There was a corner case when the shape of the predictions y of the
dataset would not be correct, due to the fact that the number of batches
was miscalculated.

This happened when `batch_len` was exactly divisible by `seq_len`, since
the predictions, which are simply the text shifted once to the right,
would not have that extra column at the end.

Fix the above issue by decrementing the number of available batches with
1 when `batch_len` exactly divides by `seq_len`.

Visual change. This only changes what the trainer reports as the final
training loss.

Not quite sure if the value before was accurate anyway, since gradient
accumulation would not let the optimizer step every batch anyway.

For a big enough dataset, this should not have any impact at all.

The final loss value will be reported based on the last calculation of
the loss, correctly taking into consideration gradient accumulation as
well.

Fix problems with some types. This enables Python's static type checks
to correctly identify some issues before runtime.

Add model constructor arguments (n_layers, n_heads, dim etc.) as pytorch
buffers. This packs them together with the model weights when calling
`torch.save()`, and loads them back in when calling `torch.load()`.

Eventually, these should be saved separately, however this will do for
now.

Visual change.

Add note in the README file about using the environment variable
CUDA_VISIBLE_DEVICES, which lets the user choose which GPU to run
training on.

The normalization layer returned float32 tensors, instead of fp16
tensors, which should've been the case when training with mixed
precision. This raised a runtime error of incompatible types. Rearrange
the operations to properly compute the norm in float32, but return the
value in fp16.

Total loss wasn't properly initialized, leading to a runtime error.

Fix tokenizer typos, add newlines

See merge request !8

Gradient accumulation

See merge request !6

32K vocab -> 16K vocab

Fix some typos in the tokenizer file.

Add newlines and whitespaces to the tokenizer model. Previously, all the
whitespace was stripped and joined into a single blank. This allows for
better tokenization for things like wikitext103, which has articles
containing newlines, with relevance.

Fix small typos in the model architecture

See merge request !7

Add gradient accumulation to the training loop. The number of gradient
accumulation steps is exposed by the trainer.

Make the gradient clipping norm value a parameter fed to the trainer.

Fix some issues with the wikitext103 dataset

See merge request !4

Add tinystories dataset

See merge request !3

Add progress bar display for training

See merge request !2

Couple of things:
- rewrite code to better check when the dataset is downloaded
- better cleanup after download + unzip
- more aggresive exit on checksum mismatch
- rewrite __main__

Couple of things, mostly for code consistency and clarity:
- reorganize imports
- reorganize initial global variables (URL, MD5 etc.)
- rename class to contain "Dataset"
- fix comments

There are also a few things which I added / replaced / removed, upon
re-consideration of how datasets should work:
- add additional folder "tinystories" where to download the .txt files
- remove the pandas DataFrame
- rewrite __main__ example
- be more aggresive when checksums for downloaded files don't match

Rewrite training loop in PyTorch

See merge request !1

Use fastai's fastprogress package to display a progress bar while
training, with useful information such as loss, estimated time of
training, current learning rate, estimated ms/batch.

Print end of epoch stats when finishing an epoch.

Add a relevant parameter for the trainer to enable/disable the progress
bar display.

Add an example of what training using the current code would look like.
Most of this script can be copied and adapted for other datasets, or for
evaluating/testing different Transformer models etc.

The model is mostly modeled after the LLama 2 transformer, though it
misses a couple of things (grouped-query attention, KV cache for
inference, and rotational encodings). These will eventually make it into
Optimus code. At that point, the model might as well be called LLoptimus.

Add Llama's 32K vocab tokenizer, as well as 2 Optimus variants trained
on WikiText103 data: a 32K vocab tokenizer, and a 60K vocab tokenizer.
Both Optimus tokenizers are unigram models.

Add WikiText103 as an example of what a Dataset needs to look like, for
us to be able to use it in the training loop. Other Dataset's can
probably directly copy most of the code, and modify small parts of it as
needed.

Add a few common functions that can be used by whatever dataset we need.

Add a training loop, written from scratch. Currently, it is quite
bare-bones (trains in FP32, no gradient accumulation, no parallel
training etc.), but eventually this will be improved with other
must-have things.

This is a custom dataloader class, similar to pytorch's DataLoader, but
specialized in NLP tasks. Right now, it is pretty much written from
scratch, but eventually we want to use the built-in DataLoader, since it
has some nice goodies attached to it (like data
prefetching/preprocesing, serving data for parallel training etc.).

Add parallel training on multiple GPU's through pytorch's
DistributedDataParallel (Pytorch DDP).

Some changes:
- label smoothing
- root-mean square norm instead of layer norm
- move norm to before layer instead of after, and add final norm layer
- remove attention for-loop, instead do a big matrix multiplication
- remove bias terms from linear layers
- add dropout
- remove and rename model parameters (easier to use in code)
- add weight tying
- add gradient accumulation (change to lower batch size and higher
  sequence length)
- add model checkpoint
- add gradient clipping
- move warmup steps to 15% of total steps, and change learning rate
  accordingly
- move to floating point 16 bits (fp16); faster training for nvidia
  GPU's
- plot final loss and learning rate scheduling