Dockerfile.sglang

Rama main · réplica de lectura del repositorio original.

# Runtime image: SGLang engine + fornax entrypoint (specs/003).
# Engine version pinned per specs/001 — bump deliberately.
#
# SGLANG_IMAGE selects the base. Two images are built from this one file
# because they differ only in that base:
#
#   default  lmsysorg/sglang:v0.5.19-cu129        -> fornax-runtime-sglang
#   K3       lmsysorg/sglang:kimi-k3-...          -> fornax-runtime-sglang-k3
#
# They stay separate *images* on purpose (specs/018 AC4): K3 support lives
# in a CUDA 13 branch build needing an r580+ driver, and folding it into
# the shared image would impose that driver on the h200/b200 pools for a
# model that cannot run there anyway. That is a base-image difference, not
# a Dockerfile difference — hence one file, two --build-arg invocations.
#
# v0.5.16 is the floor for DeepSeek-V4-Flash-0731's DSpark path
# (specs/017). The previous pin, v0.5.15.post1-cu126, was never a real
# upstream tag — cu126 variants stop before v0.5.14.
#
# This ARG must stay above the first FROM: an ARG declared after a stage's
# instructions is scoped to that stage, and the later FROM cannot see it.
ARG SGLANG_IMAGE=lmsysorg/sglang:v0.5.19-cu129

FROM golang:1.27 AS build
WORKDIR /src
COPY go.mod go.sum ./
RUN go mod download
COPY . .
RUN CGO_ENABLED=0 go build -o /out/fornax ./cmd/fornax

FROM ${SGLANG_IMAGE}
COPY --from=build /out/fornax /usr/local/bin/fornax
# s5cmd for nvme-cache weight sync from S3.
RUN pip install --no-cache-dir huggingface_hub[hf_transfer] \
    && curl -fsSL https://github.com/peak/s5cmd/releases/download/v2.3.0/s5cmd_2.3.0_Linux-64bit.tar.gz \
       | tar -xz -C /usr/local/bin s5cmd
EXPOSE 8000
ENTRYPOINT ["fornax", "serve"]
CMD ["--manifest", "/etc/fornax/model.yaml"]