---
layout: '@/layouts/Doc.astro'
title: '🏔️ Spike Skipper'
date: 2024-09-17
description: 'Implementation of a mechanism to skip bad-data training steps that cause loss spikes during LLM training.'
---

- [📝 Example](#pencil-example)
- [🧪 Implementation](#test_tube-implementation)
- [✅ Sanity Check](#white_check_mark-sanity-check)
  - [🔍 Details](#mag-details)

> [!CAUTION] Details
>
> We describe below our implementation for skipping individual steps
> during training.

## 📝 Example

Suppose we observe a large spike in our loss curve, as shown below:

<img src="/assets/spike-skipper.png" alt="spike-skipper" />

Seemingly, this spike is being caused by a batch of “bad data”. In order
to prevent this “bad data” sample from corrupting our training, we would
like to “skip” that particular training step.

This can be accomplished by passing the keyword argument
`--train-range-to-skip` and specifying the endpoints of the ranges to be
skipped.

e.g., if you would like to skip all steps from `[10, 20]` and from
`[25, 30]`, we would specify:

``` bash
PBS_O_WORKDIR=$(pwd) bash train_aGPT_7B.sh \
    --train-range-to-skip 10 20 25 30
```

## 🧪 Implementation

We discuss below the details of the implementation, and provide some
simple results to confirm things are behaving how we expect.

1.  Check if `args.train_range_to_skip is not None`
    \[[here](https://github.com/argonne-lcf/Megatron-DeepSpeed/blob/295fcb3d57a40ec513a521aa8814d99a5c8827b8/megatron/training.py#L1032)\]

    - Assert `len(args.train_range_to_skip) % 2 == 0`
      \[[here](https://github.com/argonne-lcf/Megatron-DeepSpeed/blob/295fcb3d57a40ec513a521aa8814d99a5c8827b8/megatron/training.py#L1033-L1037)\]

      <div class="aside">

      Must be even since we’re specifying the *endpoint**s*** of
      intervals to skip

      </div>

    - Zip these up into pairs
      \[[here](https://github.com/argonne-lcf/Megatron-DeepSpeed/blob/295fcb3d57a40ec513a521aa8814d99a5c8827b8/megatron/training.py#L1038-L1043)\]:

      ``` python
      ranges_to_skip = list(
          zip(
              args.train_range_to_skip[::2],
              args.train_range_to_skip[1::2]
          )
      )
      ```

2.  If current iteration is in any of these pairs
    \[[here](https://github.com/argonne-lcf/Megatron-DeepSpeed/blob/295fcb3d57a40ec513a521aa8814d99a5c8827b8/megatron/training.py#L1066-L1085)\]

    - For each `micro_step` in `range(gradient_accumulation_steps)`,
      explicitly:

      - draw a new batch of data
        \[[here](https://github.com/argonne-lcf/Megatron-DeepSpeed/blob/295fcb3d57a40ec513a521aa8814d99a5c8827b8/megatron/training.py#L1074-L1076)\]
        from our `train_data_iterator`
      - immediately discard it (instead of propagating it through the
        network)

    - Increment
      \[[here](https://github.com/argonne-lcf/Megatron-DeepSpeed/blob/295fcb3d57a40ec513a521aa8814d99a5c8827b8/megatron/training.py#L1087-L1094)\]:

      - [`DeepSpeedEngine.skipped_steps`](https://github.com/argonne-lcf/Megatron-DeepSpeed/blob/295fcb3d57a40ec513a521aa8814d99a5c8827b8/megatron/training.py#L1090)
      - [`DeepSpeedEngine.global_steps`](https://github.com/argonne-lcf/Megatron-DeepSpeed/blob/295fcb3d57a40ec513a521aa8814d99a5c8827b8/megatron/training.py#L1091)
      - [`DeepSpeedEngine.micro_steps`](https://github.com/argonne-lcf/Megatron-DeepSpeed/blob/295fcb3d57a40ec513a521aa8814d99a5c8827b8/megatron/training.py#L1092)
      - [`DeepSpeedEngine.global_samples`](https://github.com/argonne-lcf/Megatron-DeepSpeed/blob/295fcb3d57a40ec513a521aa8814d99a5c8827b8/megatron/training.py#L1093)
      - [`lr_scheduler`](https://github.com/argonne-lcf/Megatron-DeepSpeed/blob/295fcb3d57a40ec513a521aa8814d99a5c8827b8/megatron/training.py#L1094)

## ✅ Sanity Check

In order to confirm things are behaving as expected, we can explicitly
look at the tokens drawn for each step, and ensure that they are the
same regardless of whether or not that iteration was skipped.

- In particular, we see that:

  - test 1:

    ``` python
    # [2024-09-16 23:09:09.059118][INFO][training:1083] - iteration=2 [0/8]: (torch.Size([4, 4097]))
    _tokens[:10]=tensor(
        [[ 1858,  3851, 29889,  ...,   500,    13,    13],
         [  349,  6156,  1650,  ...,  5806, 28557,  3519],
         [16554,   304,  1653,  ...,   322,  6934, 14722],
         [ 4955,   310, 10465,  ...,  1438,  3841, 29892]]
    )
    # [2024-09-16 23:09:09.061999][INFO][training:1083] - iteration=2 [1/8]: (torch.Size([4, 4097]))
    _tokens[:10]=tensor(
        [[  363,  1302, 16453,  ...,  7967, 29891,   484],
         [  367,   766,  4752,  ...,     1, 29871, 30143],
         [29899,   855,  1503,  ...,  3786, 29892,  5100],
         [  465,  1974,   289,  ..., 21588,   533,   304]]
    )
    ```

  - test 2:

    ``` python
    # [2024-09-16 22:59:27.752277][INFO][pretrain_gpt_alcf:198] - args.iteration=2:
    data['text'][:10]=tensor(
        [[ 1858,  3851, 29889,  ...,   500,    13,    13],
         [  349,  6156,  1650,  ...,  5806, 28557,  3519],
         [16554,   304,  1653,  ...,   322,  6934, 14722],
         [ 4955,   310, 10465,  ...,  1438,  3841, 29892]]
    )
    # [2024-09-16 22:59:27,755] [INFO] [profiler.py:81:start_profile] Flops profiler started
    # [2024-09-16 22:59:28.568805][INFO][pretrain_gpt_alcf:198] - args.iteration=2:
    data['text'][:10]=tensor(
        [[363,  1302, 16453,  ...,  7967, 29891,   484],
         [  367,   766,  4752,  ...,     1, 29871, 30143],
         [29899,   855,  1503,  ...,  3786, 29892,  5100],
         [  465,  1974,   289,  ..., 21588,   533,   304]]
    )
    ```

  as expected.

### 🔍 Details

- First 4 steps:

  <details is-="accordion">

  <summary><code>tokens</code>:</summary>

  - Iteration 0:

    ``` sh
    [2024-09-16 22:58:50.168667][INFO][pretrain_gpt_alcf:198] - args.iteration=0: data['text'][:10]=tensor([[  304,  7344,  5146,  ...,  9776, 29914, 26419],
            [29889,    13,  4706,  ...,  9280, 30004,    13],
            [29943, 20774, 29908,  ...,   304, 27391,   322],
            [ 2645,   445, 29871,  ..., 16888,  4656, 10070]])
    [2024-09-16 22:58:58.866409][INFO][pretrain_gpt_alcf:198] - args.iteration=0: data['text'][:10]=tensor([[ 2768,   596,  1788,  ..., 27274,   393, 30010],
            [  278,  5613,  4192,  ...,   362,   310,  1950],
            [28038, 29892,  2022,  ...,  3160,   278,  2087],
            [ 4149,   907, 29888,  ..., 29896, 29892, 29896]])
    [2024-09-16 22:59:02.043059][INFO][pretrain_gpt_alcf:198] - args.iteration=0: data['text'][:10]=tensor([[  424,   322, 16232,  ...,   366,   748,   467],
            [   13,   462,  1678,  ...,  2084, 29892,  3497],
            [ 7562,   310, 19320,  ...,  8973, 22684,   358],
            [ 2089,  3633,   292,  ..., 13774,   269,  2375]])
    [2024-09-16 22:59:03.456919][INFO][pretrain_gpt_alcf:198] - args.iteration=0: data['text'][:10]=tensor([[21411,   322,  3896,  ...,  2610, 29889,   319],
            [ 8003, 29898, 29900,  ...,    12,  6658,   529],
            [  278,  4148,   310,  ...,   263, 12212,   282],
            [ 5977, 29871, 29906,  ..., 15332,   310,  1749]])
    [2024-09-16 22:59:04.596630][INFO][pretrain_gpt_alcf:198] - args.iteration=0: data['text'][:10]=tensor([[  278,  1473, 24987,  ...,   263,  2217,  3804],
            [ 2973,   263, 18778,  ...,   263,  4642,  6673],
            [  309,   323,   804,  ...,  1063, 15296,   327],
            [  278,  5864,   322,  ...,  9409, 29889,  2178]])
    [2024-09-16 22:59:05.486913][INFO][pretrain_gpt_alcf:198] - args.iteration=0: data['text'][:10]=tensor([[29892, 13731,  6617,  ..., 29871, 29896, 29946],
            [ 2892,  1012,  1266,  ...,  4036,  7512,  2068],
            [ 1473,  1556,  3619,  ...,  3762,   338,   263],
            [23353, 29918,  2177,  ...,   501,   567,   814]])
    [2024-09-16 22:59:06.361333][INFO][pretrain_gpt_alcf:198] - args.iteration=0: data['text'][:10]=tensor([[ 5400, 14378,  4768,  ...,  2107, 18677, 29889],
            [ 9200, 29887, 29914,  ...,   293, 24235,   322],
            [30143,  4746,  2184,  ..., 11891, 29974, 25760],
            [19263, 29914,   303,  ...,   358, 29889,    13]])
    [2024-09-16 22:59:07.230671][INFO][pretrain_gpt_alcf:198] - args.iteration=0: data['text'][:10]=tensor([[  309,  1306,   681,  ...,   310, 23186, 21809],
            [29896, 29929,    13,  ..., 29871, 29900,    13],
            [ 9558,   964,   263,  ...,   322,   282,   682],
            [  278, 23904, 21767,  ...,   313, 29929, 29889]])
    ```

  - Iteration 1:

    ``` bash
    [2024-09-16 22:59:19.287338][INFO][training_log:661] -  iteration=       1/  635782 | consumed_samples=         768 | consumed_tokens=     3145728 | elapsed_time_per_iteration_ms=29570.0 | learning_rate=9.4372e-09 | global_batch_size=768 | lm loss=11.167250 | loss_scale=1.0 | grad_norm=6.363 | actual_seqlen= 4096 | number_of_skipped_iterations=  0 | number_of_nan_iterations=  0 | samples_per_second=25.972 | tokens_per_gpu_per_second_tgs=4432.597 | [LM]TFLOPs=20.30 | [DS]TFLOPs=26.18 |
    [2024-09-16 22:59:19.289582][INFO][utils:207] - [Rank 0] (after 1 iterations) memory (MB) | allocated: 1894.57666015625 | max allocated: 9752.35498046875 | reserved: 11342.0 | max reserved: 11342.0
    (min, max) time across ranks (ms):
      forward-backward ...............................: (26094.39, 26095.09)
      optimizer ......................................: (3407.56, 3409.92)
    [2024-09-16 22:59:19.297183][INFO][pretrain_gpt_alcf:198] - args.iteration=1: data['text'][:10]=tensor([[ 1472, 29892,   408,  ..., 29892,  1584,   363],
          [  967, 19475,  6593,  ...,  8093, 29899, 11249],
          [ 1006,  2218, 13326,  ...,  2355,  1304,   304],
          [29900, 29916, 29947,  ...,   353,  1870, 29936]])
    [2024-09-16 22:59:20.104352][INFO][pretrain_gpt_alcf:198] - args.iteration=1: data['text'][:10]=tensor([[ 2354,   274,  1041,  ..., 29892, 13049,  9098],
          [ 8798,  9547, 10353,  ...,   303,  3143, 29889],
          [ 1373,  4056,  7236,  ...,  3186,   297,  5837],
          [ 1738, 29920,  7355,  ...,    13, 29871,  3776]])
    [2024-09-16 22:59:20.977036][INFO][utils:326] -  >> building dataset for /flare/Aurora_deployment/AuroraGPT/datasets/dolma/data_v1.7_Llama2Tokenizer/c4-0000_text_document
    [2024-09-16 22:59:20.977877][INFO][utils:326] -  > building dataset index ...
    [2024-09-16 22:59:20.977147][INFO][pretrain_gpt_alcf:198] - args.iteration=1: data['text'][:10]=tensor([[ 2020,   306,  1016,  ...,   322,   920,   372],
          [ 5921,  1749,  7306,  ..., 19252,   297,  5664],
          [  970,   770, 28547,  ...,   970,   894,  2577],
          [ 1907,   363, 14188,  ...,   756,  3646,   287]])
    [2024-09-16 22:59:21.851620][INFO][pretrain_gpt_alcf:198] - args.iteration=1: data['text'][:10]=tensor([[  715, 25392,  3104,  ...,   289,  5761,   616],
          [  426,    13,  9651,  ...,  9651,  1815, 22603],
          [ 7714,  1213,    13,  ...,    13, 29876,   457],
          [29889, 28663,  1230,  ...,  1546,   278,  6586]])
    [2024-09-16 22:59:22.720945][INFO][pretrain_gpt_alcf:198] - args.iteration=1: data['text'][:10]=tensor([[29929,    13,    13,  ..., 10739,  4770, 11277],
          [ 4528,   304,  2367,  ...,  2501,   385,  4203],
          [  869,   319,   794,  ...,  3158, 29889,  3115],
          [  592,   260,  4125,  ...,   284,  1135, 18655]])
    [2024-09-16 22:59:23.590149][INFO][pretrain_gpt_alcf:198] - args.iteration=1: data['text'][:10]=tensor([[14338, 25323,  3321,  ...,  5607,  1806,  1164],
          [  322,   278, 15352,  ...,  6462,   313,  1552],
          [25738,   714, 29889,  ..., 29915, 29879, 24842],
          [ 5122,   399, 29889,  ..., 29947,  7284,  2305]])
    [2024-09-16 22:59:24.457646][INFO][pretrain_gpt_alcf:198] - args.iteration=1: data['text'][:10]=tensor([[  367, 19310,  1891,  ...,  2408,   292,   263],
          [  470,  3307,  5713,  ...,   568,  2594, 19385],
          [29953, 29905,  1631,  ...,  1118,   343, 29897],
          [10261,   373,  5490,  ...,   511,   297,  1760]])
    [2024-09-16 22:59:25.326699][INFO][pretrain_gpt_alcf:198] - args.iteration=1: data['text'][:10]=tensor([[ 1006,   326, 29901,  ..., 14834,  6694,  9595],
          [12058,  5446, 29892,  ..., 29889,  8246,  3310],
          [ 7483,   310,   278,  ...,   402,  9851,  4423],
          [ 8041,   813,   322,  ...,  3303,  3900,   393]])
    ```

  - Iteration 2:

    ``` bash
    [2024-09-16 22:59:27.744603][INFO][training_log:661] -  iteration=       2/  635782 | consumed_samples=        1536 | consumed_tokens=     6291456 | elapsed_time_per_iteration_ms=8457.2 | learning_rate=1.88744e-08 | global_batch_size=768 | lm loss=11.164009 | loss_scale=1.0 | grad_norm=6.271 | actual_seqlen= 4096 | number_of_skipped_iterations=  0 | number_of_nan_iterations=  0 | samples_per_second=90.810 | tokens_per_gpu_per_second_tgs=15498.234 | [LM]TFLOPs=70.98| [DS]TFLOPs=91.53 |
    (min, max) time across ranks (ms):
      forward-backward ...............................: (8384.83, 8385.57)
      optimizer ......................................: (55.03, 55.61)
    [2024-09-16 22:59:27.752277][INFO][pretrain_gpt_alcf:198] - args.iteration=2: data['text'][:10]=tensor([[ 1858,  3851, 29889,  ...,   500,    13,    13],
          [  349,  6156,  1650,  ...,  5806, 28557,  3519],
          [16554,   304,  1653,  ...,   322,  6934, 14722],
          [ 4955,   310, 10465,  ...,  1438,  3841, 29892]])
    [2024-09-16 22:59:27,755] [INFO] [profiler.py:81:start_profile] Flops profiler started
    [2024-09-16 22:59:28.568805][INFO][pretrain_gpt_alcf:198] - args.iteration=2: data['text'][:10]=tensor([[  363,  1302, 16453,  ...,  7967, 29891,   484],
          [  367,   766,  4752,  ...,     1, 29871, 30143],
          [29899,   855,  1503,  ...,  3786, 29892,  5100],
          [  465,  1974,   289,  ..., 21588,   533,   304]])
    [2024-09-16 22:59:28,571] [INFO] [profiler.py:81:start_profile] Flops profiler started
    [2024-09-16 22:59:29.440843][INFO][pretrain_gpt_alcf:198] - args.iteration=2: data['text'][:10]=tensor([[29889,    13,  4806,  ...,  3086, 26040,  9220],
          [  293,  7207,   355,  ..., 18131,   520,  1247],
          [ 8619, 29889, 29871,  ...,   304, 10029,   266],
          [  363, 15202, 29892,  ...,   482, 17162, 19104]])
    [2024-09-16 22:59:29,443] [INFO] [profiler.py:81:start_profile] Flops profiler started
    [2024-09-16 22:59:30.313403][INFO][pretrain_gpt_alcf:198] - args.iteration=2: data['text'][:10]=tensor([[25561,   411,   278,  ...,   297,  2898, 26163],
          [22574,  2607, 18134,  ...,    13,  4706,   500],
          [20190, 24820,  1623,  ...,   310,   901, 29892],
          [29892,  1951,  4486,  ...,   869,   887, 30010]])
    [2024-09-16 22:59:30,316] [INFO] [profiler.py:81:start_profile] Flops profiler started
    [2024-09-16 22:59:31.185339][INFO][pretrain_gpt_alcf:198] - args.iteration=2: data['text'][:10]=tensor([[ 5371, 22417, 29892,  ...,    13,  6716,   901],
          [  353,  1565, 29936,  ..., 29878,  3567,  7196],
          [17296,   338,  1985,  ...,  3741,  9089,   422],
          [  694, 13331,   310,  ..., 21180, 29892,   607]])
    [2024-09-16 22:59:31,188] [INFO] [profiler.py:81:start_profile] Flops profiler started
    [2024-09-16 22:59:32.057207][INFO][pretrain_gpt_alcf:198] - args.iteration=2: data['text'][:10]=tensor([[  292,  8818,   267,  ..., 29892, 11275,  7407],
          [ 1870, 29897,    13,  ...,  2697, 29901,    13],
          [29913,   338,   263,  ..., 29892,   591,  3394],
          [ 2253,   472,  1554,  ...,   982,   304,   376]])
    [2024-09-16 22:59:32,060] [INFO] [profiler.py:81:start_profile] Flops profiler started
    [2024-09-16 22:59:32.930293][INFO][pretrain_gpt_alcf:198] - args.iteration=2: data['text'][:10]=tensor([[  391,  2598, 29883,  ..., 22629,   346,   440],
          [29871, 29896, 29906,  ...,   407,   583,  2833],
          [ 4262,  1836,    13,  ...,   310,   263, 10608],
          [ 1199,   411, 24770,  ...,   272,  2153, 29889]])
    [2024-09-16 22:59:32,932] [INFO] [profiler.py:81:start_profile] Flops profiler started
    [2024-09-16 22:59:33.803567][INFO][pretrain_gpt_alcf:198] - args.iteration=2: data['text'][:10]=tensor([[  620, 20503,   428,  ...,   297,  1009,  9443],
          [  950, 25078,   892,  ...,   408, 10636,   284],
          [ 1012,  2003,   364,  ...,  7313, 29912, 19303],
          [29906, 29892, 29945,  ...,   967, 26414,   472]])
    ```

  - Iteration 3:

    ``` bash
    [2024-09-16 22:59:34.881265][INFO][training_log:661] -  iteration=       3/  635782 | consumed_samples=        2304 | consumed_tokens=     9437184 | elapsed_time_per_iteration_ms=7136.5 | learning_rate=2.83116e-08 | global_batch_size=768 | lm loss=11.164038 | loss_scale=1.0 | grad_norm=6.279 | actual_seqlen= 4096 | number_of_skipped_iterations=  0 | number_of_nan_iterations=  0 | samples_per_second=107.615 | tokens_per_gpu_per_second_tgs=18366.372 | [LM]TFLOPs=84.12 | [DS]TFLOPs=108.46 |
    (min, max) time across ranks (ms):
      forward-backward ...............................: (7078.48, 7079.28)
      optimizer ......................................: (38.62, 43.08)
    [2024-09-16 22:59:34.888870][INFO][pretrain_gpt_alcf:198] - args.iteration=3: data['text'][:10]=tensor([[  496,   313, 29941,  ...,  1316,   408,  4857],
          [29899,  3204, 29889,  ...,  1074,   330,  2547],
          [29916, 29900, 29946,  ..., 18455, 29889,  4002],
          [26406,   338,  1641,  ...,   670,  1914,  6900]])
    [2024-09-16 22:59:35.719630][INFO][pretrain_gpt_alcf:198] - args.iteration=3: data['text'][:10]=tensor([[29945, 29900,   867,  ...,  7601, 12091,   310],
          [  975, 29871, 29896,  ...,  3573,   825,   306],
          [29906, 29900,  4638,  ..., 29227, 23145, 29892],
          [  278, 14368,   322,  ..., 14909, 29936, 25913]])
    [2024-09-16 22:59:36.591343][INFO][pretrain_gpt_alcf:198] - args.iteration=3: data['text'][:10]=tensor([[  988,   306,  1033,  ...,   437,   408,  1532],
          [  450, 10317,   310,  ...,   322,   752, 13036],
          [11405,  8020, 29889,  ...,   471, 18096,   287],
          [  288,  3594, 19284,  ...,   910,   338,   385]])
    [2024-09-16 22:59:37.463941][INFO][pretrain_gpt_alcf:198] - args.iteration=3: data['text'][:10]=tensor([[  322, 15151, 29946,  ..., 11648,  1497, 29889],
          [24233,   362,   467,  ...,  4513,  1353,   322],
          [ 3311, 13605, 29912,  ...,   945, 29899,  4181],
          [ 1951,   366,   508,  ...,  6589,   491,   777]])
    [2024-09-16 22:59:38.343307][INFO][pretrain_gpt_alcf:198] - args.iteration=3: data['text'][:10]=tensor([[29889, 29900,    13,  ...,  6017,   424,  1711],
          [  297,  5500,  1489,  ...,   310,  3802,  7875],
          [ 8078,  5314,   515,  ...,   373,   278,  6991],
          [13763,  6204,  6359,  ...,  4706,  2024,  1347]])
    [2024-09-16 22:59:39.214871][INFO][pretrain_gpt_alcf:198] - args.iteration=3: data['text'][:10]=tensor([[   13,  4806,  3512,  ...,   278,  7824,  6438],
          [ 2294,   938,   903,  ...,  4537,  3047,   449],
          [ 1230,  4123,   767,  ...,   310,   963, 21003],
          [ 1152,  2319, 10365,  ...,   367, 14040,   363]])
    [2024-09-16 22:59:40.085368][INFO][utils:326] -  >> building dataset for /flare/Aurora_deployment/AuroraGPT/datasets/dolma/data_v1.7_Llama2Tokenizer/tulu_flan-0000_text_document
    [2024-09-16 22:59:40.086224][INFO][utils:326] -  > building dataset index ...
    [2024-09-16 22:59:40.085475][INFO][pretrain_gpt_alcf:198] - args.iteration=3: data['text'][:10]=tensor([[27297, 29924,   801,  ..., 28947, 29892,   470],
          [12542,  5568,   703,  ...,   426,    13,  4706],
          [ 6907,   800,   322,  ..., 29892,  1661, 30304],
          [29900, 13630,   293,  ..., 26552,   363,   975]])
    [2024-09-16 22:59:40.958071][INFO][pretrain_gpt_alcf:198] - args.iteration=3: data['text'][:10]=tensor([[   13, 29946, 29953,  ..., 29953, 29945, 29871],
          [ 1283, 16578,  1156,  ...,   408,  2215,   408],
          [29906,  4229,  7671,  ...,    13,  1576,  1014],
          [  526,  2898,   304,  ...,   471,  4802, 29991]])
    ```

  - Iteration 4:

    ``` bash
    [2024-09-16 22:59:42.028460][INFO][training_log:661] -  iteration=       4/  635782 | consumed_samples=        3072 | consumed_tokens=    12582912 | elapsed_time_per_iteration_ms=7147.0 | learning_rate=3.77488e-08 | global_batch_size=768 | lm loss=11.171233 | loss_scale=1.0 | grad_norm=6.272 | actual_seqlen= 4096 | number_of_skipped_iterations=  0 | number_of_nan_iterations=  0 | samples_per_second=107.458 | tokens_per_gpu_per_second_tgs=18339.524 | [LM]TFLOPs=84.00 | [DS]TFLOPs=108.31 |
    (min, max) time across ranks (ms):
      forward-backward ...............................: (7091.77, 7092.56)
      optimizer ......................................: (39.21, 40.11)
    [2024-09-16 22:59:42.035716][INFO][pretrain_gpt_alcf:198] - args.iteration=4: data['text'][:10]=tensor([[  443,   666, 10170,  ...,   278,   619,  7323],
          [   13, 11008,   338,  ...,  2472,   363, 22049],
          [29871,    13, 29938,  ..., 29962,  8521, 29896],
          [ 1165,  2280,   304,  ...,   306,   471,  2086]])
    [2024-09-16 22:59:42.860756][INFO][pretrain_gpt_alcf:198] - args.iteration=4: data['text'][:10]=tensor([[  304,   679,   304,  ...,  1475, 29889,  1570],
          [ 2184, 29936,    13,  ...,  4706,   970,  1780],
          [29872,   352, 29901,  ..., 29905,  4915, 29912],
          [16809,   304,  1438,  ..., 13457, 29889,    13]])
    [2024-09-16 22:59:43.731208][INFO][pretrain_gpt_alcf:198] - args.iteration=4: data['text'][:10]=tensor([[12015, 29901, 20549,  ...,   322, 10752, 17906],
          [  372, 30010, 29879,  ..., 29892, 14595,   653],
          [18280, 29958,    13,  ..., 18884,   736,  6251],
          [29889,    13,    13,  ...,   599,   373, 17097]])
    [2024-09-16 22:59:44.605047][INFO][pretrain_gpt_alcf:198] - args.iteration=4: data['text'][:10]=tensor([[ 3367,   567,   964,  ...,  3353, 24870,  2181],
          [ 1262,  2609,   367,  ..., 29974, 29896,  7570],
          [29871, 29941, 29900,  ...,   341,   555,   265],
          [ 4225,   526,  6041,  ...,  1925,  1623,  2748]])
    [2024-09-16 22:59:45.479433][INFO][pretrain_gpt_alcf:198] - args.iteration=4: data['text'][:10]=tensor([[ 2283, 10162,  1496,  ..., 30656, 30317, 30605],
          [29879,  9228,   292,  ...,  7968, 29899,  7052],
          [  884,   599,   367,  ..., 29892,   278,  6054],
          [29879,   411,   278,  ...,   367,  5019,  1183]])
    [2024-09-16 22:59:46.351707][INFO][pretrain_gpt_alcf:198] - args.iteration=4: data['text'][:10]=tensor([[ 3867,   281,   761,  ..., 11949,   338,  4922],
          [  297,  1432,  2586,  ...,  5414,   278, 29811],
          [29892,   278, 15562,  ..., 10296,   310,   394],
          [ 1451,  2960,  3505,  ...,   657, 14346,  8003]])
    [2024-09-16 22:59:47.222016][INFO][pretrain_gpt_alcf:198] - args.iteration=4: data['text'][:10]=tensor([[ 6601,  2874,   414,  ...,   302,   317,  7390],
          [16415,   297,  5146,  ...,   763,   372,   471],
          [29941, 29906,  1118,  ..., 29900, 29889, 29953],
          [ 4893,   304,  4808,  ...,  2284,  2164, 18690]])
    [2024-09-16 22:59:48.094752][INFO][pretrain_gpt_alcf:198] - args.iteration=4: data['text'][:10]=tensor([[  901,   310,  2994,  ..., 29873,  1641,   766],
          [  304,  1716,  2562,  ...,  3489,   304,   367],
          [ 1949,  6736, 29871,  ..., 29965, 29909,   353],
          [   13,    13, 29930,  ..., 16497,   316,   474]])
    ```

  </details>

- Skipping steps `[2, 3]`:

  <details is-="accordion">

  <summary><code>tokens</code>:</summary>

  - Iteration 0:

    ``` bash
    [2024-09-16 23:08:47.749839][INFO][pretrain_gpt_alcf:198] - args.iteration=0: data['text'][:10]=tensor([[  304,  7344,  5146,  ...,  9776, 29914, 26419],
      [29889,    13,  4706,  ...,  9280, 30004,    13],
      [29943, 20774, 29908,  ...,   304, 27391,   322],
      [ 2645,   445, 29871,  ..., 16888,  4656, 10070]])
    [2024-09-16 23:08:51.451183][INFO][pretrain_gpt_alcf:198] - args.iteration=0: data['text'][:10]=tensor([[ 2768,   596,  1788,  ..., 27274,   393, 30010],
      [  278,  5613,  4192,  ...,   362,   310,  1950],
      [28038, 29892,  2022,  ...,  3160,   278,  2087],
      [ 4149,   907, 29888,  ..., 29896, 29892, 29896]])
    [2024-09-16 23:08:54.073597][INFO][pretrain_gpt_alcf:198] - args.iteration=0: data['text'][:10]=tensor([[  424,   322, 16232,  ...,   366,   748,   467],
      [   13,   462,  1678,  ...,  2084, 29892,  3497],
      [ 7562,   310, 19320,  ...,  8973, 22684,   358],
      [ 2089,  3633,   292,  ..., 13774,   269,  2375]])
    [2024-09-16 23:08:56.212476][INFO][pretrain_gpt_alcf:198] - args.iteration=0: data['text'][:10]=tensor([[21411,   322,  3896,  ...,  2610, 29889,   319],
      [ 8003, 29898, 29900,  ...,    12,  6658,   529],
      [  278,  4148,   310,  ...,   263, 12212,   282],
      [ 5977, 29871, 29906,  ..., 15332,   310,  1749]])
    [2024-09-16 23:08:57.207940][INFO][pretrain_gpt_alcf:198] - args.iteration=0: data['text'][:10]=tensor([[  278,  1473, 24987,  ...,   263,  2217,  3804],
      [ 2973,   263, 18778,  ...,   263,  4642,  6673],
      [  309,   323,   804,  ...,  1063, 15296,   327],
      [  278,  5864,   322,  ...,  9409, 29889,  2178]])
    [2024-09-16 23:08:58.083935][INFO][pretrain_gpt_alcf:198] - args.iteration=0: data['text'][:10]=tensor([[29892, 13731,  6617,  ..., 29871, 29896, 29946],
      [ 2892,  1012,  1266,  ...,  4036,  7512,  2068],
      [ 1473,  1556,  3619,  ...,  3762,   338,   263],
      [23353, 29918,  2177,  ...,   501,   567,   814]])
    [2024-09-16 23:08:58.951793][INFO][pretrain_gpt_alcf:198] - args.iteration=0: data['text'][:10]=tensor([[ 5400, 14378,  4768,  ...,  2107, 18677, 29889],
      [ 9200, 29887, 29914,  ...,   293, 24235,   322],
      [30143,  4746,  2184,  ..., 11891, 29974, 25760],
      [19263, 29914,   303,  ...,   358, 29889,    13]])
    [2024-09-16 23:08:59.820234][INFO][pretrain_gpt_alcf:198] - args.iteration=0: data['text'][:10]=tensor([[  309,  1306,   681,  ...,   310, 23186, 21809],
      [29896, 29929,    13,  ..., 29871, 29900,    13],
      [ 9558,   964,   263,  ...,   322,   282,   682],
      [  278, 23904, 21767,  ...,   313, 29929, 29889]])
    ```

  - Iteration 1:

    ``` bash
    [2024-09-16 23:09:08.943867][INFO][training_log:661] -  iteration=       1/  635782 | consumed_samples=         768 | consumed_tokens=     3145728 | elapsed_time_per_iteration_ms=21224.4 | learning_rate=9.4372e-09 | global_batch_size=768 | lm loss=11.167250 | loss_scale=1.0 | grad_norm=6.363 | actual_seqlen= 4096 | number_of_skipped_iterations=  0 | number_of_nan_iterations=  0 | samples_per_second=36.185 | tokens_per_gpu_per_second_tgs=6175.523 | [LM]TFLOPs=28.29 | [DS]TFLOPs=36.47 |
    [2024-09-16 23:09:08.953432][INFO][training:1083] - iteration=1 [0/8]: (torch.Size([4, 4097]))
    _tokens[:10]=tensor([[ 1472, 29892,   408,  ..., 29892,  1584,   363],
      [  967, 19475,  6593,  ...,  8093, 29899, 11249],
      [ 1006,  2218, 13326,  ...,  2355,  1304,   304],
      [29900, 29916, 29947,  ...,   353,  1870, 29936]])
    [2024-09-16 23:09:08.957524][INFO][training:1083] - iteration=1 [1/8]: (torch.Size([4, 4097]))
    _tokens[:10]=tensor([[ 2354,   274,  1041,  ..., 29892, 13049,  9098],
      [ 8798,  9547, 10353,  ...,   303,  3143, 29889],
      [ 1373,  4056,  7236,  ...,  3186,   297,  5837],
      [ 1738, 29920,  7355,  ...,    13, 29871,  3776]])
    [2024-09-16 23:09:08.966648][INFO][training:1083] - iteration=1 [2/8]: (torch.Size([4, 4097]))
    _tokens[:10]=tensor([[ 2020,   306,  1016,  ...,   322,   920,   372],
      [ 5921,  1749,  7306,  ..., 19252,   297,  5664],
      [  970,   770, 28547,  ...,   970,   894,  2577],
      [ 1907,   363, 14188,  ...,   756,  3646,   287]])
    [2024-09-16 23:09:08.969989][INFO][training:1083] - iteration=1 [3/8]: (torch.Size([4, 4097]))
    _tokens[:10]=tensor([[  715, 25392,  3104,  ...,   289,  5761,   616],
      [  426,    13,  9651,  ...,  9651,  1815, 22603],
      [ 7714,  1213,    13,  ...,    13, 29876,   457],
      [29889, 28663,  1230,  ...,  1546,   278,  6586]])
    [2024-09-16 23:09:08.990736][INFO][training:1083] - iteration=1 [4/8]: (torch.Size([4, 4097]))
    _tokens[:10]=tensor([[29929,    13,    13,  ..., 10739,  4770, 11277],
      [ 4528,   304,  2367,  ...,  2501,   385,  4203],
      [  869,   319,   794,  ...,  3158, 29889,  3115],
      [  592,   260,  4125,  ...,   284,  1135, 18655]])
    [2024-09-16 23:09:08.993101][INFO][training:1083] - iteration=1 [5/8]: (torch.Size([4, 4097]))
    _tokens[:10]=tensor([[14338, 25323,  3321,  ...,  5607,  1806,  1164],
      [  322,   278, 15352,  ...,  6462,   313,  1552],
      [25738,   714, 29889,  ..., 29915, 29879, 24842],
      [ 5122,   399, 29889,  ..., 29947,  7284,  2305]])
    [2024-09-16 23:09:09.036896][INFO][training:1083] - iteration=1 [6/8]: (torch.Size([4, 4097]))
    _tokens[:10]=tensor([[  367, 19310,  1891,  ...,  2408,   292,   263],
      [  470,  3307,  5713,  ...,   568,  2594, 19385],
      [29953, 29905,  1631,  ...,  1118,   343, 29897],
      [10261,   373,  5490,  ...,   511,   297,  1760]])
    [2024-09-16 23:09:09.039401][INFO][training:1083] - iteration=1 [7/8]: (torch.Size([4, 4097]))
    _tokens[:10]=tensor([[ 1006,   326, 29901,  ..., 14834,  6694,  9595],
      [12058,  5446, 29892,  ..., 29889,  8246,  3310],
      [ 7483,   310,   278,  ...,   402,  9851,  4423],
      [ 8041,   813,   322,  ...,  3303,  3900,   393]])
    ```

  - Iteration 2:

    ``` bash
    [2024-09-16 23:09:09.050766][INFO][training_log:661] -  iteration=       2/  635782 | consumed_samples=        1536 | consumed_tokens=     6291456 | elapsed_time_per_iteration_ms=106.8 | learning_rate=1.88744e-08 | global_batch_size=  768 | loss_scale=1.0 | grad_norm=6.363 | actual_seqlen= 4096 | number_of_skipped_iterations=  1 | number_of_nan_iterations=  0 | samples_per_second=7190.781 | tokens_per_gpu_per_second_tgs=1227226.651 | [LM]TFLOPs=5620.92 | [DS]TFLOPs=7247.49 |
    [2024-09-16 23:09:09.055864][INFO][training:1069] - Caught 3 in 'ranges_to_skip', skipping!
    [2024-09-16 23:09:09.057929][INFO][training:1082] - torch.Size([4, 4097]), len(train_data_iterator)=490723200
    [2024-09-16 23:09:09.059118][INFO][training:1083] - iteration=2 [0/8]: (torch.Size([4, 4097]))
    _tokens[:10]=tensor([[ 1858,  3851, 29889,  ...,   500,    13,    13],
    [  349,  6156,  1650,  ...,  5806, 28557,  3519],
    [16554,   304,  1653,  ...,   322,  6934, 14722],
    [ 4955,   310, 10465,  ...,  1438,  3841, 29892]])
    [2024-09-16 23:09:09.061999][INFO][training:1083] - iteration=2 [1/8]: (torch.Size([4, 4097]))
    _tokens[:10]=tensor([[  363,  1302, 16453,  ...,  7967, 29891,   484],
    [  367,   766,  4752,  ...,     1, 29871, 30143],
    [29899,   855,  1503,  ...,  3786, 29892,  5100],
    [  465,  1974,   289,  ..., 21588,   533,   304]])
    [2024-09-16 23:09:09.065494][INFO][training:1083] - iteration=2 [2/8]: (torch.Size([4, 4097]))
    _tokens[:10]=tensor([[29889,    13,  4806,  ...,  3086, 26040,  9220],
    [  293,  7207,   355,  ..., 18131,   520,  1247],
    [ 8619, 29889, 29871,  ...,   304, 10029,   266],
    [  363, 15202, 29892,  ...,   482, 17162, 19104]])
    [2024-09-16 23:09:09.069035][INFO][training:1083] - iteration=2 [3/8]: (torch.Size([4, 4097]))
    _tokens[:10]=tensor([[25561,   411,   278,  ...,   297,  2898, 26163],
    [22574,  2607, 18134,  ...,    13,  4706,   500],
    [20190, 24820,  1623,  ...,   310,   901, 29892],
    [29892,  1951,  4486,  ...,   869,   887, 30010]])
    [2024-09-16 23:09:09.072577][INFO][training:1083] - iteration=2 [4/8]: (torch.Size([4, 4097]))
    _tokens[:10]=tensor([[ 5371, 22417, 29892,  ...,    13,  6716,   901],
    [  353,  1565, 29936,  ..., 29878,  3567,  7196],
    [17296,   338,  1985,  ...,  3741,  9089,   422],
    [  694, 13331,   310,  ..., 21180, 29892,   607]])
    [2024-09-16 23:09:09.075789][INFO][training:1083] - iteration=2 [5/8]: (torch.Size([4, 4097]))
    _tokens[:10]=tensor([[  292,  8818,   267,  ..., 29892, 11275,  7407],
    [ 1870, 29897,    13,  ...,  2697, 29901,    13],
    [29913,   338,   263,  ..., 29892,   591,  3394],
    [ 2253,   472,  1554,  ...,   982,   304,   376]])
    [2024-09-16 23:09:09.079052][INFO][training:1083] - iteration=2 [6/8]: (torch.Size([4, 4097]))
    _tokens[:10]=tensor([[  391,  2598, 29883,  ..., 22629,   346,   440],
    [29871, 29896, 29906,  ...,   407,   583,  2833],
    [ 4262,  1836,    13,  ...,   310,   263, 10608],
    [ 1199,   411, 24770,  ...,   272,  2153, 29889]])
    [2024-09-16 23:09:09.082739][INFO][training:1083] - iteration=2 [7/8]: (torch.Size([4, 4097]))
    _tokens[:10]=tensor([[  620, 20503,   428,  ...,   297,  1009,  9443],
    [  950, 25078,   892,  ...,   408, 10636,   284],
    [ 1012,  2003,   364,  ...,  7313, 29912, 19303],
    [29906, 29892, 29945,  ...,   967, 26414,   472]])
    ```

  - Iteration 3:

    ``` bash
    [2024-09-16 23:09:09.135651][INFO][training_log:661] - iteration= 3/ 635782 | consumed_samples= 2304 | consumed_tokens= 9437184 | elapsed_time_per_iteration_ms=84.7 | learning_rate=2.83116e-08 | global_batch_size= 768 | loss_scale=1.0 | grad_norm=6.363 | actual_seqlen= 4096 | number_of_skipped_iterations= 1 | number_of_nan_iterations= 0 | samples_per_second=9070.783 | tokens_per_gpu_per_second_tgs=1548080.271 | [LM]TFLOPs=7090.49 | [DS]TFLOPs=9142.31 |
    [2024-09-16 23:09:09.143511][INFO][pretrain_gpt_alcf:198] - args.iteration=3: data['text'][:10]=tensor([[496,   313, 29941,  ...,  1316,   408,  4857],
      [29899,  3204, 29889,  ...,  1074,   330,  2547],
      [29916, 29900, 29946,  ..., 18455, 29889,  4002],
      [26406,   338,  1641,  ...,   670,  1914,  6900]])
    [2024-09-16 23:09:09.971988][INFO][pretrain_gpt_alcf:198] - args.iteration=3: data['text'][:10]=tensor([[29945, 29900,   867,  ...,  7601, 12091,   310],
      [  975, 29871, 29896,  ...,  3573,   825,   306],
      [29906, 29900,  4638,  ..., 29227, 23145, 29892],
      [  278, 14368,   322,  ..., 14909, 29936, 25913]])
    [2024-09-16 23:09:10.843966][INFO][pretrain_gpt_alcf:198] - args.iteration=3: data['text'][:10]=tensor([[988,   306,  1033,  ...,   437,   408,  1532],
      [  450, 10317,   310,  ...,   322,   752, 13036],
      [11405,  8020, 29889,  ...,   471, 18096,   287],
      [  288,  3594, 19284,  ...,   910,   338,   385]])
    [2024-09-16 23:09:11.715513][INFO][pretrain_gpt_alcf:198] - args.iteration=3: data['text'][:10]=tensor([[322, 15151, 29946,  ..., 11648,  1497, 29889],
      [24233,   362,   467,  ...,  4513,  1353,   322],
      [ 3311, 13605, 29912,  ...,   945, 29899,  4181],
      [ 1951,   366,   508,  ...,  6589,   491,   777]])
    [2024-09-16 23:09:12.584136][INFO][pretrain_gpt_alcf:198] - args.iteration=3: data['text'][:10]=tensor([[29889, 29900,    13,  ...,  6017,   424,  1711],
      [  297,  5500,  1489,  ...,   310,  3802,  7875],
      [ 8078,  5314,   515,  ...,   373,   278,  6991],
      [13763,  6204,  6359,  ...,  4706,  2024,  1347]])
    [2024-09-16 23:09:13.450767][INFO][pretrain_gpt_alcf:198] - args.iteration=3: data['text'][:10]=tensor([[13,  4806,  3512,  ...,   278,  7824,  6438],
      [ 2294,   938,   903,  ...,  4537,  3047,   449],
      [ 1230,  4123,   767,  ...,   310,   963, 21003],
      [ 1152,  2319, 10365,  ...,   367, 14040,   363]])
    [2024-09-16 23:09:14.317517][INFO][pretrain_gpt_alcf:198] - args.iteration=3: data['text'][:10]=tensor([[27297, 29924,   801,  ..., 28947, 29892,   470],
      [12542,  5568,   703,  ...,   426,    13,  4706],
      [ 6907,   800,   322,  ..., 29892,  1661, 30304],
      [29900, 13630,   293,  ..., 26552,   363,   975]])
    [2024-09-16 23:09:15.187191][INFO][pretrain_gpt_alcf:198] - args.iteration=3: data['text'][:10]=tensor([[13, 29946, 29953,  ..., 29953, 29945, 29871],
      [ 1283, 16578,  1156,  ...,   408,  2215,   408],
      [29906,  4229,  7671,  ...,    13,  1576,  1014],
      [  526,  2898,   304,  ...,   471,  4802, 29991]])
    ```

  </details>