From e68fdbdaad6220e51dae78774e0bff162d1bce4f Mon Sep 17 00:00:00 2001 From: BadStorm Developer Date: Sat, 8 Aug 2026 21:12:50 +0200 Subject: [PATCH] containers/llamacpp: stop router from downloading the baked-in single model --- containers/llamacpp/llamacpp-rocm.Containerfile | 6 +++++- containers/llamacpp/llamacpp.container | 4 ++++ 2 files changed, 9 insertions(+), 1 deletion(-) diff --git a/containers/llamacpp/llamacpp-rocm.Containerfile b/containers/llamacpp/llamacpp-rocm.Containerfile index 2801854..8ed31f4 100644 --- a/containers/llamacpp/llamacpp-rocm.Containerfile +++ b/containers/llamacpp/llamacpp-rocm.Containerfile @@ -73,7 +73,11 @@ ENV LD_LIBRARY_PATH=/app:$LD_LIBRARY_PATH ENV HF_XET_HIGH_PERFORMANCE=1 ENV LLAMA_ARG_HOST=0.0.0.0 ENV LLAMA_ARG_PORT=8090 -ENV LLAMA_ARG_HF_REPO=unsloth/Qwen3.5-35B-A3B-GGUF:Q2_K_XL +# Default single-model rimosso: in router mode il modello lo decide il preset +# (/config/preset.ini via LLAMA_ARG_MODELS_PRESET). Se un'immagine prodotta +# prima di questa modifica viene usata, il .container deve svuotare +# LLAMA_ARG_HF_REPO per evitare download non richiesti. +#ENV LLAMA_ARG_HF_REPO=unsloth/Qwen3.5-35B-A3B-GGUF:Q2_K_XL ENV LLAMA_ARG_LOAD_MODE=none ENV LLAMA_ARG_CTX_SIZE=128000 #ENV LLAMA_ARG_MMPROJ_URL= diff --git a/containers/llamacpp/llamacpp.container b/containers/llamacpp/llamacpp.container index 81f0750..e3e4cf2 100644 --- a/containers/llamacpp/llamacpp.container +++ b/containers/llamacpp/llamacpp.container @@ -22,6 +22,10 @@ Environment=LLAMA_ARG_HOST=0.0.0.0 Environment=LLAMA_ARG_PORT=8090 Environment=LLAMA_ARG_MODELS_PRESET=/config/preset.ini +# Svuota l'eventuale LLAMA_ARG_HF_REPO baked-in nell'immagine (default +# single-model) per non scaricare modelli non richiesti in router mode. +Environment=LLAMA_ARG_HF_REPO= + # RPC — worker ggml-rpc-server (llamacpp-rpc.container), lista host:port; backend insicuro, solo su internal.network #Environment=LLAMA_ARG_RPC=llamacpp-rpc:50052 # Con RPC attivo, ripartisce i layer tra GPU locale e worker remoto (es. "1,1" = meta' e meta'); senza, tutto in locale