Introducing Cloudy, a research tool for long-running experiments in the cloud. Your experiments move between CPUs and GPUs as needed, with terabytes of storage each. Pause anytime and resume right where you left off.

I ran a simple experiment on Cloudy last week. It cut the kernels in my LLM inference library (SimpleLLM) from 5,762 lines to 297 and made it decode 11-15% faster than vLLM 0.32.0 (latest). (using a single prompt) Blog: Code:

Building good research systems takes a lot of time, iteration and money. Agent swarms, subagents, the newest autoresearch ideas, the best models across labs: we'll benchmark them all and bring you what works best, so you can focus on what matters. More to come soon :)

hi @DavidSHolz, i made the thing you needed <3

@cloudysoftwares woah it looks so good!

@cloudysoftwares thank you!! excited for you to try it :’)

@cloudysoftwares Looks promising, great job!

@cloudysoftwares thank you :’)

@cloudysoftwares lessgooo ❤️

@cloudysoftwares ❤️❤️ would love for you to try it out, send me feedback!!!

@cloudysoftwares FINALLY!!! 🤩 I WANT TO CHECK THIS OUT

@cloudysoftwares yess!! i was going to dm you, excited for you to try cloudy :")

@cloudysoftwares 🔥🔥

@cloudysoftwares AAAAHH ty

@cloudysoftwares aye w!

@cloudysoftwares ty!!

@cloudysoftwares This is so cool.

@cloudysoftwares thank you!!

@cloudysoftwares lovely

@cloudysoftwares <3

@cloudysoftwares goat is back

@cloudysoftwares <3

@cloudysoftwares Lffgggg!

@cloudysoftwares yess

@cloudysoftwares Gorgeous

@cloudysoftwares <3

@cloudysoftwares This looks cool! Can you BYOA and BYOC (agent/compute)?

@cloudysoftwares you can bring your own claude code or codex sub, not yet with compute but you could temporarily set secrets of your gpu cloud in a chat settings, if you have excess compute somewhere else.

@cloudysoftwares letsgoo finally 🐐

@cloudysoftwares you should give it a spin for some of your work at @lossfunk!!

@cloudysoftwares @lossfunk on it boss

@cloudysoftwares @lossfunk might be stupid but i can bring other gpu renting platforms in the play too right?

@cloudysoftwares @lossfunk for now you can set secrets in chat settings, dming you

@cloudysoftwares how does pause work on the GPU side, do you checkpoint the process or do experiments need to be restartable from disk?

@cloudysoftwares How fast is resume when state has to reload from the storage tier instead of staying in memory?

@cloudysoftwares cloudy launch, tried to dm you about an open-source ml platform co-founder i can intro you to but it didn't go through

@cloudysoftwares any plans to open source it?

@cloudysoftwares pause and resume a long run without losing state? finally

@cloudysoftwares Pausing a GPU run without losing the setup is the sleeper feature.