Reverse-engineering Nvidia's CUDA-checkpoint for faster cold starts
20 points
1/21/1970
a day ago
by ilreb
Comments
zoobab
a day ago
fergus00
https://blog.doubleword.ai/fast-sglang-starts yeah, this is part of the idea. If you get cold starts down to seconds or subseconds, then you can run many models multiplexed on the same GPUs
a day ago
zoobab
I have read a Linkedin post in french recently with a university lab (one big machine serving hundreds of users) mentioning that they could swap models on demand. Will try to find back the link.
a day ago
Would it be able to swap models on demand with this?