# Flint serving image: SGLang v0.5.16 + the flint_sglang model package + the FA4 learned-sink fix.
# Build from the directory that contains this file, install.sh, the flint_sglang wheel and the patch:
#   docker build -t flint-sglang .
# Host requirement: NVIDIA driver >= 580 (CUDA 13), or CUDA forward compatibility on older
# datacenter drivers (the image ships it in /usr/local/cuda/compat).
# lmsysorg/sglang:v0.5.16-cu130 (multi-arch index digest the kit was validated with)
FROM lmsysorg/sglang@sha256:7b6a35df9839fd593a94a1eaee82d7777f472225d9f3ad1f8a2e0cb2bd1785d0

COPY install.sh fa4-learned-sink-fix.patch flint_sglang-*.whl /tmp/flint-build/
RUN bash /tmp/flint-build/install.sh /tmp/flint-build && rm -rf /tmp/flint-build

ENV FLASHINFER_VERSION=0.6.16.post1 \
    SGLANG_EXTERNAL_MODEL_PACKAGE=flint_sglang \
    FLASHINFER_USE_CUDA_NORM=1