# Patch image: reuse all Python/CUDA deps from an existing LMDeploy image, only
# overlay this repo and reinstall the lmdeploy package in editable mode.
#
# Build (from repository root); default BACKEND=pytorch (no TurboMind compile):
#   docker build -f docker/Dockerfile_patch \
#     --build-arg BASE_IMAGE=openmmlab/lmdeploy:v0.10.0-cu12 \
#     --build-arg BACKEND=pytorch \
#     --build-arg http_proxy=$http_proxy \
#     --build-arg https_proxy=$https_proxy \
#     --build-arg no_proxy=$no_proxy \
#     -t lmdeploy:test .
#
# TurboMind extension:
#   --build-arg BACKEND=turbomind
#
# BACKEND=pytorch: skip TurboMind C++ build (PyTorch engine only).
# Any other value (e.g. turbomind): build with TurboMind extension.
#
ARG BASE_IMAGE=openmmlab/lmdeploy:latest

FROM ${BASE_IMAGE}

ARG BACKEND=pytorch

WORKDIR /opt/lmdeploy

COPY . .

# --no-deps: do not pull install_requires (torch, transformers, ...).
# --no-build-isolation: use cmake_build_extension etc. already present in the base image.
RUN if [ "$BACKEND" = "pytorch" ]; then \
        DISABLE_TURBOMIND=1 pip install --no-cache-dir --no-build-isolation -e . --no-deps; \
    else \
        pip install --no-cache-dir --no-build-isolation -e . --no-deps; \
    fi