ReaLLMAsic aims to bridge the gap between theoretical model design and practical hardware implementation, ensuring efficient, scalable, and robust ML model development.
Our project stands out for its extensive exploration of various model configurations and modules, catering to a diverse range of use cases.
Key exploration features include:
Module Variation
: Explore with different module types -- e.g. Softmax, Softermax, ConSmax, and SigSoftmax -- discover which is best suited (PPA) to your application.Flexible Tokenization
: Explore different tokenization: tiktoken, sentencepiece, phonemization, character level, custom tokenization, etc.Diverse Dataset Performance Testing
: Evaluate model efficacy across various languages and datasets including: csv-timeseries, mathematics, music, lyrics, literature, and webtext.Standard and Custom Hyperparameters
: Fine-tune models using conventional hyperparameters and explore the impact of custom settings on model performance and PPA impacts.
Key analysis features:
Exploration scripts
: Are encapsulated into bash scripts which loop over the train.py's argparse parameters.Logging with automatic timestamps & labels
: run a suite of experiments and have the repo automatically organize and label them by timestamp and description
Hardware Related
Training with Hardware Emulation
: Implement different operations for forward and backward passes for hardware-implementation aware training.PPA Implications Analysis
: Understand the power, performance, and area (PPA) implications of different model designs, guiding efficient hardware-software integration.
This section contains installation locally with GPU acceleration.
(If you do not have a GPU, check out this colab, which has a T4 GPU runtime (at time of writing) for ML acceleration.)
We recommend creating a virtual env or conda environment before starting:
For venv:
python3 -m venv venv
source venv/bin/activate
Or for conda:
conda create -n nanogpt
conda activate nanogpt
If you are compatible with cu11.8, then use the following:
python3 -m pip install --upgrade pip
python3 -m pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
python3 -m pip install numpy transformers datasets tiktoken wandb tqdm tensorboard rich torchinfo
If unsure, visit the pytorch page and subtitute the appropriate line for the torch
installation line above: https://pytorch.org/get-started/locally/
This downloads and parses a literature dataset into train.bin
and val.bin
files.
python3 data/shakespeare_char/prepare.py
Training with a GPU is highly recommended, to do this now run (should take around 3-20 minutes depending on one's GPU):
python3 train.py --compile
Highly recommend setting --max_sample_tokens
which generates and shows outputs
at each new saved checkpoint.
python3 train.py --max_sample_tokens 100 --compile
minutes and the best validation loss is 1.4697. Based on the configuration, the
model checkpoints are being written into the --out_dir
directory, that
defaults to ./out
. So once the training finishes we can sample from the
best model by pointing the sampling script at this directory:
python3 sample.py
This generates a few samples, for example:
ANGELO:
And cowards it be strawn to my bed,
And thrust the gates of my threats,
Because he that ale away, and hang'd
An one with him.
DUKE VINCENTIO:
I thank your eyes against it.
This looks pretty good for a model which just learned how to spell from scratch. Keeping an eye on inference is very important, however, usually one can infer levels from validation losses.
The next section goes over how to do a massive exploration of different models
and quickly compare their quality using the validation loss
as a proxy.
The explorations directory is intended to be have a set of fully encapsulated replicable sweeps.
Using these, one can quickly and visually compare ultimate quality of
checkpoints created from training using validation loss
as a figure of merit.
To run the experiment create or modify an existing json file in the explorations
folder:
python3 run_experiments.py -c explorations/config.json
This will create logs in the following directories:
csv_logs/
logs/
This also saves timestamped and labelled folders within the output_dir
(which
defaults to out/
subdirectories)
Often for large explorations with run_experiments
one wants to monitor the
the best validation losses so far (a metric for how well the model does on next
token prediction on the current dataset).
The included inspect_ckpts.py
script reports the best valiation loss and
associated iteration number for all ckpt.pt files recursivel for a specified
parent directory.
Example usage:
python3 inspect_ckpts.py --directory ./out --sort loss
This can be wrapped with color via the watch command for a realtime dashboard.
For example to look at all checkpoint files in the out directory:
watch --color 'python3 inspect_ckpts.py --directory ./out --sort loss'
As with remainder of the repo, this script is provided as a base to open up for additional community contributions.
If using tensorboard for logging, we have provided a convenience script:
bash start_tensorboard.sh
You can view live validation loss updates on url: http://localhost:6006
Note: Only one tensorboard process can grab port 6006 at time, try closing other processes (e.g. other tensorboards) using this port, or choose an alternative port if new tensorboard isn't showing.
TODO: Add links and descriptions to other Readme's and Demos.
This repo is under active development and accepting PR's, please see the
See the Contributing_Features.md for details on how to add new features and explorations.
- Original nanoGPT Repo
- NanoGPT Discord Channel
- Zero To Hero series
- GPT video