Training in progress, step 8000

checkpoint-8000/config.json

@@ -0,0 +1,41 @@
+{
+  "_name_or_path": "emilios/whisper-medium-el-n2",
+  "activation_dropout": 0.0,
+  "activation_function": "gelu",
+  "architectures": [
+    "WhisperForConditionalGeneration"
+  ],
+  "attention_dropout": 0.1,
+  "begin_suppress_tokens": [
+    220,
+    50257
+  ],
+  "bos_token_id": 50257,
+  "d_model": 1024,
+  "decoder_attention_heads": 16,
+  "decoder_ffn_dim": 4096,
+  "decoder_layerdrop": 0.0,
+  "decoder_layers": 24,
+  "decoder_start_token_id": 50258,
+  "dropout": 0.1,
+  "encoder_attention_heads": 16,
+  "encoder_ffn_dim": 4096,
+  "encoder_layerdrop": 0.0,
+  "encoder_layers": 24,
+  "eos_token_id": 50257,
+  "forced_decoder_ids": null,
+  "init_std": 0.02,
+  "is_encoder_decoder": true,
+  "max_length": 448,
+  "max_source_positions": 1500,
+  "max_target_positions": 448,
+  "model_type": "whisper",
+  "num_hidden_layers": 24,
+  "num_mel_bins": 80,
+  "pad_token_id": 50257,
+  "scale_embedding": false,
+  "torch_dtype": "float16",
+  "transformers_version": "4.26.0.dev0",
+  "use_cache": false,
+  "vocab_size": 51865
+}

checkpoint-8000/global_step8000/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:33ee1bb88de1fcdb956b0991f37770c6cb0d0815e17f282fb3c027d24f6e83c9
+size 1527967899

checkpoint-8000/global_step8000/zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:24e3de5842b9966a86879441bf950784e7f57eab3b6e77b493b3ec6f5a5da7ee
+size 9166378846

checkpoint-8000/latest

@@ -0,0 +1 @@
+global_step8000

checkpoint-8000/pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:292ed5692b25c5ef882949ab1088ccf2a62e2935100cd3609d1745d6a0afdc10
+size 1527847357

checkpoint-8000/rng_state.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fbe3e21d07855346d756206e2eaf0dad10616d6ffdbc37c3cc6df32741adb354
+size 14575

checkpoint-8000/trainer_state.json

@@ -0,0 +1,2008 @@
+{
+  "best_metric": 9.899702823179792,
+  "best_model_checkpoint": "./checkpoint-8000",
+  "epoch": 469.94117647058823,
+  "global_step": 8000,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 1.47,
+      "learning_rate": 1.5136083400296205e-06,
+      "loss": 0.0024,
+      "step": 25
+    },
+    {
+      "epoch": 2.94,
+      "learning_rate": 1.8687587131475301e-06,
+      "loss": 0.0024,
+      "step": 50
+    },
+    {
+      "epoch": 4.41,
+      "learning_rate": 2.0711488350670174e-06,
+      "loss": 0.0023,
+      "step": 75
+    },
+    {
+      "epoch": 5.88,
+      "learning_rate": 2.213317753617305e-06,
+      "loss": 0.0023,
+      "step": 100
+    },
+    {
+      "epoch": 7.35,
+      "learning_rate": 2.3230029693718747e-06,
+      "loss": 0.002,
+      "step": 125
+    },
+    {
+      "epoch": 8.82,
+      "learning_rate": 2.412322158351148e-06,
+      "loss": 0.002,
+      "step": 150
+    },
+    {
+      "epoch": 10.29,
+      "learning_rate": 2.4876668872198717e-06,
+      "loss": 0.0022,
+      "step": 175
+    },
+    {
+      "epoch": 11.76,
+      "learning_rate": 2.552824062407326e-06,
+      "loss": 0.0021,
+      "step": 200
+    },
+    {
+      "epoch": 13.24,
+      "learning_rate": 2.610223373296667e-06,
+      "loss": 0.0034,
+      "step": 225
+    },
+    {
+      "epoch": 14.71,
+      "learning_rate": 2.661517182828361e-06,
+      "loss": 0.0019,
+      "step": 250
+    },
+    {
+      "epoch": 16.18,
+      "learning_rate": 2.7078803874740543e-06,
+      "loss": 0.0018,
+      "step": 275
+    },
+    {
+      "epoch": 17.65,
+      "learning_rate": 2.750178319990197e-06,
+      "loss": 0.0023,
+      "step": 300
+    },
+    {
+      "epoch": 19.12,
+      "learning_rate": 2.7890667754365044e-06,
+      "loss": 0.0019,
+      "step": 325
+    },
+    {
+      "epoch": 20.59,
+      "learning_rate": 2.8250546392106077e-06,
+      "loss": 0.0021,
+      "step": 350
+    },
+    {
+      "epoch": 22.06,
+      "learning_rate": 2.8585447348549113e-06,
+      "loss": 0.0023,
+      "step": 375
+    },
+    {
+      "epoch": 23.53,
+      "learning_rate": 2.889861392935294e-06,
+      "loss": 0.0021,
+      "step": 400
+    },
+    {
+      "epoch": 25.0,
+      "learning_rate": 2.9192696063561725e-06,
+      "loss": 0.0016,
+      "step": 425
+    },
+    {
+      "epoch": 26.47,
+      "learning_rate": 2.946988676871634e-06,
+      "loss": 0.0018,
+      "step": 450
+    },
+    {
+      "epoch": 27.94,
+      "learning_rate": 2.973202150939645e-06,
+      "loss": 0.0022,
+      "step": 475
+    },
+    {
+      "epoch": 29.41,
+      "learning_rate": 2.998065193492142e-06,
+      "loss": 0.0018,
+      "step": 500
+    },
+    {
+      "epoch": 30.88,
+      "learning_rate": 2.9811428571428573e-06,
+      "loss": 0.0018,
+      "step": 525
+    },
+    {
+      "epoch": 32.35,
+      "learning_rate": 2.959714285714286e-06,
+      "loss": 0.0021,
+      "step": 550
+    },
+    {
+      "epoch": 33.82,
+      "learning_rate": 2.9382857142857143e-06,
+      "loss": 0.0017,
+      "step": 575
+    },
+    {
+      "epoch": 35.29,
+      "learning_rate": 2.916857142857143e-06,
+      "loss": 0.0018,
+      "step": 600
+    },
+    {
+      "epoch": 36.76,
+      "learning_rate": 2.895428571428572e-06,
+      "loss": 0.0016,
+      "step": 625
+    },
+    {
+      "epoch": 38.24,
+      "learning_rate": 2.874e-06,
+      "loss": 0.0017,
+      "step": 650
+    },
+    {
+      "epoch": 39.71,
+      "learning_rate": 2.8525714285714285e-06,
+      "loss": 0.0016,
+      "step": 675
+    },
+    {
+      "epoch": 41.18,
+      "learning_rate": 2.8311428571428574e-06,
+      "loss": 0.0013,
+      "step": 700
+    },
+    {
+      "epoch": 42.65,
+      "learning_rate": 2.809714285714286e-06,
+      "loss": 0.0014,
+      "step": 725
+    },
+    {
+      "epoch": 44.12,
+      "learning_rate": 2.788285714285714e-06,
+      "loss": 0.0012,
+      "step": 750
+    },
+    {
+      "epoch": 45.59,
+      "learning_rate": 2.766857142857143e-06,
+      "loss": 0.0014,
+      "step": 775
+    },
+    {
+      "epoch": 47.06,
+      "learning_rate": 2.7454285714285716e-06,
+      "loss": 0.0014,
+      "step": 800
+    },
+    {
+      "epoch": 48.53,
+      "learning_rate": 2.724e-06,
+      "loss": 0.0013,
+      "step": 825
+    },
+    {
+      "epoch": 50.0,
+      "learning_rate": 2.7025714285714286e-06,
+      "loss": 0.0011,
+      "step": 850
+    },
+    {
+      "epoch": 51.47,
+      "learning_rate": 2.681142857142857e-06,
+      "loss": 0.0013,
+      "step": 875
+    },
+    {
+      "epoch": 52.94,
+      "learning_rate": 2.6597142857142857e-06,
+      "loss": 0.0014,
+      "step": 900
+    },
+    {
+      "epoch": 54.41,
+      "learning_rate": 2.6382857142857142e-06,
+      "loss": 0.0011,
+      "step": 925
+    },
+    {
+      "epoch": 55.88,
+      "learning_rate": 2.616857142857143e-06,
+      "loss": 0.0021,
+      "step": 950
+    },
+    {
+      "epoch": 57.35,
+      "learning_rate": 2.5954285714285713e-06,
+      "loss": 0.0014,
+      "step": 975
+    },
+    {
+      "epoch": 58.82,
+      "learning_rate": 2.574e-06,
+      "loss": 0.0014,
+      "step": 1000
+    },
+    {
+      "epoch": 58.82,
+      "eval_loss": 0.4951171875,
+      "eval_runtime": 168.7172,
+      "eval_samples_per_second": 1.612,
+      "eval_steps_per_second": 0.101,
+      "eval_wer": 10.36404160475483,
+      "step": 1000
+    },
+    {
+      "epoch": 60.29,
+      "learning_rate": 2.5542857142857143e-06,
+      "loss": 0.0013,
+      "step": 1025
+    },
+    {
+      "epoch": 61.76,
+      "learning_rate": 2.532857142857143e-06,
+      "loss": 0.0009,
+      "step": 1050
+    },
+    {
+      "epoch": 63.24,
+      "learning_rate": 2.5114285714285718e-06,
+      "loss": 0.0012,
+      "step": 1075
+    },
+    {
+      "epoch": 64.71,
+      "learning_rate": 2.49e-06,
+      "loss": 0.0009,
+      "step": 1100
+    },
+    {
+      "epoch": 66.18,
+      "learning_rate": 2.4685714285714284e-06,
+      "loss": 0.0014,
+      "step": 1125
+    },
+    {
+      "epoch": 67.65,
+      "learning_rate": 2.4471428571428574e-06,
+      "loss": 0.0012,
+      "step": 1150
+    },
+    {
+      "epoch": 69.12,
+      "learning_rate": 2.425714285714286e-06,
+      "loss": 0.0009,
+      "step": 1175
+    },
+    {
+      "epoch": 70.59,
+      "learning_rate": 2.404285714285714e-06,
+      "loss": 0.0008,
+      "step": 1200
+    },
+    {
+      "epoch": 72.06,
+      "learning_rate": 2.382857142857143e-06,
+      "loss": 0.0011,
+      "step": 1225
+    },
+    {
+      "epoch": 73.53,
+      "learning_rate": 2.3614285714285715e-06,
+      "loss": 0.0015,
+      "step": 1250
+    },
+    {
+      "epoch": 75.0,
+      "learning_rate": 2.34e-06,
+      "loss": 0.0008,
+      "step": 1275
+    },
+    {
+      "epoch": 76.47,
+      "learning_rate": 2.3185714285714286e-06,
+      "loss": 0.001,
+      "step": 1300
+    },
+    {
+      "epoch": 77.94,
+      "learning_rate": 2.297142857142857e-06,
+      "loss": 0.0009,
+      "step": 1325
+    },
+    {
+      "epoch": 79.41,
+      "learning_rate": 2.2757142857142856e-06,
+      "loss": 0.0009,
+      "step": 1350
+    },
+    {
+      "epoch": 80.88,
+      "learning_rate": 2.2542857142857146e-06,
+      "loss": 0.0011,
+      "step": 1375
+    },
+    {
+      "epoch": 82.35,
+      "learning_rate": 2.232857142857143e-06,
+      "loss": 0.0007,
+      "step": 1400
+    },
+    {
+      "epoch": 83.82,
+      "learning_rate": 2.2114285714285712e-06,
+      "loss": 0.001,
+      "step": 1425
+    },
+    {
+      "epoch": 85.29,
+      "learning_rate": 2.19e-06,
+      "loss": 0.001,
+      "step": 1450
+    },
+    {
+      "epoch": 86.76,
+      "learning_rate": 2.1685714285714287e-06,
+      "loss": 0.0008,
+      "step": 1475
+    },
+    {
+      "epoch": 88.24,
+      "learning_rate": 2.1471428571428573e-06,
+      "loss": 0.001,
+      "step": 1500
+    },
+    {
+      "epoch": 89.71,
+      "learning_rate": 2.125714285714286e-06,
+      "loss": 0.0008,
+      "step": 1525
+    },
+    {
+      "epoch": 91.18,
+      "learning_rate": 2.1042857142857143e-06,
+      "loss": 0.0008,
+      "step": 1550
+    },
+    {
+      "epoch": 92.65,
+      "learning_rate": 2.082857142857143e-06,
+      "loss": 0.0006,
+      "step": 1575
+    },
+    {
+      "epoch": 94.12,
+      "learning_rate": 2.0614285714285714e-06,
+      "loss": 0.0008,
+      "step": 1600
+    },
+    {
+      "epoch": 95.59,
+      "learning_rate": 2.0400000000000004e-06,
+      "loss": 0.0009,
+      "step": 1625
+    },
+    {
+      "epoch": 97.06,
+      "learning_rate": 2.0185714285714285e-06,
+      "loss": 0.0009,
+      "step": 1650
+    },
+    {
+      "epoch": 98.53,
+      "learning_rate": 1.997142857142857e-06,
+      "loss": 0.0008,
+      "step": 1675
+    },
+    {
+      "epoch": 100.0,
+      "learning_rate": 1.975714285714286e-06,
+      "loss": 0.0008,
+      "step": 1700
+    },
+    {
+      "epoch": 101.47,
+      "learning_rate": 1.9542857142857145e-06,
+      "loss": 0.0007,
+      "step": 1725
+    },
+    {
+      "epoch": 102.94,
+      "learning_rate": 1.9328571428571426e-06,
+      "loss": 0.0005,
+      "step": 1750
+    },
+    {
+      "epoch": 104.41,
+      "learning_rate": 1.9114285714285716e-06,
+      "loss": 0.0008,
+      "step": 1775
+    },
+    {
+      "epoch": 105.88,
+      "learning_rate": 1.8900000000000001e-06,
+      "loss": 0.0006,
+      "step": 1800
+    },
+    {
+      "epoch": 107.35,
+      "learning_rate": 1.8685714285714287e-06,
+      "loss": 0.0007,
+      "step": 1825
+    },
+    {
+      "epoch": 108.82,
+      "learning_rate": 1.8471428571428574e-06,
+      "loss": 0.0008,
+      "step": 1850
+    },
+    {
+      "epoch": 110.29,
+      "learning_rate": 1.8257142857142857e-06,
+      "loss": 0.0008,
+      "step": 1875
+    },
+    {
+      "epoch": 111.76,
+      "learning_rate": 1.8042857142857143e-06,
+      "loss": 0.0005,
+      "step": 1900
+    },
+    {
+      "epoch": 113.24,
+      "learning_rate": 1.782857142857143e-06,
+      "loss": 0.0007,
+      "step": 1925
+    },
+    {
+      "epoch": 114.71,
+      "learning_rate": 1.7614285714285715e-06,
+      "loss": 0.0006,
+      "step": 1950
+    },
+    {
+      "epoch": 116.18,
+      "learning_rate": 1.7399999999999999e-06,
+      "loss": 0.0005,
+      "step": 1975
+    },
+    {
+      "epoch": 117.65,
+      "learning_rate": 1.7185714285714286e-06,
+      "loss": 0.0006,
+      "step": 2000
+    },
+    {
+      "epoch": 117.65,
+      "eval_loss": 0.51806640625,
+      "eval_runtime": 168.3992,
+      "eval_samples_per_second": 1.615,
+      "eval_steps_per_second": 0.101,
+      "eval_wer": 10.280460624071322,
+      "step": 2000
+    },
+    {
+      "epoch": 118.47,
+      "learning_rate": 2.4642352941176473e-06,
+      "loss": 0.0007,
+      "step": 2025
+    },
+    {
+      "epoch": 119.94,
+      "learning_rate": 2.4554117647058827e-06,
+      "loss": 0.0006,
+      "step": 2050
+    },
+    {
+      "epoch": 121.41,
+      "learning_rate": 2.4465882352941176e-06,
+      "loss": 0.0003,
+      "step": 2075
+    },
+    {
+      "epoch": 122.88,
+      "learning_rate": 2.437764705882353e-06,
+      "loss": 0.0007,
+      "step": 2100
+    },
+    {
+      "epoch": 124.35,
+      "learning_rate": 2.4289411764705882e-06,
+      "loss": 0.0005,
+      "step": 2125
+    },
+    {
+      "epoch": 125.82,
+      "learning_rate": 2.4201176470588236e-06,
+      "loss": 0.0006,
+      "step": 2150
+    },
+    {
+      "epoch": 127.29,
+      "learning_rate": 2.411294117647059e-06,
+      "loss": 0.0007,
+      "step": 2175
+    },
+    {
+      "epoch": 128.76,
+      "learning_rate": 2.4024705882352942e-06,
+      "loss": 0.0006,
+      "step": 2200
+    },
+    {
+      "epoch": 130.24,
+      "learning_rate": 2.3936470588235295e-06,
+      "loss": 0.0011,
+      "step": 2225
+    },
+    {
+      "epoch": 131.71,
+      "learning_rate": 2.384823529411765e-06,
+      "loss": 0.0007,
+      "step": 2250
+    },
+    {
+      "epoch": 133.18,
+      "learning_rate": 2.376e-06,
+      "loss": 0.0006,
+      "step": 2275
+    },
+    {
+      "epoch": 134.65,
+      "learning_rate": 2.3671764705882355e-06,
+      "loss": 0.0005,
+      "step": 2300
+    },
+    {
+      "epoch": 136.12,
+      "learning_rate": 2.3583529411764704e-06,
+      "loss": 0.0007,
+      "step": 2325
+    },
+    {
+      "epoch": 137.59,
+      "learning_rate": 2.3495294117647058e-06,
+      "loss": 0.0008,
+      "step": 2350
+    },
+    {
+      "epoch": 139.06,
+      "learning_rate": 2.340705882352941e-06,
+      "loss": 0.0036,
+      "step": 2375
+    },
+    {
+      "epoch": 140.53,
+      "learning_rate": 2.3318823529411764e-06,
+      "loss": 0.0006,
+      "step": 2400
+    },
+    {
+      "epoch": 142.0,
+      "learning_rate": 2.3230588235294118e-06,
+      "loss": 0.0006,
+      "step": 2425
+    },
+    {
+      "epoch": 143.47,
+      "learning_rate": 2.314235294117647e-06,
+      "loss": 0.0004,
+      "step": 2450
+    },
+    {
+      "epoch": 144.94,
+      "learning_rate": 2.3054117647058824e-06,
+      "loss": 0.0018,
+      "step": 2475
+    },
+    {
+      "epoch": 146.41,
+      "learning_rate": 2.2965882352941177e-06,
+      "loss": 0.0007,
+      "step": 2500
+    },
+    {
+      "epoch": 147.88,
+      "learning_rate": 2.287764705882353e-06,
+      "loss": 0.0005,
+      "step": 2525
+    },
+    {
+      "epoch": 149.35,
+      "learning_rate": 2.2789411764705884e-06,
+      "loss": 0.0006,
+      "step": 2550
+    },
+    {
+      "epoch": 150.82,
+      "learning_rate": 2.2701176470588237e-06,
+      "loss": 0.0007,
+      "step": 2575
+    },
+    {
+      "epoch": 152.29,
+      "learning_rate": 2.2612941176470586e-06,
+      "loss": 0.0007,
+      "step": 2600
+    },
+    {
+      "epoch": 153.76,
+      "learning_rate": 2.252470588235294e-06,
+      "loss": 0.0008,
+      "step": 2625
+    },
+    {
+      "epoch": 155.24,
+      "learning_rate": 2.2436470588235293e-06,
+      "loss": 0.0005,
+      "step": 2650
+    },
+    {
+      "epoch": 156.71,
+      "learning_rate": 2.2348235294117646e-06,
+      "loss": 0.0005,
+      "step": 2675
+    },
+    {
+      "epoch": 158.18,
+      "learning_rate": 2.226e-06,
+      "loss": 0.0004,
+      "step": 2700
+    },
+    {
+      "epoch": 159.65,
+      "learning_rate": 2.2171764705882353e-06,
+      "loss": 0.0008,
+      "step": 2725
+    },
+    {
+      "epoch": 161.12,
+      "learning_rate": 2.2083529411764706e-06,
+      "loss": 0.0005,
+      "step": 2750
+    },
+    {
+      "epoch": 162.59,
+      "learning_rate": 2.199529411764706e-06,
+      "loss": 0.0005,
+      "step": 2775
+    },
+    {
+      "epoch": 164.06,
+      "learning_rate": 2.1907058823529413e-06,
+      "loss": 0.0005,
+      "step": 2800
+    },
+    {
+      "epoch": 165.53,
+      "learning_rate": 2.1818823529411766e-06,
+      "loss": 0.0006,
+      "step": 2825
+    },
+    {
+      "epoch": 167.0,
+      "learning_rate": 2.1730588235294115e-06,
+      "loss": 0.0006,
+      "step": 2850
+    },
+    {
+      "epoch": 168.47,
+      "learning_rate": 2.164235294117647e-06,
+      "loss": 0.0006,
+      "step": 2875
+    },
+    {
+      "epoch": 169.94,
+      "learning_rate": 2.155411764705882e-06,
+      "loss": 0.0008,
+      "step": 2900
+    },
+    {
+      "epoch": 171.41,
+      "learning_rate": 2.1465882352941175e-06,
+      "loss": 0.0005,
+      "step": 2925
+    },
+    {
+      "epoch": 172.88,
+      "learning_rate": 2.137764705882353e-06,
+      "loss": 0.0005,
+      "step": 2950
+    },
+    {
+      "epoch": 174.35,
+      "learning_rate": 2.128941176470588e-06,
+      "loss": 0.0004,
+      "step": 2975
+    },
+    {
+      "epoch": 175.82,
+      "learning_rate": 2.1201176470588235e-06,
+      "loss": 0.0007,
+      "step": 3000
+    },
+    {
+      "epoch": 175.82,
+      "eval_loss": 0.53173828125,
+      "eval_runtime": 153.6387,
+      "eval_samples_per_second": 1.77,
+      "eval_steps_per_second": 0.111,
+      "eval_wer": 10.11329866270431,
+      "step": 3000
+    },
+    {
+      "epoch": 177.29,
+      "learning_rate": 2.112e-06,
+      "loss": 0.0005,
+      "step": 3025
+    },
+    {
+      "epoch": 178.76,
+      "learning_rate": 2.103176470588235e-06,
+      "loss": 0.0005,
+      "step": 3050
+    },
+    {
+      "epoch": 180.24,
+      "learning_rate": 2.0943529411764705e-06,
+      "loss": 0.0007,
+      "step": 3075
+    },
+    {
+      "epoch": 181.71,
+      "learning_rate": 2.085529411764706e-06,
+      "loss": 0.0004,
+      "step": 3100
+    },
+    {
+      "epoch": 183.18,
+      "learning_rate": 2.076705882352941e-06,
+      "loss": 0.0004,
+      "step": 3125
+    },
+    {
+      "epoch": 184.65,
+      "learning_rate": 2.0678823529411765e-06,
+      "loss": 0.0005,
+      "step": 3150
+    },
+    {
+      "epoch": 186.12,
+      "learning_rate": 2.059058823529412e-06,
+      "loss": 0.0005,
+      "step": 3175
+    },
+    {
+      "epoch": 187.59,
+      "learning_rate": 2.050235294117647e-06,
+      "loss": 0.0004,
+      "step": 3200
+    },
+    {
+      "epoch": 189.06,
+      "learning_rate": 2.0414117647058825e-06,
+      "loss": 0.0005,
+      "step": 3225
+    },
+    {
+      "epoch": 190.53,
+      "learning_rate": 2.032588235294118e-06,
+      "loss": 0.0003,
+      "step": 3250
+    },
+    {
+      "epoch": 192.0,
+      "learning_rate": 2.0237647058823527e-06,
+      "loss": 0.0006,
+      "step": 3275
+    },
+    {
+      "epoch": 193.47,
+      "learning_rate": 2.014941176470588e-06,
+      "loss": 0.0005,
+      "step": 3300
+    },
+    {
+      "epoch": 194.94,
+      "learning_rate": 2.0061176470588234e-06,
+      "loss": 0.0004,
+      "step": 3325
+    },
+    {
+      "epoch": 196.41,
+      "learning_rate": 1.9972941176470587e-06,
+      "loss": 0.0004,
+      "step": 3350
+    },
+    {
+      "epoch": 197.88,
+      "learning_rate": 1.988470588235294e-06,
+      "loss": 0.0004,
+      "step": 3375
+    },
+    {
+      "epoch": 199.35,
+      "learning_rate": 1.9796470588235294e-06,
+      "loss": 0.0006,
+      "step": 3400
+    },
+    {
+      "epoch": 200.82,
+      "learning_rate": 1.9708235294117647e-06,
+      "loss": 0.0005,
+      "step": 3425
+    },
+    {
+      "epoch": 202.29,
+      "learning_rate": 1.962e-06,
+      "loss": 0.0004,
+      "step": 3450
+    },
+    {
+      "epoch": 203.76,
+      "learning_rate": 1.9531764705882353e-06,
+      "loss": 0.0005,
+      "step": 3475
+    },
+    {
+      "epoch": 205.24,
+      "learning_rate": 1.9443529411764707e-06,
+      "loss": 0.0006,
+      "step": 3500
+    },
+    {
+      "epoch": 206.71,
+      "learning_rate": 1.9355294117647056e-06,
+      "loss": 0.0003,
+      "step": 3525
+    },
+    {
+      "epoch": 208.18,
+      "learning_rate": 1.926705882352941e-06,
+      "loss": 0.0004,
+      "step": 3550
+    },
+    {
+      "epoch": 209.65,
+      "learning_rate": 1.9178823529411762e-06,
+      "loss": 0.0005,
+      "step": 3575
+    },
+    {
+      "epoch": 211.12,
+      "learning_rate": 1.9090588235294116e-06,
+      "loss": 0.0004,
+      "step": 3600
+    },
+    {
+      "epoch": 212.59,
+      "learning_rate": 1.9002352941176471e-06,
+      "loss": 0.0004,
+      "step": 3625
+    },
+    {
+      "epoch": 214.06,
+      "learning_rate": 1.8914117647058824e-06,
+      "loss": 0.0003,
+      "step": 3650
+    },
+    {
+      "epoch": 215.53,
+      "learning_rate": 1.8825882352941178e-06,
+      "loss": 0.0004,
+      "step": 3675
+    },
+    {
+      "epoch": 217.0,
+      "learning_rate": 1.873764705882353e-06,
+      "loss": 0.0005,
+      "step": 3700
+    },
+    {
+      "epoch": 218.47,
+      "learning_rate": 1.8649411764705884e-06,
+      "loss": 0.0004,
+      "step": 3725
+    },
+    {
+      "epoch": 219.94,
+      "learning_rate": 1.8561176470588238e-06,
+      "loss": 0.0005,
+      "step": 3750
+    },
+    {
+      "epoch": 221.41,
+      "learning_rate": 1.847294117647059e-06,
+      "loss": 0.0003,
+      "step": 3775
+    },
+    {
+      "epoch": 222.88,
+      "learning_rate": 1.838470588235294e-06,
+      "loss": 0.0004,
+      "step": 3800
+    },
+    {
+      "epoch": 224.35,
+      "learning_rate": 1.8296470588235293e-06,
+      "loss": 0.0003,
+      "step": 3825
+    },
+    {
+      "epoch": 225.82,
+      "learning_rate": 1.8208235294117646e-06,
+      "loss": 0.0003,
+      "step": 3850
+    },
+    {
+      "epoch": 227.29,
+      "learning_rate": 1.812e-06,
+      "loss": 0.0003,
+      "step": 3875
+    },
+    {
+      "epoch": 228.76,
+      "learning_rate": 1.8031764705882353e-06,
+      "loss": 0.0005,
+      "step": 3900
+    },
+    {
+      "epoch": 230.24,
+      "learning_rate": 1.7943529411764706e-06,
+      "loss": 0.0004,
+      "step": 3925
+    },
+    {
+      "epoch": 231.71,
+      "learning_rate": 1.785529411764706e-06,
+      "loss": 0.0004,
+      "step": 3950
+    },
+    {
+      "epoch": 233.18,
+      "learning_rate": 1.7767058823529413e-06,
+      "loss": 0.0003,
+      "step": 3975
+    },
+    {
+      "epoch": 234.65,
+      "learning_rate": 1.7678823529411766e-06,
+      "loss": 0.0004,
+      "step": 4000
+    },
+    {
+      "epoch": 234.65,
+      "eval_loss": 0.53955078125,
+      "eval_runtime": 161.7321,
+      "eval_samples_per_second": 1.682,
+      "eval_steps_per_second": 0.105,
+      "eval_wer": 10.12258543833581,
+      "step": 4000
+    },
+    {
+      "epoch": 236.12,
+      "learning_rate": 1.759764705882353e-06,
+      "loss": 0.0003,
+      "step": 4025
+    },
+    {
+      "epoch": 237.59,
+      "learning_rate": 1.7509411764705883e-06,
+      "loss": 0.0003,
+      "step": 4050
+    },
+    {
+      "epoch": 239.06,
+      "learning_rate": 1.7421176470588237e-06,
+      "loss": 0.0002,
+      "step": 4075
+    },
+    {
+      "epoch": 240.53,
+      "learning_rate": 1.733294117647059e-06,
+      "loss": 0.0004,
+      "step": 4100
+    },
+    {
+      "epoch": 242.0,
+      "learning_rate": 1.7244705882352943e-06,
+      "loss": 0.0005,
+      "step": 4125
+    },
+    {
+      "epoch": 243.47,
+      "learning_rate": 1.7156470588235296e-06,
+      "loss": 0.0003,
+      "step": 4150
+    },
+    {
+      "epoch": 244.94,
+      "learning_rate": 1.706823529411765e-06,
+      "loss": 0.0003,
+      "step": 4175
+    },
+    {
+      "epoch": 246.41,
+      "learning_rate": 1.6979999999999999e-06,
+      "loss": 0.0004,
+      "step": 4200
+    },
+    {
+      "epoch": 247.88,
+      "learning_rate": 1.6891764705882352e-06,
+      "loss": 0.0003,
+      "step": 4225
+    },
+    {
+      "epoch": 249.35,
+      "learning_rate": 1.6803529411764705e-06,
+      "loss": 0.0003,
+      "step": 4250
+    },
+    {
+      "epoch": 250.82,
+      "learning_rate": 1.6715294117647059e-06,
+      "loss": 0.0003,
+      "step": 4275
+    },
+    {
+      "epoch": 252.29,
+      "learning_rate": 1.6627058823529412e-06,
+      "loss": 0.0004,
+      "step": 4300
+    },
+    {
+      "epoch": 253.76,
+      "learning_rate": 1.6538823529411765e-06,
+      "loss": 0.0004,
+      "step": 4325
+    },
+    {
+      "epoch": 255.24,
+      "learning_rate": 1.6450588235294119e-06,
+      "loss": 0.0003,
+      "step": 4350
+    },
+    {
+      "epoch": 256.71,
+      "learning_rate": 1.6362352941176472e-06,
+      "loss": 0.0003,
+      "step": 4375
+    },
+    {
+      "epoch": 258.18,
+      "learning_rate": 1.6274117647058825e-06,
+      "loss": 0.0003,
+      "step": 4400
+    },
+    {
+      "epoch": 259.65,
+      "learning_rate": 1.6185882352941178e-06,
+      "loss": 0.0007,
+      "step": 4425
+    },
+    {
+      "epoch": 261.12,
+      "learning_rate": 1.6097647058823532e-06,
+      "loss": 0.0003,
+      "step": 4450
+    },
+    {
+      "epoch": 262.59,
+      "learning_rate": 1.600941176470588e-06,
+      "loss": 0.0004,
+      "step": 4475
+    },
+    {
+      "epoch": 264.06,
+      "learning_rate": 1.5921176470588234e-06,
+      "loss": 0.0003,
+      "step": 4500
+    },
+    {
+      "epoch": 265.53,
+      "learning_rate": 1.5832941176470587e-06,
+      "loss": 0.0004,
+      "step": 4525
+    },
+    {
+      "epoch": 267.0,
+      "learning_rate": 1.574470588235294e-06,
+      "loss": 0.0002,
+      "step": 4550
+    },
+    {
+      "epoch": 268.47,
+      "learning_rate": 1.5656470588235294e-06,
+      "loss": 0.0004,
+      "step": 4575
+    },
+    {
+      "epoch": 269.94,
+      "learning_rate": 1.5568235294117647e-06,
+      "loss": 0.0003,
+      "step": 4600
+    },
+    {
+      "epoch": 271.41,
+      "learning_rate": 1.548e-06,
+      "loss": 0.0002,
+      "step": 4625
+    },
+    {
+      "epoch": 272.88,
+      "learning_rate": 1.5391764705882354e-06,
+      "loss": 0.0002,
+      "step": 4650
+    },
+    {
+      "epoch": 274.35,
+      "learning_rate": 1.5303529411764707e-06,
+      "loss": 0.0003,
+      "step": 4675
+    },
+    {
+      "epoch": 275.82,
+      "learning_rate": 1.521529411764706e-06,
+      "loss": 0.0003,
+      "step": 4700
+    },
+    {
+      "epoch": 277.29,
+      "learning_rate": 1.5127058823529411e-06,
+      "loss": 0.0003,
+      "step": 4725
+    },
+    {
+      "epoch": 278.76,
+      "learning_rate": 1.5038823529411765e-06,
+      "loss": 0.0002,
+      "step": 4750
+    },
+    {
+      "epoch": 280.24,
+      "learning_rate": 1.4950588235294118e-06,
+      "loss": 0.0002,
+      "step": 4775
+    },
+    {
+      "epoch": 281.71,
+      "learning_rate": 1.486235294117647e-06,
+      "loss": 0.0003,
+      "step": 4800
+    },
+    {
+      "epoch": 283.18,
+      "learning_rate": 1.4774117647058823e-06,
+      "loss": 0.0003,
+      "step": 4825
+    },
+    {
+      "epoch": 284.65,
+      "learning_rate": 1.4685882352941176e-06,
+      "loss": 0.0004,
+      "step": 4850
+    },
+    {
+      "epoch": 286.12,
+      "learning_rate": 1.459764705882353e-06,
+      "loss": 0.0002,
+      "step": 4875
+    },
+    {
+      "epoch": 287.59,
+      "learning_rate": 1.4509411764705882e-06,
+      "loss": 0.0002,
+      "step": 4900
+    },
+    {
+      "epoch": 289.06,
+      "learning_rate": 1.4421176470588236e-06,
+      "loss": 0.0012,
+      "step": 4925
+    },
+    {
+      "epoch": 290.53,
+      "learning_rate": 1.433294117647059e-06,
+      "loss": 0.0003,
+      "step": 4950
+    },
+    {
+      "epoch": 292.0,
+      "learning_rate": 1.4244705882352942e-06,
+      "loss": 0.0002,
+      "step": 4975
+    },
+    {
+      "epoch": 293.47,
+      "learning_rate": 1.4156470588235296e-06,
+      "loss": 0.0004,
+      "step": 5000
+    },
+    {
+      "epoch": 293.47,
+      "eval_loss": 0.55322265625,
+      "eval_runtime": 164.7504,
+      "eval_samples_per_second": 1.651,
+      "eval_steps_per_second": 0.103,
+      "eval_wer": 10.104011887072808,
+      "step": 5000
+    },
+    {
+      "epoch": 294.94,
+      "learning_rate": 1.407529411764706e-06,
+      "loss": 0.0002,
+      "step": 5025
+    },
+    {
+      "epoch": 296.41,
+      "learning_rate": 1.3987058823529413e-06,
+      "loss": 0.0004,
+      "step": 5050
+    },
+    {
+      "epoch": 297.88,
+      "learning_rate": 1.3898823529411764e-06,
+      "loss": 0.0003,
+      "step": 5075
+    },
+    {
+      "epoch": 299.35,
+      "learning_rate": 1.3810588235294117e-06,
+      "loss": 0.0003,
+      "step": 5100
+    },
+    {
+      "epoch": 300.82,
+      "learning_rate": 1.372235294117647e-06,
+      "loss": 0.0004,
+      "step": 5125
+    },
+    {
+      "epoch": 302.29,
+      "learning_rate": 1.3634117647058824e-06,
+      "loss": 0.0004,
+      "step": 5150
+    },
+    {
+      "epoch": 303.76,
+      "learning_rate": 1.3545882352941177e-06,
+      "loss": 0.0002,
+      "step": 5175
+    },
+    {
+      "epoch": 305.24,
+      "learning_rate": 1.3457647058823528e-06,
+      "loss": 0.0003,
+      "step": 5200
+    },
+    {
+      "epoch": 306.71,
+      "learning_rate": 1.3369411764705881e-06,
+      "loss": 0.0005,
+      "step": 5225
+    },
+    {
+      "epoch": 308.18,
+      "learning_rate": 1.3281176470588235e-06,
+      "loss": 0.0002,
+      "step": 5250
+    },
+    {
+      "epoch": 309.65,
+      "learning_rate": 1.3192941176470588e-06,
+      "loss": 0.0002,
+      "step": 5275
+    },
+    {
+      "epoch": 311.12,
+      "learning_rate": 1.3104705882352941e-06,
+      "loss": 0.0002,
+      "step": 5300
+    },
+    {
+      "epoch": 312.59,
+      "learning_rate": 1.3016470588235295e-06,
+      "loss": 0.0003,
+      "step": 5325
+    },
+    {
+      "epoch": 314.06,
+      "learning_rate": 1.2928235294117648e-06,
+      "loss": 0.0003,
+      "step": 5350
+    },
+    {
+      "epoch": 315.53,
+      "learning_rate": 1.284e-06,
+      "loss": 0.0003,
+      "step": 5375
+    },
+    {
+      "epoch": 317.0,
+      "learning_rate": 1.2751764705882352e-06,
+      "loss": 0.0003,
+      "step": 5400
+    },
+    {
+      "epoch": 318.47,
+      "learning_rate": 1.2663529411764706e-06,
+      "loss": 0.0003,
+      "step": 5425
+    },
+    {
+      "epoch": 319.94,
+      "learning_rate": 1.2575294117647059e-06,
+      "loss": 0.0002,
+      "step": 5450
+    },
+    {
+      "epoch": 321.41,
+      "learning_rate": 1.2487058823529412e-06,
+      "loss": 0.0002,
+      "step": 5475
+    },
+    {
+      "epoch": 322.88,
+      "learning_rate": 1.2398823529411765e-06,
+      "loss": 0.0002,
+      "step": 5500
+    },
+    {
+      "epoch": 324.35,
+      "learning_rate": 1.2310588235294119e-06,
+      "loss": 0.0001,
+      "step": 5525
+    },
+    {
+      "epoch": 325.82,
+      "learning_rate": 1.2222352941176472e-06,
+      "loss": 0.0002,
+      "step": 5550
+    },
+    {
+      "epoch": 327.29,
+      "learning_rate": 1.2134117647058825e-06,
+      "loss": 0.0002,
+      "step": 5575
+    },
+    {
+      "epoch": 328.76,
+      "learning_rate": 1.2045882352941177e-06,
+      "loss": 0.0003,
+      "step": 5600
+    },
+    {
+      "epoch": 330.24,
+      "learning_rate": 1.195764705882353e-06,
+      "loss": 0.0003,
+      "step": 5625
+    },
+    {
+      "epoch": 331.71,
+      "learning_rate": 1.1869411764705883e-06,
+      "loss": 0.0003,
+      "step": 5650
+    },
+    {
+      "epoch": 333.18,
+      "learning_rate": 1.1781176470588236e-06,
+      "loss": 0.0003,
+      "step": 5675
+    },
+    {
+      "epoch": 334.65,
+      "learning_rate": 1.169294117647059e-06,
+      "loss": 0.0004,
+      "step": 5700
+    },
+    {
+      "epoch": 336.12,
+      "learning_rate": 1.1604705882352943e-06,
+      "loss": 0.0002,
+      "step": 5725
+    },
+    {
+      "epoch": 337.59,
+      "learning_rate": 1.1516470588235294e-06,
+      "loss": 0.0003,
+      "step": 5750
+    },
+    {
+      "epoch": 339.06,
+      "learning_rate": 1.1428235294117647e-06,
+      "loss": 0.0002,
+      "step": 5775
+    },
+    {
+      "epoch": 340.53,
+      "learning_rate": 1.134e-06,
+      "loss": 0.0002,
+      "step": 5800
+    },
+    {
+      "epoch": 342.0,
+      "learning_rate": 1.1251764705882354e-06,
+      "loss": 0.0003,
+      "step": 5825
+    },
+    {
+      "epoch": 343.47,
+      "learning_rate": 1.1163529411764707e-06,
+      "loss": 0.0002,
+      "step": 5850
+    },
+    {
+      "epoch": 344.94,
+      "learning_rate": 1.1075294117647058e-06,
+      "loss": 0.0002,
+      "step": 5875
+    },
+    {
+      "epoch": 346.41,
+      "learning_rate": 1.0987058823529412e-06,
+      "loss": 0.0002,
+      "step": 5900
+    },
+    {
+      "epoch": 347.88,
+      "learning_rate": 1.0898823529411765e-06,
+      "loss": 0.0002,
+      "step": 5925
+    },
+    {
+      "epoch": 349.35,
+      "learning_rate": 1.0810588235294118e-06,
+      "loss": 0.0003,
+      "step": 5950
+    },
+    {
+      "epoch": 350.82,
+      "learning_rate": 1.0722352941176472e-06,
+      "loss": 0.0002,
+      "step": 5975
+    },
+    {
+      "epoch": 352.29,
+      "learning_rate": 1.0634117647058823e-06,
+      "loss": 0.0013,
+      "step": 6000
+    },
+    {
+      "epoch": 352.29,
+      "eval_loss": 0.564453125,
+      "eval_runtime": 167.4651,
+      "eval_samples_per_second": 1.624,
+      "eval_steps_per_second": 0.102,
+      "eval_wer": 10.085438335809807,
+      "step": 6000
+    },
+    {
+      "epoch": 353.76,
+      "learning_rate": 1.0552941176470589e-06,
+      "loss": 0.0002,
+      "step": 6025
+    },
+    {
+      "epoch": 355.24,
+      "learning_rate": 1.0464705882352942e-06,
+      "loss": 0.0002,
+      "step": 6050
+    },
+    {
+      "epoch": 356.71,
+      "learning_rate": 1.0376470588235295e-06,
+      "loss": 0.0002,
+      "step": 6075
+    },
+    {
+      "epoch": 358.18,
+      "learning_rate": 1.0288235294117649e-06,
+      "loss": 0.0002,
+      "step": 6100
+    },
+    {
+      "epoch": 359.65,
+      "learning_rate": 1.0200000000000002e-06,
+      "loss": 0.0002,
+      "step": 6125
+    },
+    {
+      "epoch": 361.12,
+      "learning_rate": 1.0111764705882353e-06,
+      "loss": 0.0002,
+      "step": 6150
+    },
+    {
+      "epoch": 362.59,
+      "learning_rate": 1.0023529411764706e-06,
+      "loss": 0.0002,
+      "step": 6175
+    },
+    {
+      "epoch": 364.06,
+      "learning_rate": 9.93529411764706e-07,
+      "loss": 0.0004,
+      "step": 6200
+    },
+    {
+      "epoch": 365.53,
+      "learning_rate": 9.847058823529413e-07,
+      "loss": 0.0002,
+      "step": 6225
+    },
+    {
+      "epoch": 367.0,
+      "learning_rate": 9.758823529411766e-07,
+      "loss": 0.0002,
+      "step": 6250
+    },
+    {
+      "epoch": 368.47,
+      "learning_rate": 9.670588235294117e-07,
+      "loss": 0.0002,
+      "step": 6275
+    },
+    {
+      "epoch": 369.94,
+      "learning_rate": 9.58235294117647e-07,
+      "loss": 0.0002,
+      "step": 6300
+    },
+    {
+      "epoch": 371.41,
+      "learning_rate": 9.494117647058824e-07,
+      "loss": 0.0002,
+      "step": 6325
+    },
+    {
+      "epoch": 372.88,
+      "learning_rate": 9.405882352941177e-07,
+      "loss": 0.0003,
+      "step": 6350
+    },
+    {
+      "epoch": 374.35,
+      "learning_rate": 9.31764705882353e-07,
+      "loss": 0.0002,
+      "step": 6375
+    },
+    {
+      "epoch": 375.82,
+      "learning_rate": 9.229411764705882e-07,
+      "loss": 0.0002,
+      "step": 6400
+    },
+    {
+      "epoch": 377.29,
+      "learning_rate": 9.141176470588235e-07,
+      "loss": 0.0002,
+      "step": 6425
+    },
+    {
+      "epoch": 378.76,
+      "learning_rate": 9.052941176470588e-07,
+      "loss": 0.0002,
+      "step": 6450
+    },
+    {
+      "epoch": 380.24,
+      "learning_rate": 8.964705882352942e-07,
+      "loss": 0.0002,
+      "step": 6475
+    },
+    {
+      "epoch": 381.71,
+      "learning_rate": 8.876470588235295e-07,
+      "loss": 0.0002,
+      "step": 6500
+    },
+    {
+      "epoch": 383.18,
+      "learning_rate": 8.788235294117648e-07,
+      "loss": 0.0002,
+      "step": 6525
+    },
+    {
+      "epoch": 384.65,
+      "learning_rate": 8.699999999999999e-07,
+      "loss": 0.0003,
+      "step": 6550
+    },
+    {
+      "epoch": 386.12,
+      "learning_rate": 8.611764705882353e-07,
+      "loss": 0.0003,
+      "step": 6575
+    },
+    {
+      "epoch": 387.59,
+      "learning_rate": 8.523529411764706e-07,
+      "loss": 0.0002,
+      "step": 6600
+    },
+    {
+      "epoch": 389.06,
+      "learning_rate": 8.435294117647059e-07,
+      "loss": 0.0002,
+      "step": 6625
+    },
+    {
+      "epoch": 390.53,
+      "learning_rate": 8.347058823529412e-07,
+      "loss": 0.0001,
+      "step": 6650
+    },
+    {
+      "epoch": 392.0,
+      "learning_rate": 8.258823529411765e-07,
+      "loss": 0.0002,
+      "step": 6675
+    },
+    {
+      "epoch": 393.47,
+      "learning_rate": 8.170588235294118e-07,
+      "loss": 0.0001,
+      "step": 6700
+    },
+    {
+      "epoch": 394.94,
+      "learning_rate": 8.082352941176471e-07,
+      "loss": 0.0001,
+      "step": 6725
+    },
+    {
+      "epoch": 396.41,
+      "learning_rate": 7.994117647058823e-07,
+      "loss": 0.0002,
+      "step": 6750
+    },
+    {
+      "epoch": 397.88,
+      "learning_rate": 7.905882352941177e-07,
+      "loss": 0.0003,
+      "step": 6775
+    },
+    {
+      "epoch": 399.35,
+      "learning_rate": 7.817647058823529e-07,
+      "loss": 0.0004,
+      "step": 6800
+    },
+    {
+      "epoch": 400.82,
+      "learning_rate": 7.729411764705882e-07,
+      "loss": 0.0003,
+      "step": 6825
+    },
+    {
+      "epoch": 402.29,
+      "learning_rate": 7.641176470588236e-07,
+      "loss": 0.0002,
+      "step": 6850
+    },
+    {
+      "epoch": 403.76,
+      "learning_rate": 7.552941176470589e-07,
+      "loss": 0.0002,
+      "step": 6875
+    },
+    {
+      "epoch": 405.24,
+      "learning_rate": 7.464705882352941e-07,
+      "loss": 0.0002,
+      "step": 6900
+    },
+    {
+      "epoch": 406.71,
+      "learning_rate": 7.376470588235294e-07,
+      "loss": 0.0002,
+      "step": 6925
+    },
+    {
+      "epoch": 408.18,
+      "learning_rate": 7.288235294117648e-07,
+      "loss": 0.0001,
+      "step": 6950
+    },
+    {
+      "epoch": 409.65,
+      "learning_rate": 7.2e-07,
+      "loss": 0.0001,
+      "step": 6975
+    },
+    {
+      "epoch": 411.12,
+      "learning_rate": 7.111764705882353e-07,
+      "loss": 0.0002,
+      "step": 7000
+    },
+    {
+      "epoch": 411.12,
+      "eval_loss": 0.56689453125,
+      "eval_runtime": 168.4232,
+      "eval_samples_per_second": 1.615,
+      "eval_steps_per_second": 0.101,
+      "eval_wer": 10.11329866270431,
+      "step": 7000
+    },
+    {
+      "epoch": 412.59,
+      "learning_rate": 7.030588235294118e-07,
+      "loss": 0.0002,
+      "step": 7025
+    },
+    {
+      "epoch": 414.06,
+      "learning_rate": 6.94235294117647e-07,
+      "loss": 0.0003,
+      "step": 7050
+    },
+    {
+      "epoch": 415.53,
+      "learning_rate": 6.854117647058824e-07,
+      "loss": 0.0002,
+      "step": 7075
+    },
+    {
+      "epoch": 417.0,
+      "learning_rate": 6.765882352941177e-07,
+      "loss": 0.0001,
+      "step": 7100
+    },
+    {
+      "epoch": 418.47,
+      "learning_rate": 6.677647058823529e-07,
+      "loss": 0.0003,
+      "step": 7125
+    },
+    {
+      "epoch": 419.94,
+      "learning_rate": 6.589411764705882e-07,
+      "loss": 0.0002,
+      "step": 7150
+    },
+    {
+      "epoch": 421.41,
+      "learning_rate": 6.501176470588235e-07,
+      "loss": 0.0003,
+      "step": 7175
+    },
+    {
+      "epoch": 422.88,
+      "learning_rate": 6.412941176470588e-07,
+      "loss": 0.0003,
+      "step": 7200
+    },
+    {
+      "epoch": 424.35,
+      "learning_rate": 6.324705882352941e-07,
+      "loss": 0.0003,
+      "step": 7225
+    },
+    {
+      "epoch": 425.82,
+      "learning_rate": 6.236470588235294e-07,
+      "loss": 0.0002,
+      "step": 7250
+    },
+    {
+      "epoch": 427.29,
+      "learning_rate": 6.148235294117648e-07,
+      "loss": 0.0002,
+      "step": 7275
+    },
+    {
+      "epoch": 428.76,
+      "learning_rate": 6.060000000000001e-07,
+      "loss": 0.0003,
+      "step": 7300
+    },
+    {
+      "epoch": 430.24,
+      "learning_rate": 5.971764705882353e-07,
+      "loss": 0.0002,
+      "step": 7325
+    },
+    {
+      "epoch": 431.71,
+      "learning_rate": 5.883529411764707e-07,
+      "loss": 0.0003,
+      "step": 7350
+    },
+    {
+      "epoch": 433.18,
+      "learning_rate": 5.795294117647059e-07,
+      "loss": 0.0004,
+      "step": 7375
+    },
+    {
+      "epoch": 434.65,
+      "learning_rate": 5.707058823529412e-07,
+      "loss": 0.0002,
+      "step": 7400
+    },
+    {
+      "epoch": 436.12,
+      "learning_rate": 5.618823529411765e-07,
+      "loss": 0.0001,
+      "step": 7425
+    },
+    {
+      "epoch": 437.59,
+      "learning_rate": 5.530588235294118e-07,
+      "loss": 0.0002,
+      "step": 7450
+    },
+    {
+      "epoch": 439.06,
+      "learning_rate": 5.442352941176471e-07,
+      "loss": 0.0002,
+      "step": 7475
+    },
+    {
+      "epoch": 440.53,
+      "learning_rate": 5.354117647058824e-07,
+      "loss": 0.0001,
+      "step": 7500
+    },
+    {
+      "epoch": 442.0,
+      "learning_rate": 5.265882352941176e-07,
+      "loss": 0.0003,
+      "step": 7525
+    },
+    {
+      "epoch": 443.47,
+      "learning_rate": 5.17764705882353e-07,
+      "loss": 0.0008,
+      "step": 7550
+    },
+    {
+      "epoch": 444.94,
+      "learning_rate": 5.089411764705882e-07,
+      "loss": 0.0002,
+      "step": 7575
+    },
+    {
+      "epoch": 446.41,
+      "learning_rate": 5.001176470588235e-07,
+      "loss": 0.0001,
+      "step": 7600
+    },
+    {
+      "epoch": 447.88,
+      "learning_rate": 4.912941176470589e-07,
+      "loss": 0.0001,
+      "step": 7625
+    },
+    {
+      "epoch": 449.35,
+      "learning_rate": 4.824705882352941e-07,
+      "loss": 0.0001,
+      "step": 7650
+    },
+    {
+      "epoch": 450.82,
+      "learning_rate": 4.7364705882352946e-07,
+      "loss": 0.0001,
+      "step": 7675
+    },
+    {
+      "epoch": 452.29,
+      "learning_rate": 4.648235294117648e-07,
+      "loss": 0.0001,
+      "step": 7700
+    },
+    {
+      "epoch": 453.76,
+      "learning_rate": 4.56e-07,
+      "loss": 0.0003,
+      "step": 7725
+    },
+    {
+      "epoch": 455.24,
+      "learning_rate": 4.4717647058823534e-07,
+      "loss": 0.0001,
+      "step": 7750
+    },
+    {
+      "epoch": 456.71,
+      "learning_rate": 4.3835294117647056e-07,
+      "loss": 0.0003,
+      "step": 7775
+    },
+    {
+      "epoch": 458.18,
+      "learning_rate": 4.295294117647059e-07,
+      "loss": 0.0001,
+      "step": 7800
+    },
+    {
+      "epoch": 459.65,
+      "learning_rate": 4.207058823529412e-07,
+      "loss": 0.0002,
+      "step": 7825
+    },
+    {
+      "epoch": 461.12,
+      "learning_rate": 4.1188235294117644e-07,
+      "loss": 0.0002,
+      "step": 7850
+    },
+    {
+      "epoch": 462.59,
+      "learning_rate": 4.0305882352941177e-07,
+      "loss": 0.0001,
+      "step": 7875
+    },
+    {
+      "epoch": 464.06,
+      "learning_rate": 3.942352941176471e-07,
+      "loss": 0.0002,
+      "step": 7900
+    },
+    {
+      "epoch": 465.53,
+      "learning_rate": 3.8541176470588237e-07,
+      "loss": 0.0002,
+      "step": 7925
+    },
+    {
+      "epoch": 467.0,
+      "learning_rate": 3.765882352941177e-07,
+      "loss": 0.0002,
+      "step": 7950
+    },
+    {
+      "epoch": 468.47,
+      "learning_rate": 3.67764705882353e-07,
+      "loss": 0.0003,
+      "step": 7975
+    },
+    {
+      "epoch": 469.94,
+      "learning_rate": 3.5894117647058825e-07,
+      "loss": 0.0001,
+      "step": 8000
+    },
+    {
+      "epoch": 469.94,
+      "eval_loss": 0.56689453125,
+      "eval_runtime": 163.3464,
+      "eval_samples_per_second": 1.665,
+      "eval_steps_per_second": 0.104,
+      "eval_wer": 9.899702823179792,
+      "step": 8000
+    }
+  ],
+  "max_steps": 9000,
+  "num_train_epochs": 530,
+  "total_flos": 2.4643627644348688e+20,
+  "trial_name": null,
+  "trial_params": null
+}

checkpoint-8000/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:405e2f7c140f1d634318230fb26d9b2a3efb9f11f42a90f24dfd6e792f539418
+size 4795

checkpoint-8000/zero_to_fp32.py

@@ -0,0 +1,482 @@
+#!/usr/bin/env python
+
+# This script extracts fp32 consolidated weights from a zero 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
+
+# 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)
+
+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_optim_files(checkpoint_dir):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    optim_files = sorted(glob.glob(os.path.join(checkpoint_dir,
+                                                "*_optim_states.pt")),
+                         key=natural_keys)
+
+    if len(optim_files) == 0:
+        raise FileNotFoundError(
+            f"can't find '*_optim_states.pt' files in directory '{checkpoint_dir}'")
+
+    return optim_files
+
+
+def parse_model_state(file):
+    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]
+
+    ds_version = state_dict.get(DS_VERSION, None)
+
+    return buffers, param_shapes, ds_version
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+
+    total_files = len(files)
+    state_dicts = []
+    for f in files:
+        state_dicts.append(torch.load(f, map_location=device))
+
+    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_file = get_model_state_file(ds_checkpoint_dir, zero_stage)
+    buffers, param_shapes, ds_version = parse_model_state(model_file)
+    print(f'Parsing checkpoint created by deepspeed=={ds_version}')
+
+    if zero_stage == 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size,
+                                                          param_shapes,
+                                                          fp32_flat_groups,
+                                                          buffers)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size,
+                                                          param_shapes,
+                                                          fp32_flat_groups,
+                                                          buffers)
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size,
+                                               param_shapes,
+                                               fp32_flat_groups,
+                                               buffers):
+
+    # 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.")
+
+    state_dict = OrderedDict()
+
+    # buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    # 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"
+    )
+
+    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 _get_fp32_state_dict_from_zero3_checkpoint(world_size,
+                                               param_shapes,
+                                               fp32_flat_groups,
+                                               buffers):
+
+    # 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
+
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # 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
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    state_dict = OrderedDict()
+
+    # buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    # 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"{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 fp32 state dict with {total_params} params {total_numel} elements"
+    )
+
+    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("-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)

pytorch_model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:73391e01ae78440c548bd3f6d327e43048c43c15a1986a7fe6e3899aab9ecb2d
+oid sha256:292ed5692b25c5ef882949ab1088ccf2a62e2935100cd3609d1745d6a0afdc10
 size 1527847357

runs/Dec22_23-06-37_129-146-176-120/events.out.tfevents.1671750444.129-146-176-120.850347.0

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1707c83a45bcd245bc15e615f1fea52cd399cd069ca212fdeeea365121551ddc
-size 37305
+oid sha256:31b7ca987f7c5eb5ba0d19094c91d71052cfc66ee22b568f590621e95e2d17e2
+size 43903