build(container): backport EFA build fixes (#9705, #9727) to release/1.2.1 - #10425
Conversation
…#9705) Signed-off-by: Yifan Jiang <19356972+yifjiang@users.noreply.github.com> Signed-off-by: Jie Hao <jihao@nvidia.com>
…#9727) Signed-off-by: Yifan Jiang <yifjiang@users.noreply.github.com> Co-authored-by: Yifan Jiang <yifjiang@users.noreply.github.com> Signed-off-by: Jie Hao <jihao@nvidia.com>
Signed-off-by: Erez Zarum <erezz@amazon.com> Signed-off-by: Jie Hao <jihao@nvidia.com>
4945715 to
6a3b62a
Compare
dynamo-ops
left a comment
There was a problem hiding this comment.
Review: backport of EFA build fixes to release/1.2.1
The diff is clean. Cherry-picks of #9705, #9727, and #10047 applied correctly to the release branch. Conflict resolution in wheel_builder.Dockerfile (keeping yasm + adding --setopt=tsflags=nocontexts) is correct.
Two minor notes:
-
PR description lists #9705 + #9727 only, but the actual diff includes #10047 (
libfabric2.5.1 for EFA and TRTLLM) as well — therm -rf /opt/amazon/efa && cp -Pfrreplacement approach and the trtllm NIXL plugin block are from that PR. Consider updating the description for traceability. -
Dead exports in trtllm RUN block (cosmetic, non-blocking):
export LD_PRELOAD=/opt/nvidia/nvda_nixl/lib64/libnixl.so && \ export NIXL_PLUGIN_DIR=/opt/nvidia/nvda_nixl/lib64/plugins && \
These have no effect inside
RUN(theENVstatements below handle persistence). Could be removed to reduce confusion, but not a real issue since they match the subsequentENVvalues.
dynamo-ops
left a comment
There was a problem hiding this comment.
Previous review comments have been addressed. Approving.
|
Test failures are unrelated to change. |
Summary
Cherry-pick of two EFA container build fixes from
maintorelease/1.2.1:build(container): bump nixl_gdrcopy_ref to v2.5.2 (kernel >=6.15 fix)feat(container): configurable libfabric repo + v2.5.1 overlay for EFATogether these fix the Dynamo EFA container build (GDRCopy kernel ≥6.15 compatibility, configurable libfabric repo, and the libfabric v2.5.1 overlay needed for the EFA RDMA path).
Original PRs
fffea8cb4e0916fb35fc8491839cba893353180cd515c5f8c78cbacdda1e67616cf4836ce0a29968Conflicts resolved
Cherry-pick of #9727 conflicted in two files; resolved as follows:
container/context.yaml— took the incoming side (configurablenixl_libfabric_repo+nixl_libfabric_ref: v2.5.1); this is the core of the fix.container/templates/wheel_builder.Dockerfile— combined both sides: kept release branch'sgit yasm(needed for the ffmpeg/libvpx build present on 1.2.x) and added feat(container): configurable libfabric repo + v2.5.1 overlay for EFA #9727's--setopt=tsflags=nocontexts.#9705applied cleanly.Test plan
🤖 Generated with Claude Code