LlamaIndex reports nearly 10x faster OCRia-1 token generation on a single B200

LlamaIndex says it made its OCRia-1 model generate tokens nearly 10 times faster on a single B200 GPU, using 17 inference optimizations.

Some individual changes improved speed by only 2–5%, which the company says added up. LlamaIndex says each optimization was tested for both speed and accuracy, and plans to contribute some of them back to the open-source vLLM project.