# Training image for the Otaku expert LoRA, run as a Cloud Run Job on an L4.
#
# Why Cloud Run and not Vertex AI: the project has ZERO Vertex custom-training
# GPU quota in every region (every quota bucket comes back empty), and a quota
# increase is a multi-day review. Cloud Run GPU quota, on the other hand, is
# already proven -- animetix-brain runs on an nvidia-l4 there today.
#
# Built from a minimal context (this file + the script), not the repo root: the
# 674 MB repo context makes `gcloud builds submit` die on Windows with a
# temp-file lock (WinError 32), and none of it is needed here anyway.
# Pinned by digest (tag 2.5.1-cuda12.1-cudnn9-runtime, amd64, fetched
# 2026-07-19): a mutable tag rebuilt upstream would silently change the env.
FROM pytorch/pytorch:2.5.1-cuda12.1-cudnn9-runtime@sha256:831247999fbf7e08f61b3e39f6d77ee434f38f6f07f769d00db451e853878067

ENV PYTHONUNBUFFERED=1 \
    HF_XET_HIGH_PERFORMANCE=1 \
    HF_HOME=/tmp/hf

# trl/peft/transformers/accelerate/datasets/bitsandbytes are kept in lockstep
# with requirements.txt (the serving env) — enforced by
# tests/deploy/test_dockerfile_train.py; bump both together.
# trackio/hf_transfer are training-only (absent from the locks).
RUN pip install --no-cache-dir \
    "trl==0.12.1" \
    "peft==0.17.1" \
    "transformers==4.57.6" \
    "accelerate==1.13.0" \
    "datasets==5.0.0" \
    "trackio==0.31.5" \
    "bitsandbytes==0.49.2" \
    "hf_transfer==0.1.9"

WORKDIR /app
COPY hf_train_qwen35.py /app/train.py

CMD ["python", "/app/train.py"]