50% prune?
#1
by Green-eyedDevil - opened
I like this model a lot, but unfortunately it is just ever so slightly too big for my setup. Would it be possible to make a 50% prune version? I tried making one myself but REAP didn't seem to like me for whatever reason.
thank you.
I tried pruning by 50%, but the performance degradation was unacceptable.(mostly Vision performance. even 40% not great)
I needed post training, so I postponed it for now...
The original REAP does not yet support Qwen3-VL.
Also, one node of mi300x is required to run REAP.
I see. Thank you.
Green-eyedDevil changed discussion status to closed