yujiepan-work / MoEBERT

This PyTorch package implements MoEBERT: from BERT to Mixture-of-Experts via Importance-Guided Adaptation (NAACL 2022).

MoEBERT

This PyTorch package implements MoEBERT: from BERT to Mixture-of-Experts via Importance-Guided Adaptation (NAACL 2022).

Installation

• Create and activate conda environment.

conda env create -f environment.yml

• Install Transformers locally.

pip install -e .

• Note: The code is adapted from this codebase. Arguments regarding LoRA and adapter can be safely ignored.

Instructions

MoEBERT targets task-specific distillation. Before running any distillation code, a pre-trained BERT model should be fine-tuned on the target task. Path to the fine-tuned model should be passed to --model_name_or_path.

Importance Score Computation

• Use bert_base_mnli_example.sh to compute the importance scores, add a --preprocess_importance argument, remove the --do_train argument.
• If multiple GPUs are used to compute the importance scores, a importance_[rank].pkl file will be saved for each GPU. Use merge_importance.py to merge these files.
• To use the pre-computed importance scores, pass the file name to --moebert_load_importance.

Knowledge Distillation

• For GLUE tasks, see examples/text-classification/run_glue.py.
• For question answering tasks, see examples/question-answering/run_qa.py.
• Run bash bert_base_mnli_example.sh as an example.
• The codebase supports different routing strategies: gate-token, gate-sentence, hash-random and hash-balance. Choices should be passed to --moebert_route_method.
  • To use hash-balance, a balanced hash list needs to be pre-computed using hash_balance.py. Path to the saved hash list should be passed to --moebert_route_hash_list.
  • Add a load balancing loss by setting --moebert_load_balance when using trainable gating mechanisms.
  • The sentence-based gating mechanism (gate-sentence) is advantageous for inference because it induces significantly less communication overhead compared with token-level routing methods.