Cutting Edge Isn’t Plug-and-Play: Customizing FlashAttention-4 on B300
How I customized FlashAttention-4 for B300 and brought it to video language model training.
Notes and essays
Writing
A collection of technical notes, research reflections, and ideas I wanted to work through in writing.
The Rise of RNN? Review of “Retentive Network: A Successor to Transformer for Large Language Models”
A review of RetNet’s parallel, recurrent, and chunkwise retention, and its promise of parallel training with constant-time inference.
GZIP + KNN > BERT for Text Classification??
A look at compressor-based text classification with GZIP, normalized compression distance, and k-nearest neighbors.
Compositional Generalization
What is compositional generalization problem and why is this interesting?
Big NLP Trends from 2016 to 2022
Recent trends in NLP from Attention to ChatGPT
AINize + Gradio: Easiest Automatic ML model deployment
Serving a simple demo page (along with REST API) for you ML model on the web.
Introducing easy HW acceleration for PyTorch
Multi-GPU / TPU setting with PyTorch made easy with huggingface accelerate.
How to serve Pororo Models
How to serve pretrained language models provided from Pororo library by Kakao Brain Corp. with torchserve.