New 1.66-bit TQ1_0 quant that is 162GB

#14
by shimmyshimmer - opened

We added a new TQ1_0 quant that is 1.66-bit and 162GB in size. For those who want it to fit exactly on some setups and some more lenient combinations

Made for setups with 192GB RAM and Ollama

shimmyshimmer changed discussion title from New TQ1_0 quant that is 162GB to New 1.66-bit TQ1_0 quant that is 162GB
shimmyshimmer pinned discussion

What’s the difference compared to IQ1_S? What layers are compressed more?

could you please share how to run it with llama cpp? Thanks.
Edit: it seems working

./llama.cpp/llama-cli \
    -hf unsloth/DeepSeek-R1-0528-GGUF:TQ1_0 \
    --cache-type-k q4_0 \
    --threads -1 \
    --n-gpu-layers 99 \
    --prio 3 \
    --temp 0.6 \
    --top_p 0.95 \
    --min_p 0.01 \
    --ctx-size 16384 \
    --seed 3407 \
    -ot ".ffn_.*_exps.=CPU"
Unsloth AI org

What’s the difference compared to IQ1_S? What layers are compressed more?

Correct but the correct layers

Is there a comparison of performance (benchmark score) between each quantized version and the original checkpoint? Thanks!

Sign up or log in to comment