whisper.cpp: add latest

.github/workflows/docker-build.yml

@@ -11,6 +11,7 @@ on:
       - psql-cli/**/*
       - llama.cpp-base/**/*
       - llama.cpp/**/*
+      - whisper.cpp/**/*
   workflow_dispatch:  # enables manual runs
 
 jobs:
@@ -129,3 +130,33 @@ jobs:
           cd $IMAGE
           docker build --pull -t "${{ env.CI_REGISTRY_PATH }}/${{ env.IMAGE }}" .
           docker push "${{ env.CI_REGISTRY_PATH }}/${{ env.IMAGE }}"
+
+  build_whisper:
+            name: Build whisper.cpp Image
+            runs-on: ubuntu-latest
+            needs: validate
+
+            services:
+              docker:
+                image: docker:20.10.16
+                options: --privileged
+
+            env:
+              IMAGE: whisper.cpp
+              CI_REGISTRY_PATH: index.docker.io/ahoylabs
+              CI_REGISTRY_USER: ${{ secrets.CI_REGISTRY_USER }}
+              CI_REGISTRY_PASSWORD: ${{ secrets.CI_REGISTRY_PASSWORD }}
+
+            steps:
+              - name: Check out repository
+                uses: actions/checkout@v4
+
+              - name: Log in to Docker Hub
+                run: docker login -u "${{ secrets.CI_REGISTRY_USER }}" -p "${{ secrets.CI_REGISTRY_PASSWORD }}"
+
+              - name: Build whisper.cpp Docker image
+                run: |
+                  cd $IMAGE
+                  docker build --pull -t "${{ env.CI_REGISTRY_PATH }}/${{ env.IMAGE }}" .
+                  docker push "${{ env.CI_REGISTRY_PATH }}/${{ env.IMAGE }}"
+        

whisper.cpp/Dockerfile

@@ -0,0 +1,37 @@
+ARG UBUNTU_VERSION=22.04
+# This needs to generally match the container host's environment.
+# We use a slightly older version for greater compatibility
+ARG CUDA_VERSION=12.3.2
+# CUDA build image
+ARG CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
+# CUDA runtime image
+ARG CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
+# CUDA base image (excludes cublas)
+ARG CUDA_BASE_CONTAINER=nvidia/cuda:${CUDA_VERSION}-base-ubuntu${UBUNTU_VERSION}
+
+FROM ghcr.io/ggerganov/whisper.cpp:main-cuda AS whisper
+
+FROM ${CUDA_BASE_CONTAINER} AS runtime
+
+COPY --from=whisper /app/server /app/models/download-ggml-model.sh /app/
+
+RUN /bin/echo -e '#!/bin/bash\nDEBIAN_FRONTEND=noninteractive\napt-get update && apt-get install -y $@ --no-install-recommends && apt-get clean autoclean && apt-get autoremove --yes && rm -rf /var/lib/apt/lists/*' \
+    > /usr/local/sbin/apt_install_clean.sh && \
+    chmod a+x /usr/local/sbin/apt_install_clean.sh
+RUN /bin/echo -e '#!/bin/bash\nDEBIAN_FRONTEND=noninteractive\napt-get update && apt-get remove -y $@ && apt-get clean autoclean && apt-get autoremove --yes && rm -rf /var/lib/apt/lists/*' \
+    > /usr/local/sbin/apt_remove_clean.sh && \
+    chmod a+x /usr/local/sbin/apt_remove_clean.sh
+
+# we need just CUDA and CUBLAS
+# this saves ~1GB vs the -runtime image
+RUN /usr/local/sbin/apt_install_clean.sh libcublas-12-3 curl wget ffmpeg
+
+ENV MODEL="small-q5_1"
+ENV PORT="8200"
+ENV ADDITIONAL_ARGS="--host 0.0.0.0 --convert --inference-path /whisper"
+
+# mount volume here
+ENV WORKSPACE="/workspace"
+
+COPY entry.sh /
+ENTRYPOINT [ "/entry.sh" ]

whisper.cpp/entry.sh

@@ -0,0 +1,28 @@
+#!/bin/sh
+
+# we expect the following env variables
+# MODEL="small-q5_1"
+# PORT="8080"
+# ADDITIONAL_ARGS="--metrics --host 0.0.0.0--convert --inference-path /whisper"
+
+# also handle create /public and populate
+
+# set WORKSPACE to "/" if $WORKSPACE not a directory/doesn't exist
+if [ ! -d "$WORKSPACE" ]; then
+  WORKSPACE="/"
+fi
+
+cd $WORKSPACE
+mkdir -p models
+
+if [ ! -f "models/ggml-${MODEL}.bin" ]; then
+  /app/download-ggml-model.sh $MODEL /models
+fi
+MODEL_ARG="-m $WORKSPACE/models/ggml-${MODEL}.bin"
+
+# this is the path we use for health checks
+mkdir -p public/v1
+echo "alive" > public/v1/models
+PUBLIC_ARG="--public $WORKSPACE/public"
+
+/app/server $MODEL_ARG --port $PORT $ADDITIONAL_ARGS $PUBLIC_ARG