803466618715,
+      "learning_rate": 4.771212547196624e-06,
+      "loss": 7.0102,
+      "step": 3
+    },
+    {
+      "epoch": 0.034133333333333335,
+      "grad_norm": 95.14837816372646,
+      "learning_rate": 6.020599913279623e-06,
+      "loss": 6.5295,
+      "step": 4
+    },
+    {
+      "epoch": 0.042666666666666665,
+      "grad_norm": 91.76544275692784,
+      "learning_rate": 6.989700043360187e-06,
+      "loss": 6.4806,
+      "step": 5
+    },
+    {
+      "epoch": 0.0512,
+      "grad_norm": 84.4494318688335,
+      "learning_rate": 7.781512503836437e-06,
+      "loss": 6.4194,
+      "step": 6
+    },
+    {
+      "epoch": 0.05973333333333333,
+      "grad_norm": 71.37977490595638,
+      "learning_rate": 8.450980400142568e-06,
+      "loss": 5.4953,
+      "step": 7
+    },
+    {
+      "epoch": 0.06826666666666667,
+      "grad_norm": 49.31153456754566,
+      "learning_rate": 9.030899869919434e-06,
+      "loss": 5.4123,
+      "step": 8
+    },
+    {
+      "epoch": 0.0768,
+      "grad_norm": 20.37296364560341,
+      "learning_rate": 9.542425094393249e-06,
+      "loss": 5.2334,
+      "step": 9
+    },
+    {
+      "epoch": 0.08533333333333333,
+      "grad_norm": 7.968467079076881,
+      "learning_rate": 9.999999999999999e-06,
+      "loss": 5.0282,
+      "step": 10
+    },
+    {
+      "epoch": 0.09386666666666667,
+      "grad_norm": 3.559446532055649,
+      "learning_rate": 1.041392685158225e-05,
+      "loss": 4.612,
+      "step": 11
+    },
+    {
+      "epoch": 0.1024,
+      "grad_norm": 3.5528846947995674,
+      "learning_rate": 1.0791812460476248e-05,
+      "loss": 4.9475,
+      "step": 12
+    },
+    {
+      "epoch": 0.11093333333333333,
+      "grad_norm": 3.541968897471334,
+      "learning_rate": 1.1139433523068365e-05,
+      "loss": 4.2777,
+      "step": 13
+    },
+    {
+      "epoch": 0.11946666666666667,
+      "grad_norm": 3.54718070036198,
+      "learning_rate": 1.1461280356782378e-05,
+      "loss": 4.3507,
+      "step": 14
+    },
+    {
+      "epoch": 0.128,
+      "grad_norm": 3.8632334830606747,
+      "learning_rate": 1.1760912590556813e-05,
+      "loss": 4.5364,
+      "step": 15
+    },
+    {
+      "epoch": 0.13653333333333334,
+      "grad_norm": 3.6637424744054004,
+      "learning_rate": 1.2041199826559246e-05,
+      "loss": 3.9672,
+      "step": 16
+    },
+    {
+      "epoch": 0.14506666666666668,
+      "grad_norm": 3.919802904818311,
+      "learning_rate": 1.230448921378274e-05,
+      "loss": 4.0618,
+      "step": 17
+    },
+    {
+      "epoch": 0.1536,
+      "grad_norm": 4.71904950738746,
+      "learning_rate": 1.2552725051033058e-05,
+      "loss": 4.6656,
+      "step": 18
+    },
+    {
+      "epoch": 0.16213333333333332,
+      "grad_norm": 4.6656317698690835,
+      "learning_rate": 1.2787536009528288e-05,
+      "loss": 4.1131,
+      "step": 19
+    },
+    {
+      "epoch": 0.17066666666666666,
+      "grad_norm": 5.145138692367417,
+      "learning_rate": 1.301029995663981e-05,
+      "loss": 4.0989,
+      "step": 20
+    },
+    {
+      "epoch": 0.1792,
+      "grad_norm": 5.923538014759818,
+      "learning_rate": 1.3222192947339192e-05,
+      "loss": 4.4991,
+      "step": 21
+    },
+    {
+      "epoch": 0.18773333333333334,
+      "grad_norm": 5.941056962941364,
+      "learning_rate": 1.3424226808222062e-05,
+      "loss": 4.0836,
+      "step": 22
+    },
+    {
+      "epoch": 0.19626666666666667,
+      "grad_norm": 6.171026012117947,
+      "learning_rate": 1.3617278360175927e-05,
+      "loss": 3.6861,
+      "step": 23
+    },
+    {
+      "epoch": 0.2048,
+      "grad_norm": 7.130542138930838,
+      "learning_rate": 1.380211241711606e-05,
+      "loss": 4.0958,
+      "step": 24
+    },
+    {
+      "epoch": 0.21333333333333335,
+      "grad_norm": 7.328837606110418,
+      "learning_rate": 1.3979400086720374e-05,
+      "loss": 3.9524,
+      "step": 25
+    },
+    {
+      "epoch": 0.22186666666666666,
+      "grad_norm": 6.923489005711429,
+      "learning_rate": 1.4149733479708178e-05,
+      "loss": 3.6062,
+      "step": 26
+    },
+    {
+      "epoch": 0.2304,
+      "grad_norm": 6.409498926059221,
+      "learning_rate": 1.4313637641589872e-05,
+      "loss": 3.2034,
+      "step": 27
+    },
+    {
+      "epoch": 0.23893333333333333,
+      "grad_norm": 5.781628405584682,
+      "learning_rate": 1.4471580313422191e-05,
+      "loss": 2.8158,
+      "step": 28
+    },
+    {
+      "epoch": 0.24746666666666667,
+      "grad_norm": 6.2927590068243315,
+      "learning_rate": 1.4623979978989559e-05,
+      "loss": 2.9803,
+      "step": 29
+    },
+    {
+      "epoch": 0.256,
+      "grad_norm": 6.103844678630006,
+      "learning_rate": 1.4771212547196623e-05,
+      "loss": 2.847,
+      "step": 30
+    },
+    {
+      "epoch": 0.26453333333333334,
+      "grad_norm": 7.656341925867144,
+      "learning_rate": 1.4913616938342726e-05,
+      "loss": 3.0907,
+      "step": 31
+    },
+    {
+      "epoch": 0.2730666666666667,
+      "grad_norm": 6.324242877501844,
+      "learning_rate": 1.5051499783199059e-05,
+      "loss": 2.3467,
+      "step": 32
+    },
+    {
+      "epoch": 0.2816,
+      "grad_norm": 7.606313101162739,
+      "learning_rate": 1.5185139398778874e-05,
+      "loss": 2.5292,
+      "step": 33
+    },
+    {
+      "epoch": 0.29013333333333335,
+      "grad_norm": 8.553792493849265,
+      "learning_rate": 1.531478917042255e-05,
+      "loss": 2.4547,
+      "step": 34
+    },
+    {
+      "epoch": 0.2986666666666667,
+      "grad_norm": 8.483368703272543,
+      "learning_rate": 1.5440680443502753e-05,
+      "loss": 2.1956,
+      "step": 35
+    },
+    {
+      "epoch": 0.3072,
+      "grad_norm": 8.364739352838077,
+      "learning_rate": 1.5563025007672873e-05,
+      "loss": 1.8552,
+      "step": 36
+    },
+    {
+      "epoch": 0.3157333333333333,
+      "grad_norm": 9.37663682000104,
+      "learning_rate": 1.5682017240669948e-05,
+      "loss": 1.9228,
+      "step": 37
+    },
+    {
+      "epoch": 0.32426666666666665,
+      "grad_norm": 9.823047193440066,
+      "learning_rate": 1.57978359661681e-05,
+      "loss": 1.7033,
+      "step": 38
+    },
+    {
+      "epoch": 0.3328,
+      "grad_norm": 9.692618512955894,
+      "learning_rate": 1.591064607026499e-05,
+      "loss": 1.3768,
+      "step": 39
+    },
+    {
+      "epoch": 0.3413333333333333,
+      "grad_norm": 9.08889021911031,
+      "learning_rate": 1.6020599913279622e-05,
+      "loss": 1.3015,
+      "step": 40
+    },
+    {
+      "epoch": 0.34986666666666666,
+      "grad_norm": 8.081534221516058,
+      "learning_rate": 1.6127838567197353e-05,
+      "loss": 0.9228,
+      "step": 41
+    },
+    {
+      "epoch": 0.3584,
+      "grad_norm": 6.238638048950311,
+      "learning_rate": 1.6232492903979005e-05,
+      "loss": 0.7267,
+      "step": 42
+    },
+    {
+      "epoch": 0.36693333333333333,
+      "grad_norm": 3.4058036861773604,
+      "learning_rate": 1.6334684555795865e-05,
+      "loss": 0.5875,
+      "step": 43
+    },
+    {
+      "epoch": 0.37546666666666667,
+      "grad_norm": 2.079163829467713,
+      "learning_rate": 1.6434526764861872e-05,
+      "loss": 0.6355,
+      "step": 44
+    },
+    {
+      "epoch": 0.384,
+      "grad_norm": 1.5597487833024746,
+      "learning_rate": 1.6532125137753435e-05,
+      "loss": 0.5106,
+      "step": 45
+    },
+    {
+      "epoch": 0.39253333333333335,
+      "grad_norm": 2.491689602375256,
+      "learning_rate": 1.662757831681574e-05,
+      "loss": 0.6454,
+      "step": 46
+    },
+    {
+      "epoch": 0.4010666666666667,
+      "grad_norm": 2.010880438195854,
+      "learning_rate": 1.672097857935717e-05,
+      "loss": 0.4757,
+      "step": 47
+    },
+    {
+      "epoch": 0.4096,
+      "grad_norm": 1.9452805114322096,
+      "learning_rate": 1.681241237375587e-05,
+      "loss": 0.4133,
+      "step": 48
+    },
+    {
+      "epoch": 0.41813333333333336,
+      "grad_norm": 1.7620848552306103,
+      "learning_rate": 1.6901960800285137e-05,
+      "loss": 0.4004,
+      "step": 49
+    },
+    {
+      "epoch": 0.4266666666666667,
+      "grad_norm": 1.278224489774809,
+      "learning_rate": 1.6989700043360187e-05,
+      "loss": 0.3523,
+      "step": 50
+    },
+    {
+      "epoch": 0.4352,
+      "grad_norm": 1.6151354758303231,
+      "learning_rate": 1.7075701760979363e-05,
+      "loss": 0.4317,
+      "step": 51
+    },
+    {
+      "epoch": 0.4437333333333333,
+      "grad_norm": 1.3451396055695035,
+      "learning_rate": 1.716003343634799e-05,
+      "loss": 0.3474,
+      "step": 52
+    },
+    {
+      "epoch": 0.45226666666666665,
+      "grad_norm": 1.6814977782362666,
+      "learning_rate": 1.724275869600789e-05,
+      "loss": 0.3706,
+      "step": 53
+    },
+    {
+      "epoch": 0.4608,
+      "grad_norm": 1.1682442432688667,
+      "learning_rate": 1.7323937598229687e-05,
+      "loss": 0.3488,
+      "step": 54
+    },
+    {
+      "epoch": 0.4693333333333333,
+      "grad_norm": 0.8839814540462471,
+      "learning_rate": 1.7403626894942437e-05,
+      "loss": 0.293,
+      "step": 55
+    },
+    {
+      "epoch": 0.47786666666666666,
+      "grad_norm": 0.7974304806295485,
+      "learning_rate": 1.7481880270062003e-05,
+      "loss": 0.2717,
+      "step": 56
+    },
+    {
+      "epoch": 0.4864,
+      "grad_norm": 0.6232203657585239,
+      "learning_rate": 1.7558748556724913e-05,
+      "loss": 0.1741,
+      "step": 57
+    },
+    {
+      "epoch": 0.49493333333333334,
+      "grad_norm": 0.6850688604563008,
+      "learning_rate": 1.763427993562937e-05,
+      "loss": 0.228,
+      "step": 58
+    },
+    {
+      "epoch": 0.5034666666666666,
+      "grad_norm": 0.5923826384300431,
+      "learning_rate": 1.7708520116421443e-05,
+      "loss": 0.2131,
+      "step": 59
+    },
+    {
+      "epoch": 0.512,
+      "grad_norm": 0.489754430485032,
+      "learning_rate": 1.7781512503836432e-05,
+      "loss": 0.165,
+      "step": 60
+    },
+    {
+      "epoch": 0.5205333333333333,
+      "grad_norm": 0.5280940052395061,
+      "learning_rate": 1.7853298350107667e-05,
+      "loss": 0.1658,
+      "step": 61
+    },
+    {
+      "epoch": 0.5290666666666667,
+      "grad_norm": 0.4750905992036739,
+      "learning_rate": 1.7923916894982537e-05,
+      "loss": 0.1438,
+      "step": 62
+    },
+    {
+      "epoch": 0.5376,
+      "grad_norm": 0.4162998392722401,
+      "learning_rate": 1.7993405494535815e-05,
+      "loss": 0.1555,
+      "step": 63
+    },
+    {
+      "epoch": 0.5461333333333334,
+      "grad_norm": 0.26863266246370443,
+      "learning_rate": 1.806179973983887e-05,
+      "loss": 0.1323,
+      "step": 64
+    },
+    {
+      "epoch": 0.5546666666666666,
+      "grad_norm": 0.26534221125601215,
+      "learning_rate": 1.8129133566428553e-05,
+      "loss": 0.1671,
+      "step": 65
+    },
+    {
+      "epoch": 0.5632,
+      "grad_norm": 0.2548662962257576,
+      "learning_rate": 1.8195439355418686e-05,
+      "loss": 0.1308,
+      "step": 66
+    },
+    {
+      "epoch": 0.5717333333333333,
+      "grad_norm": 0.18045394638578796,
+      "learning_rate": 1.8260748027008263e-05,
+      "loss": 0.1262,
+      "step": 67
+    },
+    {
+      "epoch": 0.5802666666666667,
+      "grad_norm": 0.17070388073612064,
+      "learning_rate": 1.8325089127062364e-05,
+      "loss": 0.1192,
+      "step": 68
+    },
+    {
+      "epoch": 0.5888,
+      "grad_norm": 0.1531381679244776,
+      "learning_rate": 1.8388490907372553e-05,
+      "loss": 0.1274,
+      "step": 69
+    },
+    {
+      "epoch": 0.5973333333333334,
+      "grad_norm": 0.18196434993687946,
+      "learning_rate": 1.8450980400142568e-05,
+      "loss": 0.1375,
+      "step": 70
+    },
+    {
+      "epoch": 0.6058666666666667,
+      "grad_norm": 0.15324416972951205,
+      "learning_rate": 1.8512583487190752e-05,
+      "loss": 0.1599,
+      "step": 71
+    },
+    {
+      "epoch": 0.6144,
+      "grad_norm": 0.10884462064503801,
+      "learning_rate": 1.857332496431268e-05,
+      "loss": 0.1041,
+      "step": 72
+    },
+    {
+      "epoch": 0.6229333333333333,
+      "grad_norm": 0.12915133528192668,
+      "learning_rate": 1.8633228601204555e-05,
+      "loss": 0.1406,
+      "step": 73
+    },
+    {
+      "epoch": 0.6314666666666666,
+      "grad_norm": 0.12553425699952878,
+      "learning_rate": 1.8692317197309763e-05,
+      "loss": 0.1256,
+      "step": 74
+    },
+    {
+      "epoch": 0.64,
+      "grad_norm": 0.11976960918968543,
+      "learning_rate": 1.8750612633916997e-05,
+      "loss": 0.1144,
+      "step": 75
+    },
+    {
+      "epoch": 0.6485333333333333,
+      "grad_norm": 0.115805998298789,
+      "learning_rate": 1.8808135922807914e-05,
+      "loss": 0.1528,
+      "step": 76
+    },
+    {
+      "epoch": 0.6570666666666667,
+      "grad_norm": 0.10325948496697443,
+      "learning_rate": 1.8864907251724818e-05,
+      "loss": 0.1044,
+      "step": 77
+    },
+    {
+      "epoch": 0.6656,
+      "grad_norm": 0.09595064346541006,
+      "learning_rate": 1.8920946026904802e-05,
+      "loss": 0.1534,
+      "step": 78
+    },
+    {
+      "epoch": 0.6741333333333334,
+      "grad_norm": 0.08796742845240496,
+      "learning_rate": 1.8976270912904414e-05,
+      "loss": 0.1155,
+      "step": 79
+    },
+    {
+      "epoch": 0.6826666666666666,
+      "grad_norm": 0.08218991738379527,
+      "learning_rate": 1.9030899869919434e-05,
+      "loss": 0.1311,
+      "step": 80
+    },
+    {
+      "epoch": 0.6912,
+      "grad_norm": 0.08290815261109215,
+      "learning_rate": 1.9084850188786497e-05,
+      "loss": 0.11,
+      "step": 81
+    },
+    {
+      "epoch": 0.6997333333333333,
+      "grad_norm": 0.0794132180176064,
+      "learning_rate": 1.9138138523837165e-05,
+      "loss": 0.1135,
+      "step": 82
+    },
+    {
+      "epoch": 0.7082666666666667,
+      "grad_norm": 0.06934410705255296,
+      "learning_rate": 1.919078092376074e-05,
+      "loss": 0.109,
+      "step": 83
+    },
+    {
+      "epoch": 0.7168,
+      "grad_norm": 0.09000563031870593,
+      "learning_rate": 1.9242792860618813e-05,
+      "loss": 0.12,
+      "step": 84
+    },
+    {
+      "epoch": 0.7253333333333334,
+      "grad_norm": 0.1134042277879818,
+      "learning_rate": 1.929418925714293e-05,
+      "loss": 0.1223,
+      "step": 85
+    },
+    {
+      "epoch": 0.7338666666666667,
+      "grad_norm": 0.09118764690233076,
+      "learning_rate": 1.9344984512435673e-05,
+      "loss": 0.1459,
+      "step": 86
+    },
+    {
+      "epoch": 0.7424,
+      "grad_norm": 0.07873016754353963,
+      "learning_rate": 1.9395192526186183e-05,
+      "loss": 0.1422,
+      "step": 87
+    },
+    {
+      "epoch": 0.7509333333333333,
+      "grad_norm": 0.1796495874463076,
+      "learning_rate": 1.9444826721501687e-05,
+      "loss": 0.1291,
+      "step": 88
+    },
+    {
+      "epoch": 0.7594666666666666,
+      "grad_norm": 0.0679589944174269,
+      "learning_rate": 1.9493900066449125e-05,
+      "loss": 0.108,
+      "step": 89
+    },
+    {
+      "epoch": 0.768,
+      "grad_norm": 0.08174688574235538,
+      "learning_rate": 1.9542425094393246e-05,
+      "loss": 0.1081,
+      "step": 90
+    },
+    {
+      "epoch": 0.7765333333333333,
+      "grad_norm": 0.057137370501406756,
+      "learning_rate": 1.9590413923210934e-05,
+      "loss": 0.0934,
+      "step": 91
+    },
+    {
+      "epoch": 0.7850666666666667,
+      "grad_norm": 0.06578111924908255,
+      "learning_rate": 1.9637878273455555e-05,
+      "loss": 0.1085,
+      "step": 92
+    },
+    {
+      "epoch": 0.7936,
+      "grad_norm": 0.08945990540906254,
+      "learning_rate": 1.968482948553935e-05,
+      "loss": 0.1747,
+      "step": 93
+    },
+    {
+      "epoch": 0.8021333333333334,
+      "grad_norm": 0.06183863311044229,
+      "learning_rate": 1.9731278535996986e-05,
+      "loss": 0.1136,
+      "step": 94
+    },
+    {
+      "epoch": 0.8106666666666666,
+      "grad_norm": 0.05777899602544702,
+      "learning_rate": 1.9777236052888476e-05,
+      "loss": 0.0984,
+      "step": 95
+    },
+    {
+      "epoch": 0.8192,
+      "grad_norm": 0.08130851607693534,
+      "learning_rate": 1.9822712330395683e-05,
+      "loss": 0.187,
+      "step": 96
+    },
+    {
+      "epoch": 0.8277333333333333,
+      "grad_norm": 0.06426546202002927,
+      "learning_rate": 1.986771734266245e-05,
+      "loss": 0.1296,
+      "step": 97
+    },
+    {
+      "epoch": 0.8362666666666667,
+      "grad_norm": 0.069692313707994,
+      "learning_rate": 1.991226075692495e-05,
+      "loss": 0.1404,
+      "step": 98
+    },
+    {
+      "epoch": 0.8448,
+      "grad_norm": 0.05494542266886729,
+      "learning_rate": 1.9956351945975496e-05,
+      "loss": 0.116,
+      "step": 99
+    },
+    {
+      "epoch": 0.8533333333333334,
+      "grad_norm": 0.07571686966840627,
+      "learning_rate": 1.9999999999999998e-05,
+      "loss": 0.1539,
+      "step": 100
+    },
+    {
+      "epoch": 0.8618666666666667,
+      "grad_norm": 0.054351059117603705,
+      "learning_rate": 2e-05,
+      "loss": 0.1037,
+      "step": 101
+    },
+    {
+      "epoch": 0.8704,
+      "grad_norm": 0.06531899611551092,
+      "learning_rate": 2e-05,
+      "loss": 0.0827,
+      "step": 102
+    },
+    {
+      "epoch": 0.8789333333333333,
+      "grad_norm": 0.06131678646504652,
+      "learning_rate": 2e-05,
+      "loss": 0.1266,
+      "step": 103
+    },
+    {
+      "epoch": 0.8874666666666666,
+      "grad_norm": 0.06850220540661824,
+      "learning_rate": 2e-05,
+      "loss": 0.1456,
+      "step": 104
+    },
+    {
+      "epoch": 0.896,
+      "grad_norm": 0.05806908951252483,
+      "learning_rate": 2e-05,
+      "loss": 0.0954,
+      "step": 105
+    },
+    {
+      "epoch": 0.9045333333333333,
+      "grad_norm": 0.06503642452033717,
+      "learning_rate": 2e-05,
+      "loss": 0.1417,
+      "step": 106
+    },
+    {
+      "epoch": 0.9130666666666667,
+      "grad_norm": 0.050486271853277066,
+      "learning_rate": 2e-05,
+      "loss": 0.0959,
+      "step": 107
+    },
+    {
+      "epoch": 0.9216,
+      "grad_norm": 0.07746063813802379,
+      "learning_rate": 2e-05,
+      "loss": 0.1256,
+      "step": 108
+    },
+    {
+      "epoch": 0.9301333333333334,
+      "grad_norm": 0.051231172380840004,
+      "learning_rate": 2e-05,
+      "loss": 0.1116,
+      "step": 109
+    },
+    {
+      "epoch": 0.9386666666666666,
+      "grad_norm": 0.056296443557859455,
+      "learning_rate": 2e-05,
+      "loss": 0.1056,
+      "step": 110
+    },
+    {
+      "epoch": 0.9472,
+      "grad_norm": 0.05058663240713958,
+      "learning_rate": 2e-05,
+      "loss": 0.0971,
+      "step": 111
+    },
+    {
+      "epoch": 0.9557333333333333,
+      "grad_norm": 0.05532886570130611,
+      "learning_rate": 2e-05,
+      "loss": 0.1086,
+      "step": 112
+    },
+    {
+      "epoch": 0.9642666666666667,
+      "grad_norm": 0.05327811326654907,
+      "learning_rate": 2e-05,
+      "loss": 0.0989,
+      "step": 113
+    },
+    {
+      "epoch": 0.9728,
+      "grad_norm": 0.05663279364147864,
+      "learning_rate": 2e-05,
+      "loss": 0.0958,
+      "step": 114
+    },
+    {
+      "epoch": 0.9813333333333333,
+      "grad_norm": 0.04930904541225805,
+      "learning_rate": 2e-05,
+      "loss": 0.0887,
+      "step": 115
+    },
+    {
+      "epoch": 0.9898666666666667,
+      "grad_norm": 0.06096947951115022,
+      "learning_rate": 2e-05,
+      "loss": 0.106,
+      "step": 116
+    },
+    {
+      "epoch": 0.9984,
+      "grad_norm": 0.050092322361182495,
+      "learning_rate": 2e-05,
+      "loss": 0.0931,
+      "step": 117
+    },
+    {
+      "epoch": 1.0069333333333332,
+      "grad_norm": 0.04980408443758999,
+      "learning_rate": 2e-05,
+      "loss": 0.0955,
+      "step": 118
+    },
+    {
+      "epoch": 1.0154666666666667,
+      "grad_norm": 0.051183082721834305,
+      "learning_rate": 2e-05,
+      "loss": 0.1049,
+      "step": 119
+    },
+    {
+      "epoch": 1.024,
+      "grad_norm": 0.04332220265802814,
+      "learning_rate": 2e-05,
+      "loss": 0.0983,
+      "step": 120
+    },
+    {
+      "epoch": 1.0325333333333333,
+      "grad_norm": 0.07211731499677299,
+      "learning_rate": 2e-05,
+      "loss": 0.1386,
+      "step": 121
+    },
+    {
+      "epoch": 1.0410666666666666,
+      "grad_norm": 0.06550870223740553,
+      "learning_rate": 2e-05,
+      "loss": 0.1334,
+      "step": 122
+    },
+    {
+      "epoch": 1.0496,
+      "grad_norm": 0.05331839690767287,
+      "learning_rate": 2e-05,
+      "loss": 0.1014,
+      "step": 123
+    },
+    {
+      "epoch": 1.0581333333333334,
+      "grad_norm": 0.05227685628767905,
+      "learning_rate": 2e-05,
+      "loss": 0.1098,
+      "step": 124
+    },
+    {
+      "epoch": 1.0666666666666667,
+      "grad_norm": 0.07641693882491171,
+      "learning_rate": 2e-05,
+      "loss": 0.1127,
+      "step": 125
+    },
+    {
+      "epoch": 1.0752,
+      "grad_norm": 0.052835367770791786,
+      "learning_rate": 2e-05,
+      "loss": 0.1231,
+      "step": 126
+    },
+    {
+      "epoch": 1.0837333333333334,
+      "grad_norm": 0.07520329755025788,
+      "learning_rate": 2e-05,
+      "loss": 0.085,
+      "step": 127
+    },
+    {
+      "epoch": 1.0922666666666667,
+      "grad_norm": 0.07670066152157425,
+      "learning_rate": 2e-05,
+      "loss": 0.1071,
+      "step": 128
+    },
+    {
+      "epoch": 1.1008,
+      "grad_norm": 0.052832906560645154,
+      "learning_rate": 2e-05,
+      "loss": 0.1093,
+      "step": 129
+    },
+    {
+      "epoch": 1.1093333333333333,
+      "grad_norm": 0.06573889037311398,
+      "learning_rate": 2e-05,
+      "loss": 0.1193,
+      "step": 130
+    },
+    {
+      "epoch": 1.1178666666666666,
+      "grad_norm": 0.05175471296566334,
+      "learning_rate": 2e-05,
+      "loss": 0.1184,
+      "step": 131
+    },
+    {
+      "epoch": 1.1264,
+      "grad_norm": 0.05912231419793496,
+      "learning_rate": 2e-05,
+      "loss": 0.1154,
+      "step": 132
+    },
+    {
+      "epoch": 1.1349333333333333,
+      "grad_norm": 0.04899140475981105,
+      "learning_rate": 2e-05,
+      "loss": 0.0957,
+      "step": 133
+    },
+    {
+      "epoch": 1.1434666666666666,
+      "grad_norm": 0.05939916939142137,
+      "learning_rate": 2e-05,
+      "loss": 0.0979,
+      "step": 134
+    },
+    {
+      "epoch": 1.152,
+      "grad_norm": 0.0516819217599706,
+      "learning_rate": 2e-05,
+      "loss": 0.0834,
+      "step": 135
+    },
+    {
+      "epoch": 1.1605333333333334,
+      "grad_norm": 0.05456440346454737,
+      "learning_rate": 2e-05,
+      "loss": 0.1183,
+      "step": 136
+    },
+    {
+      "epoch": 1.1690666666666667,
+      "grad_norm": 0.059906464476343235,
+      "learning_rate": 2e-05,
+      "loss": 0.1048,
+      "step": 137
+    },
+    {
+      "epoch": 1.1776,
+      "grad_norm": 0.0720112680204319,
+      "learning_rate": 2e-05,
+      "loss": 0.1168,
+      "step": 138
+    },
+    {
+      "epoch": 1.1861333333333333,
+      "grad_norm": 0.04940202805828527,
+      "learning_rate": 2e-05,
+      "loss": 0.0948,
+      "step": 139
+    },
+    {
+      "epoch": 1.1946666666666665,
+      "grad_norm": 0.060088609545130046,
+      "learning_rate": 2e-05,
+      "loss": 0.0952,
+      "step": 140
+    },
+    {
+      "epoch": 1.2032,
+      "grad_norm": 0.04694761423612446,
+      "learning_rate": 2e-05,
+      "loss": 0.0717,
+      "step": 141
+    },
+    {
+      "epoch": 1.2117333333333333,
+      "grad_norm": 0.05628581562512457,
+      "learning_rate": 2e-05,
+      "loss": 0.1062,
+      "step": 142
+    },
+    {
+      "epoch": 1.2202666666666666,
+      "grad_norm": 0.06876420990437652,
+      "learning_rate": 2e-05,
+      "loss": 0.1218,
+      "step": 143
+    },
+    {
+      "epoch": 1.2288000000000001,
+      "grad_norm": 0.058774700501610655,
+      "learning_rate": 2e-05,
+      "loss": 0.1125,
+      "step": 144
+    },
+    {
+      "epoch": 1.2373333333333334,
+      "grad_norm": 0.061855922064341186,
+      "learning_rate": 2e-05,
+      "loss": 0.1295,
+      "step": 145
+    },
+    {
+      "epoch": 1.2458666666666667,
+      "grad_norm": 0.0813047704730138,
+      "learning_rate": 2e-05,
+      "loss": 0.1165,
+      "step": 146
+    },
+    {
+      "epoch": 1.2544,
+      "grad_norm": 0.061374000305305446,
+      "learning_rate": 2e-05,
+      "loss": 0.1094,
+      "step": 147
+    },
+    {
+      "epoch": 1.2629333333333332,
+      "grad_norm": 0.055537169110833,
+      "learning_rate": 2e-05,
+      "loss": 0.1054,
+      "step": 148
+    },
+    {
+      "epoch": 1.2714666666666667,
+      "grad_norm": 0.04423248714119304,
+      "learning_rate": 2e-05,
+      "loss": 0.0841,
+      "step": 149
+    },
+    {
+      "epoch": 1.28,
+      "grad_norm": 0.049931966607835034,
+      "learning_rate": 2e-05,
+      "loss": 0.0961,
+      "step": 150
+    },
+    {
+      "epoch": 1.2885333333333333,
+      "grad_norm": 0.06178656953298769,
+      "learning_rate": 2e-05,
+      "loss": 0.0854,
+      "step": 151
+    },
+    {
+      "epoch": 1.2970666666666666,
+      "grad_norm": 0.05783812343287897,
+      "learning_rate": 2e-05,
+      "loss": 0.1141,
+      "step": 152
+    },
+    {
+      "epoch": 1.3056,
+      "grad_norm": 0.048955120400167584,
+      "learning_rate": 2e-05,
+      "loss": 0.0947,
+      "step": 153
+    },
+    {
+      "epoch": 1.3141333333333334,
+      "grad_norm": 0.12270174746806978,
+      "learning_rate": 2e-05,
+      "loss": 0.1553,
+      "step": 154
+    },
+    {
+      "epoch": 1.3226666666666667,
+      "grad_norm": 0.06928026959973474,
+      "learning_rate": 2e-05,
+      "loss": 0.1274,
+      "step": 155
+    },
+    {
+      "epoch": 1.3312,
+      "grad_norm": 0.04756100666105405,
+      "learning_rate": 2e-05,
+      "loss": 0.0893,
+      "step": 156
+    },
+    {
+      "epoch": 1.3397333333333332,
+      "grad_norm": 0.056054951338196934,
+      "learning_rate": 2e-05,
+      "loss": 0.0831,
+      "step": 157
+    },
+    {
+      "epoch": 1.3482666666666667,
+      "grad_norm": 0.0516990471964239,
+      "learning_rate": 2e-05,
+      "loss": 0.0883,
+      "step": 158
+    },
+    {
+      "epoch": 1.3568,
+      "grad_norm": 0.06011650542069954,
+      "learning_rate": 2e-05,
+      "loss": 0.0938,
+      "step": 159
+    },
+    {
+      "epoch": 1.3653333333333333,
+      "grad_norm": 0.051831307951873976,
+      "learning_rate": 2e-05,
+      "loss": 0.1019,
+      "step": 160
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 351,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 3,
+  "save_steps": 20,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 2.273910921652863e+18,
+  "train_batch_size": 16,
+  "trial_name": null,
+  "trial_params": null
+}
diff --git a/checkpoint-160/training_args.bin b/checkpoint-160/training_args.bin
new file mode 100644
index 0000000000000000000000000000000000000000..8ac842d29fc6ad3e132cb7208dda5b61dd1079ee
--- /dev/null
+++ b/checkpoint-160/training_args.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aa4edad1da5557fcd52a6da980443588016cead6f0444a3562cfa68029c66a04
+size 6840
diff --git a/checkpoint-160/zero_to_fp32.py b/checkpoint-160/zero_to_fp32.py
new file mode 100644
index 0000000000000000000000000000000000000000..49b846633d6eb1e836e34681e44033581f4edb7b
--- /dev/null
+++ b/checkpoint-160/zero_to_fp32.py
@@ -0,0 +1,592 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example: python zero_to_fp32.py . pytorch_model.bin
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+
+    total_files = len(files)
+    state_dicts = []
+    for f in files:
+        state_dict = torch.load(f, map_location=device)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    if zero_stage <= 2:
+        fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    elif zero_stage == 3:
+        # if there is more than one param group, there will be multiple flattened tensors - one
+        # flattened tensor per group - for simplicity merge them into a single tensor
+        #
+        # XXX: could make the script more memory efficient for when there are multiple groups - it
+        # will require matching the sub-lists of param_shapes for each param group flattened tensor
+
+        fp32_flat_groups = [
+            torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
+        ]
+
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    for name, shape in param_shapes.items():
+
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # XXX: memory usage doubles here
+        state_dict[name] = torch.cat(
+            tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
+            0).narrow(0, 0, unpartitioned_numel).view(shape)
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    Note: this approach may not work if your application doesn't have sufficient free CPU memory and
+    you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+    """
+
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+    print(f"Saving fp32 state dict to {output_file}")
+    torch.save(state_dict, output_file)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument(
+        "output_file",
+        type=str,
+        help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file, tag=args.tag)
diff --git a/checkpoint-20/README.md b/checkpoint-20/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..e7d4f706aeefc6bb2aa84e963870f345a14c4af9
--- /dev/null
+++ b/checkpoint-20/README.md
@@ -0,0 +1,202 @@
+---
+library_name: peft
+base_model: ../ckpts/Meta-Llama-3-8B-Instruct
+---
+
+# 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 0.11.1
\ No newline at end of file
diff --git a/checkpoint-20/adapter_config.json b/checkpoint-20/adapter_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..ad16c073ed8946d227350e8fe4e99d689a8ede54
--- /dev/null
+++ b/checkpoint-20/adapter_config.json
@@ -0,0 +1,35 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "../ckpts/Meta-Llama-3-8B-Instruct",
+  "bias": "none",
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 16,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "down_proj",
+    "gate_proj",
+    "q_proj",
+    "lm_head",
+    "o_proj",
+    "v_proj",
+    "up_proj",
+    "k_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "use_dora": false,
+  "use_rslora": false
+}
\ No newline at end of file
diff --git a/checkpoint-20/adapter_model.safetensors b/checkpoint-20/adapter_model.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..a5751695ace327ef7df9be2a6b3f20d53644e8d0
--- /dev/null
+++ b/checkpoint-20/adapter_model.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:181727761583eeba365bc29814e3196c50f90d02ae039c8a529d15c6b58813d4
+size 1138856856
diff --git a/checkpoint-20/global_step20/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt b/checkpoint-20/global_step20/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..3b21aa8cad85e685e006ce3bc2075c0fe2ecff03
--- /dev/null
+++ b/checkpoint-20/global_step20/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:fc02bac7488ab5bf8abdc7c2b5344d7b8dfb0afbb5b445938a1d48306aa2f091
+size 528781328
diff --git a/checkpoint-20/global_step20/mp_rank_00_model_states.pt b/checkpoint-20/global_step20/mp_rank_00_model_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..d507f268adaf97bdeb412c5e835a44e56ae43a83
--- /dev/null
+++ b/checkpoint-20/global_step20/mp_rank_00_model_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:92ae839a372c83084c52ed7acb3f7c30fafafaf5b2d3a11194adddf87b919f14
+size 199905337
diff --git a/checkpoint-20/latest b/checkpoint-20/latest
new file mode 100644
index 0000000000000000000000000000000000000000..11e5c63223cdf01f44f9f3129915f9de3d647f31
--- /dev/null
+++ b/checkpoint-20/latest
@@ -0,0 +1 @@
+global_step20
\ No newline at end of file
diff --git a/checkpoint-20/rng_state.pth b/checkpoint-20/rng_state.pth
new file mode 100644
index 0000000000000000000000000000000000000000..e8254dbcbd4c6875b8d6617755cc9afd84a43b3c
--- /dev/null
+++ b/checkpoint-20/rng_state.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:926d305e96380bc96ae95066ad252aa5ad89a958f613cc36a658033c5409657d
+size 14244
diff --git a/checkpoint-20/trainer_state.json b/checkpoint-20/trainer_state.json
new file mode 100644
index 0000000000000000000000000000000000000000..198f2a79d1ff6a247e3846966433fe7cb967dca2
--- /dev/null
+++ b/checkpoint-20/trainer_state.json
@@ -0,0 +1,173 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.17066666666666666,
+  "eval_steps": 500,
+  "global_step": 20,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.008533333333333334,
+      "grad_norm": 100.21848203113535,
+      "learning_rate": 0.0,
+      "loss": 7.1962,
+      "step": 1
+    },
+    {
+      "epoch": 0.017066666666666667,
+      "grad_norm": 96.43006188910957,
+      "learning_rate": 3.0102999566398115e-06,
+      "loss": 6.9414,
+      "step": 2
+    },
+    {
+      "epoch": 0.0256,
+      "grad_norm": 97.35803466618715,
+      "learning_rate": 4.771212547196624e-06,
+      "loss": 7.0102,
+      "step": 3
+    },
+    {
+      "epoch": 0.034133333333333335,
+      "grad_norm": 95.14837816372646,
+      "learning_rate": 6.020599913279623e-06,
+      "loss": 6.5295,
+      "step": 4
+    },
+    {
+      "epoch": 0.042666666666666665,
+      "grad_norm": 91.76544275692784,
+      "learning_rate": 6.989700043360187e-06,
+      "loss": 6.4806,
+      "step": 5
+    },
+    {
+      "epoch": 0.0512,
+      "grad_norm": 84.4494318688335,
+      "learning_rate": 7.781512503836437e-06,
+      "loss": 6.4194,
+      "step": 6
+    },
+    {
+      "epoch": 0.05973333333333333,
+      "grad_norm": 71.37977490595638,
+      "learning_rate": 8.450980400142568e-06,
+      "loss": 5.4953,
+      "step": 7
+    },
+    {
+      "epoch": 0.06826666666666667,
+      "grad_norm": 49.31153456754566,
+      "learning_rate": 9.030899869919434e-06,
+      "loss": 5.4123,
+      "step": 8
+    },
+    {
+      "epoch": 0.0768,
+      "grad_norm": 20.37296364560341,
+      "learning_rate": 9.542425094393249e-06,
+      "loss": 5.2334,
+      "step": 9
+    },
+    {
+      "epoch": 0.08533333333333333,
+      "grad_norm": 7.968467079076881,
+      "learning_rate": 9.999999999999999e-06,
+      "loss": 5.0282,
+      "step": 10
+    },
+    {
+      "epoch": 0.09386666666666667,
+      "grad_norm": 3.559446532055649,
+      "learning_rate": 1.041392685158225e-05,
+      "loss": 4.612,
+      "step": 11
+    },
+    {
+      "epoch": 0.1024,
+      "grad_norm": 3.5528846947995674,
+      "learning_rate": 1.0791812460476248e-05,
+      "loss": 4.9475,
+      "step": 12
+    },
+    {
+      "epoch": 0.11093333333333333,
+      "grad_norm": 3.541968897471334,
+      "learning_rate": 1.1139433523068365e-05,
+      "loss": 4.2777,
+      "step": 13
+    },
+    {
+      "epoch": 0.11946666666666667,
+      "grad_norm": 3.54718070036198,
+      "learning_rate": 1.1461280356782378e-05,
+      "loss": 4.3507,
+      "step": 14
+    },
+    {
+      "epoch": 0.128,
+      "grad_norm": 3.8632334830606747,
+      "learning_rate": 1.1760912590556813e-05,
+      "loss": 4.5364,
+      "step": 15
+    },
+    {
+      "epoch": 0.13653333333333334,
+      "grad_norm": 3.6637424744054004,
+      "learning_rate": 1.2041199826559246e-05,
+      "loss": 3.9672,
+      "step": 16
+    },
+    {
+      "epoch": 0.14506666666666668,
+      "grad_norm": 3.919802904818311,
+      "learning_rate": 1.230448921378274e-05,
+      "loss": 4.0618,
+      "step": 17
+    },
+    {
+      "epoch": 0.1536,
+      "grad_norm": 4.71904950738746,
+      "learning_rate": 1.2552725051033058e-05,
+      "loss": 4.6656,
+      "step": 18
+    },
+    {
+      "epoch": 0.16213333333333332,
+      "grad_norm": 4.6656317698690835,
+      "learning_rate": 1.2787536009528288e-05,
+      "loss": 4.1131,
+      "step": 19
+    },
+    {
+      "epoch": 0.17066666666666666,
+      "grad_norm": 5.145138692367417,
+      "learning_rate": 1.301029995663981e-05,
+      "loss": 4.0989,
+      "step": 20
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 351,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 3,
+  "save_steps": 20,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 2.924957640079442e+17,
+  "train_batch_size": 16,
+  "trial_name": null,
+  "trial_params": null
+}
diff --git a/checkpoint-20/training_args.bin b/checkpoint-20/training_args.bin
new file mode 100644
index 0000000000000000000000000000000000000000..8ac842d29fc6ad3e132cb7208dda5b61dd1079ee
--- /dev/null
+++ b/checkpoint-20/training_args.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aa4edad1da5557fcd52a6da980443588016cead6f0444a3562cfa68029c66a04
+size 6840
diff --git a/checkpoint-20/zero_to_fp32.py b/checkpoint-20/zero_to_fp32.py
new file mode 100644
index 0000000000000000000000000000000000000000..49b846633d6eb1e836e34681e44033581f4edb7b
--- /dev/null
+++ b/checkpoint-20/zero_to_fp32.py
@@ -0,0 +1,592 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example: python zero_to_fp32.py . pytorch_model.bin
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+
+    total_files = len(files)
+    state_dicts = []
+    for f in files:
+        state_dict = torch.load(f, map_location=device)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    if zero_stage <= 2:
+        fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    elif zero_stage == 3:
+        # if there is more than one param group, there will be multiple flattened tensors - one
+        # flattened tensor per group - for simplicity merge them into a single tensor
+        #
+        # XXX: could make the script more memory efficient for when there are multiple groups - it
+        # will require matching the sub-lists of param_shapes for each param group flattened tensor
+
+        fp32_flat_groups = [
+            torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
+        ]
+
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    for name, shape in param_shapes.items():
+
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # XXX: memory usage doubles here
+        state_dict[name] = torch.cat(
+            tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
+            0).narrow(0, 0, unpartitioned_numel).view(shape)
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    Note: this approach may not work if your application doesn't have sufficient free CPU memory and
+    you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+    """
+
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+    print(f"Saving fp32 state dict to {output_file}")
+    torch.save(state_dict, output_file)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument(
+        "output_file",
+        type=str,
+        help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file, tag=args.tag)
diff --git a/checkpoint-40/README.md b/checkpoint-40/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..e7d4f706aeefc6bb2aa84e963870f345a14c4af9
--- /dev/null
+++ b/checkpoint-40/README.md
@@ -0,0 +1,202 @@
+---
+library_name: peft
+base_model: ../ckpts/Meta-Llama-3-8B-Instruct
+---
+
+# 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 0.11.1
\ No newline at end of file
diff --git a/checkpoint-40/adapter_config.json b/checkpoint-40/adapter_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..ad16c073ed8946d227350e8fe4e99d689a8ede54
--- /dev/null
+++ b/checkpoint-40/adapter_config.json
@@ -0,0 +1,35 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "../ckpts/Meta-Llama-3-8B-Instruct",
+  "bias": "none",
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 16,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "down_proj",
+    "gate_proj",
+    "q_proj",
+    "lm_head",
+    "o_proj",
+    "v_proj",
+    "up_proj",
+    "k_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "use_dora": false,
+  "use_rslora": false
+}
\ No newline at end of file
diff --git a/checkpoint-40/adapter_model.safetensors b/checkpoint-40/adapter_model.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..3b5879d19cd4f8dde7e26054ed1062a9ea0c49ef
--- /dev/null
+++ b/checkpoint-40/adapter_model.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:df22fc3523779ffcc3c8161bd924866943914d0fe3f077b47683ce0a44bbd378
+size 1138856856
diff --git a/checkpoint-40/global_step40/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt b/checkpoint-40/global_step40/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..7c10a2861b3f1e14617b012a49b35ac6e2c73c36
--- /dev/null
+++ b/checkpoint-40/global_step40/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:d1659655a3d80ff77b190ea02ca19985a67df95eb8d71a71bee18bc73f17802e
+size 528781328
diff --git a/checkpoint-40/global_step40/mp_rank_00_model_states.pt b/checkpoint-40/global_step40/mp_rank_00_model_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..49b525d1793eaa3d4e4a8bc0afe5613a0906676c
--- /dev/null
+++ b/checkpoint-40/global_step40/mp_rank_00_model_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:22800478617c1bd7dfacaacc129c28a57125fff961281628b3488d45734cfe67
+size 199905337
diff --git a/checkpoint-40/latest b/checkpoint-40/latest
new file mode 100644
index 0000000000000000000000000000000000000000..8631ab8ddebf60eb3e7f5f2c2b1a2da8298a43c3
--- /dev/null
+++ b/checkpoint-40/latest
@@ -0,0 +1 @@
+global_step40
\ No newline at end of file
diff --git a/checkpoint-40/rng_state.pth b/checkpoint-40/rng_state.pth
new file mode 100644
index 0000000000000000000000000000000000000000..7b9e0578d84b49cb235d98e0d8879a37a434e8c5
--- /dev/null
+++ b/checkpoint-40/rng_state.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a49aafe79c0893003242520587ee35071de0249e66c0af69e715f46e1600841d
+size 14244
diff --git a/checkpoint-40/trainer_state.json b/checkpoint-40/trainer_state.json
new file mode 100644
index 0000000000000000000000000000000000000000..d7551f1e8a47468d64544f9bf3e05f95baea0ad6
--- /dev/null
+++ b/checkpoint-40/trainer_state.json
@@ -0,0 +1,313 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.3413333333333333,
+  "eval_steps": 500,
+  "global_step": 40,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.008533333333333334,
+      "grad_norm": 100.21848203113535,
+      "learning_rate": 0.0,
+      "loss": 7.1962,
+      "step": 1
+    },
+    {
+      "epoch": 0.017066666666666667,
+      "grad_norm": 96.43006188910957,
+      "learning_rate": 3.0102999566398115e-06,
+      "loss": 6.9414,
+      "step": 2
+    },
+    {
+      "epoch": 0.0256,
+      "grad_norm": 97.35803466618715,
+      "learning_rate": 4.771212547196624e-06,
+      "loss": 7.0102,
+      "step": 3
+    },
+    {
+      "epoch": 0.034133333333333335,
+      "grad_norm": 95.14837816372646,
+      "learning_rate": 6.020599913279623e-06,
+      "loss": 6.5295,
+      "step": 4
+    },
+    {
+      "epoch": 0.042666666666666665,
+      "grad_norm": 91.76544275692784,
+      "learning_rate": 6.989700043360187e-06,
+      "loss": 6.4806,
+      "step": 5
+    },
+    {
+      "epoch": 0.0512,
+      "grad_norm": 84.4494318688335,
+      "learning_rate": 7.781512503836437e-06,
+      "loss": 6.4194,
+      "step": 6
+    },
+    {
+      "epoch": 0.05973333333333333,
+      "grad_norm": 71.37977490595638,
+      "learning_rate": 8.450980400142568e-06,
+      "loss": 5.4953,
+      "step": 7
+    },
+    {
+      "epoch": 0.06826666666666667,
+      "grad_norm": 49.31153456754566,
+      "learning_rate": 9.030899869919434e-06,
+      "loss": 5.4123,
+      "step": 8
+    },
+    {
+      "epoch": 0.0768,
+      "grad_norm": 20.37296364560341,
+      "learning_rate": 9.542425094393249e-06,
+      "loss": 5.2334,
+      "step": 9
+    },
+    {
+      "epoch": 0.08533333333333333,
+      "grad_norm": 7.968467079076881,
+      "learning_rate": 9.999999999999999e-06,
+      "loss": 5.0282,
+      "step": 10
+    },
+    {
+      "epoch": 0.09386666666666667,
+      "grad_norm": 3.559446532055649,
+      "learning_rate": 1.041392685158225e-05,
+      "loss": 4.612,
+      "step": 11
+    },
+    {
+      "epoch": 0.1024,
+      "grad_norm": 3.5528846947995674,
+      "learning_rate": 1.0791812460476248e-05,
+      "loss": 4.9475,
+      "step": 12
+    },
+    {
+      "epoch": 0.11093333333333333,
+      "grad_norm": 3.541968897471334,
+      "learning_rate": 1.1139433523068365e-05,
+      "loss": 4.2777,
+      "step": 13
+    },
+    {
+      "epoch": 0.11946666666666667,
+      "grad_norm": 3.54718070036198,
+      "learning_rate": 1.1461280356782378e-05,
+      "loss": 4.3507,
+      "step": 14
+    },
+    {
+      "epoch": 0.128,
+      "grad_norm": 3.8632334830606747,
+      "learning_rate": 1.1760912590556813e-05,
+      "loss": 4.5364,
+      "step": 15
+    },
+    {
+      "epoch": 0.13653333333333334,
+      "grad_norm": 3.6637424744054004,
+      "learning_rate": 1.2041199826559246e-05,
+      "loss": 3.9672,
+      "step": 16
+    },
+    {
+      "epoch": 0.14506666666666668,
+      "grad_norm": 3.919802904818311,
+      "learning_rate": 1.230448921378274e-05,
+      "loss": 4.0618,
+      "step": 17
+    },
+    {
+      "epoch": 0.1536,
+      "grad_norm": 4.71904950738746,
+      "learning_rate": 1.2552725051033058e-05,
+      "loss": 4.6656,
+      "step": 18
+    },
+    {
+      "epoch": 0.16213333333333332,
+      "grad_norm": 4.6656317698690835,
+      "learning_rate": 1.2787536009528288e-05,
+      "loss": 4.1131,
+      "step": 19
+    },
+    {
+      "epoch": 0.17066666666666666,
+      "grad_norm": 5.145138692367417,
+      "learning_rate": 1.301029995663981e-05,
+      "loss": 4.0989,
+      "step": 20
+    },
+    {
+      "epoch": 0.1792,
+      "grad_norm": 5.923538014759818,
+      "learning_rate": 1.3222192947339192e-05,
+      "loss": 4.4991,
+      "step": 21
+    },
+    {
+      "epoch": 0.18773333333333334,
+      "grad_norm": 5.941056962941364,
+      "learning_rate": 1.3424226808222062e-05,
+      "loss": 4.0836,
+      "step": 22
+    },
+    {
+      "epoch": 0.19626666666666667,
+      "grad_norm": 6.171026012117947,
+      "learning_rate": 1.3617278360175927e-05,
+      "loss": 3.6861,
+      "step": 23
+    },
+    {
+      "epoch": 0.2048,
+      "grad_norm": 7.130542138930838,
+      "learning_rate": 1.380211241711606e-05,
+      "loss": 4.0958,
+      "step": 24
+    },
+    {
+      "epoch": 0.21333333333333335,
+      "grad_norm": 7.328837606110418,
+      "learning_rate": 1.3979400086720374e-05,
+      "loss": 3.9524,
+      "step": 25
+    },
+    {
+      "epoch": 0.22186666666666666,
+      "grad_norm": 6.923489005711429,
+      "learning_rate": 1.4149733479708178e-05,
+      "loss": 3.6062,
+      "step": 26
+    },
+    {
+      "epoch": 0.2304,
+      "grad_norm": 6.409498926059221,
+      "learning_rate": 1.4313637641589872e-05,
+      "loss": 3.2034,
+      "step": 27
+    },
+    {
+      "epoch": 0.23893333333333333,
+      "grad_norm": 5.781628405584682,
+      "learning_rate": 1.4471580313422191e-05,
+      "loss": 2.8158,
+      "step": 28
+    },
+    {
+      "epoch": 0.24746666666666667,
+      "grad_norm": 6.2927590068243315,
+      "learning_rate": 1.4623979978989559e-05,
+      "loss": 2.9803,
+      "step": 29
+    },
+    {
+      "epoch": 0.256,
+      "grad_norm": 6.103844678630006,
+      "learning_rate": 1.4771212547196623e-05,
+      "loss": 2.847,
+      "step": 30
+    },
+    {
+      "epoch": 0.26453333333333334,
+      "grad_norm": 7.656341925867144,
+      "learning_rate": 1.4913616938342726e-05,
+      "loss": 3.0907,
+      "step": 31
+    },
+    {
+      "epoch": 0.2730666666666667,
+      "grad_norm": 6.324242877501844,
+      "learning_rate": 1.5051499783199059e-05,
+      "loss": 2.3467,
+      "step": 32
+    },
+    {
+      "epoch": 0.2816,
+      "grad_norm": 7.606313101162739,
+      "learning_rate": 1.5185139398778874e-05,
+      "loss": 2.5292,
+      "step": 33
+    },
+    {
+      "epoch": 0.29013333333333335,
+      "grad_norm": 8.553792493849265,
+      "learning_rate": 1.531478917042255e-05,
+      "loss": 2.4547,
+      "step": 34
+    },
+    {
+      "epoch": 0.2986666666666667,
+      "grad_norm": 8.483368703272543,
+      "learning_rate": 1.5440680443502753e-05,
+      "loss": 2.1956,
+      "step": 35
+    },
+    {
+      "epoch": 0.3072,
+      "grad_norm": 8.364739352838077,
+      "learning_rate": 1.5563025007672873e-05,
+      "loss": 1.8552,
+      "step": 36
+    },
+    {
+      "epoch": 0.3157333333333333,
+      "grad_norm": 9.37663682000104,
+      "learning_rate": 1.5682017240669948e-05,
+      "loss": 1.9228,
+      "step": 37
+    },
+    {
+      "epoch": 0.32426666666666665,
+      "grad_norm": 9.823047193440066,
+      "learning_rate": 1.57978359661681e-05,
+      "loss": 1.7033,
+      "step": 38
+    },
+    {
+      "epoch": 0.3328,
+      "grad_norm": 9.692618512955894,
+      "learning_rate": 1.591064607026499e-05,
+      "loss": 1.3768,
+      "step": 39
+    },
+    {
+      "epoch": 0.3413333333333333,
+      "grad_norm": 9.08889021911031,
+      "learning_rate": 1.6020599913279622e-05,
+      "loss": 1.3015,
+      "step": 40
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 351,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 3,
+  "save_steps": 20,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 5.708931636490404e+17,
+  "train_batch_size": 16,
+  "trial_name": null,
+  "trial_params": null
+}
diff --git a/checkpoint-40/training_args.bin b/checkpoint-40/training_args.bin
new file mode 100644
index 0000000000000000000000000000000000000000..8ac842d29fc6ad3e132cb7208dda5b61dd1079ee
--- /dev/null
+++ b/checkpoint-40/training_args.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aa4edad1da5557fcd52a6da980443588016cead6f0444a3562cfa68029c66a04
+size 6840
diff --git a/checkpoint-40/zero_to_fp32.py b/checkpoint-40/zero_to_fp32.py
new file mode 100644
index 0000000000000000000000000000000000000000..49b846633d6eb1e836e34681e44033581f4edb7b
--- /dev/null
+++ b/checkpoint-40/zero_to_fp32.py
@@ -0,0 +1,592 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example: python zero_to_fp32.py . pytorch_model.bin
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+
+    total_files = len(files)
+    state_dicts = []
+    for f in files:
+        state_dict = torch.load(f, map_location=device)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    if zero_stage <= 2:
+        fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    elif zero_stage == 3:
+        # if there is more than one param group, there will be multiple flattened tensors - one
+        # flattened tensor per group - for simplicity merge them into a single tensor
+        #
+        # XXX: could make the script more memory efficient for when there are multiple groups - it
+        # will require matching the sub-lists of param_shapes for each param group flattened tensor
+
+        fp32_flat_groups = [
+            torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
+        ]
+
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    for name, shape in param_shapes.items():
+
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # XXX: memory usage doubles here
+        state_dict[name] = torch.cat(
+            tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
+            0).narrow(0, 0, unpartitioned_numel).view(shape)
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    Note: this approach may not work if your application doesn't have sufficient free CPU memory and
+    you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+    """
+
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+    print(f"Saving fp32 state dict to {output_file}")
+    torch.save(state_dict, output_file)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument(
+        "output_file",
+        type=str,
+        help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file, tag=args.tag)
diff --git a/checkpoint-60/README.md b/checkpoint-60/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..e7d4f706aeefc6bb2aa84e963870f345a14c4af9
--- /dev/null
+++ b/checkpoint-60/README.md
@@ -0,0 +1,202 @@
+---
+library_name: peft
+base_model: ../ckpts/Meta-Llama-3-8B-Instruct
+---
+
+# 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 0.11.1
\ No newline at end of file
diff --git a/checkpoint-60/adapter_config.json b/checkpoint-60/adapter_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..ad16c073ed8946d227350e8fe4e99d689a8ede54
--- /dev/null
+++ b/checkpoint-60/adapter_config.json
@@ -0,0 +1,35 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "../ckpts/Meta-Llama-3-8B-Instruct",
+  "bias": "none",
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 16,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "down_proj",
+    "gate_proj",
+    "q_proj",
+    "lm_head",
+    "o_proj",
+    "v_proj",
+    "up_proj",
+    "k_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "use_dora": false,
+  "use_rslora": false
+}
\ No newline at end of file
diff --git a/checkpoint-60/adapter_model.safetensors b/checkpoint-60/adapter_model.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..1e35e584361c6261e9c4af6ea5e70c0462bb4cbe
--- /dev/null
+++ b/checkpoint-60/adapter_model.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:77335a2ac39e530a5aa962884d40c39cb11419d44e8052c02325a25871cb3493
+size 1138856856
diff --git a/checkpoint-60/global_step60/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt b/checkpoint-60/global_step60/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..d25b204332c531a51ec954290d139c575888edb6
--- /dev/null
+++ b/checkpoint-60/global_step60/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:72fe210e9583468ebc447e663cdb5e604bb48429b5370b6ebe2ac0eef1c8b4be
+size 528781328
diff --git a/checkpoint-60/global_step60/mp_rank_00_model_states.pt b/checkpoint-60/global_step60/mp_rank_00_model_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..d44ca31b5158f431652974b2f8656cf3edf7da8f
--- /dev/null
+++ b/checkpoint-60/global_step60/mp_rank_00_model_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:96748f130193979fdb318f39bbcc7a68656cd18dae9c2703108db3b979b02fcc
+size 199905337
diff --git a/checkpoint-60/latest b/checkpoint-60/latest
new file mode 100644
index 0000000000000000000000000000000000000000..6dac34b840ecfb636ba8ab1e4da79fa1bdc8c3d4
--- /dev/null
+++ b/checkpoint-60/latest
@@ -0,0 +1 @@
+global_step60
\ No newline at end of file
diff --git a/checkpoint-60/rng_state.pth b/checkpoint-60/rng_state.pth
new file mode 100644
index 0000000000000000000000000000000000000000..dd3ffd279d461273ef27d2b6bf5825c51e430c25
--- /dev/null
+++ b/checkpoint-60/rng_state.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:59a5d7a69796d62bbddf60556572e2039646e271b96568851e851a9d2963d37e
+size 14244
diff --git a/checkpoint-60/trainer_state.json b/checkpoint-60/trainer_state.json
new file mode 100644
index 0000000000000000000000000000000000000000..90f15707e6ea9190026f4a15be797c7e71f842d4
--- /dev/null
+++ b/checkpoint-60/trainer_state.json
@@ -0,0 +1,453 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.512,
+  "eval_steps": 500,
+  "global_step": 60,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.008533333333333334,
+      "grad_norm": 100.21848203113535,
+      "learning_rate": 0.0,
+      "loss": 7.1962,
+      "step": 1
+    },
+    {
+      "epoch": 0.017066666666666667,
+      "grad_norm": 96.43006188910957,
+      "learning_rate": 3.0102999566398115e-06,
+      "loss": 6.9414,
+      "step": 2
+    },
+    {
+      "epoch": 0.0256,
+      "grad_norm": 97.35803466618715,
+      "learning_rate": 4.771212547196624e-06,
+      "loss": 7.0102,
+      "step": 3
+    },
+    {
+      "epoch": 0.034133333333333335,
+      "grad_norm": 95.14837816372646,
+      "learning_rate": 6.020599913279623e-06,
+      "loss": 6.5295,
+      "step": 4
+    },
+    {
+      "epoch": 0.042666666666666665,
+      "grad_norm": 91.76544275692784,
+      "learning_rate": 6.989700043360187e-06,
+      "loss": 6.4806,
+      "step": 5
+    },
+    {
+      "epoch": 0.0512,
+      "grad_norm": 84.4494318688335,
+      "learning_rate": 7.781512503836437e-06,
+      "loss": 6.4194,
+      "step": 6
+    },
+    {
+      "epoch": 0.05973333333333333,
+      "grad_norm": 71.37977490595638,
+      "learning_rate": 8.450980400142568e-06,
+      "loss": 5.4953,
+      "step": 7
+    },
+    {
+      "epoch": 0.06826666666666667,
+      "grad_norm": 49.31153456754566,
+      "learning_rate": 9.030899869919434e-06,
+      "loss": 5.4123,
+      "step": 8
+    },
+    {
+      "epoch": 0.0768,
+      "grad_norm": 20.37296364560341,
+      "learning_rate": 9.542425094393249e-06,
+      "loss": 5.2334,
+      "step": 9
+    },
+    {
+      "epoch": 0.08533333333333333,
+      "grad_norm": 7.968467079076881,
+      "learning_rate": 9.999999999999999e-06,
+      "loss": 5.0282,
+      "step": 10
+    },
+    {
+      "epoch": 0.09386666666666667,
+      "grad_norm": 3.559446532055649,
+      "learning_rate": 1.041392685158225e-05,
+      "loss": 4.612,
+      "step": 11
+    },
+    {
+      "epoch": 0.1024,
+      "grad_norm": 3.5528846947995674,
+      "learning_rate": 1.0791812460476248e-05,
+      "loss": 4.9475,
+      "step": 12
+    },
+    {
+      "epoch": 0.11093333333333333,
+      "grad_norm": 3.541968897471334,
+      "learning_rate": 1.1139433523068365e-05,
+      "loss": 4.2777,
+      "step": 13
+    },
+    {
+      "epoch": 0.11946666666666667,
+      "grad_norm": 3.54718070036198,
+      "learning_rate": 1.1461280356782378e-05,
+      "loss": 4.3507,
+      "step": 14
+    },
+    {
+      "epoch": 0.128,
+      "grad_norm": 3.8632334830606747,
+      "learning_rate": 1.1760912590556813e-05,
+      "loss": 4.5364,
+      "step": 15
+    },
+    {
+      "epoch": 0.13653333333333334,
+      "grad_norm": 3.6637424744054004,
+      "learning_rate": 1.2041199826559246e-05,
+      "loss": 3.9672,
+      "step": 16
+    },
+    {
+      "epoch": 0.14506666666666668,
+      "grad_norm": 3.919802904818311,
+      "learning_rate": 1.230448921378274e-05,
+      "loss": 4.0618,
+      "step": 17
+    },
+    {
+      "epoch": 0.1536,
+      "grad_norm": 4.71904950738746,
+      "learning_rate": 1.2552725051033058e-05,
+      "loss": 4.6656,
+      "step": 18
+    },
+    {
+      "epoch": 0.16213333333333332,
+      "grad_norm": 4.6656317698690835,
+      "learning_rate": 1.2787536009528288e-05,
+      "loss": 4.1131,
+      "step": 19
+    },
+    {
+      "epoch": 0.17066666666666666,
+      "grad_norm": 5.145138692367417,
+      "learning_rate": 1.301029995663981e-05,
+      "loss": 4.0989,
+      "step": 20
+    },
+    {
+      "epoch": 0.1792,
+      "grad_norm": 5.923538014759818,
+      "learning_rate": 1.3222192947339192e-05,
+      "loss": 4.4991,
+      "step": 21
+    },
+    {
+      "epoch": 0.18773333333333334,
+      "grad_norm": 5.941056962941364,
+      "learning_rate": 1.3424226808222062e-05,
+      "loss": 4.0836,
+      "step": 22
+    },
+    {
+      "epoch": 0.19626666666666667,
+      "grad_norm": 6.171026012117947,
+      "learning_rate": 1.3617278360175927e-05,
+      "loss": 3.6861,
+      "step": 23
+    },
+    {
+      "epoch": 0.2048,
+      "grad_norm": 7.130542138930838,
+      "learning_rate": 1.380211241711606e-05,
+      "loss": 4.0958,
+      "step": 24
+    },
+    {
+      "epoch": 0.21333333333333335,
+      "grad_norm": 7.328837606110418,
+      "learning_rate": 1.3979400086720374e-05,
+      "loss": 3.9524,
+      "step": 25
+    },
+    {
+      "epoch": 0.22186666666666666,
+      "grad_norm": 6.923489005711429,
+      "learning_rate": 1.4149733479708178e-05,
+      "loss": 3.6062,
+      "step": 26
+    },
+    {
+      "epoch": 0.2304,
+      "grad_norm": 6.409498926059221,
+      "learning_rate": 1.4313637641589872e-05,
+      "loss": 3.2034,
+      "step": 27
+    },
+    {
+      "epoch": 0.23893333333333333,
+      "grad_norm": 5.781628405584682,
+      "learning_rate": 1.4471580313422191e-05,
+      "loss": 2.8158,
+      "step": 28
+    },
+    {
+      "epoch": 0.24746666666666667,
+      "grad_norm": 6.2927590068243315,
+      "learning_rate": 1.4623979978989559e-05,
+      "loss": 2.9803,
+      "step": 29
+    },
+    {
+      "epoch": 0.256,
+      "grad_norm": 6.103844678630006,
+      "learning_rate": 1.4771212547196623e-05,
+      "loss": 2.847,
+      "step": 30
+    },
+    {
+      "epoch": 0.26453333333333334,
+      "grad_norm": 7.656341925867144,
+      "learning_rate": 1.4913616938342726e-05,
+      "loss": 3.0907,
+      "step": 31
+    },
+    {
+      "epoch": 0.2730666666666667,
+      "grad_norm": 6.324242877501844,
+      "learning_rate": 1.5051499783199059e-05,
+      "loss": 2.3467,
+      "step": 32
+    },
+    {
+      "epoch": 0.2816,
+      "grad_norm": 7.606313101162739,
+      "learning_rate": 1.5185139398778874e-05,
+      "loss": 2.5292,
+      "step": 33
+    },
+    {
+      "epoch": 0.29013333333333335,
+      "grad_norm": 8.553792493849265,
+      "learning_rate": 1.531478917042255e-05,
+      "loss": 2.4547,
+      "step": 34
+    },
+    {
+      "epoch": 0.2986666666666667,
+      "grad_norm": 8.483368703272543,
+      "learning_rate": 1.5440680443502753e-05,
+      "loss": 2.1956,
+      "step": 35
+    },
+    {
+      "epoch": 0.3072,
+      "grad_norm": 8.364739352838077,
+      "learning_rate": 1.5563025007672873e-05,
+      "loss": 1.8552,
+      "step": 36
+    },
+    {
+      "epoch": 0.3157333333333333,
+      "grad_norm": 9.37663682000104,
+      "learning_rate": 1.5682017240669948e-05,
+      "loss": 1.9228,
+      "step": 37
+    },
+    {
+      "epoch": 0.32426666666666665,
+      "grad_norm": 9.823047193440066,
+      "learning_rate": 1.57978359661681e-05,
+      "loss": 1.7033,
+      "step": 38
+    },
+    {
+      "epoch": 0.3328,
+      "grad_norm": 9.692618512955894,
+      "learning_rate": 1.591064607026499e-05,
+      "loss": 1.3768,
+      "step": 39
+    },
+    {
+      "epoch": 0.3413333333333333,
+      "grad_norm": 9.08889021911031,
+      "learning_rate": 1.6020599913279622e-05,
+      "loss": 1.3015,
+      "step": 40
+    },
+    {
+      "epoch": 0.34986666666666666,
+      "grad_norm": 8.081534221516058,
+      "learning_rate": 1.6127838567197353e-05,
+      "loss": 0.9228,
+      "step": 41
+    },
+    {
+      "epoch": 0.3584,
+      "grad_norm": 6.238638048950311,
+      "learning_rate": 1.6232492903979005e-05,
+      "loss": 0.7267,
+      "step": 42
+    },
+    {
+      "epoch": 0.36693333333333333,
+      "grad_norm": 3.4058036861773604,
+      "learning_rate": 1.6334684555795865e-05,
+      "loss": 0.5875,
+      "step": 43
+    },
+    {
+      "epoch": 0.37546666666666667,
+      "grad_norm": 2.079163829467713,
+      "learning_rate": 1.6434526764861872e-05,
+      "loss": 0.6355,
+      "step": 44
+    },
+    {
+      "epoch": 0.384,
+      "grad_norm": 1.5597487833024746,
+      "learning_rate": 1.6532125137753435e-05,
+      "loss": 0.5106,
+      "step": 45
+    },
+    {
+      "epoch": 0.39253333333333335,
+      "grad_norm": 2.491689602375256,
+      "learning_rate": 1.662757831681574e-05,
+      "loss": 0.6454,
+      "step": 46
+    },
+    {
+      "epoch": 0.4010666666666667,
+      "grad_norm": 2.010880438195854,
+      "learning_rate": 1.672097857935717e-05,
+      "loss": 0.4757,
+      "step": 47
+    },
+    {
+      "epoch": 0.4096,
+      "grad_norm": 1.9452805114322096,
+      "learning_rate": 1.681241237375587e-05,
+      "loss": 0.4133,
+      "step": 48
+    },
+    {
+      "epoch": 0.41813333333333336,
+      "grad_norm": 1.7620848552306103,
+      "learning_rate": 1.6901960800285137e-05,
+      "loss": 0.4004,
+      "step": 49
+    },
+    {
+      "epoch": 0.4266666666666667,
+      "grad_norm": 1.278224489774809,
+      "learning_rate": 1.6989700043360187e-05,
+      "loss": 0.3523,
+      "step": 50
+    },
+    {
+      "epoch": 0.4352,
+      "grad_norm": 1.6151354758303231,
+      "learning_rate": 1.7075701760979363e-05,
+      "loss": 0.4317,
+      "step": 51
+    },
+    {
+      "epoch": 0.4437333333333333,
+      "grad_norm": 1.3451396055695035,
+      "learning_rate": 1.716003343634799e-05,
+      "loss": 0.3474,
+      "step": 52
+    },
+    {
+      "epoch": 0.45226666666666665,
+      "grad_norm": 1.6814977782362666,
+      "learning_rate": 1.724275869600789e-05,
+      "loss": 0.3706,
+      "step": 53
+    },
+    {
+      "epoch": 0.4608,
+      "grad_norm": 1.1682442432688667,
+      "learning_rate": 1.7323937598229687e-05,
+      "loss": 0.3488,
+      "step": 54
+    },
+    {
+      "epoch": 0.4693333333333333,
+      "grad_norm": 0.8839814540462471,
+      "learning_rate": 1.7403626894942437e-05,
+      "loss": 0.293,
+      "step": 55
+    },
+    {
+      "epoch": 0.47786666666666666,
+      "grad_norm": 0.7974304806295485,
+      "learning_rate": 1.7481880270062003e-05,
+      "loss": 0.2717,
+      "step": 56
+    },
+    {
+      "epoch": 0.4864,
+      "grad_norm": 0.6232203657585239,
+      "learning_rate": 1.7558748556724913e-05,
+      "loss": 0.1741,
+      "step": 57
+    },
+    {
+      "epoch": 0.49493333333333334,
+      "grad_norm": 0.6850688604563008,
+      "learning_rate": 1.763427993562937e-05,
+      "loss": 0.228,
+      "step": 58
+    },
+    {
+      "epoch": 0.5034666666666666,
+      "grad_norm": 0.5923826384300431,
+      "learning_rate": 1.7708520116421443e-05,
+      "loss": 0.2131,
+      "step": 59
+    },
+    {
+      "epoch": 0.512,
+      "grad_norm": 0.489754430485032,
+      "learning_rate": 1.7781512503836432e-05,
+      "loss": 0.165,
+      "step": 60
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 351,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 3,
+  "save_steps": 20,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 8.473824188421898e+17,
+  "train_batch_size": 16,
+  "trial_name": null,
+  "trial_params": null
+}
diff --git a/checkpoint-60/training_args.bin b/checkpoint-60/training_args.bin
new file mode 100644
index 0000000000000000000000000000000000000000..8ac842d29fc6ad3e132cb7208dda5b61dd1079ee
--- /dev/null
+++ b/checkpoint-60/training_args.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aa4edad1da5557fcd52a6da980443588016cead6f0444a3562cfa68029c66a04
+size 6840
diff --git a/checkpoint-60/zero_to_fp32.py b/checkpoint-60/zero_to_fp32.py
new file mode 100644
index 0000000000000000000000000000000000000000..49b846633d6eb1e836e34681e44033581f4edb7b
--- /dev/null
+++ b/checkpoint-60/zero_to_fp32.py
@@ -0,0 +1,592 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example: python zero_to_fp32.py . pytorch_model.bin
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+
+    total_files = len(files)
+    state_dicts = []
+    for f in files:
+        state_dict = torch.load(f, map_location=device)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    if zero_stage <= 2:
+        fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    elif zero_stage == 3:
+        # if there is more than one param group, there will be multiple flattened tensors - one
+        # flattened tensor per group - for simplicity merge them into a single tensor
+        #
+        # XXX: could make the script more memory efficient for when there are multiple groups - it
+        # will require matching the sub-lists of param_shapes for each param group flattened tensor
+
+        fp32_flat_groups = [
+            torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
+        ]
+
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    for name, shape in param_shapes.items():
+
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # XXX: memory usage doubles here
+        state_dict[name] = torch.cat(
+            tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
+            0).narrow(0, 0, unpartitioned_numel).view(shape)
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    Note: this approach may not work if your application doesn't have sufficient free CPU memory and
+    you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+    """
+
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+    print(f"Saving fp32 state dict to {output_file}")
+    torch.save(state_dict, output_file)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument(
+        "output_file",
+        type=str,
+        help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file, tag=args.tag)
diff --git a/checkpoint-80/README.md b/checkpoint-80/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..e7d4f706aeefc6bb2aa84e963870f345a14c4af9
--- /dev/null
+++ b/checkpoint-80/README.md
@@ -0,0 +1,202 @@
+---
+library_name: peft
+base_model: ../ckpts/Meta-Llama-3-8B-Instruct
+---
+
+# 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 0.11.1
\ No newline at end of file
diff --git a/checkpoint-80/adapter_config.json b/checkpoint-80/adapter_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..ad16c073ed8946d227350e8fe4e99d689a8ede54
--- /dev/null
+++ b/checkpoint-80/adapter_config.json
@@ -0,0 +1,35 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "../ckpts/Meta-Llama-3-8B-Instruct",
+  "bias": "none",
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 16,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "down_proj",
+    "gate_proj",
+    "q_proj",
+    "lm_head",
+    "o_proj",
+    "v_proj",
+    "up_proj",
+    "k_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "use_dora": false,
+  "use_rslora": false
+}
\ No newline at end of file
diff --git a/checkpoint-80/adapter_model.safetensors b/checkpoint-80/adapter_model.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..e5b220be87d770efa22c9a3bd0b89f8013ff3677
--- /dev/null
+++ b/checkpoint-80/adapter_model.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:10625511bb4a701e8f8e64bc3f7db95052839ef791bc49bc5a2d6b06919ef7b5
+size 1138856856
diff --git a/checkpoint-80/global_step80/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt b/checkpoint-80/global_step80/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..f20ca11895ff276ee101dec8924bb8886fb5c438
--- /dev/null
+++ b/checkpoint-80/global_step80/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3295f7e9af2a8907fb77c532e880faab2d3a9dcd2339152e00274dc3270d6d6e
+size 528781328
diff --git a/checkpoint-80/global_step80/mp_rank_00_model_states.pt b/checkpoint-80/global_step80/mp_rank_00_model_states.pt
new file mode 100644
index 0000000000000000000000000000000000000000..4597b6784e892de36b1d03a24211e18025f489da
--- /dev/null
+++ b/checkpoint-80/global_step80/mp_rank_00_model_states.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:40a127a519a4ea92e0fdc77e5cb7af8a1ba57fc56493947402fdaf9ea88074e0
+size 199905337
diff --git a/checkpoint-80/latest b/checkpoint-80/latest
new file mode 100644
index 0000000000000000000000000000000000000000..75eab498d0366633484ab40334e4b8fb92b16dad
--- /dev/null
+++ b/checkpoint-80/latest
@@ -0,0 +1 @@
+global_step80
\ No newline at end of file
diff --git a/checkpoint-80/rng_state.pth b/checkpoint-80/rng_state.pth
new file mode 100644
index 0000000000000000000000000000000000000000..8e517b54bddc9a584fd365a8263b41bc4db34a77
--- /dev/null
+++ b/checkpoint-80/rng_state.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3cc159e4a5aac19e6e964e22a137b013ca1c8ce5f2b961284169db4c58ac7f68
+size 14244
diff --git a/checkpoint-80/trainer_state.json b/checkpoint-80/trainer_state.json
new file mode 100644
index 0000000000000000000000000000000000000000..15e164c1bb266a4e40a02fdf6b068c5df922b0b9
--- /dev/null
+++ b/checkpoint-80/trainer_state.json
@@ -0,0 +1,593 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.6826666666666666,
+  "eval_steps": 500,
+  "global_step": 80,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.008533333333333334,
+      "grad_norm": 100.21848203113535,
+      "learning_rate": 0.0,
+      "loss": 7.1962,
+      "step": 1
+    },
+    {
+      "epoch": 0.017066666666666667,
+      "grad_norm": 96.43006188910957,
+      "learning_rate": 3.0102999566398115e-06,
+      "loss": 6.9414,
+      "step": 2
+    },
+    {
+      "epoch": 0.0256,
+      "grad_norm": 97.35803466618715,
+      "learning_rate": 4.771212547196624e-06,
+      "loss": 7.0102,
+      "step": 3
+    },
+    {
+      "epoch": 0.034133333333333335,
+      "grad_norm": 95.14837816372646,
+      "learning_rate": 6.020599913279623e-06,
+      "loss": 6.5295,
+      "step": 4
+    },
+    {
+      "epoch": 0.042666666666666665,
+      "grad_norm": 91.76544275692784,
+      "learning_rate": 6.989700043360187e-06,
+      "loss": 6.4806,
+      "step": 5
+    },
+    {
+      "epoch": 0.0512,
+      "grad_norm": 84.4494318688335,
+      "learning_rate": 7.781512503836437e-06,
+      "loss": 6.4194,
+      "step": 6
+    },
+    {
+      "epoch": 0.05973333333333333,
+      "grad_norm": 71.37977490595638,
+      "learning_rate": 8.450980400142568e-06,
+      "loss": 5.4953,
+      "step": 7
+    },
+    {
+      "epoch": 0.06826666666666667,
+      "grad_norm": 49.31153456754566,
+      "learning_rate": 9.030899869919434e-06,
+      "loss": 5.4123,
+      "step": 8
+    },
+    {
+      "epoch": 0.0768,
+      "grad_norm": 20.37296364560341,
+      "learning_rate": 9.542425094393249e-06,
+      "loss": 5.2334,
+      "step": 9
+    },
+    {
+      "epoch": 0.08533333333333333,
+      "grad_norm": 7.968467079076881,
+      "learning_rate": 9.999999999999999e-06,
+      "loss": 5.0282,
+      "step": 10
+    },
+    {
+      "epoch": 0.09386666666666667,
+      "grad_norm": 3.559446532055649,
+      "learning_rate": 1.041392685158225e-05,
+      "loss": 4.612,
+      "step": 11
+    },
+    {
+      "epoch": 0.1024,
+      "grad_norm": 3.5528846947995674,
+      "learning_rate": 1.0791812460476248e-05,
+      "loss": 4.9475,
+      "step": 12
+    },
+    {
+      "epoch": 0.11093333333333333,
+      "grad_norm": 3.541968897471334,
+      "learning_rate": 1.1139433523068365e-05,
+      "loss": 4.2777,
+      "step": 13
+    },
+    {
+      "epoch": 0.11946666666666667,
+      "grad_norm": 3.54718070036198,
+      "learning_rate": 1.1461280356782378e-05,
+      "loss": 4.3507,
+      "step": 14
+    },
+    {
+      "epoch": 0.128,
+      "grad_norm": 3.8632334830606747,
+      "learning_rate": 1.1760912590556813e-05,
+      "loss": 4.5364,
+      "step": 15
+    },
+    {
+      "epoch": 0.13653333333333334,
+      "grad_norm": 3.6637424744054004,
+      "learning_rate": 1.2041199826559246e-05,
+      "loss": 3.9672,
+      "step": 16
+    },
+    {
+      "epoch": 0.14506666666666668,
+      "grad_norm": 3.919802904818311,
+      "learning_rate": 1.230448921378274e-05,
+      "loss": 4.0618,
+      "step": 17
+    },
+    {
+      "epoch": 0.1536,
+      "grad_norm": 4.71904950738746,
+      "learning_rate": 1.2552725051033058e-05,
+      "loss": 4.6656,
+      "step": 18
+    },
+    {
+      "epoch": 0.16213333333333332,
+      "grad_norm": 4.6656317698690835,
+      "learning_rate": 1.2787536009528288e-05,
+      "loss": 4.1131,
+      "step": 19
+    },
+    {
+      "epoch": 0.17066666666666666,
+      "grad_norm": 5.145138692367417,
+      "learning_rate": 1.301029995663981e-05,
+      "loss": 4.0989,
+      "step": 20
+    },
+    {
+      "epoch": 0.1792,
+      "grad_norm": 5.923538014759818,
+      "learning_rate": 1.3222192947339192e-05,
+      "loss": 4.4991,
+      "step": 21
+    },
+    {
+      "epoch": 0.18773333333333334,
+      "grad_norm": 5.941056962941364,
+      "learning_rate": 1.3424226808222062e-05,
+      "loss": 4.0836,
+      "step": 22
+    },
+    {
+      "epoch": 0.19626666666666667,
+      "grad_norm": 6.171026012117947,
+      "learning_rate": 1.3617278360175927e-05,
+      "loss": 3.6861,
+      "step": 23
+    },
+    {
+      "epoch": 0.2048,
+      "grad_norm": 7.130542138930838,
+      "learning_rate": 1.380211241711606e-05,
+      "loss": 4.0958,
+      "step": 24
+    },
+    {
+      "epoch": 0.21333333333333335,
+      "grad_norm": 7.328837606110418,
+      "learning_rate": 1.3979400086720374e-05,
+      "loss": 3.9524,
+      "step": 25
+    },
+    {
+      "epoch": 0.22186666666666666,
+      "grad_norm": 6.923489005711429,
+      "learning_rate": 1.4149733479708178e-05,
+      "loss": 3.6062,
+      "step": 26
+    },
+    {
+      "epoch": 0.2304,
+      "grad_norm": 6.409498926059221,
+      "learning_rate": 1.4313637641589872e-05,
+      "loss": 3.2034,
+      "step": 27
+    },
+    {
+      "epoch": 0.23893333333333333,
+      "grad_norm": 5.781628405584682,
+      "learning_rate": 1.4471580313422191e-05,
+      "loss": 2.8158,
+      "step": 28
+    },
+    {
+      "epoch": 0.24746666666666667,
+      "grad_norm": 6.2927590068243315,
+      "learning_rate": 1.4623979978989559e-05,
+      "loss": 2.9803,
+      "step": 29
+    },
+    {
+      "epoch": 0.256,
+      "grad_norm": 6.103844678630006,
+      "learning_rate": 1.4771212547196623e-05,
+      "loss": 2.847,
+      "step": 30
+    },
+    {
+      "epoch": 0.26453333333333334,
+      "grad_norm": 7.656341925867144,
+      "learning_rate": 1.4913616938342726e-05,
+      "loss": 3.0907,
+      "step": 31
+    },
+    {
+      "epoch": 0.2730666666666667,
+      "grad_norm": 6.324242877501844,
+      "learning_rate": 1.5051499783199059e-05,
+      "loss": 2.3467,
+      "step": 32
+    },
+    {
+      "epoch": 0.2816,
+      "grad_norm": 7.606313101162739,
+      "learning_rate": 1.5185139398778874e-05,
+      "loss": 2.5292,
+      "step": 33
+    },
+    {
+      "epoch": 0.29013333333333335,
+      "grad_norm": 8.553792493849265,
+      "learning_rate": 1.531478917042255e-05,
+      "loss": 2.4547,
+      "step": 34
+    },
+    {
+      "epoch": 0.2986666666666667,
+      "grad_norm": 8.483368703272543,
+      "learning_rate": 1.5440680443502753e-05,
+      "loss": 2.1956,
+      "step": 35
+    },
+    {
+      "epoch": 0.3072,
+      "grad_norm": 8.364739352838077,
+      "learning_rate": 1.5563025007672873e-05,
+      "loss": 1.8552,
+      "step": 36
+    },
+    {
+      "epoch": 0.3157333333333333,
+      "grad_norm": 9.37663682000104,
+      "learning_rate": 1.5682017240669948e-05,
+      "loss": 1.9228,
+      "step": 37
+    },
+    {
+      "epoch": 0.32426666666666665,
+      "grad_norm": 9.823047193440066,
+      "learning_rate": 1.57978359661681e-05,
+      "loss": 1.7033,
+      "step": 38
+    },
+    {
+      "epoch": 0.3328,
+      "grad_norm": 9.692618512955894,
+      "learning_rate": 1.591064607026499e-05,
+      "loss": 1.3768,
+      "step": 39
+    },
+    {
+      "epoch": 0.3413333333333333,
+      "grad_norm": 9.08889021911031,
+      "learning_rate": 1.6020599913279622e-05,
+      "loss": 1.3015,
+      "step": 40
+    },
+    {
+      "epoch": 0.34986666666666666,
+      "grad_norm": 8.081534221516058,
+      "learning_rate": 1.6127838567197353e-05,
+      "loss": 0.9228,
+      "step": 41
+    },
+    {
+      "epoch": 0.3584,
+      "grad_norm": 6.238638048950311,
+      "learning_rate": 1.6232492903979005e-05,
+      "loss": 0.7267,
+      "step": 42
+    },
+    {
+      "epoch": 0.36693333333333333,
+      "grad_norm": 3.4058036861773604,
+      "learning_rate": 1.6334684555795865e-05,
+      "loss": 0.5875,
+      "step": 43
+    },
+    {
+      "epoch": 0.37546666666666667,
+      "grad_norm": 2.079163829467713,
+      "learning_rate": 1.6434526764861872e-05,
+      "loss": 0.6355,
+      "step": 44
+    },
+    {
+      "epoch": 0.384,
+      "grad_norm": 1.5597487833024746,
+      "learning_rate": 1.6532125137753435e-05,
+      "loss": 0.5106,
+      "step": 45
+    },
+    {
+      "epoch": 0.39253333333333335,
+      "grad_norm": 2.491689602375256,
+      "learning_rate": 1.662757831681574e-05,
+      "loss": 0.6454,
+      "step": 46
+    },
+    {
+      "epoch": 0.4010666666666667,
+      "grad_norm": 2.010880438195854,
+      "learning_rate": 1.672097857935717e-05,
+      "loss": 0.4757,
+      "step": 47
+    },
+    {
+      "epoch": 0.4096,
+      "grad_norm": 1.9452805114322096,
+      "learning_rate": 1.681241237375587e-05,
+      "loss": 0.4133,
+      "step": 48
+    },
+    {
+      "epoch": 0.41813333333333336,
+      "grad_norm": 1.7620848552306103,
+      "learning_rate": 1.6901960800285137e-05,
+      "loss": 0.4004,
+      "step": 49
+    },
+    {
+      "epoch": 0.4266666666666667,
+      "grad_norm": 1.278224489774809,
+      "learning_rate": 1.6989700043360187e-05,
+      "loss": 0.3523,
+      "step": 50
+    },
+    {
+      "epoch": 0.4352,
+      "grad_norm": 1.6151354758303231,
+      "learning_rate": 1.7075701760979363e-05,
+      "loss": 0.4317,
+      "step": 51
+    },
+    {
+      "epoch": 0.4437333333333333,
+      "grad_norm": 1.3451396055695035,
+      "learning_rate": 1.716003343634799e-05,
+      "loss": 0.3474,
+      "step": 52
+    },
+    {
+      "epoch": 0.45226666666666665,
+      "grad_norm": 1.6814977782362666,
+      "learning_rate": 1.724275869600789e-05,
+      "loss": 0.3706,
+      "step": 53
+    },
+    {
+      "epoch": 0.4608,
+      "grad_norm": 1.1682442432688667,
+      "learning_rate": 1.7323937598229687e-05,
+      "loss": 0.3488,
+      "step": 54
+    },
+    {
+      "epoch": 0.4693333333333333,
+      "grad_norm": 0.8839814540462471,
+      "learning_rate": 1.7403626894942437e-05,
+      "loss": 0.293,
+      "step": 55
+    },
+    {
+      "epoch": 0.47786666666666666,
+      "grad_norm": 0.7974304806295485,
+      "learning_rate": 1.7481880270062003e-05,
+      "loss": 0.2717,
+      "step": 56
+    },
+    {
+      "epoch": 0.4864,
+      "grad_norm": 0.6232203657585239,
+      "learning_rate": 1.7558748556724913e-05,
+      "loss": 0.1741,
+      "step": 57
+    },
+    {
+      "epoch": 0.49493333333333334,
+      "grad_norm": 0.6850688604563008,
+      "learning_rate": 1.763427993562937e-05,
+      "loss": 0.228,
+      "step": 58
+    },
+    {
+      "epoch": 0.5034666666666666,
+      "grad_norm": 0.5923826384300431,
+      "learning_rate": 1.7708520116421443e-05,
+      "loss": 0.2131,
+      "step": 59
+    },
+    {
+      "epoch": 0.512,
+      "grad_norm": 0.489754430485032,
+      "learning_rate": 1.7781512503836432e-05,
+      "loss": 0.165,
+      "step": 60
+    },
+    {
+      "epoch": 0.5205333333333333,
+      "grad_norm": 0.5280940052395061,
+      "learning_rate": 1.7853298350107667e-05,
+      "loss": 0.1658,
+      "step": 61
+    },
+    {
+      "epoch": 0.5290666666666667,
+      "grad_norm": 0.4750905992036739,
+      "learning_rate": 1.7923916894982537e-05,
+      "loss": 0.1438,
+      "step": 62
+    },
+    {
+      "epoch": 0.5376,
+      "grad_norm": 0.4162998392722401,
+      "learning_rate": 1.7993405494535815e-05,
+      "loss": 0.1555,
+      "step": 63
+    },
+    {
+      "epoch": 0.5461333333333334,
+      "grad_norm": 0.26863266246370443,
+      "learning_rate": 1.806179973983887e-05,
+      "loss": 0.1323,
+      "step": 64
+    },
+    {
+      "epoch": 0.5546666666666666,
+      "grad_norm": 0.26534221125601215,
+      "learning_rate": 1.8129133566428553e-05,
+      "loss": 0.1671,
+      "step": 65
+    },
+    {
+      "epoch": 0.5632,
+      "grad_norm": 0.2548662962257576,
+      "learning_rate": 1.8195439355418686e-05,
+      "loss": 0.1308,
+      "step": 66
+    },
+    {
+      "epoch": 0.5717333333333333,
+      "grad_norm": 0.18045394638578796,
+      "learning_rate": 1.8260748027008263e-05,
+      "loss": 0.1262,
+      "step": 67
+    },
+    {
+      "epoch": 0.5802666666666667,
+      "grad_norm": 0.17070388073612064,
+      "learning_rate": 1.8325089127062364e-05,
+      "loss": 0.1192,
+      "step": 68
+    },
+    {
+      "epoch": 0.5888,
+      "grad_norm": 0.1531381679244776,
+      "learning_rate": 1.8388490907372553e-05,
+      "loss": 0.1274,
+      "step": 69
+    },
+    {
+      "epoch": 0.5973333333333334,
+      "grad_norm": 0.18196434993687946,
+      "learning_rate": 1.8450980400142568e-05,
+      "loss": 0.1375,
+      "step": 70
+    },
+    {
+      "epoch": 0.6058666666666667,
+      "grad_norm": 0.15324416972951205,
+      "learning_rate": 1.8512583487190752e-05,
+      "loss": 0.1599,
+      "step": 71
+    },
+    {
+      "epoch": 0.6144,
+      "grad_norm": 0.10884462064503801,
+      "learning_rate": 1.857332496431268e-05,
+      "loss": 0.1041,
+      "step": 72
+    },
+    {
+      "epoch": 0.6229333333333333,
+      "grad_norm": 0.12915133528192668,
+      "learning_rate": 1.8633228601204555e-05,
+      "loss": 0.1406,
+      "step": 73
+    },
+    {
+      "epoch": 0.6314666666666666,
+      "grad_norm": 0.12553425699952878,
+      "learning_rate": 1.8692317197309763e-05,
+      "loss": 0.1256,
+      "step": 74
+    },
+    {
+      "epoch": 0.64,
+      "grad_norm": 0.11976960918968543,
+      "learning_rate": 1.8750612633916997e-05,
+      "loss": 0.1144,
+      "step": 75
+    },
+    {
+      "epoch": 0.6485333333333333,
+      "grad_norm": 0.115805998298789,
+      "learning_rate": 1.8808135922807914e-05,
+      "loss": 0.1528,
+      "step": 76
+    },
+    {
+      "epoch": 0.6570666666666667,
+      "grad_norm": 0.10325948496697443,
+      "learning_rate": 1.8864907251724818e-05,
+      "loss": 0.1044,
+      "step": 77
+    },
+    {
+      "epoch": 0.6656,
+      "grad_norm": 0.09595064346541006,
+      "learning_rate": 1.8920946026904802e-05,
+      "loss": 0.1534,
+      "step": 78
+    },
+    {
+      "epoch": 0.6741333333333334,
+      "grad_norm": 0.08796742845240496,
+      "learning_rate": 1.8976270912904414e-05,
+      "loss": 0.1155,
+      "step": 79
+    },
+    {
+      "epoch": 0.6826666666666666,
+      "grad_norm": 0.08218991738379527,
+      "learning_rate": 1.9030899869919434e-05,
+      "loss": 0.1311,
+      "step": 80
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 351,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 3,
+  "save_steps": 20,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 1.1373018660449812e+18,
+  "train_batch_size": 16,
+  "trial_name": null,
+  "trial_params": null
+}
diff --git a/checkpoint-80/training_args.bin b/checkpoint-80/training_args.bin
new file mode 100644
index 0000000000000000000000000000000000000000..8ac842d29fc6ad3e132cb7208dda5b61dd1079ee
--- /dev/null
+++ b/checkpoint-80/training_args.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aa4edad1da5557fcd52a6da980443588016cead6f0444a3562cfa68029c66a04
+size 6840
diff --git a/checkpoint-80/zero_to_fp32.py b/checkpoint-80/zero_to_fp32.py
new file mode 100644
index 0000000000000000000000000000000000000000..49b846633d6eb1e836e34681e44033581f4edb7b
--- /dev/null
+++ b/checkpoint-80/zero_to_fp32.py
@@ -0,0 +1,592 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example: python zero_to_fp32.py . pytorch_model.bin
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+
+    total_files = len(files)
+    state_dicts = []
+    for f in files:
+        state_dict = torch.load(f, map_location=device)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    if zero_stage <= 2:
+        fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    elif zero_stage == 3:
+        # if there is more than one param group, there will be multiple flattened tensors - one
+        # flattened tensor per group - for simplicity merge them into a single tensor
+        #
+        # XXX: could make the script more memory efficient for when there are multiple groups - it
+        # will require matching the sub-lists of param_shapes for each param group flattened tensor
+
+        fp32_flat_groups = [
+            torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
+        ]
+
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    for name, shape in param_shapes.items():
+
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # XXX: memory usage doubles here
+        state_dict[name] = torch.cat(
+            tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
+            0).narrow(0, 0, unpartitioned_numel).view(shape)
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    Note: this approach may not work if your application doesn't have sufficient free CPU memory and
+    you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+    """
+
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+    print(f"Saving fp32 state dict to {output_file}")
+    torch.save(state_dict, output_file)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument(
+        "output_file",
+        type=str,
+        help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file, tag=args.tag)
diff --git a/config.yaml b/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e10622613c4b9b525db646e9c45db6eda2d898bf
--- /dev/null
+++ b/config.yaml
@@ -0,0 +1,74 @@
+data:
+  prompt: llama3_formal
+  train: ../data/susgen/FINAL/PER_3500/FINAL_PER3500_30k.json
+  val: null
+  val_split_ratio: 0.005
+device: cuda
+instruct_mask: true
+local_rank: 0
+model:
+  acceleration: null
+  int4_config:
+    bnb_4bit_compute_dtype: bfloat16
+    bnb_4bit_quant_type: nf4
+    bnb_4bit_use_double_quant: true
+    load_in_4bit: true
+    load_in_8bit: false
+  int8_config:
+    load_in_4bit: false
+    load_in_8bit: true
+  lora:
+    bias: none
+    inference_mode: false
+    lora_alpha: 32
+    lora_dropout: 0.1
+    r: 16
+    target_modules:
+    - q_proj
+    - k_proj
+    - v_proj
+    - o_proj
+    - gate_proj
+    - up_proj
+    - down_proj
+    - lm_head
+    task_type: CAUSAL_LM
+  lora_path: false
+  model_path: ../ckpts/Meta-Llama-3-8B-Instruct
+  quantization: int4
+  seed: 2024
+  show_config: false
+  use_lora: true
+name: 30k-Llama3-8B-Instruct
+output_dir: ../results/
+tokenizer:
+  add_bos_token: true
+  add_eos_token: false
+  add_prefix_space: false
+  encode:
+    max_length: 2048
+    return_tensors: pt
+    truncation: true
+  model_max_length: 2048
+  padding_side: left
+  pretrained_model_name_or_path: ../ckpts/Meta-Llama-3-8B-Instruct
+  truncation_side: right
+  use_fast: true
+trainer: NewTrainer
+training:
+  bf16: true
+  deepspeed: ./configs/ds_configs/ds_config_stage_2.json
+  gradient_accumulation_steps: 16
+  learning_rate: 2.0e-05
+  logging_steps: 1
+  lr_scheduler_type: cosine
+  num_train_epochs: 3
+  optim: paged_adamw_32bit
+  per_device_train_batch_size: 16
+  remove_unused_columns: false
+  report_to: wandb
+  resume_from_checkpoint: null
+  save_steps: 20
+  save_strategy: steps
+  warmup_steps: 100
+  weight_decay: 0.01