开启时会模糊预览图，关闭后正常显示
4/4 Physical Infrastructure for Cursor's Composer model training. They claim to have trained (and continue to train) across thousands of GPUs. They train models in low precision, and they use asynchronous RL (next tweet to explain what it is). Quote: "We built custom training infrastructure leveraging PyTorch and Ray to power asynchronous reinforcement learning at scale. We natively train our models at low precision by combining our MXFP8 MoE kernels with expert parallelism and hybrid sharded data parallelism, allowing us to scale training to thousands of NVIDIA GPUs with minimal communication cost. Additionally, training with MXFP8 allows us to deliver faster inference speeds without requiring post-training quantisation."
5/5 What is async RL that Customer Composer model training uses? It uses asynchronous execution at multiple levels to avoid waiting on slow operations e.g. a long roll-out generation. As you know, for a given problem, in RL like GRPO we generate multiple trajectorier. However, some trajectories can take too long to complete. So, once they have enough trajectories, they run the training. Partial samples/roll-outs are resumed later with updated model. This causes a situation where some tokens are generated by the old model/policy and some by new. However, this is acceptable. If you want to understand more about Async RL, please read APRIL - a project for Async RL.
i have a bunch of markov chain monte carlo stuff lying around from grad school i can throw at this, just have to port it from matlab
back in the prehistory of 3d computer vision (2016) we would use probabilistic / ebm models to fit shape templates to street scenes
RT @tibo_maker: SuperX was an "obvious product" my thought: "if I grew Tweet Hunter to $250k MRR, it should be easy with SuperX" yet,…
Built Tweet Hunter, Taplio (sold $8m) Growing https://t.co/OyNJ8ZUyOh - https://t.co/jS9GQJ5Ps8 - https://t.co/EFUcKeBbpU - https://t.co/JkVOl1O0S1 - https://t.co/KG9PgxJabg Sharing weekly tips about growth: https://t.co/ereQodN3Ov
now that the simulator is out i'm very keep to try this out!
i have a bunch of markov chain monte carlo stuff lying around from grad school i can throw at this, just have to port it from matlab
Morgan Stanley to buy private shares platform EquityZen
Top and breaking news, pictures and videos from Reuters. For breaking business news, follow @ReutersBiz. Our daily podcast is here: https://t.co/KO0QFy0d3a
Hurricane Melissa unleashed devastation in Jamaica as the strongest storm on record ever to hit the Caribbean island nation, and roared later into eastern Cuba, smashing the city of Santiago and flooding rural land
Top and breaking news, pictures and videos from Reuters. For breaking business news, follow @ReutersBiz. Our daily podcast is here: https://t.co/KO0QFy0d3a