dan

llama-cpp-rocm-lemonade-dsv4 (b1305-gfx1151-swap-hipblaslt-pathfix)

Published 2026-08-10 05:39:29 +00:00 by dan

Installation

docker pull forge.coffee-anon.com/dan/llama-cpp-rocm-lemonade-dsv4:b1305-gfx1151-swap-hipblaslt-pathfix
sha256:f0073e85d2470f3172658afa57716cef4843aa27aee07cd3342d03365edf5bbd

Image layers

ARG RELEASE
ARG LAUNCHPAD_BUILD_ARCH
LABEL org.opencontainers.image.version=24.04
ADD file:d938ff3d4eee15d8600de84bf85eac6ecd0f20bc92dfe305dafbff0bdc974c0f in /
CMD ["/bin/bash"]
ARG LLAMACPP_ROCM_BUILD=b1305
ARG GFX_TARGET=gfx1151
LABEL maintainer=citizendaniel
LABEL description=llama.cpp ROCm/HIP (AMD lemonade prebuilt) + llama-swap — DeepSeek V4 Flash + DSpark
LABEL backend=rocm-hip-lemonade
LABEL gpu.target=gfx1151
LABEL llamacpp.rocm.build=b1305
LABEL llamacpp.sha=3581ba0c
LABEL llamacpp.dspark.pr=https://github.com/ggml-org/llama.cpp/pull/25784
LABEL llamacpp.dspark.specpr=https://github.com/ggml-org/llama.cpp/pull/25173
LABEL llamacpp.lightningindexer.pr=https://github.com/ggml-org/llama.cpp/pull/24231
LABEL llamaswap.version=197
ARG DEBIAN_FRONTEND=noninteractive
RUN |3 LLAMACPP_ROCM_BUILD=b1305 GFX_TARGET=gfx1151 DEBIAN_FRONTEND=noninteractive /bin/sh -c apt-get update && apt-get install -y --no-install-recommends ca-certificates curl libgomp1 libnuma1 libelf1 libdrm2 libdrm-amdgpu1 libatomic1 libtinfo6 zlib1g libzstd1 pciutils && rm -rf /var/lib/apt/lists/* # buildkit
COPY /opt/llamacpp /opt/llamacpp # buildkit
COPY /tmp/llama-swap /usr/bin/llama-swap # buildkit
RUN |3 LLAMACPP_ROCM_BUILD=b1305 GFX_TARGET=gfx1151 DEBIAN_FRONTEND=noninteractive /bin/sh -c set -eux; for b in /opt/llamacpp/llama-*; do [ -x "$b" ] && ln -sf "$b" "/usr/bin/$(basename "$b")"; done; echo /opt/llamacpp > /etc/ld.so.conf.d/llamacpp-rocm.conf; ldconfig # buildkit
ENV ROCBLAS_TENSILE_LIBPATH=/opt/llamacpp/rocblas/library/gfx1151
ENV HIPBLASLT_TENSILE_LIBPATH=/opt/llamacpp/hipblaslt/library/gfx1151
ENV LD_LIBRARY_PATH=/opt/llamacpp
ENV HIP_VISIBLE_DEVICES=0
ENV GPU_DEVICE_ORDINAL=0
ENV ROCBLAS_USE_HIPBLASLT=1
ENV HSA_ENABLE_SDMA=0
ARG UNIFIED_MEMORY=0
ENV GGML_CUDA_ENABLE_UNIFIED_MEMORY=0
RUN |4 LLAMACPP_ROCM_BUILD=b1305 GFX_TARGET=gfx1151 DEBIAN_FRONTEND=noninteractive UNIFIED_MEMORY=0 /bin/sh -c set -eu; build="${LLAMACPP_ROCM_BUILD:-unknown}"; /usr/bin/llama-server --help >/tmp/h.txt 2>&1 || true; grep -q 'draft-dspark' /tmp/h.txt || { echo "FATAL: --spec-type lacks draft-dspark — build ${build} predates llama.cpp PR #25784"; exit 1; }; grep -q 'ngram-mod' /tmp/h.txt || { echo "FATAL: --spec-type lacks ngram-mod — cannot chain drafters"; exit 1; }; test -x /usr/bin/llama-swap; test -f /opt/llamacpp/hipblaslt/library/gfx1151/TensileLibrary_lazy_gfx1151.dat.zlib; test -f /opt/llamacpp/hipblaslt/library/gfx1151/Kernels.so-000-gfx1151.hsaco; rm -f /tmp/h.txt; echo "OK: draft-dspark + ngram-mod + gfx1151 hipBLASLt kernels present in ${build}" # buildkit
EXPOSE [8080/tcp]
ENTRYPOINT ["/usr/bin/llama-swap"]
CMD ["--config" "/config/llama-swap/config.yaml" "--listen" "0.0.0.0:8080"]

Labels

Key Value
backend rocm-hip-lemonade
description llama.cpp ROCm/HIP (AMD lemonade prebuilt) + llama-swap — DeepSeek V4 Flash + DSpark
gpu.target gfx1151
llamacpp.dspark.pr https://github.com/ggml-org/llama.cpp/pull/25784
llamacpp.dspark.specpr https://github.com/ggml-org/llama.cpp/pull/25173
llamacpp.lightningindexer.pr https://github.com/ggml-org/llama.cpp/pull/24231
llamacpp.rocm.build b1305
llamacpp.sha 3581ba0c
llamaswap.version 197
maintainer citizendaniel
org.opencontainers.image.version 24.04
Details
Container
2026-08-10 05:39:29 +00:00
2
OCI / Docker
linux/amd64
424 MiB
Versions (3) View all