Dockerfile.gfx906

# viiwork docker image built with the gfx906-stripped llama.cpp fork.
#
# This Dockerfile is the cutover target for Phase 1 of the gfx906 fork
# work. The original Dockerfile is left untouched as the rollback path.
#
# Build via:
#
#   docker build -t viiwork:gfx906 \
#     --build-context fork=/home/janit/gfx906-work/llama.cpp-gfx906 \
#     -f Dockerfile.gfx906 .
#
# (or `make docker-gfx906`).
#
# The named build context `fork` is required - it points at the local
# clone of llama.cpp-gfx906 that holds the Phase 1 commits. The fork
# tree is COPYd in instead of git-cloned because we don't push it to a
# remote during local validation.

# Stage 1: Build llama.cpp-gfx906 with ROCm/gfx906
FROM rocm/dev-ubuntu-24.04:6.2.4-complete AS llama-build
RUN apt-get update && apt-get install -y cmake git
COPY --from=fork . /llama.cpp
WORKDIR /llama.cpp
# The fork already carries the gfx906 fp8 sed as a real commit
# (`fix(hip): disable __hip_fp8_e4m3 path for gfx906`) so no in-place
# patching is needed here.
RUN cmake -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx906 \
    && cmake --build build --target llama-server llama-cli -j$(nproc)
# llama-cli is included alongside llama-server for `rocprof`-based
# kernel profiling (Phase 2 of the gfx906 fork plan). rocprof v1 in
# ROCm 6.2 cannot cleanly wrap llama-server through a driver subshell
# but wraps single-shot llama-cli directly. Including the second
# binary permanently is ~10 MB and saves us from maintaining a
# parallel "-prof" image variant.
# Smoke test: both binaries run and can find HIP runtime
RUN ldd /llama.cpp/build/bin/llama-server && /llama.cpp/build/bin/llama-server --help > /dev/null
RUN ldd /llama.cpp/build/bin/llama-cli    && /llama.cpp/build/bin/llama-cli    --help > /dev/null

# Stage 2: Build viiwork (unchanged from the original Dockerfile)
FROM golang:1.23.6 AS go-build
WORKDIR /src
COPY go.mod go.sum ./
RUN go mod download
COPY . .
ARG VERSION=dev
RUN CGO_ENABLED=0 go build -ldflags "-X main.version=${VERSION}" -o /viiwork ./cmd/viiwork

# Stage 3: Runtime (unchanged from the original Dockerfile)
FROM rocm/dev-ubuntu-24.04:6.2.4-complete
RUN apt-get update && apt-get install -y --no-install-recommends ipmitool && rm -rf /var/lib/apt/lists/*
COPY --from=llama-build /llama.cpp/build/bin/llama-server /usr/local/bin/
COPY --from=llama-build /llama.cpp/build/bin/llama-cli    /usr/local/bin/
COPY --from=llama-build /llama.cpp/build /tmp/llama-build/
RUN find /tmp/llama-build -name '*.so*' -exec cp {} /usr/local/lib/ \; && rm -rf /tmp/llama-build && ldconfig
COPY --from=go-build /viiwork /usr/local/bin/
ENV HSA_OVERRIDE_GFX_VERSION=9.0.6
EXPOSE 8080
ENTRYPOINT ["viiwork"]
CMD ["--config", "/etc/viiwork/viiwork.yaml"]