@iamurskyHinode: a GPU cloud with a persistent Linux workspace
Hi everyone, I'm Ilya. I'm building Hinode, a GPU cloud with a persistent Linux workspace for people building and running AI.
You set up apps and dependencies in your workspace, save a snapshot, and launch GPU machines from that setup. Models, datasets and outputs live in a shared folder that all your machines can access. Pausing a machine preserves its environment, so you can pick up where you left off.
Hinode supports workflows such as ComfyUI, model training and batch inference. You can work through a desktop, SSH or VS Code, or connect an AI assistant through MCP to start machines and run jobs.
If you're building with open models, which part of your GPU workflow takes the most effort today: environment setup, moving data, or managing compute?
Here's Hinode: hinode.run
Comments2
Welcome to Lounge, Ilya. The shared folder across machines is the part I'd poke at: with big checkpoints and datasets, how do you handle read latency on a cold machine, and does a snapshot include model weights or only the environment?
Thanks, Ryan! Each machine keeps a local cache of what it reads from the shared folder. The first load of a big checkpoint takes a moment, and after that it loads at local disk speed. Lots of tiny files are the slow case on a first read, so packed datasets work best. A snapshot holds only the environment: apps, packages and settings. Weights and datasets stay in the shared folder, so you download a model once and every machine has it.
2 comments
Welcome to Lounge, Ilya. The shared folder across machines is the part I'd poke at: with big checkpoints and datasets, how do you handle read latency on a cold machine, and does a snapshot include model weights or only the environment?
Thanks, Ryan! Each machine keeps a local cache of what it reads from the shared folder. The first load of a big checkpoint takes a moment, and after that it loads at local disk speed. Lots of tiny files are the slow case on a first read, so packed datasets work best. A snapshot holds only the environment: apps, packages and settings. Weights and datasets stay in the shared folder, so you download a model once and every machine has it.