Slim simulator image (multi-stage + runtime-only deps)

.dockerignore

@@ -0,0 +1,67 @@
+# Keeps the Docker build context small for scripts/sim.Dockerfile.
+# .git is intentionally NOT excluded: scripts/install-sim.sh calls
+# `git submodule update` in the builder stage; the runtime stage drops
+# all .git/ subtrees so they never reach the final image.
+
+# Caches / generated artefacts
+**/__pycache__
+**/*.pyc
+**/*.pyo
+.pytest_cache
+.ruff_cache
+.mypy_cache
+
+# Local venvs (install-vllm.sh / install-vllm-cpu.sh)
+.venv
+.venv-cpu
+
+# Module outputs that can be hundreds of MB
+profiler/perf
+bench/results
+outputs
+
+# Non-runtime artefacts: tests, planning docs, the docs site. The
+# image is meant for running `python -m serving`, not for hacking
+# inside; iteration is expected via the dev container (docker-sim.sh)
+# with the repo bind-mounted.
+tests
+agent_plan
+docs
+
+# Host-side build artefacts. The builder stage compiles ASTRA-Sim fresh
+# in /build inside the container, so any host-side outputs would only
+# bloat the build context (and astra-sim/build/ contains the build.sh
+# entry point, so the exclusion has to be scoped to the CMake output
+# subtree only — NOT all of astra-sim/build/).
+astra-sim/build/astra_analytical/build/
+**/CMakeCache.txt
+**/CMakeFiles/
+**/cmake_install.cmake
+**/Makefile.cmake
+**/*.o
+**/*.a
+**/*.so
+**/*.dylib
+
+# Python packaging artefacts (root-level only — pip install -e for a
+# subdirectory submodule creates its own egg-info we want to keep).
+/build/
+/dist/
+*.egg-info
+
+# Notebook checkpoints
+**/.ipynb_checkpoints
+
+# Docs site (Docusaurus build deps + node_modules pull a lot)
+docs/node_modules
+docs/build
+docs/.docusaurus
+
+# Editor / OS junk
+.DS_Store
+.idea
+.vscode
+*.swp
+
+# CI metadata that's irrelevant inside the image
+.github

scripts/README.md

@@ -14,7 +14,7 @@ live with their module — only setup and build helpers are here.
 | `install-vllm-cpu.sh` | Bare-metal vLLM **CPU** install. Builds vLLM 0.19.0 from source with `VLLM_TARGET_DEVICE=cpu` into `.venv-cpu` (separate from the GPU venv). Works on x86_64 and aarch64 (NVIDIA Grace). Used by `profiler/profile-cpu.sh`. |
 | `install-sim.sh`  | Bare-metal simulator install for fresh Docker containers / minimal Linux hosts (root or non-root). Installs apt build deps, inits submodules, pip-installs sim Python deps, builds ASTRA-Sim's analytical backend, and installs the Chakra converter (`--no-deps`). Skips vLLM — profiler/bench need `install-vllm.sh` and a GPU. Idempotent. |
 | `compile.sh`      | Build ASTRA-Sim's analytical backend and install the Chakra trace converter. |
-| `sim.Dockerfile`  | Ubuntu 24.04 base + `install-sim.sh`. Built and published to GHCR by `.github/workflows/build-sim-image.yml` on every push to `main` touching the install / build paths. Pull with `docker pull ghcr.io/psal-postech/llmservingsimspec/sim:latest`. |
+| `sim.Dockerfile`  | Multi-stage Ubuntu 24.04 image. Stage 1 compiles ASTRA-Sim with the C++ toolchain; stage 2 carries only the simulator runtime — python + the deps `serving/` actually imports + the compiled binary + chakra. No compilers, no `.git`, no `transformers / datasets / xgboost / matplotlib / scikit-learn` (those live in the vLLM image, `docker-vllm.sh`). `.dockerignore` keeps the build context lean (no `perf/`, `results/`, venv, CMake outputs). Built and published to GHCR by `.github/workflows/build-sim-image.yml` on every push to `main` touching the install / build paths. Pull with `docker pull ghcr.io/psal-postech/llmservingsimspec/sim:latest`. |
 
 ## Typical first-time setup
 

scripts/sim.Dockerfile

@@ -1,45 +1,87 @@
-# LLMServingSim simulator image.
+# LLMServingSim simulator image (slim, multi-stage).
 #
 # Built by .github/workflows/build-sim-image.yml on every push to main
 # that touches scripts/install-sim.sh, the astra-sim submodule, or this
 # file. Published to ghcr.io/<owner>/<repo>/sim.
 #
+# Stage 1 (builder) carries the C++ toolchain to compile ASTRA-Sim;
+# Stage 2 (runtime) carries only the python interpreter, the
+# simulator's runtime python deps, the compiled ASTRA-Sim binary, the
+# chakra package, and the repo source — no compilers, no headers, no
+# .git history. .dockerignore at the repo root strips perf/results/
+# venv/CMake artefacts from the context so neither stage sees them.
+#
+# install-sim.sh also pulls workload-generator / bench / power-model
+# training deps (transformers, datasets, scikit-learn, xgboost,
+# matplotlib). Those are NOT installed here: the sim image is for
+# `python -m serving` only, and workload generation lives in the vLLM
+# docker image (scripts/docker-vllm.sh).
+#
 # Local build:
 #   docker build -f scripts/sim.Dockerfile -t llmservingsim-sim .
-#
-# Local run (mount your working copy over /workspace if you want live
-# edits — the image carries the repo state at build time):
-#   docker run -it --rm -v "$(pwd)":/workspace llmservingsim-sim bash
 
+# ============================================================================
+# Stage 1: builder — toolchain for ASTRA-Sim native build
+# ============================================================================
+FROM ubuntu:24.04 AS builder
+
+ENV DEBIAN_FRONTEND=noninteractive
+
+RUN apt-get update -qq \
+ && apt-get install -y -qq --no-install-recommends \
+        git ca-certificates \
+        build-essential cmake \
+        protobuf-compiler libprotobuf-dev \
+ && rm -rf /var/lib/apt/lists/*
+
+WORKDIR /build
+COPY . /build
+
+# Init submodules, build the analytical backend, then aggressively
+# prune everything we don't need at runtime: CMake build trees,
+# downloaded cmake _deps, object files, every .git directory. The
+# runtime stage copies the leftover tree as-is.
+RUN git config --global --add safe.directory '*' \
+ && git submodule update --init --recursive --depth 1 \
+ && bash astra-sim/build/astra_analytical/build.sh \
+ && rm -rf astra-sim/build/astra_analytical/build/CMakeFiles \
+           astra-sim/build/astra_analytical/build/_deps \
+           astra-sim/build/astra_analytical/build/CMakeCache.txt \
+           astra-sim/build/astra_analytical/build/cmake_install.cmake \
+           astra-sim/build/astra_analytical/build/Makefile \
+ && find astra-sim -name '*.o' -delete \
+ && find astra-sim -name '*.a' -delete \
+ && find . -type d -name .git -prune -exec rm -rf {} + \
+ && find . -type d -name __pycache__ -prune -exec rm -rf {} +
+
+# ============================================================================
+# Stage 2: runtime — slim, only what's needed to RUN the simulator
+# ============================================================================
 FROM ubuntu:24.04
 
 ENV DEBIAN_FRONTEND=noninteractive \
-    PIP_BREAK_SYSTEM_PACKAGES=1
+    PIP_BREAK_SYSTEM_PACKAGES=1 \
+    PIP_NO_CACHE_DIR=1 \
+    PIP_DISABLE_PIP_VERSION_CHECK=1
 
-# Pre-install the apt packages install-sim.sh expects. Doing this in a
-# dedicated layer lets Docker cache them independently of the script's
-# pip / submodule / build steps further down, which is what dominates
-# rebuild time when the source changes.
+# Runtime apt: python + pip + tls roots. No compilers, no git, no
+# protobuf C++ runtime — the python protobuf wheel ships its own.
 RUN apt-get update -qq \
  && apt-get install -y -qq --no-install-recommends \
-        git build-essential cmake \
-        protobuf-compiler libprotobuf-dev \
         python3 python3-pip python-is-python3 \
-        ca-certificates curl \
- && rm -rf /var/lib/apt/lists/*
+        ca-certificates \
+ && rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*
 
 WORKDIR /workspace
+COPY --from=builder /build /workspace
 
-# Copy the repo (including submodules — the workflow's checkout step
-# fetches them on the runner so they're in the build context).
-COPY . /workspace
-
-# install-sim.sh is idempotent and detects pre-installed apt packages,
-# so its apt phase is a no-op here; the work it actually does in this
-# layer is: python deps, ASTRA-Sim analytical-backend build, chakra
-# install. See the script for the full sequence.
-RUN bash scripts/install-sim.sh
+# Simulator runtime python deps. Narrower than scripts/install-sim.sh
+# on purpose — see the header comment.
+RUN pip3 install --quiet --no-input \
+        pyyaml pyinstrument msgspec \
+        pandas numpy rich protobuf \
+ && pip3 install --quiet --no-input --no-deps \
+        /workspace/astra-sim/extern/graph_frontend/chakra \
+ && find / -type d -name __pycache__ -prune -exec rm -rf {} + 2>/dev/null || true
 
-# Default entry: interactive shell at the repo root. Match docker-sim.sh
-# behaviour so the two surfaces (CI image / dev script) feel the same.
 CMD ["/bin/bash"]