ybelkada / switch-base-8-xsum

Model Card for Switch Transformers Base - 8 experts

Table of Contents

1. TL;DR
2. Model Details
3. Usage
4. Uses
5. Bias, Risks, and Limitations
6. Training Details
7. Evaluation
8. Environmental Impact
9. Citation
10. Model Card Authors

TL;DR

Switch Transformers is a Mixture of Experts (MoE) model trained on Masked Language Modeling (MLM) task. The model architecture is similar to the classic T5, but with the Feed Forward layers replaced by the Sparse MLP layers containing "experts" MLP. According to the original paper the model enables faster training (scaling properties) while being better than T5 on fine-tuned tasks. As mentioned in the first few lines of the abstract :

we advance the current scale of language models by pre-training up to trillion parameter models on the “Colossal Clean Crawled Corpus”, and achieve a 4x speedup over the T5-XXL model.

Disclaimer : Content from this model card has been written by the Hugging Face team, and parts of it were copy pasted from the original paper .

Model Details

Model Description

• Model type: Language model
• Language(s) (NLP): English
• License: Apache 2.0
• Related Models: All FLAN-T5 Checkpoints
• Original Checkpoints: All Original FLAN-T5 Checkpoints
• Resources for more information:
  • Research paper
  • GitHub Repo
  • Hugging Face Switch Transformers Docs (Similar to T5)

Usage

Note that these checkpoints has been trained on Masked-Language Modeling (MLM) task. Therefore the checkpoints are not "ready-to-use" for downstream tasks. You may want to check FLAN-T5 for running fine-tuned weights or fine-tune your own MoE following this notebook

Find below some example scripts on how to use the model in transformers :

Using the Pytorch model

Running the model on a CPU

from transformers import AutoTokenizer, SwitchTransformersConditionalGeneration

tokenizer = AutoTokenizer.from_pretrained("google/switch-base-8")
model = SwitchTransformersConditionalGeneration.from_pretrained("google/switch-base-8")

input_text = "A <extra_id_0> walks into a bar a orders a <extra_id_1> with <extra_id_2> pinch of <extra_id_3>."
input_ids = tokenizer(input_text, return_tensors="pt").input_ids

outputs = model.generate(input_ids)
print(tokenizer.decode(outputs[0]))
>>> <pad> <extra_id_0> man<extra_id_1> beer<extra_id_2> a<extra_id_3> salt<extra_id_4>.</s>

Running the model on a GPU

# pip install accelerate
from transformers import AutoTokenizer, SwitchTransformersConditionalGeneration

tokenizer = AutoTokenizer.from_pretrained("google/switch-base-8")
model = SwitchTransformersConditionalGeneration.from_pretrained("google/switch-base-8", device_map="auto")

input_text = "A <extra_id_0> walks into a bar a orders a <extra_id_1> with <extra_id_2> pinch of <extra_id_3>."
input_ids = tokenizer(input_text, return_tensors="pt").input_ids.to(0)

outputs = model.generate(input_ids)
print(tokenizer.decode(outputs[0]))
>>> <pad> <extra_id_0> man<extra_id_1> beer<extra_id_2> a<extra_id_3> salt<extra_id_4>.</s>

Running the model on a GPU using different precisions

FP16

# pip install accelerate
from transformers import AutoTokenizer, SwitchTransformersConditionalGeneration

tokenizer = AutoTokenizer.from_pretrained("google/switch-base-8")
model = SwitchTransformersConditionalGeneration.from_pretrained("google/switch-base-8", device_map="auto", torch_dtype=torch.float16)

input_text = "A <extra_id_0> walks into a bar a orders a <extra_id_1> with <extra_id_2> pinch of <extra_id_3>."
input_ids = tokenizer(input_text, return_tensors="pt").input_ids.to(0)

outputs = model.generate(input_ids)
print(tokenizer.decode(outputs[0]))
>>> <pad> <extra_id_0> man<extra_id_1> beer<extra_id_2> a<extra_id_3> salt<extra_id_4>.</s>

INT8

# pip install bitsandbytes accelerate
from transformers import AutoTokenizer, SwitchTransformersConditionalGeneration

tokenizer = AutoTokenizer.from_pretrained("google/switch-base-8")
model = SwitchTransformersConditionalGeneration.from_pretrained("google/switch-base-8", device_map="auto")

input_text = "A <extra_id_0> walks into a bar a orders a <extra_id_1> with <extra_id_2> pinch of <extra_id_3>."
input_ids = tokenizer(input_text, return_tensors="pt").input_ids.to(0)

outputs = model.generate(input_ids)
print(tokenizer.decode(outputs[0]))
>>> <pad> <extra_id_0> man<extra_id_1> beer<extra_id_2> a<extra_id_3> salt<extra_id_4>.</s>

Uses

Direct Use and Downstream Use

The authors write in the original paper's model card that:

The primary use is research on language models, including: research on zero-shot NLP tasks and in-context few-shot learning NLP tasks, such as reasoning, and question answering; advancing fairness and safety research, and understanding limitations of current large language models

See the research paper for further details.

Out-of-Scope Use

More information needed.

Bias, Risks, and Limitations

More information needed.

Ethical considerations and risks

More information needed.

Known Limitations

More information needed.

Sensitive Use:

SwitchTransformers should not be applied for any unacceptable use cases, e.g., generation of abusive speech.

Training Details

Training Data

The model was trained on a Masked Language Modeling task, on Colossal Clean Crawled Corpus (C4) dataset, following the same procedure as T5 .

Training Procedure

According to the model card from the original paper :

These models are based on pretrained SwitchTransformers and are not fine-tuned. It is normal if they perform well on zero-shot tasks.

The model has been trained on TPU v3 or TPU v4 pods, using t5x codebase together with jax .

Evaluation

Testing Data, Factors & Metrics

The authors evaluated the model on various tasks and compared the results against T5. See the table below for some quantitative evaluation: For full details, please check the research paper .

Results

For full results for Switch Transformers, see the research paper , Table 5.

Environmental Impact

Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019) .

• Hardware Type: Google Cloud TPU Pods - TPU v3 or TPU v4 | Number of chips ≥ 4.
• Hours used: More information needed
• Cloud Provider: GCP
• Compute Region: More information needed
• Carbon Emitted: More information needed

Citation

BibTeX:

@misc{https://doi.org/10.48550/arxiv.2101.03961,
  doi = {10.48550/ARXIV.2101.03961},
  
  url = {https://arxiv.org/abs/2101.03961},
  
  author = {Fedus, William and Zoph, Barret and Shazeer, Noam},
  
  keywords = {Machine Learning (cs.LG), Artificial Intelligence (cs.AI), FOS: Computer and information sciences, FOS: Computer and information sciences},
  
  title = {Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity},
  
  publisher = {arXiv},
  
  year = {2021},
  
  copyright = {arXiv.org perpetual, non-exclusive license}
}