FasterTransformer VS transformers

https://docs.nvidia.com/ai-enterprise/overview/0.1.0/platfor...

RIVA: NVIDIA® Riva, a premium edition of NVIDIA AI Enterprise software, is a GPU-accelerated speech and translation AI SDK

FasterTransformer: https://github.com/NVIDIA/FasterTransformer an

There could be some spectrum of this expertise. For instance, https://github.com/NVIDIA/FasterTransformer, https://github.com/microsoft/DeepSpeed

Have anyone tried to optimize the forward and backward using custom Cuda code or fused kernel to speed up the training time of current LLMs? I only have seen FasterTransformer ( NVIDIA/FasterTransformer) and other similar tools but they're only focusing on inference.

Replit built Ghostwriter on the open source scene based on Salesforce’s Codegen, using Nvidia’s FasterTransformer and Triton server for highly optimized decoders, and the knowledge distillation process of the CodeGen model from two billion parameters to a faster model of one billion parameters.

Nvidia FasterTransformer is a mix of Pytorch and CUDA/C++ dedicated code. The performance boost is huge on T5, they report a 10X speedup like TensorRT. However, the speedup is computed on a translation task where sequences are 25 tokens long on average. In our experience, improvement on very short sequences tend to decrease by large margins on longer ones. Still we plan to dig deeper into this project as it implements very interesting ideas.

On the other side of the spectrum, there is Nvidia demos (here or there) showing us how to build manually a full Transformer graph (operator by operator) in TensorRT to get best performance from their hardware. It’s out of reach for many NLP practitioners and it’s time consuming to debug/maintain/adapt to a slightly different architecture (I tried). Plus, there is a secret: the very optimized model only works for specific sequence lengths and batch sizes. Truth is that, so far (and it will improve soon), it’s mainly for MLPerf benchmark (the one used to compare DL hardware), marketing content, and very specialized engineers.

Is t5x an encoder/decoder architecture?

Some more general options.

The Flax ecosystem

https://github.com/google/flax?tab=readme-ov-file

or dm-haiku

https://github.com/google-deepmind/dm-haiku

were some of the best developed communities in the Jax AI field

Perhaps the “trax” repo? https://github.com/google/trax

Some HF examples https://github.com/huggingface/transformers/tree/main/exampl...

Sadly it seems much of the work is proprietary these days, but one example could be Grok-1, if you customize the details. https://github.com/xai-org/grok-1/blob/main/run.py

The HuggingFace transformers library already has support for a similar method called prompt lookup decoding that uses the existing context to generate an ngram model: https://github.com/huggingface/transformers/issues/27722

I don't think it would be that hard to switch it out for a pretrained ngram model.

Most of this tutorial is based on Hugging Face course about Transformers and on Niels Rogge's Transformers tutorials: make sure to check their work and give them a star on GitHub, if you please ❤️

* Superconvergence + LR range finder + Fast AI's Ranger21 optimizer was the goto optimizer for CNNs, and worked fabulously well, but on transformers, the learning rate range finder sadi 1e-3 was the best, whilst 1e-5 was better. However, the 1 cycle learning rate stuck. https://github.com/huggingface/transformers/issues/16013

Thanks! :) I'm pushing them into transformers, pytorch-gemma and collabing with the Gemma team to resolve all the issues :)

The RoPE fix should already be in transformers 4.38.2: https://github.com/huggingface/transformers/pull/29285

My main PR for transformers which fixes most of the issues (some still left): https://github.com/huggingface/transformers/pull/29402

If you want to tinker with the architecture Hugging Face has a FOSS implementation in transformers: https://github.com/huggingface/transformers/blob/main/src/tr...

If you want to reproduce the training pipeline, you couldn't do that even if you wanted to because you don't have access to thousands of A100s.

I am aware that using mixed precision reduces the stability of weight and there will be little consistency but don't expect it to be this much. I have attached the graph of evaluation metrics. If someone can give me some insight into this issue, that would be great.