36 lines
1.3 KiB
Docker
36 lines
1.3 KiB
Docker
# llama-server für Qwen3.6 Vision — NVIDIA RTX PRO 6000 Blackwell (sm_120)
|
|
#
|
|
# WICHTIG:
|
|
# - Muss aus aktuellem Source gebaut werden: Qwen3.6 nutzt ein neues
|
|
# Rope-Encoding (rope.dimension_sections 3 statt 4), alte Images/Builds
|
|
# brechen mit "wrong array length".
|
|
# - Blackwell braucht CUDA >= 12.8. KEIN CUDA 13.2 verwenden — erzeugt
|
|
# mit Qwen3.6 Gibberish (bekannter NVIDIA-Bug).
|
|
|
|
FROM nvidia/cuda:12.8.1-devel-ubuntu24.04 AS build
|
|
|
|
RUN apt-get update && apt-get install -y --no-install-recommends \
|
|
git cmake build-essential libcurl4-openssl-dev ca-certificates \
|
|
&& rm -rf /var/lib/apt/lists/*
|
|
|
|
RUN git clone --depth 1 https://github.com/ggml-org/llama.cpp /src
|
|
|
|
# 120 = Blackwell (RTX PRO 6000). Für andere Karten anpassen.
|
|
RUN cmake /src -B /build \
|
|
-DGGML_CUDA=ON \
|
|
-DCMAKE_CUDA_ARCHITECTURES=120 \
|
|
-DGGML_CUDA_FORCE_CUBLAS=OFF \
|
|
-DBUILD_SHARED_LIBS=OFF \
|
|
-DLLAMA_CURL=ON \
|
|
&& cmake --build /build --config Release -j --target llama-server
|
|
|
|
FROM nvidia/cuda:12.8.1-runtime-ubuntu24.04
|
|
|
|
RUN apt-get update && apt-get install -y --no-install-recommends \
|
|
libcurl4 libgomp1 curl ca-certificates \
|
|
&& rm -rf /var/lib/apt/lists/*
|
|
|
|
COPY --from=build /build/bin/llama-server /usr/local/bin/llama-server
|
|
|
|
EXPOSE 8000
|
|
ENTRYPOINT ["llama-server"] |