Mozilla-Ocho / llamafile

Distribute and run LLMs with a single file.

Home Page:https://llamafile.ai

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Llava-v1.5 llamafile seems to be leaking outputs?

jjshoots opened this issue · comments

commented

Using the following command:

llava-v1.5-7b-q4-main.llamafile --image test_images/apanese_style_buildings.jpg --grammar 'root ::= [a-z]+ (" " [a-z]+)+' -n 10 -p '### User: The image has... ### Assistant:' --silent-prompt --no-display-prompt 2>/dev/null

I get the following output:

     1 -> ''
   835 -> ' ###'
  4911 -> ' User'
 29901 -> ':'
   450 -> ' The'
  1967 -> ' image'
   756 -> ' has'
   856 -> '...'
   835 -> ' ###'
  4007 -> ' Ass'
 22137 -> 'istant'
 29901 -> ':'
    13 -> '
'
 22933 -> 'ASS'
  9047 -> 'IST'
 13566 -> 'ANT'
 29901 -> ':'
 eunaciousness towards a building

I don't think the first many lines of text are supposed to be in there?

Sorry about that! Something got lost in translation with the last sync. I'll upstream my modifications soon for better synchronicity.

Fixed in llamafile 0.7.2 release. Weights uploaded https://huggingface.co/jartine/llava-v1.5-7B-GGUF