Dockerfile.synthesis-api

# Synthesis-as-a-Service container: nsynth Rust synthesizer behind the
# stdlib-only HTTP server in ncpu/synthesis_api. Built for Cloud Run.
#
#   gcloud builds submit --config cloudbuild.synthesis-api.yaml .
#
# Stage 1 compiles the mog_synth release binary; stage 2 is a slim Python
# runtime that shells out to it. No torch, no pip installs — the server is
# stdlib-only by design.

FROM rust:1-bookworm AS build
WORKDIR /src
COPY nsynth ./nsynth
RUN cargo build --release --manifest-path nsynth/Cargo.toml --bin mog_synth

FROM python:3.12-slim
WORKDIR /app
# default_backend_path() resolves <repo>/nsynth/target/release/mog_synth
# relative to the server module, so mirror that layout.
COPY --from=build /src/nsynth/target/release/mog_synth /app/nsynth/target/release/mog_synth
COPY ncpu/__init__.py /app/ncpu/__init__.py
COPY ncpu/synthesis_api /app/ncpu/synthesis_api
# /prompt pipeline: deterministic prompt parser + cascade (stdlib-only;
# torch imports in self_optimizing are function-local and never hit on
# the /prompt verify path).
COPY ncpu/autoresearch /app/ncpu/autoresearch
COPY ncpu/self_optimizing/__init__.py /app/ncpu/self_optimizing/__init__.py
COPY ncpu/self_optimizing/humaneval_runner.py /app/ncpu/self_optimizing/humaneval_runner.py

# Memory banks land in $HOME — /tmp is the only writable fs on Cloud Run.
# Banks persist for the instance lifetime: warm instances answer repeat
# problems from the solved cache in milliseconds.
ENV HOME=/tmp
EXPOSE 8080

# Public-tier limits: 30s default solve, 60s hard ceiling, 2 concurrent
# solves (excess → 429), CORS open (read-only compute, no credentials).
CMD exec python3 -m ncpu.synthesis_api.server \
    --host 0.0.0.0 --port ${PORT:-8080} \
    --timeout 30 --max-timeout 60 --max-concurrency 2 \
    --cors-origin '*'