TorchTPU getting native PyTorch support on vLLM and SGLang is huge, having to rewrite serving logic into JAX just to test TPU economics was always the main friction point
@zhuohan123 rollout generation eats almost all the wall clock time in post-training loops, engine throughput is basically the hard ceiling on training steps per day
@tri_dao the main bottleneck with diffusion workers is kv cache reuse across turns, autoregressive drafts still hold an advantage there for multi-turn loops
258 Followers 450 FollowingBrowse IKEA to select interior design and furniture, and use the profits earned from crypto investments to create your dream independent studio.
125K Followers 1K FollowingAgents & Gemini API, MTS @GoogleDeepMind | prev: Tech Lead at @huggingface, AWS ML Hero 🤗 Sharing my own views and AI News 🧑🏻💻 https://t.co/7IosdlO6RA
53K Followers 36 FollowingA high-throughput and memory-efficient inference and serving engine for LLMs. Join https://t.co/lxJ0SfX5pJ to discuss together with the community!
11K Followers 975 Followingbuilding @vllm_project at @meta | ex-openai | cs phd @ 🌁 uc berkeley | machine learning system | the real agi is the friends we made along the way
18K Followers 204 FollowingLarge Model Systems Organization: We developed SGLang @sgl_project (https://t.co/OjwQadINKU), Chatbot Arena (now @arena), and Vicuna!