adding checkpoint 6000

checkpoint-6000/config.json

@@ -0,0 +1,41 @@
+{
+  "_name_or_path": "openai/whisper-medium",
+  "activation_dropout": 0.0,
+  "activation_function": "gelu",
+  "architectures": [
+    "WhisperForConditionalGeneration"
+  ],
+  "attention_dropout": 0.0,
+  "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": 0.0,
+  "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-6000/global_step6000/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:235a017d37438a1c9c6d9d3fe96e5a8a5b803294909de73f9ff7887322ee4b21
+size 1527967899

checkpoint-6000/global_step6000/zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9a17a4f5e0259867e1fd646c5a33eab8d0a5e98d1e61f84d0c9a2ae7a29745dc
+size 9166378846

checkpoint-6000/latest

@@ -0,0 +1 @@
+global_step6000

checkpoint-6000/pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:54cd6bbe3bb4b7e34719dfbf66d1ed965431cc33215003245869ddd3befc1620
+size 1527847357

checkpoint-6000/rng_state.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f92a27bbf0f0370921b88ccbb3b02c1f378569056abb5feff869f62853b3aac7
+size 14575

checkpoint-6000/trainer_state.json

@@ -0,0 +1,1510 @@
+{
+  "best_metric": 12.226690123146518,
+  "best_model_checkpoint": "./checkpoint-6000",
+  "epoch": 0.2,
+  "global_step": 6000,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.0,
+      "learning_rate": 5.0453611334320685e-06,
+      "loss": 1.2966,
+      "step": 25
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 6.229195710491767e-06,
+      "loss": 0.8695,
+      "step": 50
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 6.903829450223392e-06,
+      "loss": 0.7429,
+      "step": 75
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 7.377725845391017e-06,
+      "loss": 0.6562,
+      "step": 100
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 7.743343231239583e-06,
+      "loss": 0.6217,
+      "step": 125
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 8.041073861170494e-06,
+      "loss": 0.5393,
+      "step": 150
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 8.292222957399574e-06,
+      "loss": 0.5282,
+      "step": 175
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 8.509413541357755e-06,
+      "loss": 0.5062,
+      "step": 200
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 8.700744577655557e-06,
+      "loss": 0.4942,
+      "step": 225
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 8.871723942761204e-06,
+      "loss": 0.4508,
+      "step": 250
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 9.026267958246849e-06,
+      "loss": 0.4466,
+      "step": 275
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 9.16726106663399e-06,
+      "loss": 0.4026,
+      "step": 300
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 9.296889251455016e-06,
+      "loss": 0.37,
+      "step": 325
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 9.416848797368692e-06,
+      "loss": 0.3864,
+      "step": 350
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 9.528482449516371e-06,
+      "loss": 0.3801,
+      "step": 375
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 9.632871309784314e-06,
+      "loss": 0.3893,
+      "step": 400
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 9.73089868785391e-06,
+      "loss": 0.4079,
+      "step": 425
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 9.823295589572114e-06,
+      "loss": 0.3554,
+      "step": 450
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 9.910673836465484e-06,
+      "loss": 0.3197,
+      "step": 475
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 9.993550644973805e-06,
+      "loss": 0.3405,
+      "step": 500
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 9.98871794871795e-06,
+      "loss": 0.3328,
+      "step": 525
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 9.975897435897436e-06,
+      "loss": 0.3147,
+      "step": 550
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 9.963076923076925e-06,
+      "loss": 0.2954,
+      "step": 575
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 9.950256410256412e-06,
+      "loss": 0.3034,
+      "step": 600
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 9.937435897435898e-06,
+      "loss": 0.2931,
+      "step": 625
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 9.924615384615385e-06,
+      "loss": 0.2798,
+      "step": 650
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 9.911794871794874e-06,
+      "loss": 0.2619,
+      "step": 675
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 9.89897435897436e-06,
+      "loss": 0.3157,
+      "step": 700
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 9.886153846153846e-06,
+      "loss": 0.2937,
+      "step": 725
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 9.873333333333334e-06,
+      "loss": 0.2606,
+      "step": 750
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 9.860512820512821e-06,
+      "loss": 0.2482,
+      "step": 775
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 9.847692307692308e-06,
+      "loss": 0.2539,
+      "step": 800
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 9.834871794871795e-06,
+      "loss": 0.2501,
+      "step": 825
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 9.822051282051283e-06,
+      "loss": 0.241,
+      "step": 850
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 9.80923076923077e-06,
+      "loss": 0.2323,
+      "step": 875
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 9.796410256410257e-06,
+      "loss": 0.2507,
+      "step": 900
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 9.783589743589744e-06,
+      "loss": 0.2157,
+      "step": 925
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 9.770769230769232e-06,
+      "loss": 0.2356,
+      "step": 950
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 9.757948717948719e-06,
+      "loss": 0.2352,
+      "step": 975
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 9.745128205128206e-06,
+      "loss": 0.2181,
+      "step": 1000
+    },
+    {
+      "epoch": 0.05,
+      "eval_loss": 0.2607421875,
+      "eval_runtime": 228.2189,
+      "eval_samples_per_second": 4.382,
+      "eval_steps_per_second": 0.14,
+      "eval_wer": 18.384016084443328,
+      "step": 1000
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 9.733333333333334e-06,
+      "loss": 0.2279,
+      "step": 1025
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 9.720512820512822e-06,
+      "loss": 0.225,
+      "step": 1050
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 9.707692307692308e-06,
+      "loss": 0.2238,
+      "step": 1075
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 9.694871794871794e-06,
+      "loss": 0.228,
+      "step": 1100
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 9.682051282051283e-06,
+      "loss": 0.2261,
+      "step": 1125
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 9.66923076923077e-06,
+      "loss": 0.196,
+      "step": 1150
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 9.656410256410257e-06,
+      "loss": 0.1962,
+      "step": 1175
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 9.643589743589743e-06,
+      "loss": 0.1931,
+      "step": 1200
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 9.630769230769232e-06,
+      "loss": 0.1827,
+      "step": 1225
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 9.617948717948719e-06,
+      "loss": 0.1815,
+      "step": 1250
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 9.605128205128206e-06,
+      "loss": 0.1972,
+      "step": 1275
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 9.592307692307692e-06,
+      "loss": 0.1681,
+      "step": 1300
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 9.579487179487181e-06,
+      "loss": 0.151,
+      "step": 1325
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 9.566666666666668e-06,
+      "loss": 0.1743,
+      "step": 1350
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 9.553846153846155e-06,
+      "loss": 0.1808,
+      "step": 1375
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 9.541025641025641e-06,
+      "loss": 0.1984,
+      "step": 1400
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 9.52820512820513e-06,
+      "loss": 0.2098,
+      "step": 1425
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 9.515384615384617e-06,
+      "loss": 0.1741,
+      "step": 1450
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 9.502564102564103e-06,
+      "loss": 0.1489,
+      "step": 1475
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 9.48974358974359e-06,
+      "loss": 0.1655,
+      "step": 1500
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 9.476923076923079e-06,
+      "loss": 0.1643,
+      "step": 1525
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 9.464102564102566e-06,
+      "loss": 0.151,
+      "step": 1550
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 9.451282051282052e-06,
+      "loss": 0.1431,
+      "step": 1575
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 9.43846153846154e-06,
+      "loss": 0.1503,
+      "step": 1600
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 9.425641025641026e-06,
+      "loss": 0.1491,
+      "step": 1625
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 9.412820512820515e-06,
+      "loss": 0.1408,
+      "step": 1650
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 9.4e-06,
+      "loss": 0.1294,
+      "step": 1675
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 9.387179487179488e-06,
+      "loss": 0.1618,
+      "step": 1700
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 9.374358974358975e-06,
+      "loss": 0.1484,
+      "step": 1725
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 9.361538461538462e-06,
+      "loss": 0.1298,
+      "step": 1750
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 9.348717948717949e-06,
+      "loss": 0.1231,
+      "step": 1775
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 9.335897435897436e-06,
+      "loss": 0.1241,
+      "step": 1800
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 9.323076923076924e-06,
+      "loss": 0.1215,
+      "step": 1825
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 9.310256410256411e-06,
+      "loss": 0.1225,
+      "step": 1850
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 9.297435897435898e-06,
+      "loss": 0.1122,
+      "step": 1875
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 9.284615384615385e-06,
+      "loss": 0.1191,
+      "step": 1900
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 9.271794871794873e-06,
+      "loss": 0.0995,
+      "step": 1925
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 9.25897435897436e-06,
+      "loss": 0.1199,
+      "step": 1950
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 9.246153846153847e-06,
+      "loss": 0.1243,
+      "step": 1975
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 9.233333333333334e-06,
+      "loss": 0.1089,
+      "step": 2000
+    },
+    {
+      "epoch": 0.05,
+      "eval_loss": 0.2489013671875,
+      "eval_runtime": 209.1058,
+      "eval_samples_per_second": 4.782,
+      "eval_steps_per_second": 0.153,
+      "eval_wer": 16.486554410655945,
+      "step": 2000
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 9.221538461538462e-06,
+      "loss": 0.1231,
+      "step": 2025
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 9.208717948717948e-06,
+      "loss": 0.1173,
+      "step": 2050
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 9.195897435897437e-06,
+      "loss": 0.1234,
+      "step": 2075
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 9.183076923076924e-06,
+      "loss": 0.1244,
+      "step": 2100
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 9.17025641025641e-06,
+      "loss": 0.1223,
+      "step": 2125
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 9.157435897435897e-06,
+      "loss": 0.1027,
+      "step": 2150
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 9.144615384615386e-06,
+      "loss": 0.1022,
+      "step": 2175
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 9.131794871794873e-06,
+      "loss": 0.1013,
+      "step": 2200
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 9.11897435897436e-06,
+      "loss": 0.0949,
+      "step": 2225
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 9.106153846153846e-06,
+      "loss": 0.0971,
+      "step": 2250
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 9.093333333333333e-06,
+      "loss": 0.1098,
+      "step": 2275
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 9.080512820512822e-06,
+      "loss": 0.0861,
+      "step": 2300
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 9.067692307692309e-06,
+      "loss": 0.0765,
+      "step": 2325
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 9.054871794871795e-06,
+      "loss": 0.0933,
+      "step": 2350
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 9.042051282051282e-06,
+      "loss": 0.1025,
+      "step": 2375
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 9.02923076923077e-06,
+      "loss": 0.1138,
+      "step": 2400
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 9.016410256410258e-06,
+      "loss": 0.1198,
+      "step": 2425
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 9.003589743589744e-06,
+      "loss": 0.0981,
+      "step": 2450
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 8.990769230769231e-06,
+      "loss": 0.0807,
+      "step": 2475
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 8.97794871794872e-06,
+      "loss": 0.0875,
+      "step": 2500
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 8.965128205128207e-06,
+      "loss": 0.0911,
+      "step": 2525
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 8.952307692307693e-06,
+      "loss": 0.0796,
+      "step": 2550
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 8.93948717948718e-06,
+      "loss": 0.0766,
+      "step": 2575
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 8.926666666666669e-06,
+      "loss": 0.0862,
+      "step": 2600
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 8.913846153846154e-06,
+      "loss": 0.0822,
+      "step": 2625
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 8.90102564102564e-06,
+      "loss": 0.0757,
+      "step": 2650
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 8.88820512820513e-06,
+      "loss": 0.0674,
+      "step": 2675
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 8.875384615384616e-06,
+      "loss": 0.0913,
+      "step": 2700
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 8.862564102564103e-06,
+      "loss": 0.0815,
+      "step": 2725
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 8.84974358974359e-06,
+      "loss": 0.0687,
+      "step": 2750
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 8.836923076923078e-06,
+      "loss": 0.0649,
+      "step": 2775
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 8.824102564102565e-06,
+      "loss": 0.0647,
+      "step": 2800
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 8.811282051282052e-06,
+      "loss": 0.0607,
+      "step": 2825
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 8.798461538461539e-06,
+      "loss": 0.065,
+      "step": 2850
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 8.785641025641025e-06,
+      "loss": 0.058,
+      "step": 2875
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 8.772820512820514e-06,
+      "loss": 0.0618,
+      "step": 2900
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 8.76e-06,
+      "loss": 0.0491,
+      "step": 2925
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 8.747179487179488e-06,
+      "loss": 0.0646,
+      "step": 2950
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 8.734358974358974e-06,
+      "loss": 0.0673,
+      "step": 2975
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 8.721538461538463e-06,
+      "loss": 0.0564,
+      "step": 3000
+    },
+    {
+      "epoch": 0.05,
+      "eval_loss": 0.261474609375,
+      "eval_runtime": 212.2501,
+      "eval_samples_per_second": 4.711,
+      "eval_steps_per_second": 0.151,
+      "eval_wer": 16.14727318421714,
+      "step": 3000
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 8.70974358974359e-06,
+      "loss": 0.1893,
+      "step": 3025
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 8.696923076923078e-06,
+      "loss": 0.1774,
+      "step": 3050
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 8.684102564102565e-06,
+      "loss": 0.1866,
+      "step": 3075
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 8.671282051282051e-06,
+      "loss": 0.1896,
+      "step": 3100
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 8.658461538461538e-06,
+      "loss": 0.1925,
+      "step": 3125
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 8.645641025641027e-06,
+      "loss": 0.2126,
+      "step": 3150
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 8.632820512820514e-06,
+      "loss": 0.2004,
+      "step": 3175
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 8.62e-06,
+      "loss": 0.1636,
+      "step": 3200
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 8.607179487179487e-06,
+      "loss": 0.1597,
+      "step": 3225
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 8.594358974358976e-06,
+      "loss": 0.1768,
+      "step": 3250
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 8.581538461538463e-06,
+      "loss": 0.1684,
+      "step": 3275
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 8.56871794871795e-06,
+      "loss": 0.1739,
+      "step": 3300
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 8.555897435897436e-06,
+      "loss": 0.1706,
+      "step": 3325
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 8.543076923076923e-06,
+      "loss": 0.1701,
+      "step": 3350
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 8.530256410256412e-06,
+      "loss": 0.1564,
+      "step": 3375
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 8.517435897435898e-06,
+      "loss": 0.1627,
+      "step": 3400
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 8.504615384615385e-06,
+      "loss": 0.1586,
+      "step": 3425
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 8.491794871794872e-06,
+      "loss": 0.1579,
+      "step": 3450
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 8.47897435897436e-06,
+      "loss": 0.1676,
+      "step": 3475
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 8.466153846153847e-06,
+      "loss": 0.1558,
+      "step": 3500
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 8.453333333333334e-06,
+      "loss": 0.1624,
+      "step": 3525
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 8.440512820512821e-06,
+      "loss": 0.1459,
+      "step": 3550
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 8.427692307692308e-06,
+      "loss": 0.151,
+      "step": 3575
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 8.414871794871795e-06,
+      "loss": 0.1524,
+      "step": 3600
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 8.402051282051282e-06,
+      "loss": 0.1444,
+      "step": 3625
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 8.38923076923077e-06,
+      "loss": 0.1218,
+      "step": 3650
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 8.376410256410257e-06,
+      "loss": 0.1207,
+      "step": 3675
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 8.363589743589744e-06,
+      "loss": 0.1199,
+      "step": 3700
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 8.35076923076923e-06,
+      "loss": 0.1154,
+      "step": 3725
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 8.337948717948719e-06,
+      "loss": 0.1075,
+      "step": 3750
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 8.325128205128206e-06,
+      "loss": 0.1005,
+      "step": 3775
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 8.312307692307693e-06,
+      "loss": 0.1053,
+      "step": 3800
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 8.29948717948718e-06,
+      "loss": 0.1087,
+      "step": 3825
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 8.286666666666668e-06,
+      "loss": 0.1207,
+      "step": 3850
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 8.273846153846155e-06,
+      "loss": 0.1099,
+      "step": 3875
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 8.261025641025642e-06,
+      "loss": 0.1054,
+      "step": 3900
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 8.248205128205129e-06,
+      "loss": 0.1019,
+      "step": 3925
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 8.235384615384615e-06,
+      "loss": 0.0974,
+      "step": 3950
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 8.222564102564104e-06,
+      "loss": 0.0975,
+      "step": 3975
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 8.20974358974359e-06,
+      "loss": 0.1031,
+      "step": 4000
+    },
+    {
+      "epoch": 0.1,
+      "eval_loss": 0.1995849609375,
+      "eval_runtime": 214.2177,
+      "eval_samples_per_second": 4.668,
+      "eval_steps_per_second": 0.149,
+      "eval_wer": 14.224679567730586,
+      "step": 4000
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 8.197948717948719e-06,
+      "loss": 0.1122,
+      "step": 4025
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 8.185128205128206e-06,
+      "loss": 0.1038,
+      "step": 4050
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 8.172307692307692e-06,
+      "loss": 0.1172,
+      "step": 4075
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 8.15948717948718e-06,
+      "loss": 0.1251,
+      "step": 4100
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 8.146666666666668e-06,
+      "loss": 0.1306,
+      "step": 4125
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 8.133846153846155e-06,
+      "loss": 0.1043,
+      "step": 4150
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 8.121025641025641e-06,
+      "loss": 0.1095,
+      "step": 4175
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 8.108205128205128e-06,
+      "loss": 0.1194,
+      "step": 4200
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 8.095384615384617e-06,
+      "loss": 0.1209,
+      "step": 4225
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 8.082564102564104e-06,
+      "loss": 0.1108,
+      "step": 4250
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 8.06974358974359e-06,
+      "loss": 0.1059,
+      "step": 4275
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 8.056923076923077e-06,
+      "loss": 0.0923,
+      "step": 4300
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 8.044102564102566e-06,
+      "loss": 0.1027,
+      "step": 4325
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 8.031282051282053e-06,
+      "loss": 0.1008,
+      "step": 4350
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 8.01846153846154e-06,
+      "loss": 0.0885,
+      "step": 4375
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 8.005641025641026e-06,
+      "loss": 0.0916,
+      "step": 4400
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 7.992820512820515e-06,
+      "loss": 0.1045,
+      "step": 4425
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 7.980000000000002e-06,
+      "loss": 0.0954,
+      "step": 4450
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 7.967179487179488e-06,
+      "loss": 0.1013,
+      "step": 4475
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 7.954358974358975e-06,
+      "loss": 0.0983,
+      "step": 4500
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 7.941538461538462e-06,
+      "loss": 0.1021,
+      "step": 4525
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 7.928717948717949e-06,
+      "loss": 0.1024,
+      "step": 4550
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 7.915897435897436e-06,
+      "loss": 0.1068,
+      "step": 4575
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 7.903076923076922e-06,
+      "loss": 0.1057,
+      "step": 4600
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 7.890256410256411e-06,
+      "loss": 0.1144,
+      "step": 4625
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 7.877435897435898e-06,
+      "loss": 0.1401,
+      "step": 4650
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 7.864615384615385e-06,
+      "loss": 0.1234,
+      "step": 4675
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 7.851794871794871e-06,
+      "loss": 0.1364,
+      "step": 4700
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 7.83897435897436e-06,
+      "loss": 0.1367,
+      "step": 4725
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 7.826153846153847e-06,
+      "loss": 0.1399,
+      "step": 4750
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 7.813333333333334e-06,
+      "loss": 0.1387,
+      "step": 4775
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 7.80051282051282e-06,
+      "loss": 0.1429,
+      "step": 4800
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 7.787692307692309e-06,
+      "loss": 0.134,
+      "step": 4825
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 7.774871794871796e-06,
+      "loss": 0.1615,
+      "step": 4850
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 7.762051282051283e-06,
+      "loss": 0.1502,
+      "step": 4875
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 7.74923076923077e-06,
+      "loss": 0.1574,
+      "step": 4900
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 7.736410256410258e-06,
+      "loss": 0.1529,
+      "step": 4925
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 7.723589743589745e-06,
+      "loss": 0.1406,
+      "step": 4950
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 7.710769230769232e-06,
+      "loss": 0.1363,
+      "step": 4975
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 7.697948717948718e-06,
+      "loss": 0.14,
+      "step": 5000
+    },
+    {
+      "epoch": 0.15,
+      "eval_loss": 0.18212890625,
+      "eval_runtime": 217.0927,
+      "eval_samples_per_second": 4.606,
+      "eval_steps_per_second": 0.147,
+      "eval_wer": 13.005780346820808,
+      "step": 5000
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 7.686153846153846e-06,
+      "loss": 0.1421,
+      "step": 5025
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 7.673333333333333e-06,
+      "loss": 0.1507,
+      "step": 5050
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 7.660512820512822e-06,
+      "loss": 0.1445,
+      "step": 5075
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 7.647692307692309e-06,
+      "loss": 0.1517,
+      "step": 5100
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 7.634871794871795e-06,
+      "loss": 0.1363,
+      "step": 5125
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 7.622051282051282e-06,
+      "loss": 0.1548,
+      "step": 5150
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 7.60923076923077e-06,
+      "loss": 0.1403,
+      "step": 5175
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 7.596410256410257e-06,
+      "loss": 0.1543,
+      "step": 5200
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 7.5835897435897444e-06,
+      "loss": 0.1386,
+      "step": 5225
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 7.570769230769231e-06,
+      "loss": 0.1194,
+      "step": 5250
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 7.557948717948719e-06,
+      "loss": 0.1075,
+      "step": 5275
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 7.545128205128206e-06,
+      "loss": 0.1076,
+      "step": 5300
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 7.5323076923076934e-06,
+      "loss": 0.1034,
+      "step": 5325
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 7.51948717948718e-06,
+      "loss": 0.1096,
+      "step": 5350
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 7.506666666666668e-06,
+      "loss": 0.1051,
+      "step": 5375
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 7.493846153846155e-06,
+      "loss": 0.0989,
+      "step": 5400
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 7.481025641025642e-06,
+      "loss": 0.1002,
+      "step": 5425
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 7.468205128205129e-06,
+      "loss": 0.0961,
+      "step": 5450
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 7.455384615384615e-06,
+      "loss": 0.1017,
+      "step": 5475
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 7.442564102564103e-06,
+      "loss": 0.088,
+      "step": 5500
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 7.42974358974359e-06,
+      "loss": 0.0921,
+      "step": 5525
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 7.416923076923077e-06,
+      "loss": 0.0921,
+      "step": 5550
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 7.404102564102564e-06,
+      "loss": 0.0853,
+      "step": 5575
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 7.391282051282052e-06,
+      "loss": 0.0752,
+      "step": 5600
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 7.378461538461539e-06,
+      "loss": 0.0705,
+      "step": 5625
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 7.365641025641026e-06,
+      "loss": 0.0743,
+      "step": 5650
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 7.352820512820513e-06,
+      "loss": 0.0859,
+      "step": 5675
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 7.340000000000001e-06,
+      "loss": 0.0759,
+      "step": 5700
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 7.327179487179488e-06,
+      "loss": 0.0821,
+      "step": 5725
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 7.3143589743589745e-06,
+      "loss": 0.0824,
+      "step": 5750
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 7.301538461538462e-06,
+      "loss": 0.0747,
+      "step": 5775
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 7.288717948717949e-06,
+      "loss": 0.072,
+      "step": 5800
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 7.275897435897437e-06,
+      "loss": 0.0785,
+      "step": 5825
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 7.2630769230769235e-06,
+      "loss": 0.0751,
+      "step": 5850
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 7.250256410256411e-06,
+      "loss": 0.0822,
+      "step": 5875
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 7.237435897435898e-06,
+      "loss": 0.101,
+      "step": 5900
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 7.224615384615386e-06,
+      "loss": 0.0896,
+      "step": 5925
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 7.2117948717948725e-06,
+      "loss": 0.1193,
+      "step": 5950
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 7.19897435897436e-06,
+      "loss": 0.0841,
+      "step": 5975
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 7.186153846153847e-06,
+      "loss": 0.0872,
+      "step": 6000
+    },
+    {
+      "epoch": 0.2,
+      "eval_loss": 0.1734619140625,
+      "eval_runtime": 224.4062,
+      "eval_samples_per_second": 4.456,
+      "eval_steps_per_second": 0.143,
+      "eval_wer": 12.226690123146518,
+      "step": 6000
+    }
+  ],
+  "max_steps": 20000,
+  "num_train_epochs": 9223372036854775807,
+  "total_flos": 3.9191228855447716e+20,
+  "trial_name": null,
+  "trial_params": null
+}

checkpoint-6000/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a3261208c10fe629e4db911bd85a63442dc29a9f5a7d2fd10502e57c4432d708
+size 4731

checkpoint-6000/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)