William Valentin
50f2640846
Adds a custom whisper.cpp Docker image built with CMAKE_CUDA_ARCHITECTURES=120 so it actually initializes on the RTX 5070 Ti — the upstream ghcr.io/ggml-org/whisper.cpp:main-cuda only ships kernels for sm_75/80/86/90. Compose changes: - New whisper-init one-shot service downloads ggml-medium.bin and ggml-small.bin into the shared volume on first run, fixing the original crash where whisper-server tried to load a model that was never fetched. - New whisper-server-gpu service (image whisper.cpp:cuda-blackwell, built locally from ./whisper-cuda-blackwell/Dockerfile) on port 18801 — the benchmarked path (~150 ms per short clip, ~93x faster than CPU/medium with identical WER on JFK + 4 TTS samples). - Existing whisper-server (CPU/medium) moves to port 18811 as the fallback for when GPU is unavailable. Container names unchanged so monitoring and volume bindings keep working. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
92 lines
4.0 KiB
Docker
92 lines
4.0 KiB
Docker
# whisper.cpp built for NVIDIA Blackwell (compute capability 12.0 / sm_120).
|
|
#
|
|
# Why this exists: the official `ghcr.io/ggml-org/whisper.cpp:main-cuda` image
|
|
# only ships CUDA kernels for sm_75/80/86/90 (Turing -> Hopper) and includes
|
|
# no PTX, so it fails to initialize CUDA on RTX 50-series Blackwell GPUs with
|
|
# `ggml_cuda_init: failed to initialize CUDA: system has unsupported display
|
|
# driver / cuda driver combination`.
|
|
#
|
|
# Build:
|
|
# docker build -t whisper.cpp:cuda-blackwell ./whisper-cuda-blackwell
|
|
#
|
|
# Override args if needed:
|
|
# --build-arg CUDA_ARCH=120 # set to your GPU's compute capability
|
|
# --build-arg WHISPER_REF=v1.7.6 # pin to a specific whisper.cpp tag
|
|
|
|
ARG CUDA_VERSION=12.9.1
|
|
ARG UBUNTU_VERSION=24.04
|
|
|
|
# ---------- Build stage ----------
|
|
FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION} AS builder
|
|
|
|
ARG CUDA_ARCH=120
|
|
ARG WHISPER_REF=master
|
|
|
|
RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
|
|
build-essential \
|
|
cmake \
|
|
git \
|
|
ca-certificates \
|
|
&& rm -rf /var/lib/apt/lists/*
|
|
|
|
WORKDIR /src
|
|
RUN git clone --depth 1 --branch "${WHISPER_REF}" https://github.com/ggml-org/whisper.cpp.git . 2>/dev/null \
|
|
|| git clone --depth 1 https://github.com/ggml-org/whisper.cpp.git .
|
|
|
|
# Build with CUDA enabled and explicit Blackwell architecture.
|
|
# Adding "${CUDA_ARCH}-virtual" emits PTX too so JIT can target newer chips
|
|
# from the same family if Blackwell ever gets minor variants.
|
|
#
|
|
# Why the linker dance:
|
|
# `libggml-cuda.so` calls into the CUDA driver API (`cuGetErrorString`,
|
|
# `cuMemMap`, etc.) but doesn't declare a `DT_NEEDED` for libcuda. When the
|
|
# whisper-server executable links against libggml-cuda.so, ld then sees those
|
|
# unresolved symbols and refuses. We fix it two ways at once:
|
|
# 1. Symlink the driver stub to /usr/local/cuda/lib64/stubs/libcuda.so.1
|
|
# and register the dir with ldconfig so `-lcuda` resolves.
|
|
# 2. Force `-lcuda` onto the link line via CMAKE_*_LINKER_FLAGS, with
|
|
# `--no-as-needed` so ld keeps it even when ordering would normally drop it.
|
|
RUN ln -sf /usr/local/cuda/lib64/stubs/libcuda.so /usr/local/cuda/lib64/stubs/libcuda.so.1 \
|
|
&& echo "/usr/local/cuda/lib64/stubs" > /etc/ld.so.conf.d/cuda-stubs.conf \
|
|
&& ldconfig \
|
|
&& cmake -B build \
|
|
-DGGML_CUDA=ON \
|
|
-DCMAKE_CUDA_ARCHITECTURES="${CUDA_ARCH};${CUDA_ARCH}-virtual" \
|
|
-DCMAKE_BUILD_TYPE=Release \
|
|
-DCMAKE_EXE_LINKER_FLAGS="-L/usr/local/cuda/lib64/stubs -Wl,--no-as-needed -lcuda -Wl,--as-needed" \
|
|
-DCMAKE_SHARED_LINKER_FLAGS="-L/usr/local/cuda/lib64/stubs -Wl,--no-as-needed -lcuda -Wl,--as-needed" \
|
|
&& cmake --build build --config Release -j "$(nproc)" \
|
|
--target whisper-server whisper-cli whisper-bench
|
|
|
|
# ---------- Runtime stage ----------
|
|
FROM nvidia/cuda:${CUDA_VERSION}-cudnn-runtime-ubuntu${UBUNTU_VERSION}
|
|
|
|
RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
|
|
ffmpeg \
|
|
curl \
|
|
ca-certificates \
|
|
libgomp1 \
|
|
&& rm -rf /var/lib/apt/lists/*
|
|
|
|
# Binaries
|
|
COPY --from=builder /src/build/bin/whisper-server /usr/local/bin/whisper-server
|
|
COPY --from=builder /src/build/bin/whisper-cli /usr/local/bin/whisper-cli
|
|
COPY --from=builder /src/build/bin/whisper-bench /usr/local/bin/whisper-bench
|
|
|
|
# whisper.cpp builds dynamic libs by default; copy them all.
|
|
COPY --from=builder /src/build/src/libwhisper.so* /usr/local/lib/
|
|
COPY --from=builder /src/build/ggml/src/libggml.so* /usr/local/lib/
|
|
COPY --from=builder /src/build/ggml/src/libggml-base.so* /usr/local/lib/
|
|
COPY --from=builder /src/build/ggml/src/libggml-cpu.so* /usr/local/lib/
|
|
COPY --from=builder /src/build/ggml/src/ggml-cuda/libggml-cuda.so* /usr/local/lib/
|
|
|
|
# Helpful extras: the model-download script and the JFK sample.
|
|
COPY --from=builder /src/models/download-ggml-model.sh /app/models/download-ggml-model.sh
|
|
COPY --from=builder /src/samples /app/samples
|
|
|
|
RUN ldconfig
|
|
|
|
WORKDIR /app
|
|
EXPOSE 8080
|
|
ENTRYPOINT ["whisper-server"]
|