Context size? YaRN still supported?
#3
by
Thireus
- opened
What is the supported context size and is YaRN still working as expected with large context sizes?
Not the OP, but it should be the same since no retraining was done.
Also not "OP" ; experiments (and uploaded) Qwen3 8B from 64k context to 320K context.
It holds together.
Should be okay with A3Bs ; built 128k IQ1_M (of 30B-A3B) and works great.
Kalamaze's excellent project is due for "tweaking" shortly in the "lab".