Alibaba-NLP/new-impl

English | 中文

Code implementation of new GTE embeddings

This model is a BERT-like encoder with the following optimizations implemented:

1. Replacing absolute position embeddings with RoPE [^1].
2. Substituting the conventional activation functions with Gated Linear Units (GLU) [^2].
3. Setting attention dropout to 0 to use xformers and flash_attn.
4. Using unpadding to eliminate the needless computations for padding tokens [^3]. (this is off by default and should be used in conjunction with xformers for optimal acceleration).
5. Setting vocab_size as a multiple of 64.

Recommendation: Enable Unpadding and Acceleration with xformers

This code supports the acceleration of attention computations using xformers, which can automatically choose the optimal implementation based on the type of device, such as flash_attn. Therefore, we can also achieve significant acceleration on old devices like the V100.

Firstly, install xformers (with pytorch pre-installed):

if pytorch is installed using conda:
    conda install xformers -c xformers
elif pytorch is installed using pip:
    # cuda 11.8 version
    pip3 install -U xformers --index-url https://download.pytorch.org/whl/cu118
    # cuda 12.1 version
    pip3 install -U xformers --index-url https://download.pytorch.org/whl/cu121

For more information, refer to Installing xformers.

Then, when loading the model, set unpad_inputs and use_memory_efficient_attention to true, and enable fp16 mixed precision computation to achieve the fastest acceleration.

import torch
from transformers import AutoModel, AutoTokenizer

path = 'Alibaba-NLP/gte-base-en-v1.5'
device = torch.device('cuda')
tokenzier = AutoTokenizer.from_pretrained(path)
model = AutoModel.from_pretrained(
    path,
    trust_remote_code=True,
    unpad_inputs=True,
    use_memory_efficient_attention=True,
).to(device)

with torch.autocast(device_type=device.type, dtype=torch.float16):  # or bfloat16
    with torch.inference_mode():
        outputs = model(**inputs.to(device))

Alternatively, you can directly modify the unpad_inputs and use_memory_efficient_attention settings to true in the model's config.json, eliminating the need to set them in the code.

---

Clarification of Relationship with nomic-embed and nomicBERT

One may question the originality of our work and consider it a mere replication of nomicBERT. To clarify, our work is parallel but stems from the same idea as nomicBERT.

Applying RoPE and GLU to BERT to support longer texts is a straightforward idea. Our exploration of the transformer++ encoder (i.e., BERT + RoPE + GLU) began in August 2023. And by November 2023, we had completed the gte-base-en-v1.1. Then, I went on to prepare for the ACL submission of the other project...

The release of nomic-embed [^4] brought to our attention the pressure, as well as provided us with more resources, which allowed us to continue with this project. Without the outstanding work of nomicai, the release of gte-v1.5 could have been delayed much longer. Thanks!

---

[^1]: Su, Jianlin, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. "Roformer: Enhanced transformer with rotary position embedding." Neurocomputing 568 (2024): 127063.

[^2]: Shazeer, Noam. "Glu variants improve transformer." arXiv preprint arXiv:2002.05202 (2020).

[^3]: Portes, Jacob, Alexander Trott, Sam Havens, Daniel King, Abhinav Venigalla, Moin Nadeem, Nikhil Sardana, Daya Khudia, and Jonathan Frankle. "MosaicBERT: A Bidirectional Encoder Optimized for Fast Pretraining." Advances in Neural Information Processing Systems 36 (2024).

[^4]: Nussbaum, Zach, John X. Morris, Brandon Duderstadt, and Andriy Mulyar. "Nomic Embed: Training a Reproducible Long Context Text Embedder." arXiv preprint arXiv:2402.01613 (2024).