Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,7 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+pytorch_model.bin00 filter=lfs diff=lfs merge=lfs -text
+pytorch_model.bin01 filter=lfs diff=lfs merge=lfs -text
+pytorch_model.bin02 filter=lfs diff=lfs merge=lfs -text
+pytorch_model.bin03 filter=lfs diff=lfs merge=lfs -text

latest

@@ -0,0 +1 @@
+global_step400

pytorch_model.bin00

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:84688ba4575f263eb196a9ea53b74575d0f125c11a7debf0ce09b1416756fd4c
+size 20000000000

pytorch_model.bin01

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6dcdb12d08ebbda10c423f933c2127c7677d6d4c50a20ef34f9f5950fdea5543
+size 20000000000

pytorch_model.bin02

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2ee36326dfc1db3d192cc9be38c992c6fedfe256205eef81a5d726d81ec8fa60
+size 20000000000

pytorch_model.bin03

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8226804c339d0a8f3fdbb0930f8acdb18453240b3a6362074bfff7fdf6439f22
+size 8778066425

rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3325e5522dd0e76d77f51b712ce830a70a2c478865eb15a68ffe9a623451e309
+size 21687

rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0c4467a2ae97d673892c18c7cac38b25e4f8a49eb6e46b80eab84253fda3acc3
+size 21687

rng_state_2.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3703baeed556d24505246fa48122667c1e210a3e1bb2b9b466b0f66f64fc4f37
+size 21687

rng_state_3.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dd5325bc5a3ef2f0a16edfe8622ca2f8e7df14fcc0770b7bae8fbd49c7f36ea8
+size 21687

rng_state_4.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ca5fbe88c50d42730361a56567ac258872b50bbd20970599ef9787d6a3d59b15
+size 21687

rng_state_5.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:58b05009ec6cb6e915e03c2b655bb218080c7ee85bd7d0d49bd9579df367b67b
+size 21687

rng_state_6.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bb508ceb56e538bdafb465e4213db07bdcdf36d2d54b795f92df0cf54e74beb0
+size 21687

rng_state_7.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fa375275d7d4f35c68b510c479335930f40f6c6913040807ac747cd7580a1493
+size 21687

trainer_state.json

@@ -0,0 +1,2493 @@
+{
+  "best_metric": 1.0441503524780273,
+  "best_model_checkpoint": "rm-ckpt-reward-p100-s0/checkpoint-400",
+  "epoch": 0.4020504573323952,
+  "eval_steps": 50,
+  "global_step": 400,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.0,
+      "learning_rate": 0.0,
+      "loss": 1.447,
+      "step": 1
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 7.525749891599529e-06,
+      "loss": 1.4166,
+      "step": 2
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 1.192803136799156e-05,
+      "loss": 1.417,
+      "step": 3
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 1.5051499783199057e-05,
+      "loss": 1.3947,
+      "step": 4
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 1.7474250108400467e-05,
+      "loss": 1.3739,
+      "step": 5
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 1.945378125959109e-05,
+      "loss": 1.3352,
+      "step": 6
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 2.1127451000356418e-05,
+      "loss": 1.3181,
+      "step": 7
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 2.2577249674798584e-05,
+      "loss": 1.3712,
+      "step": 8
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 2.385606273598312e-05,
+      "loss": 1.3858,
+      "step": 9
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 2.4999999999999998e-05,
+      "loss": 1.4424,
+      "step": 10
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 2.6034817128955623e-05,
+      "loss": 1.3381,
+      "step": 11
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 2.6979531151190617e-05,
+      "loss": 1.2937,
+      "step": 12
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 2.7848583807670913e-05,
+      "loss": 1.3021,
+      "step": 13
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 2.8653200891955945e-05,
+      "loss": 1.274,
+      "step": 14
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 2.940228147639203e-05,
+      "loss": 1.2896,
+      "step": 15
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 3.0102999566398115e-05,
+      "loss": 1.19,
+      "step": 16
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 3.076122303445685e-05,
+      "loss": 1.1889,
+      "step": 17
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 3.1381812627582646e-05,
+      "loss": 1.2028,
+      "step": 18
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 3.1968840023820715e-05,
+      "loss": 1.2259,
+      "step": 19
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 3.2525749891599525e-05,
+      "loss": 1.1971,
+      "step": 20
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 3.305548236834798e-05,
+      "loss": 1.1972,
+      "step": 21
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 3.3560567020555153e-05,
+      "loss": 1.2193,
+      "step": 22
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 3.404319590043982e-05,
+      "loss": 1.1354,
+      "step": 23
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 3.450528104279015e-05,
+      "loss": 1.1434,
+      "step": 24
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 3.4948500216800935e-05,
+      "loss": 1.1237,
+      "step": 25
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 3.537433369927044e-05,
+      "loss": 1.1227,
+      "step": 26
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 3.578409410397468e-05,
+      "loss": 1.1554,
+      "step": 27
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 3.6178950783555475e-05,
+      "loss": 1.131,
+      "step": 28
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 3.65599499474739e-05,
+      "loss": 1.2016,
+      "step": 29
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 3.6928031367991554e-05,
+      "loss": 1.0968,
+      "step": 30
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 3.728404234585681e-05,
+      "loss": 1.0713,
+      "step": 31
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 3.762874945799765e-05,
+      "loss": 1.1537,
+      "step": 32
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 3.796284849694718e-05,
+      "loss": 1.0856,
+      "step": 33
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 3.8286972926056376e-05,
+      "loss": 1.08,
+      "step": 34
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 3.8601701108756885e-05,
+      "loss": 1.067,
+      "step": 35
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 3.890756251918218e-05,
+      "loss": 1.1298,
+      "step": 36
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 3.920504310167487e-05,
+      "loss": 1.0986,
+      "step": 37
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 3.949458991542025e-05,
+      "loss": 1.0649,
+      "step": 38
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 3.977661517566247e-05,
+      "loss": 1.0582,
+      "step": 39
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 4.005149978319905e-05,
+      "loss": 1.1037,
+      "step": 40
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 4.031959641799338e-05,
+      "loss": 1.052,
+      "step": 41
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 4.058123225994751e-05,
+      "loss": 1.1048,
+      "step": 42
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 4.0836711389489654e-05,
+      "loss": 1.0817,
+      "step": 43
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 4.108631691215468e-05,
+      "loss": 1.0712,
+      "step": 44
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 4.133031284438358e-05,
+      "loss": 1.0865,
+      "step": 45
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 4.156894579203935e-05,
+      "loss": 1.0837,
+      "step": 46
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 4.180244644839293e-05,
+      "loss": 1.0915,
+      "step": 47
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 4.203103093438968e-05,
+      "loss": 1.0845,
+      "step": 48
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 4.2254902000712836e-05,
+      "loss": 1.0425,
+      "step": 49
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 4.247425010840046e-05,
+      "loss": 1.0479,
+      "step": 50
+    },
+    {
+      "epoch": 0.05,
+      "eval_val_accuracy": 0.6191666666666666,
+      "eval_val_loss": 1.1118515729904175,
+      "eval_val_runtime": 913.0126,
+      "eval_val_samples_per_second": 1.095,
+      "eval_val_steps_per_second": 0.137,
+      "step": 50
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 4.2689254402448405e-05,
+      "loss": 1.079,
+      "step": 51
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 4.290008359086998e-05,
+      "loss": 1.0855,
+      "step": 52
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 4.310689674001973e-05,
+      "loss": 1.084,
+      "step": 53
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 4.330984399557421e-05,
+      "loss": 1.0787,
+      "step": 54
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 4.350906723735609e-05,
+      "loss": 1.1333,
+      "step": 55
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 4.370470067515501e-05,
+      "loss": 1.1016,
+      "step": 56
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 4.3896871391812285e-05,
+      "loss": 1.0716,
+      "step": 57
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 4.408569983907343e-05,
+      "loss": 1.0861,
+      "step": 58
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 4.42713002910536e-05,
+      "loss": 1.0583,
+      "step": 59
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 4.445378125959108e-05,
+      "loss": 1.0793,
+      "step": 60
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 4.463324587526917e-05,
+      "loss": 1.0761,
+      "step": 61
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 4.4809792237456346e-05,
+      "loss": 1.0995,
+      "step": 62
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 4.498351373633954e-05,
+      "loss": 1.0915,
+      "step": 63
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 4.515449934959717e-05,
+      "loss": 1.0415,
+      "step": 64
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 4.532283391607138e-05,
+      "loss": 1.0429,
+      "step": 65
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 4.548859838854671e-05,
+      "loss": 1.0886,
+      "step": 66
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 4.565187006752065e-05,
+      "loss": 1.0332,
+      "step": 67
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 4.581272281765591e-05,
+      "loss": 1.0271,
+      "step": 68
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 4.597122726843138e-05,
+      "loss": 1.0591,
+      "step": 69
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 4.612745100035642e-05,
+      "loss": 1.023,
+      "step": 70
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 4.628145871797688e-05,
+      "loss": 1.0665,
+      "step": 71
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 4.643331241078171e-05,
+      "loss": 1.064,
+      "step": 72
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 4.658307150301139e-05,
+      "loss": 1.1339,
+      "step": 73
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 4.67307929932744e-05,
+      "loss": 1.0609,
+      "step": 74
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 4.687653158479249e-05,
+      "loss": 1.1128,
+      "step": 75
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 4.702033980701978e-05,
+      "loss": 1.0733,
+      "step": 76
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 4.716226812931204e-05,
+      "loss": 1.0761,
+      "step": 77
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 4.7302365067262006e-05,
+      "loss": 1.0609,
+      "step": 78
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 4.744067728226103e-05,
+      "loss": 1.0306,
+      "step": 79
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 4.757724967479858e-05,
+      "loss": 1.0882,
+      "step": 80
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 4.771212547196624e-05,
+      "loss": 1.0556,
+      "step": 81
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 4.7845346309592914e-05,
+      "loss": 1.0455,
+      "step": 82
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 4.7976952309401844e-05,
+      "loss": 1.0653,
+      "step": 83
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 4.810698215154703e-05,
+      "loss": 1.0128,
+      "step": 84
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 4.823547314285732e-05,
+      "loss": 1.0739,
+      "step": 85
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 4.836246128108918e-05,
+      "loss": 1.161,
+      "step": 86
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 4.8487981315465456e-05,
+      "loss": 1.0686,
+      "step": 87
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 4.8612066803754214e-05,
+      "loss": 1.0785,
+      "step": 88
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 4.873475016612281e-05,
+      "loss": 1.0362,
+      "step": 89
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 4.885606273598312e-05,
+      "loss": 1.0775,
+      "step": 90
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 4.897603480802733e-05,
+      "loss": 1.0445,
+      "step": 91
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 4.909469568363888e-05,
+      "loss": 1.0452,
+      "step": 92
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 4.9212073713848375e-05,
+      "loss": 1.0687,
+      "step": 93
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 4.932819633999246e-05,
+      "loss": 1.0775,
+      "step": 94
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 4.9443090132221186e-05,
+      "loss": 1.0239,
+      "step": 95
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 4.9556780825989205e-05,
+      "loss": 1.0532,
+      "step": 96
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 4.9669293356656114e-05,
+      "loss": 1.0628,
+      "step": 97
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 4.978065189231237e-05,
+      "loss": 1.0562,
+      "step": 98
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 4.989087986493874e-05,
+      "loss": 1.0404,
+      "step": 99
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 4.9999999999999996e-05,
+      "loss": 1.0711,
+      "step": 100
+    },
+    {
+      "epoch": 0.1,
+      "eval_val_accuracy": 0.6268333333333334,
+      "eval_val_loss": 1.0909922122955322,
+      "eval_val_runtime": 906.1076,
+      "eval_val_samples_per_second": 1.104,
+      "eval_val_steps_per_second": 0.138,
+      "step": 100
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 1e-05,
+      "loss": 1.0378,
+      "step": 101
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 1e-05,
+      "loss": 1.0447,
+      "step": 102
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 1e-05,
+      "loss": 1.0323,
+      "step": 103
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 1e-05,
+      "loss": 1.0135,
+      "step": 104
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 1e-05,
+      "loss": 1.0168,
+      "step": 105
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 1e-05,
+      "loss": 1.0689,
+      "step": 106
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 1e-05,
+      "loss": 1.0389,
+      "step": 107
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 1e-05,
+      "loss": 1.0665,
+      "step": 108
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 1e-05,
+      "loss": 1.0668,
+      "step": 109
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 1e-05,
+      "loss": 1.0728,
+      "step": 110
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 1e-05,
+      "loss": 1.0747,
+      "step": 111
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 1e-05,
+      "loss": 1.0314,
+      "step": 112
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 1e-05,
+      "loss": 1.1075,
+      "step": 113
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 1e-05,
+      "loss": 1.0645,
+      "step": 114
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 1e-05,
+      "loss": 1.0528,
+      "step": 115
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 1e-05,
+      "loss": 1.0783,
+      "step": 116
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 1e-05,
+      "loss": 1.0524,
+      "step": 117
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 1e-05,
+      "loss": 1.0451,
+      "step": 118
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 1e-05,
+      "loss": 1.0281,
+      "step": 119
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 1e-05,
+      "loss": 1.0829,
+      "step": 120
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 1e-05,
+      "loss": 1.0088,
+      "step": 121
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 1e-05,
+      "loss": 1.0634,
+      "step": 122
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 1e-05,
+      "loss": 1.0836,
+      "step": 123
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 1e-05,
+      "loss": 1.08,
+      "step": 124
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 1e-05,
+      "loss": 1.0523,
+      "step": 125
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 1e-05,
+      "loss": 1.0564,
+      "step": 126
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 1e-05,
+      "loss": 1.038,
+      "step": 127
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 1e-05,
+      "loss": 1.0494,
+      "step": 128
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 1e-05,
+      "loss": 1.0329,
+      "step": 129
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 1e-05,
+      "loss": 1.0806,
+      "step": 130
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 1e-05,
+      "loss": 1.0763,
+      "step": 131
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 1e-05,
+      "loss": 1.0504,
+      "step": 132
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 1e-05,
+      "loss": 1.044,
+      "step": 133
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 1e-05,
+      "loss": 0.9895,
+      "step": 134
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 1e-05,
+      "loss": 1.0688,
+      "step": 135
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 1e-05,
+      "loss": 1.0471,
+      "step": 136
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 1e-05,
+      "loss": 1.0711,
+      "step": 137
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 1e-05,
+      "loss": 1.0506,
+      "step": 138
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 1e-05,
+      "loss": 1.0789,
+      "step": 139
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 1e-05,
+      "loss": 1.0888,
+      "step": 140
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 1e-05,
+      "loss": 1.0746,
+      "step": 141
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 1e-05,
+      "loss": 1.1229,
+      "step": 142
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 1e-05,
+      "loss": 1.0826,
+      "step": 143
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 1e-05,
+      "loss": 1.0504,
+      "step": 144
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 1e-05,
+      "loss": 1.0311,
+      "step": 145
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 1e-05,
+      "loss": 1.0456,
+      "step": 146
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 1e-05,
+      "loss": 1.0097,
+      "step": 147
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 1e-05,
+      "loss": 1.0282,
+      "step": 148
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 1e-05,
+      "loss": 1.0507,
+      "step": 149
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 1e-05,
+      "loss": 1.054,
+      "step": 150
+    },
+    {
+      "epoch": 0.15,
+      "eval_val_accuracy": 0.626,
+      "eval_val_loss": 1.0630038976669312,
+      "eval_val_runtime": 907.7978,
+      "eval_val_samples_per_second": 1.102,
+      "eval_val_steps_per_second": 0.138,
+      "step": 150
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 1e-05,
+      "loss": 1.0165,
+      "step": 151
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 1e-05,
+      "loss": 0.9992,
+      "step": 152
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 1e-05,
+      "loss": 1.045,
+      "step": 153
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 1e-05,
+      "loss": 1.0378,
+      "step": 154
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 1e-05,
+      "loss": 1.0502,
+      "step": 155
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 1e-05,
+      "loss": 0.9911,
+      "step": 156
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 1e-05,
+      "loss": 1.063,
+      "step": 157
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 1e-05,
+      "loss": 1.0383,
+      "step": 158
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 1e-05,
+      "loss": 1.0511,
+      "step": 159
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 1e-05,
+      "loss": 1.0204,
+      "step": 160
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 1e-05,
+      "loss": 1.084,
+      "step": 161
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 1e-05,
+      "loss": 1.0873,
+      "step": 162
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 1e-05,
+      "loss": 1.0304,
+      "step": 163
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 1e-05,
+      "loss": 1.0118,
+      "step": 164
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 1e-05,
+      "loss": 1.04,
+      "step": 165
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 1e-05,
+      "loss": 1.0018,
+      "step": 166
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 1e-05,
+      "loss": 1.0432,
+      "step": 167
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 1e-05,
+      "loss": 1.0116,
+      "step": 168
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 1e-05,
+      "loss": 1.0647,
+      "step": 169
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 1e-05,
+      "loss": 1.091,
+      "step": 170
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 1e-05,
+      "loss": 1.0167,
+      "step": 171
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 1e-05,
+      "loss": 1.0952,
+      "step": 172
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 1e-05,
+      "loss": 1.0423,
+      "step": 173
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 1e-05,
+      "loss": 1.0198,
+      "step": 174
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 1e-05,
+      "loss": 1.0341,
+      "step": 175
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 1e-05,
+      "loss": 1.0657,
+      "step": 176
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 1e-05,
+      "loss": 1.0084,
+      "step": 177
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 1e-05,
+      "loss": 1.0483,
+      "step": 178
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 1e-05,
+      "loss": 1.0493,
+      "step": 179
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 1e-05,
+      "loss": 1.0328,
+      "step": 180
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 1e-05,
+      "loss": 1.0585,
+      "step": 181
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 1e-05,
+      "loss": 0.9869,
+      "step": 182
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 1e-05,
+      "loss": 1.0266,
+      "step": 183
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 1e-05,
+      "loss": 1.0462,
+      "step": 184
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 1e-05,
+      "loss": 1.0306,
+      "step": 185
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 1e-05,
+      "loss": 1.0459,
+      "step": 186
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 1e-05,
+      "loss": 1.0004,
+      "step": 187
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 1e-05,
+      "loss": 1.0417,
+      "step": 188
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 1e-05,
+      "loss": 1.0712,
+      "step": 189
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 1e-05,
+      "loss": 1.0228,
+      "step": 190
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 1e-05,
+      "loss": 0.9888,
+      "step": 191
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 1e-05,
+      "loss": 0.9393,
+      "step": 192
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 1e-05,
+      "loss": 1.052,
+      "step": 193
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 1e-05,
+      "loss": 1.0751,
+      "step": 194
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 1e-05,
+      "loss": 1.0468,
+      "step": 195
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 1e-05,
+      "loss": 1.0235,
+      "step": 196
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 1e-05,
+      "loss": 1.0428,
+      "step": 197
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 1e-05,
+      "loss": 1.0408,
+      "step": 198
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 1e-05,
+      "loss": 1.0112,
+      "step": 199
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 1e-05,
+      "loss": 1.0329,
+      "step": 200
+    },
+    {
+      "epoch": 0.19,
+      "eval_val_accuracy": 0.6278333333333334,
+      "eval_val_loss": 1.0557304620742798,
+      "eval_val_runtime": 906.1331,
+      "eval_val_samples_per_second": 1.104,
+      "eval_val_steps_per_second": 0.138,
+      "step": 200
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 1e-05,
+      "loss": 1.0219,
+      "step": 201
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 1e-05,
+      "loss": 1.0479,
+      "step": 202
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 1e-05,
+      "loss": 0.9808,
+      "step": 203
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 1e-05,
+      "loss": 1.0404,
+      "step": 204
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 1e-05,
+      "loss": 1.0214,
+      "step": 205
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 1e-05,
+      "loss": 0.9991,
+      "step": 206
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 1e-05,
+      "loss": 0.9435,
+      "step": 207
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 1e-05,
+      "loss": 1.0083,
+      "step": 208
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 1e-05,
+      "loss": 1.0187,
+      "step": 209
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 1e-05,
+      "loss": 1.0493,
+      "step": 210
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 1e-05,
+      "loss": 0.9925,
+      "step": 211
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 1e-05,
+      "loss": 1.0142,
+      "step": 212
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 1e-05,
+      "loss": 1.0131,
+      "step": 213
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 1e-05,
+      "loss": 1.0041,
+      "step": 214
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 1e-05,
+      "loss": 1.0187,
+      "step": 215
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 1e-05,
+      "loss": 1.0551,
+      "step": 216
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 1e-05,
+      "loss": 1.0224,
+      "step": 217
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 1e-05,
+      "loss": 1.0654,
+      "step": 218
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 1e-05,
+      "loss": 1.0472,
+      "step": 219
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 1e-05,
+      "loss": 0.9667,
+      "step": 220
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 1e-05,
+      "loss": 0.9944,
+      "step": 221
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 1e-05,
+      "loss": 1.0283,
+      "step": 222
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 1e-05,
+      "loss": 1.0047,
+      "step": 223
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 1e-05,
+      "loss": 0.9781,
+      "step": 224
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 1e-05,
+      "loss": 1.0127,
+      "step": 225
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 1e-05,
+      "loss": 1.035,
+      "step": 226
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 1e-05,
+      "loss": 0.9573,
+      "step": 227
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 1e-05,
+      "loss": 1.0365,
+      "step": 228
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 1e-05,
+      "loss": 1.0508,
+      "step": 229
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 1e-05,
+      "loss": 1.0138,
+      "step": 230
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 1e-05,
+      "loss": 1.0037,
+      "step": 231
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 1e-05,
+      "loss": 1.013,
+      "step": 232
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 1e-05,
+      "loss": 1.0025,
+      "step": 233
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 1e-05,
+      "loss": 1.0309,
+      "step": 234
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 1e-05,
+      "loss": 0.9718,
+      "step": 235
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 1e-05,
+      "loss": 0.9968,
+      "step": 236
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 1e-05,
+      "loss": 1.024,
+      "step": 237
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 1e-05,
+      "loss": 1.0028,
+      "step": 238
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 1e-05,
+      "loss": 1.0701,
+      "step": 239
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 1e-05,
+      "loss": 1.0428,
+      "step": 240
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 1e-05,
+      "loss": 0.9651,
+      "step": 241
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 1e-05,
+      "loss": 1.0095,
+      "step": 242
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 1e-05,
+      "loss": 1.0495,
+      "step": 243
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 1e-05,
+      "loss": 1.0441,
+      "step": 244
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 1e-05,
+      "loss": 1.0412,
+      "step": 245
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 1e-05,
+      "loss": 1.0335,
+      "step": 246
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 1e-05,
+      "loss": 1.0092,
+      "step": 247
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 1e-05,
+      "loss": 1.0408,
+      "step": 248
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 1e-05,
+      "loss": 0.993,
+      "step": 249
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 1e-05,
+      "loss": 0.9546,
+      "step": 250
+    },
+    {
+      "epoch": 0.25,
+      "eval_val_accuracy": 0.6341666666666667,
+      "eval_val_loss": 1.053076148033142,
+      "eval_val_runtime": 906.4123,
+      "eval_val_samples_per_second": 1.103,
+      "eval_val_steps_per_second": 0.138,
+      "step": 250
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 1e-05,
+      "loss": 1.043,
+      "step": 251
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 1e-05,
+      "loss": 1.0082,
+      "step": 252
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 1e-05,
+      "loss": 1.0281,
+      "step": 253
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 1e-05,
+      "loss": 0.9779,
+      "step": 254
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 1e-05,
+      "loss": 1.0074,
+      "step": 255
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 1e-05,
+      "loss": 0.9639,
+      "step": 256
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 1e-05,
+      "loss": 1.0121,
+      "step": 257
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 1e-05,
+      "loss": 1.036,
+      "step": 258
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 1e-05,
+      "loss": 1.0143,
+      "step": 259
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 1e-05,
+      "loss": 1.0407,
+      "step": 260
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 1e-05,
+      "loss": 1.0202,
+      "step": 261
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 1e-05,
+      "loss": 1.0099,
+      "step": 262
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 1e-05,
+      "loss": 1.0226,
+      "step": 263
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 1e-05,
+      "loss": 1.02,
+      "step": 264
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 1e-05,
+      "loss": 1.0197,
+      "step": 265
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 1e-05,
+      "loss": 1.0621,
+      "step": 266
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 1e-05,
+      "loss": 1.0261,
+      "step": 267
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 1e-05,
+      "loss": 1.0392,
+      "step": 268
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 1e-05,
+      "loss": 0.9644,
+      "step": 269
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 1e-05,
+      "loss": 0.9943,
+      "step": 270
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 1e-05,
+      "loss": 1.0354,
+      "step": 271
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 1e-05,
+      "loss": 0.9826,
+      "step": 272
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 1e-05,
+      "loss": 0.9934,
+      "step": 273
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 1e-05,
+      "loss": 1.0222,
+      "step": 274
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 1e-05,
+      "loss": 1.0752,
+      "step": 275
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 1e-05,
+      "loss": 1.0337,
+      "step": 276
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 1e-05,
+      "loss": 0.976,
+      "step": 277
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 1e-05,
+      "loss": 1.051,
+      "step": 278
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 1e-05,
+      "loss": 1.0338,
+      "step": 279
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 1e-05,
+      "loss": 1.0026,
+      "step": 280
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 1e-05,
+      "loss": 1.0118,
+      "step": 281
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 1e-05,
+      "loss": 0.9982,
+      "step": 282
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 1e-05,
+      "loss": 1.0136,
+      "step": 283
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 1e-05,
+      "loss": 0.9613,
+      "step": 284
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 1e-05,
+      "loss": 1.0122,
+      "step": 285
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 1e-05,
+      "loss": 1.049,
+      "step": 286
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 1e-05,
+      "loss": 1.0164,
+      "step": 287
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 1e-05,
+      "loss": 1.0052,
+      "step": 288
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 1e-05,
+      "loss": 1.0244,
+      "step": 289
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 1e-05,
+      "loss": 0.9886,
+      "step": 290
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 1e-05,
+      "loss": 1.0039,
+      "step": 291
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 1e-05,
+      "loss": 1.0088,
+      "step": 292
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 1e-05,
+      "loss": 1.0003,
+      "step": 293
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 1e-05,
+      "loss": 0.9984,
+      "step": 294
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 1e-05,
+      "loss": 1.0288,
+      "step": 295
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 1e-05,
+      "loss": 0.9928,
+      "step": 296
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 1e-05,
+      "loss": 0.992,
+      "step": 297
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 1e-05,
+      "loss": 0.9863,
+      "step": 298
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 1e-05,
+      "loss": 0.9959,
+      "step": 299
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 1e-05,
+      "loss": 1.0094,
+      "step": 300
+    },
+    {
+      "epoch": 0.3,
+      "eval_val_accuracy": 0.6303333333333334,
+      "eval_val_loss": 1.0518085956573486,
+      "eval_val_runtime": 905.7847,
+      "eval_val_samples_per_second": 1.104,
+      "eval_val_steps_per_second": 0.138,
+      "step": 300
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 1e-05,
+      "loss": 1.0049,
+      "step": 301
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 1e-05,
+      "loss": 1.005,
+      "step": 302
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 1e-05,
+      "loss": 1.0033,
+      "step": 303
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 1e-05,
+      "loss": 1.054,
+      "step": 304
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 1e-05,
+      "loss": 0.9742,
+      "step": 305
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 1e-05,
+      "loss": 1.0061,
+      "step": 306
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 1e-05,
+      "loss": 1.015,
+      "step": 307
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 1e-05,
+      "loss": 1.0414,
+      "step": 308
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 1e-05,
+      "loss": 0.9995,
+      "step": 309
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 1e-05,
+      "loss": 0.9901,
+      "step": 310
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 1e-05,
+      "loss": 0.9936,
+      "step": 311
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 1e-05,
+      "loss": 1.0476,
+      "step": 312
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 1e-05,
+      "loss": 0.9987,
+      "step": 313
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 1e-05,
+      "loss": 1.0564,
+      "step": 314
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 1e-05,
+      "loss": 1.0129,
+      "step": 315
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 1e-05,
+      "loss": 0.9667,
+      "step": 316
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 1e-05,
+      "loss": 1.0264,
+      "step": 317
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 1e-05,
+      "loss": 0.9835,
+      "step": 318
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 1e-05,
+      "loss": 1.0199,
+      "step": 319
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 1e-05,
+      "loss": 0.9652,
+      "step": 320
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 1e-05,
+      "loss": 0.9874,
+      "step": 321
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 1e-05,
+      "loss": 1.0079,
+      "step": 322
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 1e-05,
+      "loss": 1.0858,
+      "step": 323
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 1e-05,
+      "loss": 1.015,
+      "step": 324
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 1e-05,
+      "loss": 0.9782,
+      "step": 325
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 1e-05,
+      "loss": 1.0589,
+      "step": 326
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 1e-05,
+      "loss": 1.0152,
+      "step": 327
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 1e-05,
+      "loss": 0.9639,
+      "step": 328
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 1e-05,
+      "loss": 1.0089,
+      "step": 329
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 1e-05,
+      "loss": 1.056,
+      "step": 330
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 1e-05,
+      "loss": 0.985,
+      "step": 331
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 1e-05,
+      "loss": 1.027,
+      "step": 332
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 1e-05,
+      "loss": 1.0565,
+      "step": 333
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 1e-05,
+      "loss": 1.0337,
+      "step": 334
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 1e-05,
+      "loss": 1.0453,
+      "step": 335
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 1e-05,
+      "loss": 1.0108,
+      "step": 336
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 1e-05,
+      "loss": 1.0289,
+      "step": 337
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 1e-05,
+      "loss": 1.0096,
+      "step": 338
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 1e-05,
+      "loss": 1.0241,
+      "step": 339
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 1e-05,
+      "loss": 0.9743,
+      "step": 340
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 1e-05,
+      "loss": 1.0387,
+      "step": 341
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 1e-05,
+      "loss": 1.0023,
+      "step": 342
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 1e-05,
+      "loss": 1.0257,
+      "step": 343
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 1e-05,
+      "loss": 0.9827,
+      "step": 344
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 1e-05,
+      "loss": 1.0269,
+      "step": 345
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 1e-05,
+      "loss": 1.0017,
+      "step": 346
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 1e-05,
+      "loss": 0.994,
+      "step": 347
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 1e-05,
+      "loss": 1.0285,
+      "step": 348
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 1e-05,
+      "loss": 1.0634,
+      "step": 349
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 1e-05,
+      "loss": 1.0224,
+      "step": 350
+    },
+    {
+      "epoch": 0.35,
+      "eval_val_accuracy": 0.6265,
+      "eval_val_loss": 1.0556230545043945,
+      "eval_val_runtime": 905.4395,
+      "eval_val_samples_per_second": 1.104,
+      "eval_val_steps_per_second": 0.138,
+      "step": 350
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 1e-05,
+      "loss": 0.9755,
+      "step": 351
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 1e-05,
+      "loss": 1.0004,
+      "step": 352
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 1e-05,
+      "loss": 1.0171,
+      "step": 353
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 1e-05,
+      "loss": 1.0,
+      "step": 354
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 1e-05,
+      "loss": 0.9887,
+      "step": 355
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 1e-05,
+      "loss": 1.0387,
+      "step": 356
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 1e-05,
+      "loss": 0.9966,
+      "step": 357
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 1e-05,
+      "loss": 1.0213,
+      "step": 358
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 1e-05,
+      "loss": 1.0168,
+      "step": 359
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 1e-05,
+      "loss": 1.0332,
+      "step": 360
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 1e-05,
+      "loss": 1.039,
+      "step": 361
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 1e-05,
+      "loss": 1.0034,
+      "step": 362
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 1e-05,
+      "loss": 0.9631,
+      "step": 363
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 1e-05,
+      "loss": 1.0406,
+      "step": 364
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 1e-05,
+      "loss": 0.9821,
+      "step": 365
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 1e-05,
+      "loss": 1.0231,
+      "step": 366
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 1e-05,
+      "loss": 1.0272,
+      "step": 367
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 1e-05,
+      "loss": 1.0231,
+      "step": 368
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 1e-05,
+      "loss": 0.9833,
+      "step": 369
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 1e-05,
+      "loss": 1.0219,
+      "step": 370
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 1e-05,
+      "loss": 0.9647,
+      "step": 371
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 1e-05,
+      "loss": 1.0044,
+      "step": 372
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 1e-05,
+      "loss": 0.9968,
+      "step": 373
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 1e-05,
+      "loss": 0.9956,
+      "step": 374
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 1e-05,
+      "loss": 1.0278,
+      "step": 375
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 1e-05,
+      "loss": 0.9908,
+      "step": 376
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 1e-05,
+      "loss": 1.0058,
+      "step": 377
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 1e-05,
+      "loss": 1.0185,
+      "step": 378
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 1e-05,
+      "loss": 0.97,
+      "step": 379
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 1e-05,
+      "loss": 0.9491,
+      "step": 380
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 1e-05,
+      "loss": 1.0639,
+      "step": 381
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 1e-05,
+      "loss": 1.0,
+      "step": 382
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 1e-05,
+      "loss": 0.9812,
+      "step": 383
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 1e-05,
+      "loss": 1.0691,
+      "step": 384
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 1e-05,
+      "loss": 1.0074,
+      "step": 385
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 1e-05,
+      "loss": 0.9924,
+      "step": 386
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 1e-05,
+      "loss": 0.9976,
+      "step": 387
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 1e-05,
+      "loss": 1.0217,
+      "step": 388
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 1e-05,
+      "loss": 0.9745,
+      "step": 389
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 1e-05,
+      "loss": 1.0014,
+      "step": 390
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 1e-05,
+      "loss": 1.0192,
+      "step": 391
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 1e-05,
+      "loss": 0.9825,
+      "step": 392
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 1e-05,
+      "loss": 1.007,
+      "step": 393
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 1e-05,
+      "loss": 1.0288,
+      "step": 394
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 1e-05,
+      "loss": 1.0342,
+      "step": 395
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 1e-05,
+      "loss": 1.0183,
+      "step": 396
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 1e-05,
+      "loss": 0.9884,
+      "step": 397
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 1e-05,
+      "loss": 0.9674,
+      "step": 398
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 1e-05,
+      "loss": 0.9795,
+      "step": 399
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 1e-05,
+      "loss": 0.9885,
+      "step": 400
+    },
+    {
+      "epoch": 0.4,
+      "eval_val_accuracy": 0.627,
+      "eval_val_loss": 1.0441503524780273,
+      "eval_val_runtime": 905.3415,
+      "eval_val_samples_per_second": 1.105,
+      "eval_val_steps_per_second": 0.138,
+      "step": 400
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 1988,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 2,
+  "save_steps": 50,
+  "total_flos": 0.0,
+  "train_batch_size": 1,
+  "trial_name": null,
+  "trial_params": null
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d8021270c95d7f97a8c735e62d90edfe50e36e0eedac76c20ba637c77a1dcd38
+size 6267

zero_to_fp32.py

@@ -0,0 +1,587 @@
+#!/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 _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()
+            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)