Tag
2 articles
Prime Intellect has launched Prime Inference, a serverless and reserved serving platform for frontier open models, leveraging NVIDIA Blackwell hardware and optimized technologies like vLLM and NVFP4 KV compression.
Researchers at UC San Diego introduce DFlash, a new speculative decoding technique that drafts whole token blocks in parallel, achieving up to 15x throughput improvement on NVIDIA Blackwell.