The simplest, fastest repository for training/finetuning medium-sized GPTs. It is a rewrite of [minGPT](https://github.com/karpathy/minGPT) that prioritizes teeth over education. Still under active development, but currently the file `train.py` reproduces GPT-2 (124M) on OpenWebText, running on a single 8XA100 40GB node in about 4 days of training. The code itself is plain and readable: `train.py` is a ~300-line boilerplate training loop and `model.py` a ~300-line GPT model definition, which can optionally load the GPT-2 weights from OpenAI. That's it.
Because the code is so simple, it is very easy to hack to your needs, train new models from scratch, or finetune pretrained checkpoints (e.g. biggest one currently available as a starting point would be the GPT-2 1.3B model from OpenAI).
To download and tokenize the [OpenWebText](https://huggingface.co/datasets/openwebtext) dataset. This will create a `train.bin` and `val.bin` which holds the GPT2 BPE token ids in one sequence, stored as raw uint16 bytes. Then we're ready to kick off training. The training script currently by default tries to reproduce the smallest GPT-2 released by OpenAI, i.e. the 124M version of GPT-2. We can train as follows on a single device, though I encourage you to read the code and see all of the settings and paths up top in the file:
If you do not have GPU also add `--device=cpu --compile=False`, though you'd have to also adjust the default network size to be much much smaller (see "i only have a macbook" section below). To train using PyTorch Distributed Data Parallel (DDP) run the script with torchrun. For example to train on a node with 4 GPUs run:
It is a good idea to benchmark your interconnect (e.g. iperf3). In particular, if you don't have Infiniband then also prepend `NCCL_IB_DISABLE=1` to the above launches. Your multinode training will work, but most likely _crawl_.
By default checkpoints are periodically written to the `--out_dir` (`./out` by default). Once we have one, we can sample from the model:
However, we have to note that GPT-2 was trained on (closed, never released) WebText, while OpenWebText is just a best-effort open reproduction of this dataset. This means there is a dataset domain gap. Indeed, taking the GPT-2 (124M) checkpoint and finetuning on OWT directly for a while reaches loss down to ~2.9. This then becomes the more appropriate baseline w.r.t. reproduction.
For an example of how to finetune a GPT on new text go to `data/shakespeare` and look at `prepare.py` to download the tiny shakespeare dataset and render it into a `train.bin` and `val.bin`. Unlike OpenWebText this will run in seconds. Finetuning takes very little time, e.g. on a single GPU just a few minutes. Run an example finetuning like:
This will load the config parameter overrides in `config/finetune_shakespeare.py` (I didn't tune them much though). Basically, we initialize from a GPT2 checkpoint with `init_from` and train as normal, except shorter and with a small learning rate. The best checkpoint (lowest validation loss) will be in the `out_dir` directory, e.g. in `out-shakespeare` by default, per the config file. You can then run the code in `sample.py` to generate infinite Shakespeare. Note that you'll have to edit it to point to the correct `out_dir`.
It's possible to play with the code if you only have a macbook or some other cheap computer. In this case it's much easier to just work with the Shakespeare dataset. Step 1 render the training data:
```
$ cd data/shakespeare
$ python prepare.py
```
Then launch the training script with a baby network, here is an example:
This creates a much smaller Transformer (4 layers, 4 heads, 64 embedding size), runs only on CPU, does not torch.compile the model (torch seems to give an error if you try), only evaluates for one iteration so you can see the training loop at work immediately, and also makes sure the context length is much smaller (e.g. 64 tokens), and the batch size is reduced to 8. On my MacBook Air (M1) this takes about 400ms per iteration. The network is still pretty expensive because the current vocabulary is hard-coded to be the GPT-2 BPE encodings of `vocab_size=50257`. So the embeddings table and the last layer are still massive.
You can now also work with tiny shakespeare on the character level, see `data/shakespeare_char` and run `prepare.py` to tokenize it on the character level. If you have a GPU you can use the decent starter settings in a provided config file, train as follows:
Finally, on Apple Silicon Macbooks you can use device `--device mps` ("Metal Performance Shaders"), which can significantly accelerate training (2-3X). You will need a specific version of PyTorch. See [Issue 28](https://github.com/karpathy/nanoGPT/issues/28).
For model benchmarking `bench.py` might be useful. It's identical to what happens in the meat of the training loop of `train.py`, but omits much of the other complexities.
Code by default now uses [PyTorch 2.0](https://pytorch.org/get-started/pytorch-2.0/). At the time of writing (Dec 29, 2022) this makes `torch.compile()` available in the nightly release. The improvement from the one line of code is noticeable, e.g. cutting down iteration time from ~250ms / iter to 135ms / iter. Nice work PyTorch team!
- I am still not 100% confident that my GPT-2 small reproduction hyperparameters are good, if someone has reproduced GPT-2 I'd be eager to exchange notes ty
- Actually reproduce GPT-2 results and have clean configs that reproduce the result. It was estimated ~3 years ago that the training cost of 1.5B model was ~$50K (?). Sounds a bit too high.
- Note that by default this repo uses PyTorch 2.0 (i.e. `torch.compile`). This is fairly new and experimental, and not yet available on all platforms (e.g. Windows). If you're running into related error messages try to disable this by adding `--compile=False` flag. This will slow down the code but at least it will run.
