[AMD] Clean patch and move script to amd/

[zyzshishui]

New Dockerfile WIP

[gemini-code-assist]

Code Review

This pull request refactors the ROCm Dockerfile for MI350/MI355, switching to a new base image and updating the build process for core components like AITER and TransformerEngine. It also updates the Qwen3-4B training script and removes several obsolete run scripts. The reviewer suggests optimizing the Dockerfile by combining apt-get commands and Python dependency installations to reduce image layers, and recommends using pattern matching instead of hardcoded line numbers in sed commands to improve robustness.

[gemini-code-assist]

It is a best practice to combine apt-get update and apt-get install into a single RUN command. This reduces the number of image layers and ensures that the package list is fresh when installing. Additionally, using apt-get is preferred for non-interactive scripts, and cleaning up /var/lib/apt/lists/* helps keep the image size small.

RUN apt-get update && apt-get install -y --no-install-recommends \
    build-essential \
    cmake \
    dnsutils \
    ethtool \
    git \
    nvtop \
    rsync \
    && rm -rf /var/lib/apt/lists/*

[gemini-code-assist]

Using a hardcoded line number (459) in sed is fragile. If the upstream file aiter/ops/triton/attention/pa_mqa_logits.py changes, this command might modify the wrong line or fail. It is safer to use a pattern match to target the specific code block.

[gemini-code-assist]

The numpy installation can be combined with the requirements.txt installation to reduce the number of layers and ensure all dependencies are resolved in a single step.

RUN rm -rf /usr/lib/python3/dist-packages/jwt /usr/lib/python3/dist-packages/PyJWT* && \
    pip install -r /tmp/requirements.txt "numpy<2"