short-transformers 0.2.0

Cutting transformers layers

:scissors: Short Transformers

• Pytorch implementation of layer pruning proposed in The Unreasonable Ineffectiveness of the Deeper Layers.
• The repository reproduces and extends original methods by offering different layer pruning criteria.

Installation:

pip install short-transformers

Required additional dependencies: transformers, datasets.

Quickstart:

from short_transformers import ShortTransformer

# load from path/hf_hub
model = ShortTransformer.from_pretrained(model_name)

# or use hf model
model = ShortTransformer.from_model(hf_model)

# remove n layers, use hf dataset to find the optimal cut
short_model = model.remove_layers(block_size=5, dataset=dataset, limit=1000)

# continue training to heal after the cut
# ...

# save as hf model
short_mdoel.save_pretrained(output_path)

Example log output:

[2024-05-04 18:39:03] [short-transformers] Running inference on 1000 samples.
1000it [00:06, ?it/s]
[2024-05-04 18:39:09] [short-transformers] Inference results saved to `model.memory`.
[2024-05-04 18:39:09] [short-transformers] Choosing optimal 5-layers block to prune.
[2024-05-04 18:39:09] [short-transformers] Best 5-layers block to prune starts at layer: 23.
[2024-05-04 18:39:09] [short-transformers] Removing layers: [23, 24, 25, 26, 27]
[2024-05-04 18:39:09] [short-transformers] Changing model config to reflect changes:
        config.num_hidden_layers: 32 -> 27
        config._name_or_path: meta-llama/Meta-Llama-3-8B -> meta-llama/Meta-Llama-3-8B-27L
[2024-05-04 18:39:09] [short-transformers] Cleaning `model.memory`.

Both short_model and the saved model are fully compatible with transformers. See examples/basic.py for a complete working example.

Pruning in steps:

Pruning can composed step-by-step and customized:

1. Analyze model layers:

from datasets import load_dataset
from short_transformers import ShortTransformer
from short_transformers.graph import draw_diagram

# load from path/hf_hub
model_name = "meta-llama/Meta-Llama-3-8B"

model = ShortTransformer.from_pretrained(model_name, device_map="auto")

dataset = load_dataset("allenai/c4", "en", split="validation", streaming=True)

# run inderence on a subset of the datsets
model.analyse_layers(
    dataset=dataset,
    key="text",
    limit=100,
    max_length=1000,
)

# after this step, results will be saved to model.memory, which can be saved
model.save_memory("memory.npz")

# and visualized as seaborh graph (see examples in the README.md)
draw_diagram("memory.npz", "memory.png")

Example output:

2. Find optimal block size and start_layer:

# finding optimial block of size 'num' to prune
start_layer = model.get_optimal_cut(block_size=5)

# evaluating all possibe block sizes to prune,
# for each block returns score 0-1
# which is averaged over samples distance between input and output to/from a block
block_score = model.get_block_score_stats(return_md=True, threshold=0.3)

Example output:

Block_size	Removed_layers	Score (avg dist)
1	25-25	0.123
2	24-25	0.155
3	25-27	0.181
4	24-27	0.204
5	23-27	0.226
6	22-27	0.248
7	22-28	0.268
8	20-27	0.291

3. Pruning layers:

# pruning 5-layers block
# this will also clean the `model.memory``
model.cut(start_layer=start_layer, block_size=5)

# or clean the model memory manually
model.clean_memory()

# save the pruned model
model.save_pretrained("model_path")

See example/prune_in_steps.py for a complete working example.

Supported pruning methods:

• based on layer input/output distances:

  • angular distance of the last token (original)
  • averaged angular distances of all tokens

• todo: based on layer linear replacement trining loss

Citing

If you use Short Transformers in your research, please cite with the following BibText

@misc{russak2024shorttransformers,
    title  = {ShortTransformers, optimal layer pruning tools},
    author = {Melisa Russak},
    url    = {https://github.com/melisa/short-transformers},
    year   = {2024}
}

@misc{gromov2024unreasonable,
      title={The Unreasonable Ineffectiveness of the Deeper Layers}, 
      author={Andrey Gromov and Kushal Tirumala and Hassan Shapourian and Paolo Glorioso and Daniel A. Roberts},
      year={2024},
      eprint={2403.17887},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}