Difference in the length of positional embeddings produce different results

Hi, I am currently experimenting with how the length of dialogue histories in one input affects the performance of dialogue models using multi-session chat data. While I am working on BlenderbotSmallForConditionalGeneration from Huggingface’s transformers with the checkpoint “blenderbot_small-90M”, I encountered results which are not understandable for me.

Since I want to put long inputs(ex. 1024, 2048, 4096…), I expanded the positional embedding matrix of the encoder since it is initialized in the size (512, 512). I copied the first 512 embeddings and appended them repeatedly to make the embedding matrix the size I want. On the other hand, I truncated the position embedding matrix of the decoder into (128, 512), since the max target length is 128.

from transformers import BlenderbotSmallForConditionalGeneration
from transformers.models.blenderbot.modeling_blenderbot import BlenderbotLearnedPositionalEmbedding
from torch import nn

import torch

src_max_len = SRC_MAX_LEN  # 4906, 2048, 1024...
trg_max_len = 128

def reset_position_embeddings():
    # Expand encoder position embedding.
    encoder_weights = model.model.encoder.embed_positions.weight.data
    model.model.encoder.embed_positions = BlenderbotLearnedPositionalEmbedding(src_max_len, model.config.d_model)
    num_repeats = src_max_len // model.config.max_position_embeddings
    new_encoder_weights = encoder_weights.repeat(num_repeats, 1)
    with torch.no_grad():
        model.model.encoder.embed_positions.weight = nn.Parameter(new_encoder_weights)
    assert torch.equal(model.model.encoder.embed_positions.weight.data, encoder_weights.repeat(num_repeats, 1))
    model.config.max_length = src_max_len
    model.config.max_position_embeddings = self.args.src_max_len
    # Truncate decoder position embedding.
    decoder_weights = model.model.decoder.embed_positions.weight.data
    model.model.decoder.embed_positions = BlenderbotLearnedPositionalEmbedding(trg_max_len, model.config.d_model)
    with torch.no_grad():
        model.model.decoder.embed_positions.weight = nn.Parameter(decoder_weights[:self.args.trg_max_len, :])
    assert torch.equal(model.model.decoder.embed_positions.weight, decoder_weights[:self.args.trg_max_len])

After modifying the model, I trained it with different lengths of source data. But the maximum length of the source inputs is shorter than 2048 and the target response is the same, the results from the 4096 and 2024 versions must be identical, even if there is a difference in the size of position embeddings. However, the results were different.


This is odd since I checked all other variables, including the model parameters except the expanded parts of the position embeddings, the preprocessed data itself, the order of batches, etc. The reproducibility was guaranteed when I tested other data and models, but the only difference is the size of position embeddings.

I thought although the max length of the embedding matrix is different, the inputs are the same and this should not affect the results. Did I understand correctly, or there is something I am missing?