# syntax=docker/dockerfile:1.3-labs

ARG FROM_IMAGE_NAME
FROM ${FROM_IMAGE_NAME} as mcore_image
ENV PIP_CONSTRAINT="" 
RUN pip3 install -U pip

FROM mcore_image as build_te
ARG TE_COMMIT=bee4649c15a79ffcb9689ca7c0c963f5febaa28a
WORKDIR /opt
COPY patches/nemo_2.3.0_te.patch .
RUN \
    git clone https://github.com/NVIDIA/TransformerEngine.git && \
    cd TransformerEngine && \
    git fetch origin ${TE_COMMIT} && \ 
    git checkout ${TE_COMMIT} && \
    patch -p1 < /opt/nemo_2.3.0_te.patch && \
    git submodule init && git submodule update && \
    rm /opt/nemo_2.3.0_te.patch && \
    pip3 wheel --no-cache-dir -v .

FROM mcore_image as build_causal_conv1d
WORKDIR /opt
RUN CAUSAL_CONV1D_FORCE_BUILD=TRUE pip3 wheel --no-cache-dir -v git+https://github.com/Dao-AILab/causal-conv1d.git@v1.2.2.post1

FROM mcore_image as build_grouped_gemm
WORKDIR /opt
RUN pip3 wheel --no-cache-dir -v git+https://github.com/fanshiqing/grouped_gemm@v1.1.2

FROM mcore_image as build_experimental_flash_attention
WORKDIR /opt
ARG EXPERIMENTAL_FLASH_ATTN_VERSION=c0f04c0b6c747914d95205867d86dd19c027d01d
RUN --mount=type=secret,id=EXPERIMENTAL_FLASH_ATTN \
    EXPERIMENTAL_FLASH_ATTN=$(cat /run/secrets/EXPERIMENTAL_FLASH_ATTN) && \
    pip uninstall -y ninja && \
    pip install --no-cache-dir ninja && \
    MAX_JOBS=4 pip wheel --no-cache-dir -v $EXPERIMENTAL_FLASH_ATTN@${EXPERIMENTAL_FLASH_ATTN_VERSION} && \
    ls -al

FROM mcore_image as build_mamba_ssm
WORKDIR /opt
RUN git clone https://github.com/state-spaces/mamba.git && \
    cd mamba && \
    git checkout v2.2.0 && \
    sed -i "/triton/d" setup.py && \
    MAMBA_FORCE_BUILD=TRUE pip3 wheel --no-cache-dir -v .

FROM mcore_image as main
ENV DEBIAN_FRONTEND=noninteractive

RUN apt-get update && \
    apt-get install -y --no-install-recommends gettext python3-venv && \
    apt-get clean && \
    python -m venv /opt/jet && \
    wget https://github.com/mikefarah/yq/releases/download/v4.44.1/yq_linux_amd64 -O /usr/local/bin/yq && \
    chmod a+x /usr/local/bin/yq

COPY --from=build_causal_conv1d /opt/causal_conv1d-*.whl ./
COPY --from=build_grouped_gemm /opt/grouped_gemm-*.whl ./
COPY --from=build_mamba_ssm /opt/mamba/mamba_ssm-*.whl ./
COPY --from=build_te /opt/TransformerEngine/transformer_engine-*.whl ./

RUN \
    --mount=type=bind,source=requirements,target=requirements \
    --mount=type=bind,source=pyproject.toml,target=pyproject.toml \
    --mount=type=bind,source=setup.py,target=setup.py \
    --mount=type=bind,source=megatron/core/package_info.py,target=megatron/core/package_info.py \
    --mount=type=bind,source=megatron/core/README.md,target=megatron/core/README.md \
    --mount=type=bind,source=megatron/core/requirements.txt,target=megatron/core/requirements.txt \
    --mount=type=bind,source=requirements_mlm.txt,target=requirements_mlm.txt \
    --mount=type=bind,source=requirements_ci.txt,target=requirements_ci.txt \
    --mount=type=bind,source=megatron/core/__init__.py,target=megatron/core/__init__.py <<"EOF" bash -ex
pip install -U pip
pip install --no-cache-dir causal_conv1d-*.whl mamba_ssm-*.whl grouped_gemm-*.whl transformer_engine*.whl
PY_ENV=pytorch_25.03 pip install --no-cache-dir . -r requirements_mlm.txt -r requirements_ci.txt
EOF

# Since megatron does not have any dependencies (and isn't a dependency to any other package), we can install it separately to make everything a bit quicker
ARG MCORE_REPO
ARG MCORE_REF
ARG MCORE_BACKWARDS_REF
RUN <<"EOF" bash -exu
# Checkout latest
cd /opt
rm -rf /opt/megatron-lm; mkdir megatron-lm; cd megatron-lm
git init
git remote add origin ${MCORE_REPO}
git fetch origin '+refs/merge-requests/*:refs/remotes/merge-requests/*'
git fetch origin $MCORE_REF
git checkout $MCORE_REF

# Checkout backwards-ref
cd /opt
rm -rf /opt/megatron-lm-legacy; mkdir megatron-lm-legacy; cd megatron-lm-legacy
git init
git remote add origin ${MCORE_REPO}
git fetch origin $MCORE_BACKWARDS_REF
git checkout $MCORE_BACKWARDS_REF
rm -rf megatron; cp -a /opt/megatron-lm/megatron ./
EOF

RUN <<"EOF" bash -ex
pip install -U pip
PY_ENV=pytorch_25.03 pip install --no-cache-dir -e /opt/megatron-lm
EOF

ENV PYTHONPATH="/opt/megatron-lm:$PYTHONPATH"

##### For NVIDIANS only #####
FROM main as jet
ARG CACHEBUST=0
# COPY --from=build_experimental_flash_attention /opt/*.whl ./experimental_flash_attention/
ARG EXPERIMENTAL_FLASH_ATTN_VERSION=c0f04c0b6c747914d95205867d86dd19c027d01d
COPY --from=build_experimental_flash_attention /opt/*.whl ./experimental_flash_attention/
RUN --mount=type=secret,id=JET_INDEX_URLS \
    --mount=type=secret,id=LOGGER_INDEX_URL \
    --mount=type=secret,id=EXPERIMENTAL_FLASH_ATTN \
    LOGGER_INDEX_URL=$(cat /run/secrets/LOGGER_INDEX_URL) && \
    JET_INDEX_URLS=$(cat /run/secrets/JET_INDEX_URLS) && \
    EXPERIMENTAL_FLASH_ATTN=$(cat /run/secrets/EXPERIMENTAL_FLASH_ATTN) && \
    pip install --no-cache-dir "jet-client~=2.0" jet-api --upgrade $JET_INDEX_URLS  && \
    pip install --no-cache-dir "one-logger" --upgrade $LOGGER_INDEX_URL && \
    pip install --no-cache-dir --no-build-isolation ./experimental_flash_attention/*flash_attn*.whl

ENV PATH="$PATH:/opt/jet/bin"
###