The docstring says

            - `"ffd"` (First Fit Decreasing): Slower but preserves sequence boundaries. Sequences are never cut in the
                middle.

but the implemented algorithm searches the bin with the least remaining capacity, which is called Best-Fit-Decreasing in the paper https://arxiv.org/abs/2404.10830.

• The algorithm before the optimization:

  trl/trl/data_utils.py

  Lines 507 to 517 in fef915e

  	# Optimized bin packing using a priority queue approach
  	bins_by_remaining = defaultdict(list) # remaining_space -> [bin_indices]
  	bins = [] # [(current_length, seq_indices)]
  	for i, (seq_len, _start, _end) in enumerate(sequences):
  	# Find bins with enough space using the dictionary
  	placed = False
  	for remaining in range(seq_len, seq_length + 1):
  	if bins_by_remaining[remaining]:
  	# Use the first available bin with this remaining space
  	bin_idx = bins_by_remaining[remaining].pop()

• In v0.19.0 https://github.com/huggingface/trl/blob/v0.19.0/trl/data_utils.py#L551-L561, it looks like Appendix B (An Illustration of optimized BFD Algorithm) of https://arxiv.org/pdf/2404.10830v2