Added gfx950a to pyt_mochi_inference.ubuntu dockerfile to enable fa

[vadseshu]

Motivation

PR to fix the pyt_mochi_video_inference enablment on Mi350/ Mi355.

Technical Details

Updated gfx arch for Mi350 on mochi inference dockerfile

Test Plan

Test Result

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[Copilot]

Pull request overview

Updates the AMD Ubuntu Mochi inference Dockerfile to include the MI350/MI355 GPU architecture in the ROCm arch list so flash-attention can be built/enabled on those GPUs.

Changes:

• Add gfx950 to PYTORCH_ROCM_ARCH in pyt_mochi_inference.ubuntu.amd.Dockerfile.

Comments suppressed due to low confidence (1)

docker/pyt_mochi_inference.ubuntu.amd.Dockerfile:40

• PYTORCH_ROCM_ARCH is a semicolon-delimited list, but it’s later expanded unquoted in the GPU_ARCHS=$(echo ${PYTORCH_ROCM_ARCH} | ...) command and then used in an env-var assignment. In sh, the semicolons will be treated as command separators (e.g., echo gfx950;gfx90a;...), which can break the Docker build. Quote the expansion and ensure the computed GPU_ARCHS value is passed/assigned in a way that preserves the semicolons (or convert the list to a delimiter that’s shell-safe before using it).

ARG PYTORCH_ROCM_ARCH=gfx950;gfx90a;gfx942;gfx1100;gfx1101;gfx1200;gfx1201
RUN git clone ${FA_REPO}
RUN cd flash-attention \
    && git submodule update --init \
    && GPU_ARCHS=$(echo ${PYTORCH_ROCM_ARCH} | sed -e 's/;gfx1[0-9]\{3\}//g') python3 setup.py bdist_wheel --dist-dir=dist \

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The PR title mentions adding gfx950a, but the Dockerfile adds gfx950. Please confirm the intended ROCm arch string for MI350/MI355 and align either the Dockerfile value or the PR title/description to avoid confusion for downstream users.