ROCm 7 CI

[henk717]

I forgot we could use docker, that eliminates the diskspace and dependency issues making it work again.
However, ROCm 7 compiling is broken. Only merge this after uvos's fix lands upstream.

Its basically as follows, AMD can't properly release drivers.
So while they fixed the 9000 with the good speeds, they broke that for some older datacenter cards which now can't compile.
uvos is planning to dynamically exclude those again if ROCm7 is detected so they will use a slower method.

Because our users are primarily using consumer cards at home it still makes sense for us to switch to 7.
Datacenters can self compile with 6 should they want faster speeds.

[LostRuins]

alright i have no idea if it works, but let's give this a shot

[LostRuins]

Only merge this after uvos's fix lands upstream.

tell me when

[henk717]

Can be tracked here : ggml-org#16153

[LostRuins]

should i merge this now? @henk717

[henk717]

No, I am now also getting disk space limits on the docker method that worked a few weeks ago.
Unless you know of a way to clean up the host image prior to the docker step we can scrap this.

[henk717]

Found a way to get enough space but its a WIP for now

[henk717]

Compiles again (I accidentally had it generate cuda builds and we have enough space now for both rocm and cuda, so should be safe). Asking people in discord if this works well for them, I already confirmed it works well on Runpod.
Based on the feedback you see there feel free to merge.

[LostRuins]

Alright, merging.