Show HN: Reame – a CPU inference server that gets faster as it runs

JackC 32 minutes ago | parent | next [-]

> What Reame is NOT for — said plainly, because trust is built here

It is so weird (or, used to be) to see an LLM's internal thought process pop up this way. Like imagine how strange it would be to read human writing that accidentally included thoughts undercutting the ongoing sentence. It's the moment you know that nothing you're reading has necessarily been seen by a human before or relates to reality.

▲

ComputerGuru an hour ago | parent | prev | next [-]

Seems the entire repo and documentation is all ai-generated?

▲

JPLeRouzic 5 hours ago | parent | prev | next [-]

Thanks for sharing your work!

Please how to select the model? I downloaded tinyLlama, put it in ./models, changed reame.conf but I get:

(No such file or directory)

Otherwise putting the model in /opt does not please me much, I fear to forget a model is there, if it is in reame folder its much easier to notice and manage.

▲

K0IN 4 hours ago | parent | prev | next [-]

Sick that you can get 2 arm cores and 12 GB ram for free at Oracle cloud, did not know that

▲

GTP 4 hours ago | parent | next [-]

You can actually get more. You have an amount of resources for ARM servers, if you allocate all of that to a single one you ger 24GB RAM and 4 vCPU.

	▲	derac 3 minutes ago \| parent [-]
		This was reduced recently to 2/12. It's also next to impossible to allocate the free stuff unless you put in a CC. If you do, you get the servers right away, though. Just be careful not to spend. You can set up a budget trigger to shut down the vms. Old servers may have been grandfathered, but mine was set to 1/1 and couldn't be reshaped.

▲

esseph 4 hours ago | parent | prev | next [-]

Yeah but it's Oracle

▲

CryptoBanker 4 hours ago | parent | prev [-]

If you have good luck perhaps. That tier of server is virtually always sold out

▲

ohadkr 7 hours ago | parent | prev | next [-]

The persistent KV cache is interesting; I’d love to see how much of the speedup remains when requests share less structure.

▲

tyzoid 6 hours ago | parent | prev | next [-]

Why qwen 2.5 everywhere? Why not 3.5?

	▲	TheGoddessInari 4 hours ago \| parent [-]
		Because llms reflect qwen 2.5 heavily in training data, and that's who is giving the recommendations.

▲

gaurang404 4 hours ago | parent | prev [-]

looks really interesting