refactor ascend Dockerfile

[yao-fengchen]

No description provided.

[windreamer]

Maybe we can use multistage build of docker buildx to simplify this Dockerfile, here is an example to illustrate:

ARG ASCEND_DEVICE=a3
ARG ASCEND_HUB=swr.cn-south-1.myhuaweicloud.com/ascendhub

FROM ${ASCEND_HUB}/cann:8.1.rc1-a3-openeuler22.03-py3.10 AS a3_base

FROM ${ASCEND_HUB}/cann:8.1.rc1-910b-ubuntu22.04-py3.10 AS a2_base

FROM ${ASCEND_HUB}/cann:8.1.rc1-310p-ubuntu22.04-py3.10 AS 300i_base

FROM ${ASCEND_DEVICE}_base AS builder
ENV LMDEPLOY_TARGET_DEVICE=ascend
WORKDIR /opt/lmdeploy
COPY . .
RUN --mount=type=cache,target=/root/.cache \
    pip config set global.index-url https://mirrors.aliyun.com/pypi/simple && \
    pip config set global.trusted-host mirrors.aliyun.com && \
    pip install --no-cache-dir -U pip build && \
    python -m build -w -o /wheels -v .

FROM ${BASE_IMAGE} AS final
RUN --mount=type=cache,target=/root/.cache \
    pip config set global.index-url  https://mirrors.aliyun.com/pypi/simple  && \
    pip config set global.trusted-host  mirrors.aliyun.com  && \
    pip install --no-cache-dir torch==2.3.1 torch-npu==2.3.1 torchvision==0.18.1

RUN --mount=type=cache,target=/wheels,from=builder,source=/wheels \
    pip install --no-cache-dir /wheels/*.whl
ENTRYPOINT []

[windreamer]

And also, we can consider to use https://hub.docker.com/r/ascendai/pytorch as base image to minimize our effort?
Not quite sure abort if the ascend device type matters.

[yao-fengchen]

Maybe we can use multistage build of docker buildx to simplify this Dockerfile, here is an example to illustrate:

ARG ASCEND_DEVICE=a3
ARG ASCEND_HUB=swr.cn-south-1.myhuaweicloud.com/ascendhub

FROM ${ASCEND_HUB}/cann:8.1.rc1-a3-openeuler22.03-py3.10 AS a3_base

FROM ${ASCEND_HUB}/cann:8.1.rc1-910b-ubuntu22.04-py3.10 AS a2_base

FROM ${ASCEND_HUB}/cann:8.1.rc1-310p-ubuntu22.04-py3.10 AS 300i_base

FROM ${ASCEND_DEVICE}_base AS builder
ENV LMDEPLOY_TARGET_DEVICE=ascend
WORKDIR /opt/lmdeploy
COPY . .
RUN --mount=type=cache,target=/root/.cache \
    pip config set global.index-url https://mirrors.aliyun.com/pypi/simple && \
    pip config set global.trusted-host mirrors.aliyun.com && \
    pip install --no-cache-dir -U pip build && \
    python -m build -w -o /wheels -v .

FROM ${BASE_IMAGE} AS final
RUN --mount=type=cache,target=/root/.cache \
    pip config set global.index-url  https://mirrors.aliyun.com/pypi/simple  && \
    pip config set global.trusted-host  mirrors.aliyun.com  && \
    pip install --no-cache-dir torch==2.3.1 torch-npu==2.3.1 torchvision==0.18.1

RUN --mount=type=cache,target=/wheels,from=builder,source=/wheels \
    pip install --no-cache-dir /wheels/*.whl
ENTRYPOINT []

ok, i will simplify this Dockerfile in the last commit

[jinminxi104]

And also, we can consider to use https://hub.docker.com/r/ascendai/pytorch as base image to minimize our effort? Not quite sure abort if the ascend device type matters.

I think Huawei's official Ascend Hub offers a more network-friendly experience for all users.

[lvhan028]

Can we install requirements_ascend.txt instead?

[yao-fengchen]

Our current stable torch-npu versions on A2 and A3 are different. A2 uses torch npu==2.3.1, while A3 uses torch npu==2.7.1rc1. This Dockerfile is for A2 and 300i, we temporarily write torch npu==2.3.1 here. We will ensure that A2 runs stably on high torch-npu versions in the future.

[windreamer]

you can have different environment variable defined accordingly, for example:

FROM ${ASCEND_HUB}/cann:8.1.rc1-910b-ubuntu22.04-py3.10 AS ascend_a2_base
ENV TORCH_VERSION=2.3.1

FROM ${ASCEND_HUB}/cann:8.1.rc1-XXX-ubuntu22.04-py3.10 AS ascend_a2_base
ENV TORCH_VERSION=SOME_VERSION_ELSE

......

RUN pip install --no-cache-dir torch==${TORCH_VERSION} torch-npu==${TORCH_VERSION}

[yao-fengchen]

you can have different environment variable defined accordingly, for example:

FROM ${ASCEND_HUB}/cann:8.1.rc1-910b-ubuntu22.04-py3.10 AS ascend_a2_base
ENV TORCH_VERSION=2.3.1

FROM ${ASCEND_HUB}/cann:8.1.rc1-XXX-ubuntu22.04-py3.10 AS ascend_a2_base
ENV TORCH_VERSION=SOME_VERSION_ELSE

......

RUN pip install --no-cache-dir torch==${TORCH_VERSION} torch-npu==${TORCH_VERSION}

This Dockerfile is for A2 and 300i, A2 and 300i are the same except for base_image. Due to some internal problems, we only provide images for A3 for download for the time being.