pytorch

AI Newsletter - October 2024 Sarvam: Breaking New Ground in Indian Language AI India’s linguistic diversity has long posed a challenge for AI development, but Sarvam-1 is changing the game. This 2-billion parameter language model specifically targets 10 major Indian languages alongside English, proving that thoughtful design can outperform brute force. The model’s innovative approach tackles two critical challenges: token efficiency and data quality. Traditional multilingual models struggle with Indic scripts, requiring up to 8 tokens per word compared to English’s 1....

AI Newsletter - October 2024 1. Impact of Quantization on Large Language Models Recent extensive testing of the Llama 3.1 series models has shed light on the critical role of quantization in optimizing LLM deployments. The evaluation, conducted across various model sizes (8B, 70B, and 405B), compared three quantization schemes against the baseline 16-bit model: W8A8-INT: 8-bit integer quantization for weights and activations W8A8-FP: 8-bit floating-point quantization for weights and activations W4A16-INT: 4-bit integer quantization for weights, 16-bit precision for activations Key findings include:...

Deep Dive part 1: PyTorch vmap vs JAX vmap Introduction: Hello everyone, I am starting a new series of blog posts called deep dives, where I will do a deep and detailed analysis of machine learning frameworks and algorithms. Here I will break down new APIs and features of frameworks such as PyTorch, JAX, and tinygrad. In addition to that, I will also write about new algorithms in RLHF and other areas of research....