You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Could the FlyDSL team please publish a v0.1.4.post1 (or similar post-release) wheel matrix with manylinux_2_27 / manylinux_2_28 platform tags, matching what v0.1.5+ already ships? The upstream PR #386 ("support glibc2.28") that landed for v0.1.5+ would just need to be backported as a maintenance patch onto the v0.1.4 line.
Why
amd-aiter v0.1.13.post1 ships with Requires-Dist: flydsl ==0.1.4 (build pin chosen at the time v0.1.13 was cut). Downstream consumers (vLLM, ATOM, SGLang) building or installing on systems with glibc < 2.35 cannot resolve flydsl==0.1.4 because the only PyPI artifact is manylinux_2_35:
$ pip install flydsl==0.1.4
ERROR: Could not find a version that satisfies the requirement flydsl==0.1.4
(from versions: 0.1.6.dev529, 0.1.6, 0.1.7.dev551, 0.1.7)
ERROR: No matching distribution found for flydsl==0.1.4
The same error is what motivated #396 ("centos install failed"), which is still OPEN even though the fix (PR #386) has been merged for newer versions only.
This is currently blocking us in two places:
AITER's manylinux_2_28 build container (glibc 2.28) — pip install -r requirements.txt fails because flydsl==0.1.4 is not installable. We have to bump our pin to 0.1.6 to make builds pass, which then propagates as the Requires-Dist of every wheel we ship.
Partner deployments on RHEL 8 / Ubuntu 20.04 (glibc 2.31 / 2.27) — same install failure on the consumer side.
Specific ask
Either:
Option A (preferred): Tag a v0.1.4.post1 from v0.1.4 plus the cherry-pick of support glibc2.28 #386 (and any minimal companion fixes), build the manylinux_2_27/2_28 wheel matrix with the same CI as v0.1.6 uses, publish to PyPI.
Option B: Host the manylinux_2_27/2_28 v0.1.4.post1 wheels on a non-PyPI location (GitHub Release artifacts, internal Artifactory, S3) — we can wire --index-url / --find-links into our build and partner install instructions.
Option A is preferred because it keeps pip install flydsl==0.1.4 working out of the box for any downstream that has the version pinned.
Bumping our local amd-aiter pin to flydsl==0.1.6 so we can build wheels in our manylinux_2_28 container, but this changes what every consumer of amd-aiter v0.1.13.post1 sees. Reverting to the v0.1.4 line as soon as a manylinux_2_27/2_28 wheel is available.
Summary
Could the FlyDSL team please publish a
v0.1.4.post1(or similar post-release) wheel matrix with manylinux_2_27 / manylinux_2_28 platform tags, matching whatv0.1.5+already ships? The upstream PR #386 ("support glibc2.28") that landed forv0.1.5+would just need to be backported as a maintenance patch onto thev0.1.4line.Why
amd-aiter v0.1.13.post1ships withRequires-Dist: flydsl ==0.1.4(build pin chosen at the timev0.1.13was cut). Downstream consumers (vLLM, ATOM, SGLang) building or installing on systems with glibc < 2.35 cannot resolveflydsl==0.1.4because the only PyPI artifact ismanylinux_2_35:The same error is what motivated #396 ("centos install failed"), which is still OPEN even though the fix (PR #386) has been merged for newer versions only.
This is currently blocking us in two places:
AITER's manylinux_2_28 build container (glibc 2.28) —
pip install -r requirements.txtfails becauseflydsl==0.1.4is not installable. We have to bump our pin to0.1.6to make builds pass, which then propagates as theRequires-Distof every wheel we ship.Partner deployments on RHEL 8 / Ubuntu 20.04 (glibc 2.31 / 2.27) — same install failure on the consumer side.
Specific ask
Either:
Option A (preferred): Tag a
v0.1.4.post1fromv0.1.4plus the cherry-pick of support glibc2.28 #386 (and any minimal companion fixes), build the manylinux_2_27/2_28 wheel matrix with the same CI asv0.1.6uses, publish to PyPI.Option B: Host the manylinux_2_27/2_28
v0.1.4.post1wheels on a non-PyPI location (GitHub Release artifacts, internal Artifactory, S3) — we can wire--index-url/--find-linksinto our build and partner install instructions.Option A is preferred because it keeps
pip install flydsl==0.1.4working out of the box for any downstream that has the version pinned.Reference
flydsl==0.1.4pin (https://github.com/ROCm/aiter/blob/v0.1.13/setup.py#L16)Workaround we are taking now
Bumping our local
amd-aiterpin toflydsl==0.1.6so we can build wheels in our manylinux_2_28 container, but this changes what every consumer ofamd-aiter v0.1.13.post1sees. Reverting to the v0.1.4 line as soon as a manylinux_2_27/2_28 wheel is available.cc @gbyu-amd @valarLip @sunway513