T5 Efficient Small Kv256

🚀 T5-Efficient-SMALL-KV256 (Deep-Narrow version)

T5-Efficient-SMALL-KV256 is a variant of Google's original T5 that adheres to the T5 model architecture. It's a pretrained-only checkpoint, released alongside the paper Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers by Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler.

In essence, the paper suggests that a Deep-Narrow model architecture is more beneficial for downstream performance compared to other model architectures with a similar parameter count.

The paper states:

We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased before considering any other forms of uniform scaling across other dimensions. This is largely due to how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise, a tall base model might also generally more efficient compared to a large model. We generally find that, regardless of size, even if absolute performance might increase as we continue to stack layers, the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36 layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e., params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params, FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to consider.

More precisely, model depth is defined as the number of transformer blocks stacked sequentially. Thus, a sequence of word embeddings is processed sequentially by each transformer block.

📚 Documentation

Details model architecture

This model checkpoint - t5-efficient-small-kv256 - is of the Small model type with the following variations:

• kv is 256

It has 117.14 million parameters, requiring ca. 468.58 MB of memory in full precision (fp32) or 234.29 MB of memory in half precision (fp16 or bf16).

A summary of the original T5 model architectures is as follows:

The following abbreviations are used:

If a model checkpoint has no specific el or dl, then both the number of encoder- and decoder layers correspond to nl.

Pre-Training

The checkpoint was pretrained on the Colossal, Cleaned version of Common Crawl (C4) for 524288 steps using the span-based masked language modeling (MLM) objective.

Fine-Tuning

⚠️ Important Note

This model is a pretrained checkpoint and needs to be fine-tuned for practical use. The checkpoint was pretrained in English, so it's only useful for English NLP tasks.

You can follow one of the following examples to fine-tune the model:

PyTorch:

• Summarization
• Question Answering
• Text Classification - Note: You'll need to slightly adapt the training example here to work with an encoder-decoder model.

Tensorflow:

• Summarization
• Text Classification - Note: You'll need to slightly adapt the training example here to work with an encoder-decoder model.

JAX/Flax:

• Summarization
• Text Classification - Note: You'll need to slightly adapt the training example here to work with an encoder-decoder model.

More information

We highly recommend that readers carefully read the original paper Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers to gain a more in-depth understanding of this model checkpoint.

As explained in the following issue, checkpoints including the sh or skv model architecture variations have not been ported to Transformers as they probably have limited practical use and lack a detailed description. These checkpoints are stored here and might be ported in the future.

📄 License

This project is licensed under the apache-2.0 license.