2026-06-17 12:53:43 +01:00
|
|
|
apiVersion: v1
|
|
|
|
|
kind: ConfigMap
|
|
|
|
|
metadata:
|
|
|
|
|
name: llamacpp-cuda-config
|
|
|
|
|
data:
|
|
|
|
|
LLAMA_CACHE: /models
|
|
|
|
|
LLAMA_ARG_HOST: 0.0.0.0
|
|
|
|
|
LLAMA_ARG_PORT: "8080"
|
2026-06-17 13:11:28 +01:00
|
|
|
LLAMA_ARG_HF_REPO: "unsloth/gemma-4-12b-it-GGUF:Q6_K"
|
2026-06-17 12:53:43 +01:00
|
|
|
LLAMA_ARG_CTX_SIZE: "32768"
|
|
|
|
|
LLAMA_ARG_FLASH_ATTN: auto
|
|
|
|
|
LLAMA_ARG_FIT: "on"
|
