My first merge
@yamatazen Awesome that you managed to create your first merge. It already got already quantized 18 hours ago to https://huggingface.co/mradermacher/Ayla-Light-12B-GGUF and imatrix quants would have followed in around 2 hours.
You can check for progress at http://hf.tst.eu/status.html or regularly check the model
summary page at https://hf.tst.eu/model#Ayla-Light-12B-GGUF for imatrix quants to appear.
@mradermacher
I tried to forcefully bump the priority from 1 to -2000 using llmc add force -2000 i https://huggingface.co/yamatazen/Ayla-Light-12B to ensure imatrix quants are getting computed immediately but it doesn't look like it worked as the priority on leia is still shown as 1.
oh
Anyway, thanks.
@yamatazen All the imatrix quants are now available under https://huggingface.co/mradermacher/Ayla-Light-12B-i1-GGUF
@mradermacher
Oh no I now see what happened. Because Ayla-Light-12B was already on leia and no longer the wait queue by the time I tried to change priority it instead added a new one with -2000 priority to marco which just completed. Now it makes sense why I had to force it because usually I don't need to use force to change the priority. I guess it is no longer possible to change the priority once a model left the wait queue. In any case I just waited for Ayla-Light-12B on marco to be completed and uploaded to HuggingFace and then used llmc nuke Ayla-Light-12B to nuke the now useless imatrix taks from leia. I guess a bit concerning is that it also showed marco in the nuke output but I really waited untill all hfu tasks for it where done. The last quant uploaded by marco was i1-IQ3_S.
nico1 ~# llmc nuke Ayla-Light-12B
Ayla-Light-12B: not found
leia leia found Ayla-Light-12B
leia leia rm -rf /llmjob/wdir/Ayla-Light-12B.gguf
leia leia rm -rf /llmjob/wdir/Ayla-Light-12B-GGUF
leia leia rm -rf /llmjob/wdir/Ayla-Light-12B-i1-GGUF
nuked Ayla-Light-12B on leia
marco marco found Ayla-Light-12B
nuked Ayla-Light-12B on marco
Indeed, changing priorities is only (easily) possible while it is in the global queue. Once it is fired off to a worker, the scheduler doesn't even know it exists anymore. I could provide a command the change a job, but the issue is that sucha command already exists (for me :), namely llmjob edit, which allows editing the json file locally on each node. I'll put a vague feature request on my virtual todo list.
I guess a bit concerning is that it also showed marco in the nuke output
Right, didn't think of that. Yes, finished jobs stay in the queue for... until a non-automatic housecleaning job archives them away. That's why it nuked it on marco, it nuked the hidden/deleted job, which is harmless.