Skip to content

[Sync][rebuild] FlashQLA/GDN (#1947) + param-buffer cpu-backup (#1952) + DistOptim step ckpt rider (#1945) #155

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
CalvinXKY merged 1 commit into from
Jun 7, 2026
Merged

[Sync][rebuild] FlashQLA/GDN (#1947) + param-buffer cpu-backup (#1952) + DistOptim step ckpt rider (#1945) #155

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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: 11 additions & 0 deletions docker/Dockerfile
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -45,6 +45,17 @@ RUN git clone https://github.com/Dao-AILab/flash-attention.git && \
RUN pip install git+https://github.com/ISEEKYAN/mbridge.git@89eb10887887bc74853f89a4de258c0702932a1c --no-deps

RUN pip install flash-linear-attention==0.4.1
# FlashQLA: optional GDN backend for Qwen3.5/Qwen3-Next (--qwen-gdn-backend flashqla; requires SM90+).
# Mirrors slime's split (x86 docker/Dockerfile installs it; arm Dockerfile.gb10 makes it opt-in):
# vime has one Dockerfile, so gate on INSTALL_FLASHQLA (justfile sets =1 for the x86 release,
# leaves =0 for arm). Verified on gb200/Blackwell: the wheel builds but `import flash_qla` raises
# "FlashQLA now support sm90 only" and the install downgrades tilelang (breaks vLLM) — so arm skips.
ARG INSTALL_FLASHQLA=0
RUN if [ "${INSTALL_FLASHQLA}" = "1" ]; then \
pip install git+https://github.com/QwenLM/FlashQLA.git --no-build-isolation; \
else \
echo "Skipping FlashQLA (INSTALL_FLASHQLA=0; sm90/Hopper-only — use --qwen-gdn-backend fla)"; \
fi
RUN pip install tilelang -f https://tile-ai.github.io/whl/nightly/cu128/

# cublas + (cu13) cuda dev headers. The arm64 (sbsa) vllm/vllm-openai base ships
Expand Down
4 changes: 2 additions & 2 deletions docker/justfile
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,7 +24,7 @@ BUILDER := "vime-builder"

# Default — cu12.9, no cu marker in the tag.
build:
ARG_TAG_SUFFIX="" ARG_BUILD_EXTRA_ARGS="" just _build-digest
ARG_TAG_SUFFIX="" ARG_BUILD_EXTRA_ARGS="--build-arg INSTALL_FLASHQLA=1" just _build-digest

# cu13 variant — vLLM default-CUDA base; ENABLE_CUDA_13 builds the CUDA-13
# TransformerEngine/Triton on top.
Expand Down Expand Up @@ -71,7 +71,7 @@ build-test:
cd ..

VERSION="$(cat docker/version.txt | tr -d '\n')"
docker build -f docker/Dockerfile . --build-arg HTTP_PROXY="$http_proxy" --build-arg HTTPS_PROXY="$https_proxy" --build-arg NO_PROXY="localhost,127.0.0.1" -t "{{IMAGE}}:vime-test-${VERSION}"
docker build -f docker/Dockerfile . --build-arg HTTP_PROXY="$http_proxy" --build-arg HTTPS_PROXY="$https_proxy" --build-arg NO_PROXY="localhost,127.0.0.1" --build-arg INSTALL_FLASHQLA=1 -t "{{IMAGE}}:vime-test-${VERSION}"
docker push "{{IMAGE}}:vime-test-${VERSION}"

docker tag "{{IMAGE}}:vime-test-${VERSION}" "{{IMAGE}}:vime-test-latest"
Expand Down
Loading
Loading