Dockerfile.py312

# The Python 3.12 image, for any model whose inference package needs a newer
# Python than the main image ships (see cluster/submit_job.py's
# resolve_k8s_image / a k8s profile's `image_overrides`). NOT a replacement
# for the main Dockerfile -- every model whose dependencies fit the main
# image's Python 3.11.10 + torch~=2.14 floor still runs from that image;
# route a model here only when it genuinely can't.
#
# Named after the Python version, not any one model, so the NEXT model that
# needs >=3.12 can potentially reuse this same image (if its dependency tree
# doesn't conflict with what's already installed below) instead of getting
# its own dedicated Dockerfile -- same axis this repo already uses for
# Dockerfile.v100 (GPU architecture, not model name). Renamed from
# Dockerfile.limix2 (2026-09-30) once a second Python-3.12-only model became
# plausible; LIMIX2 is still, as of this rename, the ONLY model actually
# built into this image -- the package-install section below (COPY/RUN pip
# install) is still LIMIX2-specific, not yet generalized, since there's only
# one real consumer to generalize FROM. If you're adding a second model here:
# check its dependencies don't conflict with LIMIX2's `torch>=2.9.1` pin (this
# image has no shared-floor protection the way the main image does -- see the
# `--no-deps`-free install below) before assuming it can share this image
# rather than needing its own.
#
# Why a separate-from-main image at all (LIMIX2's original reason, still the
# only concrete case): LIMIX2's inference package
# (github.com/limix-ldm-ai/LimiX) declares `requires-python = ">=3.12"` and
# `dependencies = ["torch>=2.9.1", ...]`. The main image is Python 3.11.10
# (inherited from its pytorch/pytorch:2.5.1-cuda12.4-cudnn9-runtime base) with
# a shared torch~=2.14 floor forced by Causilo (see pyproject.toml) -- a real,
# unconditional incompatibility (confirmed: no `pytorch/pytorch` tag pairs
# torch==2.9.1 with Python 3.12 -- pytorch/pytorch:2.9.1-cuda12.8-cudnn9-runtime
# still ships Python 3.11.14), not a workaround-able version-range overlap.
#
# Base: plain nvidia/cuda (Ubuntu 24.04 ships Python 3.12 natively, confirmed
# via `apt-get install python3.12`: Python 3.12.3), not pytorch/pytorch --
# torch itself comes from pip (the cu128 wheel bundles its own CUDA runtime
# libs, same as every other model's inference path in the main image; the
# node's NVIDIA driver + container toolkit supplies the rest, same as the main
# image's own "runtime"-not-"devel" base).
#
# Build and push (fill in your own registry -- keep this a DISTINCT tag from
# the main image's, e.g. `:py312`, never overwriting `:v1`):
#   docker build --platform linux/amd64 -f Dockerfile.py312 -t <registry>/<you>/raman-bench:py312 .
#   docker push <registry>/<you>/raman-bench:py312
# Then add `image_overrides: {"LIMIX2": "<registry>/<you>/raman-bench:py312"}`
# (plus any other model routed here) to your k8s profile (see
# cluster/profiles/k8s_example.yaml).
FROM nvidia/cuda:12.8.1-cudnn-runtime-ubuntu24.04

RUN apt-get update && apt-get install -y --no-install-recommends \
        python3.12 \
        python3.12-venv \
        python3.12-dev \
        python3-pip \
        git \
        build-essential \
    && rm -rf /var/lib/apt/lists/* \
    && ln -sf /usr/bin/python3.12 /usr/bin/python
# python3.12-venv: Ubuntu 24.04's system Python is "externally managed"
# (PEP 668) and refuses a bare `pip install` outside a venv/--break-system-packages;
# `--break-system-packages` below is fine here (this container runs nothing
# but this one pinned dependency set, unlike a dev machine's system Python).
# build-essential: same reasoning as the main Dockerfile -- torch's own triton
# JIT compilation path shells out to `cc` at call time.
# python3.12-dev: confirmed real crash without it -- triton's JIT-compiled CUDA
# utility shim (`cuda_utils.c`) `#include <Python.h>`, which build-essential's
# gcc alone doesn't provide (that only gives the compiler, not the interpreter's
# own C headers). Every LIMIX2 task failed identically until this was added:
# `fatal error: Python.h: No such file or directory`, raised during TabArena's
# mandatory warm-up step (require_warmup=True) -- fatal to the whole task even
# though the real fit code ran and produced a model afterward.

WORKDIR /app

# Install the released package WITHOUT the `[models]` extra (that extra pulls
# in causilo>=1.0.0, whose own unpinned `torch>=2.13` floor is exactly the
# conflict this dedicated image exists to avoid -- see the module-level note
# above). `[autogluon,benchmark]` covers everything LIMIX2Model
# itself needs (AbstractTorchModel lives in autogluon.tabular) plus
# run_experiment.py's own tabarena/tracking requirements, without dragging in
# every other model's third-party package (and torch constraint) along the
# way.
COPY pyproject.toml README.md requirements-limix2-git.txt ./
COPY src ./src
# --pre: same reasoning as the main Dockerfile (autogluon floor
# >=1.6.3b20260917, a prerelease).
#
# tabarena installed WITHOUT requirements-tabarena-git.txt's extras bracket
# (tabicl,ebm,search_spaces,realmlp,tabdpt,tabm,perpetualboosting,chimeraboost,
# nori) -- confirmed as a real build failure: the `nori` extra pulls
# `synthefy-nori` -> `synthefy` -> `pyvinecopulib`, which needs Eigen3 headers
# at build time (`CMake Error ... Could not find ... Eigen3`) that this image
# doesn't carry. None of those nine extras' models run in this image (it's
# LIMIX2-only), so the bare package -- same git commit pin as
# requirements-tabarena-git.txt, kept in sync by hand -- is all that's needed;
# `tabarena.models.limix_2` itself has no extra-gated dependency.
# Split into separate RUN layers (same reasoning as the main Dockerfile's own split,
# 2026-10-01): a single layer bundling torch+autogluon+tabarena+LimiX grew to several
# GB in one blob, which a registry push repeatedly stalled on indefinitely through this
# network's path. Smaller per-step layers push independently.
RUN pip install --no-cache-dir --break-system-packages --pre -e ".[autogluon,benchmark]"
# Bumped 2026-10-01: was still pinned to 3db69f9 (the commit current when this Dockerfile was
# written, 2026-09-15), two bumps behind requirements-tabarena-git.txt's actual current pin
# (9d771c3) despite this comment's own "kept in sync by hand" claim -- confirmed stale, not assumed;
# fixed when actually activating LIMIX2 for the first real cluster run. Keep this in sync with
# requirements-tabarena-git.txt's pin going forward -- there is no automated check for drift between
# the two files. Bumped with it 2026-10-09 to 742133b4 (9d771c3 + LightPFN/PrismBoost cherry-picks) and
# 2026-10-13 to 66198aba (+ many-class support for LightPFN and TabFM+).
RUN pip install --no-cache-dir --break-system-packages --pre \
    "tabarena @ git+https://github.com/mario-koddenbrock/tabarena.git@66198abac051459bdcb8e5106836b910a6a64ad4#subdirectory=packages/tabarena"
# requirements-limix2-git.txt installed WITHOUT --no-deps here, unlike the
# main image's documented install path (see that file's own comment): the
# --no-deps workaround exists only because the main image's shared
# torch~=2.14 floor (from causilo, skipped above) would otherwise get
# clobbered by LimiX's torch>=2.9.1 dependency. This image has no such floor
# to protect -- LimiX-2 is the only model running in it -- so its dependency
# tree (including its own torch>=2.9.1 pin) installs normally, letting pip's
# resolver pick a consistent version across LimiX + autogluon + tabarena
# instead of hand-pinning one.
RUN pip install --no-cache-dir --break-system-packages --pre -r requirements-limix2-git.txt
# nvtx: the one extra LimiX needs that isn't in its own declared dependency
# tree (see requirements-limix2-git.txt's own comment) -- still required even
# without --no-deps above.
RUN pip install --no-cache-dir --break-system-packages nvtx
# tabpfn-extensions: LimiX2Model routes >10-class targets through
# tabarena.utils.wrapper_utils.import_many_class_classifier (ManyClassClassifier),
# which lives in the `[models]` extra skipped above -- confirmed live: every
# cancer_cell_cooh/nh2 and rruff_mineral_raw LIMIX2 task failed with
# `ModuleNotFoundError: No module named 'tabpfn_extensions'`. torch pinned to the
# version already installed so tabpfn's own dependency tree can't move LimiX's
# torch; setuptools<80 for the same pkg_resources reason as pyproject.toml.
RUN pip install --no-cache-dir --break-system-packages --pre \
    "tabpfn-extensions>=0.4.3" "setuptools<80" \
    "torch==$(python -c 'import torch; print(torch.__version__.split("+")[0])')"

# scripts/ and cluster/ are what run_experiment.py and k8s_entrypoint.sh need
# at runtime -- same reasoning as the main Dockerfile.
COPY scripts ./scripts
COPY cluster ./cluster
COPY configs ./configs

RUN chmod +x cluster/k8s_entrypoint.sh

ENTRYPOINT ["/bin/bash", "cluster/k8s_entrypoint.sh"]