Skip to content

ci(build): bump FA4, quack and ffpa-attn #3733

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
akoumpa merged 5 commits into from
Aug 30, 2026
Merged

ci(build): bump FA4, quack and ffpa-attn #3733

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
bc096d6
build: bump FA3/FA4 pin to pick up CuTe compile-key fixes
akoumpa Aug 28, 2026
325d1ed
build: pin full flash-attention commit
akoumpa Aug 28, 2026
9fe6cfd
build: bump cutlass-dsl to 4.6.2, ffpa-attn to 0.2.3, quack-kernels t…
akoumpa Aug 29, 2026
cd642d4
build: enable FA4 (INSTALL_FA4=true) by default
akoumpa Aug 29, 2026
06bf919
build: correct the apache-tvm-ffi comment
akoumpa Aug 30, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 7 additions & 4 deletions docker/Dockerfile
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -183,9 +183,12 @@ RUN mkdir -p /wheels && \
rm -rf /tmp/torchao

# FA3 (Hopper-only, SM90a-locked, ~25 min nvcc) + FA4 cute (Blackwell, lightweight). arm64 -> skip FA3: GB200 target, QEMU cross-build impractical.
ARG FLASH_ATTN_REF=002cce0a1
# Pin >= #2762/#2507: older refs rebuilt the CuTe compile key every call when max_seqlen
# is a tensor (varlen/packed), so HF models recompiled each step. Also picks up #2745
# (fwd dynamic-shape correctness) and #2819 (faster SM100 mask compile).
ARG FLASH_ATTN_REF=ce088ab9ce0fc0434dcd8afa0a791da9fcc3a820
ARG INSTALL_FA3=true
ARG INSTALL_FA4=false
ARG INSTALL_FA4=true
ARG TARGETARCH
RUN if [ "$TARGETARCH" = "arm64" ]; then INSTALL_FA3=false; fi && \
if [ "$INSTALL_FA3" = "true" ] || [ "$INSTALL_FA4" = "true" ]; then \
Expand All @@ -210,10 +213,10 @@ WORKDIR /opt/Automodel
# torchao MXFP8 + FA3/FA4: prebuilt wheels from the wheel_builder stage, installed into system site-packages
# before uv sync — pip targets system deps, uv the venv (torchao/FA gated `never`), so uv sync won't prune or shadow them.
COPY --from=wheel_builder /wheels /tmp/wheels/
ARG INSTALL_FA4=false
ARG INSTALL_FA4=true
RUN pip install --no-cache-dir --no-deps /tmp/wheels/*.whl && \
if [ "$INSTALL_FA4" = "true" ]; then \
pip install --no-cache-dir "nvidia-cutlass-dsl[cu13]==4.6.0.dev0" && \
pip install --no-cache-dir "nvidia-cutlass-dsl[cu13]==4.6.2" && \
FA2_DIR=$(python -c "import flash_attn, os; print(os.path.dirname(flash_attn.__file__))") && \
VENV_CUTE=/opt/venv/lib/python3.12/site-packages/flash_attn/cute && \
{ [ -e "$FA2_DIR/cute" ] || ln -s "$VENV_CUTE" "$FA2_DIR/cute"; }; \
Expand Down
Loading
Loading