The model is working okay! (Temporary fix for stop token being ignored)

#1
by phakio - opened

A little slow I'm just testing CPU only right now, about 8 t/s generation but I'm not complaining about that!

I noticed that this GGUF, and a few other I tested would continue generation, stop token was ignored. I'm sharing my launch script to show one that is known working... " -r "<|im_end|>" " seems to prevent the model from over-generating.

I wouldn't say the backend or this quant are ready for prime time agentic tasks, but for testing and chatting it's fun to play with! Thanks. Sometimes it just replies in gibberish, or Chinese even though inquiry was English. I like its little quirks lol

/home/phone/llama.cpp-deepseek-v4-flash/build/bin/llama-server \
    --model /run/media/phone/SharedData/LocalModelsBIG/CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF/cyberneurova-DeepSeek-V4-Flash-abliterated.gguf \
    --alias cyberneurova/CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF \
    --ctx-size 10000 \
    -ctk q8_0 \
    -ngl 99 \
    --parallel 1 \
    --threads 48 \
    --threads-batch 56 \
    --host 0.0.0.0 \
    --port 8081 \
    --no-mmap \
    -r "<|im_end|>"
phakio changed discussion title from The model is working okay! to The model is working okay! (Temporary fix for stop token being ignored)
cyberneurova org

hey this model is a q2 model we just uploaded 2 new versions q2 and q8 check them out

Not able to download the Q8 version... Q2 link works, but is not suitable for my needs.

@pixillated thanks for the model refresh, when I'm back home in a few days I'll give them a try!

phakio changed discussion status to closed

image

400 bad request still.

Hello?

Sign up or log in to comment