28 lines
1.0 KiB
Bash
Executable File
28 lines
1.0 KiB
Bash
Executable File
#!/bin/bash
|
|
#grab container image name
|
|
MODEL_NAME="Qwen3.6-35B-A3B-Uncensored-Aggressive-noThink"
|
|
CONT_NAME="llama-$MODEL_NAME"
|
|
|
|
MODEL_FOLDER="/home/jared/.cache/huggingface/hub/models--HauhauCS--Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive"
|
|
MODEL="snapshots/f12a584fecbeb5f20001130d8ecd66c9327ae685/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-IQ4_XS.gguf"
|
|
MODEL_MM="snapshots/f12a584fecbeb5f20001130d8ecd66c9327ae685/mmproj-Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-f16.gguf"
|
|
|
|
source ../env/llama_env.sh
|
|
echo $IMAGE
|
|
docker container create --name $CONT_NAME --network llms -p 8355:8355 -v $MODEL_FOLDER:/model --user $(id -u):$(id -g) --gpus=all --restart on-failure:3 \
|
|
--health-cmd "curl -f http://$CONT_NAME:10987/v1/models || exit 1" \
|
|
--health-interval 5s \
|
|
--health-timeout 5s \
|
|
--health-start-period 20s \
|
|
$IMAGE \
|
|
/llama.cpp/build/bin/llama-server \
|
|
--model /model/$MODEL \
|
|
--mm /model/$MODEL_MM \
|
|
--alias $MODEL_NAME \
|
|
-ctk q8_0 -ctv q8_0 \
|
|
--reasoning-budget 0 \
|
|
--host $CONT_NAME \
|
|
--port 8355 --no-mmap -dio
|
|
|
|
echo "Done!"
|
|
echo $CONT_NAME |