This library contains plugins to accelerate finetuning with the following optimizations:
- Padding-Free Flash Attention Computation
- Multipack Distributed Sampling
| Plugin | Description | Depends | Loading | Augmentation | Callbacks |
|---|---|---|---|---|---|
| padding_free | Padding-Free Flash Attention Computation | flash_attn | ✅ | ||
| multipack sampler | Multipack Distributed Sampling | numba | ✅ |
Transformers natively supports padding-free from v4.44.0 see here. The padding-free plugin will use the transformers library if compatible,
otherwise if transformers < v4.44.0 the plugin will use an internal implementation instead.
To reproduce the benchmarks, simply run the following commands,
Reproduce Padding Free on A100 80GB
tox -e run-benches -- "1 2" "4 8" benchmark_outputs scenarios-orca.yaml "none"
Reproduce MultiPack on A100 80GB
tox -e run-benches -- "2 4 8" "16 32 64" benchmark_outputs scenarios-orca.yaml "padding-free"
The padding-free plugin currently only works with pre-tokenized datasets, this is because it is currently designed to replace
the data collator from SFTTrainer with a custom data collator to manipulate the input to the modified flash attention forward.
There are some cases, the data collator for SFTTrainer will handle the formatting and tokenization from raw text datasets. The plugin is currently unable to both handle the original data collation and apply its custom data collator over it as the same time. This issue will be addressed in a future commit to support this case.
In the meantime, the plugin expects the user to provide a pretokenized dataset that
- is formatted with a template for instruct-tuning cases
- is tokenized
- has template labels that are masked to exclude from loss computation
- has eos token appended
The multipack plugin currently also requires the padding-free plugin to work. This may change in the future if there is demand for multipack to work standalone without padding free.