group-beam-search

Description

Diverse beam search is a variant of beam search that produces more diverse output candidates to choose from. This strategy measures the dissimilarity of sequences and a penalty is applied if sequences are too similar. To avoid high computation costs, the number of beams is divided into groups.

Enable diverse beam search with the num_beams, num_beam_groups and diversity_penalty parameters (the num_beams parameter should be divisible by num_beam_groups).

This implementation matches the group_beam_search functionality present in transformers<4.56.0.

Base model

• Qwen/Qwen3-0.6B

Model compatibility

• Decoder-only transformer models

Additional Arguments

• num_beams (int, optional, defaults to 1): Number of beams for beam search. If not greater than num_beam_groups, will be set to num_beam_groups.

• num_beam_groups (int, optional, defaults to 1): Number of groups to divide num_beams into for beam search.

• diversity_penalty (float, optional, defaults to 0.0): Diversity penalty applied to beams.

• early_stopping (bool or str, optional, defaults to False): Whether to stop beam search when at least num_beams complete candidates are finished per batch or not. If not False, it should be an integer greater than 1 indicating the minimum number of beams required to be finished per batch.

• max_length (int, optional, defaults to 20): The maximum length of the generated sequence.

• num_return_sequences (int, optional, defaults to 1): The number of sequences to return.

• repetition_penalty (float, optional, defaults to None): Helps reduce repetition. A value of 1.2 is recommended.

Output Type changes

• The generate method output remains the same as default transformers generation, but logits are post-processed using the DoLa contrastive scoring before token selection.

Example usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, infer_device

device = infer_device()

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-0.6B")
inputs = tokenizer("Hugging Face is an open-source company", return_tensors="pt").to(device)

model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-0.6B", dtype=torch.float16).to(device)
# explicitly set to 100 because Llama2 generation length is 4096
outputs = model.generate(**inputs, max_new_tokens=50, num_beams=6, num_beam_groups=3, diversity_penalty=1.0, do_sample=False, custom_generate="transformers-community/group-beam-search", trust_remote_code=True)
tokenizer.batch_decode(outputs, skip_special_tokens=True)