12 lines
988 B
Docker
12 lines
988 B
Docker
# syntax=docker/dockerfile:1.7
|
|
# CI-only GB10 image. The regular Spark image keeps its native SDPA defaults.
|
|
ARG SPARK_BASE_IMAGE=ghcr.io/hao-ai-lab/fastvideo/fastvideo-dev:py3.12-cuda13.0.0-latest
|
|
FROM ${SPARK_BASE_IMAGE}
|
|
|
|
# VMoBA and the Wan causal gate require FA2. Install only this wheel so the
|
|
# validated torch/flashinfer/CUTLASS stack from the base image stays intact.
|
|
ARG FLASH_ATTN_WHEEL=https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.9.22/flash_attn-2.8.3+cu130torch2.12-cp312-cp312-linux_aarch64.whl
|
|
RUN /root/.local/bin/uv pip install --python /opt/venv/bin/python --no-deps "${FLASH_ATTN_WHEEL}" && \
|
|
/opt/venv/bin/python -c "import glob, shutil; [shutil.rmtree(p) for p in glob.glob('/opt/venv/lib/python*/site-packages/flash_attn/cute')]" && \
|
|
/opt/venv/bin/python -c "import flash_attn, flashinfer; assert hasattr(flash_attn, 'flash_attn_varlen_func'); print('FA2', flash_attn.__version__, 'flashinfer', flashinfer.__version__)"
|