Is the config,json correct?

#2
by HankFromResearch - opened

is "max_position_embeddings": 2048 correct?

I only ask this because I have noticed that some updates in other GPTQ repos have changed max_position_embeddings to 4096.
I just set my position embedding compression factor to 2 (max length is 4096), since I just assumed it was a particular nuance of this quant.

Yeah you're right it should be 4096. The original Llama 2 models set 2048 so the fine tunes that came out early also had 2048. But it should be 4096 for all Llama 2 models and Llama 2 fine tunes

I'll fix that now

Awesome, love your work, cheers for that.

HankFromResearch changed discussion status to closed

Sign up or log in to comment