32k or 8k context?

#1
by mclassHF2023 - opened

This shows up in oobabooga as 8192 context and also seems to generate gibberish when set to higher context. I can't test the original model, but is this meant to be genuine 32k context?

From what I can see, the mnodel has a base context size of 8192, and uses rope scaling to get to 32k. Anything more you should ask on the original model - this is just quants of it. See the model card for a link.

Just tested with f16 - seems pretty coherent at 32k context size. Make sure your tool is uptodate w.r.t. llama3.

mradermacher changed discussion status to closed

Sign up or log in to comment