Dockerfile.vllm

Rama main · réplica de lectura del repositorio original.

# Runtime image: vLLM engine + fornax entrypoint (specs/003).
# Engine version pinned per specs/001 — bump deliberately.
#
# v0.28.0 (2026-08-26) is the floor for the gb10 pool (specs/019): it
# publishes linux/arm64 alongside amd64 and builds SM121 kernels, and it
# postdates Qwen3.8-27B's release so the architecture is in it
# (specs/022).
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 vllm/vllm-openai:v0.32.0
COPY --from=build /out/fornax /usr/local/bin/fornax
# Run:ai streamer enables load: s3-stream; s5cmd covers nvme-cache.
RUN pip install --no-cache-dir "vllm[runai]" runai-model-streamer runai-model-streamer-s3 \
    && 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"]