Distily

In one command, distill an existing LLM into a smaller or different architecture.

Install

pip install -U "git+https://github.com/lapp0/distily.git"

Features

Distily allows you to distill a model with

• Quantized weights: e.g. TriLM, bitnet
• Distinct architecture: State-Space models such as Mamba, Mixture-of-Experts (MoE)
• Modified architecture: Decrease (or increase) the
  • number of layers
  • width and depth of attention heads and dense layer.
  • the number of attention and KV heads.

Usage

Minimal Example: distily_gpt2

Command to create a distilled gpt2 with only 6 layers:

python3 -m distily.run \
    --teacher_model_name_or_path gpt2 \
    --output_dir distily_gpt2 \
    --hub_model_id "distily/distily_gpt2" \
    --push_to_hub True \
    --student_model_config {"n_layers": 6} \
    --student_model_as_bitnet True

The Resulting distily_gpt2 Model has (TODO: explain metrics).

For more examples, review the Examples documentation.

Note on Hub Credentials

To push to hub, you must prepare your hub token

HF_WRITE=<your hub token> python3 -c "from huggingface_hub.hf_api import HfFolder; HfFolder.save_token('${HF_WRITE}')"

Further Reading

TODO: commit the linked docs once complete

Using Distily

• How Distillation Works: The Distily Recipe
• Quickstart / Examples
• Parameter Selection

Available Models

• Official Distily Models
• All HF Models Created With Distily

Contributing

• Contributing Guidelines

Roadmap

Improved performance / sampling efficiency:

• Standard knowledge distillation using logits.
• Distill using intermediate features including hidden states and attentions.
• Implement Value-Transfer (simply distillation loss on v of q,k,v)
• Improve sampling efficiency through synthetic data generation.
• Implement cross-entropy classification loss (traditional LLM loss function)
• Apply projector to logits (https://arxiv.org/pdf/2310.17183)
• Apply "teacher recording", run teacher inference once, use features dataset any number of times.

Distill to a different model shape / size:

• Distill to model with fewer num_hidden_layers by implementing layer mappers.
• Distill to a model with modified module dimensions and behaviors (e.g., intermediate_size, hidden_act) by employing projectors.
• Distill to a model with modified num_attention_heads and num_key_value_heads.

Distill to a different architecture:

• Distill to Bitnet (b1.58)
• Distill to State-Space / Mamba
• Distill to MoE
• Distill to Parameter Sharing (ALBERT-style) Model

Additional Techniques:

• Distill from multiple models at once
• Pruning