# cd /home/mshahidul/llama.cpp
# export CUDA_DEVICE_ORDER=PCI_BUS_ID
# export CUDA_VISIBLE_DEVICES=3

# ./build/bin/llama-server \
#   -m /home/mshahidul/readctrl/models/translategemma-27b-it-Q8_0.gguf \
#   -ngl 99 \
#   -fa on \
#   -c 16384 \
#   --port 8081 \
#   --host 0.0.0.0


cd /home/mshahidul/llama.cpp
export CUDA_DEVICE_ORDER=PCI_BUS_ID
export CUDA_VISIBLE_DEVICES=4

./build/bin/llama-server \
  -m unsloth/DeepSeek-R1-Distill-Llama-8B-GGUF/DeepSeek-R1-Distill-Llama-8B-Q4_K_M.gguf \
  -ngl 99 \
  -fa on \
  -c 16384 \
  --port 8085 \
  --host 0.0.0.0