buildNanoGPT 0.0.8

A template for nbdev-based project

buildNanoGPT

buildNanoGPT is developed based on Andrej Karpathy’s build-nanoGPT repo and Let’s reproduce GPT-2 (124M) with added notes and details for teaching purposes using nbdev, which enables package development, testing, documentation, and dissemination all in one place - Jupyter Notebook or Visual Studio Code Jupyter Notebook in my case 😄.

Literate Programming

buildNanoGPT

flowchart LR
  A(Andrej's build-nanoGPT) --> C((Combination))
  B(Jeremy's nbdev) --> C
  C -->|Literate Programming| D(buildNanoGPT)

micrograd2023

Disclaimers

buildNanoGPT is written based on Andrej Karpathy’s build-nanoGPT and his “Neural Networks: Zero to Hero” lecture series. Andrej is the man who needs no introduction in the field of Deep Learning.

Andrej released a series of lectures called Neural Network: Zero to Hero, which I found extremely educational and practical. I am reviewing the lectures and creating notes for myself and for teaching purposes.

I developed buildNanoGPT using nbdev, which was developed by Jeremy Howard, the man who also needs no introduction in the field of Deep Learning. Jeremy also created fastai Deep Learning software library and Courses that are extremely influential. I highly recommend fastai if you are interested in starting your journey and learning with ML and DL.

nbdev is a powerful tool that can be used to efficiently develop, build, test, document, and distribute software packages all in one place, Jupyter Notebook or Jupyter Notebooks in VS Code, which I am using.

If you study lectures by Andrej and Jeremy you will probably notice that they are both great educators and utilize both top-down and bottom-up approaches in their teaching, but Andrej predominantly uses bottom-up approach while Jeremy predominantly uses top-down one. I personally fascinated by both educators and found values from both of them and hope you are too!

Usage

Prepare FineWeb-Edu-10B data

from buildNanoGPT import data
import tiktoken
import numpy as np

enc = tiktoken.get_encoding("gpt2")
eot = enc._special_tokens['<|endoftext|>'] # end of text token
eot

50256

t_ref = [eot]
t_ref.extend(enc.encode("Hello, world!"))
t_ref = np.array(t_ref).astype(np.uint16)
t_ref

array([50256, 15496,    11,   995,     0], dtype=uint16)

t_ref = [eot]
t_ref.extend(enc.encode("Hello, world!"))
t_ref = np.array(t_ref).astype(np.int32)
t_ref

array([50256, 15496,    11,   995,     0], dtype=int32)

doc = {"text":"Hello, world!"}
t_test = data.tokenize(doc)
t_test

array([50256, 15496,    11,   995,     0], dtype=uint16)

assert np.all(t_ref == t_test)

# Download and Prepare the FineWeb-Edu-10B sample Data
data.edu_fineweb10B_prep(is_test=True)

Resolving data files:   0%|          | 0/1630 [00:00<?, ?it/s]

Loading dataset shards:   0%|          | 0/98 [00:00<?, ?it/s]

'Hello from `prepare_edu_fineweb10B()`! if you want to download the dataset, set is_test=False and run again.'

Prepare HellaSwag Evaluation data

data.hellaswag_val_prep(is_test=True)

'Hello from `hellaswag_val_prep()`! if you want to download the dataset, set is_test=False and run again.'

Training

# either running 03_train.ipynb or short-cut by running train script from the buildNanoGPT package
from buildNanoGPT import train

using device: cuda
total desired batch size: 524288
=> calculated gradient accumulation steps: 32
found 99 shards for split train
found 1 shards for split val
num decayed parameter tensors: 50, with 124,354,560 parameters
num non-decayed parameter tensors: 98, with 121,344 parameters
using fused AdamW: True
validation loss: 10.9834
HellaSwag accuracy: 2534/10042=0.2523
step     0 | loss: 10.981724 | lr 6.0000e-06 | norm: 15.4339 | dt: 82809.98ms | tok/sec: 6331.22
step     1 | loss: 10.655205 | lr 1.2000e-05 | norm: 12.4931 | dt: 10492.83ms | tok/sec: 49966.29
step     2 | loss: 10.274603 | lr 1.8000e-05 | norm: 7.7501 | dt: 10522.88ms | tok/sec: 49823.61
step     3 | loss: 10.004156 | lr 2.4000e-05 | norm: 5.2698 | dt: 10481.91ms | tok/sec: 50018.35
step     4 | loss: 9.833108 | lr 3.0000e-05 | norm: 3.6179 | dt: 10495.18ms | tok/sec: 49955.14
step     5 | loss: 9.711222 | lr 3.6000e-05 | norm: 2.7871 | dt: 10484.25ms | tok/sec: 50007.21
step     6 | loss: 9.642426 | lr 4.2000e-05 | norm: 2.4048 | dt: 10679.06ms | tok/sec: 49094.97
step     7 | loss: 9.612312 | lr 4.8000e-05 | norm: 2.3183 | dt: 10555.78ms | tok/sec: 49668.32
step     8 | loss: 9.558184 | lr 5.4000e-05 | norm: 2.2464 | dt: 10685.39ms | tok/sec: 49065.86
step     9 | loss: 9.526472 | lr 6.0000e-05 | norm: 2.2171 | dt: 10548.39ms | tok/sec: 49703.14
step    10 | loss: 9.463450 | lr 6.6000e-05 | norm: 2.1546 | dt: 10559.73ms | tok/sec: 49649.78
step    11 | loss: 9.413282 | lr 7.2000e-05 | norm: 2.1401 | dt: 10495.94ms | tok/sec: 49951.49
step    12 | loss: 9.340552 | lr 7.8000e-05 | norm: 2.0149 | dt: 10668.78ms | tok/sec: 49142.26
step    13 | loss: 9.278631 | lr 8.4000e-05 | norm: 1.9368 | dt: 10605.16ms | tok/sec: 49437.05
step    14 | loss: 9.159446 | lr 9.0000e-05 | norm: 1.9737 | dt: 10701.77ms | tok/sec: 48990.76
step    15 | loss: 9.111786 | lr 9.6000e-05 | norm: 3.0525 | dt: 10732.83ms | tok/sec: 48849.00
step    16 | loss: 9.029915 | lr 1.0200e-04 | norm: 1.9619 | dt: 10790.65ms | tok/sec: 48587.23
step    17 | loss: 8.937255 | lr 1.0800e-04 | norm: 1.8786 | dt: 10621.46ms | tok/sec: 49361.22
step    18 | loss: 8.955976 | lr 1.1400e-04 | norm: 2.0179 | dt: 10545.33ms | tok/sec: 49717.53
step    19 | loss: 8.888343 | lr 1.2000e-04 | norm: 1.9142 | dt: 10598.08ms | tok/sec: 49470.11
step    20 | loss: 8.672051 | lr 1.2600e-04 | norm: 1.7543 | dt: 10730.04ms | tok/sec: 48861.68
step    21 | loss: 8.556496 | lr 1.3200e-04 | norm: 1.6246 | dt: 10822.08ms | tok/sec: 48446.13
step    22 | loss: 8.463942 | lr 1.3800e-04 | norm: 1.4898 | dt: 10733.11ms | tok/sec: 48847.72
step    23 | loss: 8.389053 | lr 1.4400e-04 | norm: 1.9412 | dt: 10555.51ms | tok/sec: 49669.61
step    24 | loss: 8.257857 | lr 1.5000e-04 | norm: 2.0539 | dt: 10732.67ms | tok/sec: 48849.75
step    25 | loss: 8.128786 | lr 1.5600e-04 | norm: 1.4269 | dt: 10609.93ms | tok/sec: 49414.84
step    26 | loss: 8.098352 | lr 1.6200e-04 | norm: 2.0206 | dt: 10487.59ms | tok/sec: 49991.30
step    27 | loss: 7.961097 | lr 1.6800e-04 | norm: 1.2978 | dt: 10578.22ms | tok/sec: 49562.95
step    28 | loss: 7.884172 | lr 1.7400e-04 | norm: 1.2289 | dt: 10497.51ms | tok/sec: 49944.04
step    29 | loss: 7.765845 | lr 1.8000e-04 | norm: 1.1969 | dt: 10724.78ms | tok/sec: 48885.65
step    30 | loss: 7.821087 | lr 1.8600e-04 | norm: 1.0228 | dt: 10792.80ms | tok/sec: 48577.58
step    31 | loss: 7.689835 | lr 1.9200e-04 | norm: 0.9216 | dt: 10752.80ms | tok/sec: 48758.30
step    32 | loss: 7.641486 | lr 1.9800e-04 | norm: 0.8666 | dt: 10985.01ms | tok/sec: 47727.58
step    33 | loss: 7.572504 | lr 2.0400e-04 | norm: 0.7996 | dt: 10684.39ms | tok/sec: 49070.46
step    34 | loss: 7.429519 | lr 2.1000e-04 | norm: 0.7874 | dt: 10696.01ms | tok/sec: 49017.15
step    35 | loss: 7.414855 | lr 2.1600e-04 | norm: 0.7272 | dt: 10580.76ms | tok/sec: 49551.08
step    36 | loss: 7.393157 | lr 2.2200e-04 | norm: 0.8536 | dt: 10748.95ms | tok/sec: 48775.74
step    37 | loss: 7.287198 | lr 2.2800e-04 | norm: 0.5487 | dt: 10921.08ms | tok/sec: 48006.98
step    38 | loss: 7.252760 | lr 2.3400e-04 | norm: 0.4738 | dt: 10716.44ms | tok/sec: 48923.69
step    39 | loss: 7.292991 | lr 2.4000e-04 | norm: 0.5769 | dt: 10659.42ms | tok/sec: 49185.43
step    40 | loss: 7.251584 | lr 2.4600e-04 | norm: 0.9509 | dt: 10570.06ms | tok/sec: 49601.22
step    41 | loss: 7.209351 | lr 2.5200e-04 | norm: 1.7773 | dt: 10611.45ms | tok/sec: 49407.78
step    42 | loss: 7.140303 | lr 2.5800e-04 | norm: 0.9441 | dt: 10753.44ms | tok/sec: 48755.36
step    43 | loss: 7.216593 | lr 2.6400e-04 | norm: 2.1513 | dt: 10632.68ms | tok/sec: 49309.09
step    44 | loss: 7.155683 | lr 2.7000e-04 | norm: 1.3599 | dt: 10780.88ms | tok/sec: 48631.27
step    45 | loss: 7.159153 | lr 2.7600e-04 | norm: 1.1990 | dt: 10722.27ms | tok/sec: 48897.11
step    46 | loss: 7.126624 | lr 2.8200e-04 | norm: 0.8272 | dt: 10791.48ms | tok/sec: 48583.50
step    47 | loss: 7.190242 | lr 2.8800e-04 | norm: 0.9578 | dt: 10718.49ms | tok/sec: 48914.35
step    48 | loss: 7.194102 | lr 2.9400e-04 | norm: 0.7273 | dt: 10651.67ms | tok/sec: 49221.22
step    49 | loss: 7.113352 | lr 3.0000e-04 | norm: 1.1239 | dt: 10732.94ms | tok/sec: 48848.51
step    50 | loss: 7.169769 | lr 3.0600e-04 | norm: 1.0528 | dt: 10706.81ms | tok/sec: 48967.72
step    51 | loss: 7.103631 | lr 3.1200e-04 | norm: 1.0537 | dt: 10826.62ms | tok/sec: 48425.82
step    52 | loss: 7.092214 | lr 3.1800e-04 | norm: 0.7355 | dt: 10777.80ms | tok/sec: 48645.18
step    53 | loss: 7.021073 | lr 3.2400e-04 | norm: 0.8493 | dt: 10907.12ms | tok/sec: 48068.41
step    54 | loss: 7.030515 | lr 3.3000e-04 | norm: 0.7924 | dt: 10822.94ms | tok/sec: 48442.27
step    55 | loss: 7.027347 | lr 3.3600e-04 | norm: 0.8563 | dt: 10661.62ms | tok/sec: 49175.26
step    56 | loss: 7.007086 | lr 3.4200e-04 | norm: 1.2067 | dt: 10764.39ms | tok/sec: 48705.77
step    57 | loss: 6.978011 | lr 3.4800e-04 | norm: 0.5606 | dt: 10967.17ms | tok/sec: 47805.22
step    58 | loss: 6.919628 | lr 3.5400e-04 | norm: 1.3408 | dt: 10802.21ms | tok/sec: 48535.23
step    59 | loss: 6.887385 | lr 3.6000e-04 | norm: 1.3971 | dt: 10907.45ms | tok/sec: 48066.97
step    60 | loss: 6.879627 | lr 3.6600e-04 | norm: 0.7581 | dt: 10768.36ms | tok/sec: 48687.80
step    61 | loss: 6.906055 | lr 3.7200e-04 | norm: 0.9657 | dt: 10613.11ms | tok/sec: 49400.03
step    62 | loss: 6.795964 | lr 3.7800e-04 | norm: 0.6819 | dt: 10593.62ms | tok/sec: 49490.92
step    63 | loss: 6.780255 | lr 3.8400e-04 | norm: 0.7485 | dt: 10719.51ms | tok/sec: 48909.68
step    64 | loss: 6.767306 | lr 3.9000e-04 | norm: 0.7399 | dt: 10806.62ms | tok/sec: 48515.44
step    65 | loss: 6.801779 | lr 3.9600e-04 | norm: 0.7439 | dt: 10609.56ms | tok/sec: 49416.58
step    66 | loss: 6.721136 | lr 4.0200e-04 | norm: 0.5727 | dt: 10749.83ms | tok/sec: 48771.73
step    67 | loss: 6.750595 | lr 4.0800e-04 | norm: 0.7310 | dt: 10711.53ms | tok/sec: 48946.13
step    68 | loss: 6.730660 | lr 4.1400e-04 | norm: 0.5052 | dt: 10772.71ms | tok/sec: 48668.16
step    69 | loss: 6.631037 | lr 4.2000e-04 | norm: 0.6577 | dt: 10736.56ms | tok/sec: 48832.04
step    70 | loss: 6.612390 | lr 4.2600e-04 | norm: 0.6208 | dt: 10598.25ms | tok/sec: 49469.31
step    71 | loss: 6.643014 | lr 4.3200e-04 | norm: 0.6751 | dt: 10712.97ms | tok/sec: 48939.57
step    72 | loss: 6.602534 | lr 4.3800e-04 | norm: 0.8274 | dt: 10685.25ms | tok/sec: 49066.50
step    73 | loss: 6.606695 | lr 4.4400e-04 | norm: 1.0497 | dt: 10784.33ms | tok/sec: 48615.72
step    74 | loss: 6.532132 | lr 4.5000e-04 | norm: 0.9483 | dt: 11051.53ms | tok/sec: 47440.31
step    75 | loss: 6.571723 | lr 4.5600e-04 | norm: 0.5493 | dt: 10943.98ms | tok/sec: 47906.50
step    76 | loss: 6.519442 | lr 4.6200e-04 | norm: 0.6364 | dt: 11138.90ms | tok/sec: 47068.20
step    77 | loss: 6.553431 | lr 4.6800e-04 | norm: 0.6423 | dt: 10943.91ms | tok/sec: 47906.81
step    78 | loss: 6.525961 | lr 4.7400e-04 | norm: 0.4541 | dt: 10733.66ms | tok/sec: 48845.21
step    79 | loss: 6.474160 | lr 4.8000e-04 | norm: 0.6690 | dt: 10748.03ms | tok/sec: 48779.93
step    80 | loss: 6.481711 | lr 4.8600e-04 | norm: 0.5859 | dt: 10679.49ms | tok/sec: 49093.00
step    81 | loss: 6.486966 | lr 4.9200e-04 | norm: 0.6897 | dt: 10656.78ms | tok/sec: 49197.58
step    82 | loss: 6.430150 | lr 4.9800e-04 | norm: 0.6284 | dt: 10426.83ms | tok/sec: 50282.59
step    83 | loss: 6.387268 | lr 5.0400e-04 | norm: 0.5746 | dt: 10644.15ms | tok/sec: 49255.97
step    84 | loss: 6.405340 | lr 5.1000e-04 | norm: 0.5523 | dt: 10856.28ms | tok/sec: 48293.53
step    85 | loss: 6.371199 | lr 5.1600e-04 | norm: 0.6764 | dt: 10573.15ms | tok/sec: 49586.76
step    86 | loss: 6.367082 | lr 5.2200e-04 | norm: 0.7355 | dt: 10731.52ms | tok/sec: 48854.94
step    87 | loss: 6.404164 | lr 5.2800e-04 | norm: 0.7907 | dt: 10878.82ms | tok/sec: 48193.45
step    88 | loss: 6.383866 | lr 5.3400e-04 | norm: 0.7472 | dt: 10855.23ms | tok/sec: 48298.20

How to install

The buildNanoGPT package was uploaded to PyPI and can be easily installed using the below command.

pip install buildNanoGPT

Developer install

If you want to develop buildNanoGPT yourself, please use an editable installation.

git clone https://github.com/hdocmsu/buildNanoGPT.git

pip install -e "buildNanoGPT[dev]"

You also need to use an editable installation of nbdev, fastcore, and execnb.

Happy Coding!!!

Note: buildNanoGPT is currently Work in Progress (WIP).