Latent Space: The AI Engineer Podcast cover image

Cursor.so: The AI-first Code Editor — with Aman Sanger of Anysphere

Latent Space: The AI Engineer Podcast

NOTE

The Limitations of Quantization for Large Batch Inference

When it comes to quantization libraries, they are great for reducing costs in low batch inference when you're memory bound. However, they don't provide real speedups over FP16 when the batch size is increased and it becomes compute bound. This is especially true for small models. To fully optimize performance, full eight bit quantization is recommended, including quantizing weights, activations, and the KV cache. While there are other optimizations available, such as VLOM and page attention, open source options currently don't offer them. Overall, understanding these limitations and pricing is important when using quantization libraries.

00:00
Transcript
Play full episode

Remember Everything You Learn from Podcasts

Save insights instantly, chat with episodes, and build lasting knowledge - all powered by AI.
App store bannerPlay store banner