aidevices's picture
Sync experiment for HF Jobs Docker image
a36b814 verified
Raw History Blame Contribute Delete
5.97 kB
# Image for the ARChitects ARC Prize 2024 reproduction (SCRUM-80) on HuggingFace Jobs.
#
# NOTE: there is intentionally NO ENTRYPOINT/CMD. HF Jobs issues `python <script_path>`
# (IMAGE_SCRIPT_PATH, default /app/train.py), so an entrypoint here would conflict.
#
# BASE IMAGE β€” why a bare CUDA image and not pytorch/pytorch:
# An earlier revision used `pytorch/pytorch:2.5.1-cuda12.4-cudnn9-runtime`, which ships its
# own conda torch. `pip install unsloth` then installed a *second*, newer torch over the top
# into the same site-packages. torch._inductor got loaded twice and `import unsloth` died with
# `AssertionError: duplicate template name` before any of our code ran.
# Fix: start from a CUDA image with no torch at all and let unsloth pull the torch it wants.
# One torch, no collision.
#
# PYTHON β€” why the distro's stock python3 (3.10) and not python3.11:
# A later revision installed `python3.11`, but Ubuntu 22.04's universe only carries
# `3.11.0~rc1` (a release candidate), while `python3-pip` remains pip for 3.10. Symlinking
# python->3.11 therefore ran 3.10's pip under a 3.11 interpreter, and the build failed.
# Stock `python3` gives one interpreter with its matching pip. unsloth supports >=3.9,<3.15.
#
# GUARD β€” do NOT `import unsloth` at build time:
# `import unsloth` probes for a GPU on import (unsloth_zoo/device_type.py calls
# get_device_type() at module scope) and raises
# `NotImplementedError: Unsloth cannot find any torch accelerator? You need a GPU.`
# The HF Space builder is a CPU machine (SPACE_HARDWARE: cpu-basic), so such a guard can
# never pass there β€” it made the build fail unconditionally. Import torch (CPU-safe) and
# assert unsloth is *installed* via importlib.metadata, which reads metadata without
# importing the package.
#
# `-base` (not `-runtime`/`-cudnn-runtime`): torch's wheel bundles its own CUDA libraries as
# nvidia-* packages (cuDNN, cuBLAS, NCCL, cuSPARSE, … ~2.9 GB of wheels). A cudnn-runtime base
# ships those libs too, so we were installing CUDA twice and the image ballooned to ~11.6 GB.
# The base tag provides just the driver interface, which is all torch needs.
# unsloth/bitsandbytes ship prebuilt wheels, so no nvcc (`-devel`) is required.
#
# Build time is substantial (torch + CUDA wheels + unsloth). The HF Space builds this image, so
# SPACE_BUILD_TIMEOUT_SECONDS must be generous (>= 3600).
FROM nvidia/cuda:12.4.1-base-ubuntu22.04
ENV PYTHONUNBUFFERED=1 \
PYTHONDONTWRITEBYTECODE=1 \
PYTHONPATH=/app:/app/vendor/architects \
HF_HUB_DISABLE_TELEMETRY=1 \
HF_HOME=/app/.hf_cache \
DEBIAN_FRONTEND=noninteractive
WORKDIR /app
# The CUDA base ships no Python, so install the distro's stock one.
#
# Use `python3` (Ubuntu 22.04's stable 3.10), NOT `python3.11`. Jammy's universe repo only
# carries Python `3.11.0~rc1` β€” a release candidate β€” and `python3-pip` is pip for 3.10, so
# symlinking python->3.11 made `python -m pip` run 3.10's pip under a 3.11 interpreter. That
# mismatch broke the previous build. unsloth supports >=3.9,<3.15 and cp310 wheels exist for
# torch/triton/bitsandbytes, so 3.10 needs no source builds.
#
# `gcc` is required at RUN TIME, not build time: unsloth's kernels are Triton, and Triton
# JIT-compiles a small C shim (its `driver.c`) on the first kernel launch. Without a compiler the
# job dies mid-inference with
# RuntimeError: Failed to find C compiler. Please specify via CC environment variable.
# (triton/runtime/build.py looks for `gcc`, then `clang`.) Triton ships its own `cuda.h` and
# dlopen()s `libcuda.so.1` from the driver, so no CUDA toolkit or `-devel` base image is needed β€”
# just a compiler.
#
# git is needed by some unsloth/transformers install paths.
RUN apt-get update && apt-get install -y --no-install-recommends \
python3 python3-dev python3-pip git gcc \
&& rm -rf /var/lib/apt/lists/* \
&& ln -sf /usr/bin/python3 /usr/bin/python
# One interpreter, one pip. Fail loudly here rather than debug a mismatch later.
RUN python -c "import sys; assert sys.version_info[:2] == (3, 10), sys.version" \
&& python -m pip --version
# Assert the C compiler Triton needs, so a missing one fails this free CPU build instead of
# surfacing mid-inference on a paid GPU (which is exactly how it was found).
RUN command -v gcc >/dev/null || { echo "gcc missing: Triton JIT will fail at kernel launch"; exit 1; }
COPY requirements.txt .
# unsloth resolves and installs its own torch build here β€” do not pre-install torch.
RUN python -m pip install --no-cache-dir --upgrade pip \
&& python -m pip install --no-cache-dir -r requirements.txt \
&& python -m pip install --no-cache-dir huggingface_hub
# Sanity-check the install on the (CPU) builder, so a broken environment fails the free Space
# build rather than a paid GPU job. `import torch` is safe without a GPU and would still catch
# the torch._inductor double-install that killed the first run. unsloth is only checked for
# presence β€” importing it here would raise "You need a GPU" on the CPU builder.
RUN python -c "\
import torch, importlib.metadata as md; \
print('torch', torch.__version__, '| unsloth', md.version('unsloth'), '| cuda available:', torch.cuda.is_available())"
COPY . .
# Import the vendored modules we actually use. None of them import unsloth at module scope
# (model_tools defers it into a function), so this is CPU-safe β€” and it catches library-API
# drift for free. The 2024 code is resolved against current libraries, so imports break as
# upstream APIs move: `from trl import DataCollatorForCompletionOnlyLM` (removed in trl 0.20.0)
# failed here only after a paid GPU job had already started. Now it fails the Space build.
RUN python -c "\
import arc_loader, model_tools, inference_tools, selection, arc_downloader; \
import importlib.metadata as md; \
print('vendored modules import OK | trl', md.version('trl'), '| transformers', md.version('transformers'))"