mradermacher

/

model_requests

like 83

What you linked are gemma-3-12b-it GGUF quants and with a very specific horror themed imatrix. I assume you want imatrix quants of https://huggingface.co/google/gemma-3-12b-it which our high quality proprietary imatrix dataset.

Seams like @mradermacher never updated to the version where I fixed the return codes - this not only explains why Gemma3 is not working but also some dry-run detect errors that didn't trigger as expected:

marco    nice size (static/imatrix) -- jobs 10/6-12 maxm 150 free 16154 budget 464 uploads 49 hfd 0
        -2000   25 si gemma-3-12b-it                               error/1 ~Gemma3ForConditionalGeneration

@mradermacher I just updated my fork to latest llama.cpp again so when you update the next time you have the latest version. I recommend you update as soon you have time to do so as we can't do any Gemma3 based model with the current version.

sure, update us running, will restart the failed models. and yes, I only update when absolutely needed - I was not aware that you have an updated version that would detect more errors, only that you added a non-zero exit code (which is currently ignored due to it being in a pipe | grep - does this affect anything?)

sure, update us running, will restart the failed models. and yes, I only update when absolutely needed - I was not aware that you have an updated version that would detect more errors, only that you added a non-zero exit code (which is currently ignored due to it being in a pipe | grep - does this affect anything?)

You miss some errors like the one you complained about dry-run not catching by only checking for that specific error message so I recommend to use the exit code instead on the latest version. It's not that I added detection for more errors but that your old way of detecting errors misses some errors dry-run was already detecting. I mentioned back when I fixed the return code that I also merged latest llama.cpp into our fork to add Gemma3 support but only did so in some random sentence so it probably got missed.

old way of detecting errors misses some errors dry-run was already detecting

Ok, I'm soon back at the old way (exit status)