• llvmlite, numba, and llvm versions

    From Diane Trout@21:1/5 to All on Tue Feb 18 19:20:01 2025
    Hello,

    llvmlite barely managed to be updated to llvm-15, and then llvm-15 was
    removed from Debian.

    There's a bit of work on llvmlite to support llvm19 https://github.com/numba/llvmlite/pull/1092
    but it looks like it doesn't work yet.

    Unfortunately for a number of python tools, numba can't be installed
    without llvmlite, and I'm not sure how we can have a version of
    llvmlite available for the next release?

    It can be built against llvm-15, but we'd somehow need that to be added
    back to python, and then there's hoping that llvmlite could be updated
    to llvm-19.

    After writing this email I'm going to try building the llvm19 branch to
    see what happens but as all the reported tests had failed, and the logs
    had been deleted, I don't have high hopes.

    Do you have any ideas of what could be done to help get a version of
    llvmlite that works with numba into Debian?

    Diane

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)
  • From M. Zhou@21:1/5 to Diane Trout on Tue Feb 18 20:00:01 2025
    On Tue, 2025-02-18 at 09:50 -0800, Diane Trout wrote:

    Do you have any ideas of what could be done to help get a version of
    llvmlite that works with numba into Debian?

    No idea. I'm keeping an eye on upstream release but the only option
    I can see to make it work is to depend on the LLVM version that does
    not exist in unstable.

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)
  • From Diane Trout@21:1/5 to M. Zhou on Mon Mar 3 05:00:01 2025
    On Tue, 2025-02-18 at 13:41 -0500, M. Zhou wrote:
    On Tue, 2025-02-18 at 09:50 -0800, Diane Trout wrote:

    Do you have any ideas of what could be done to help get a version
    of
    llvmlite that works with numba into Debian?

    No idea. I'm keeping an eye on upstream release but the only option
    I can see to make it work is to depend on the LLVM version that does
    not exist in unstable.

    I found some time and energy to rebase the 1092 llvm-19 compatibility
    branch against 0.44.0
    https://github.com/detrout/llvmlite/tree/llvm19

    I reported on the PR that the rebase almost works. https://github.com/numba/llvmlite/pull/1092#issuecomment-2693122561

    2 tests fail on x86_64 and 0 tests fail on arm64, both out of 379
    tests.

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)
  • From PICCA Frederic-Emmanuel@21:1/5 to All on Tue Mar 18 14:30:02 2025
    No idea. I'm keeping an eye on upstream release but the only option
    I can see to make it work is to depend on the LLVM version that does
    not exist in unstable.

    I found some time and energy to rebase the 1092 llvm-19 compatibility
    branch against 0.44.0
    https://github.com/detrout/llvmlite/tree/llvm19

    I reported on the PR that the rebase almost works. https://github.com/numba/llvmlite/pull/1092#issuecomment-2693122561

    2 tests fail on x86_64 and 0 tests fail on arm64, both out of 379
    tests.


    some news ?

    friendly ping :))

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)
  • From Diane Trout@21:1/5 to PICCA Frederic-Emmanuel on Wed Mar 19 07:20:01 2025
    On Tue, 2025-03-18 at 14:08 +0100, PICCA Frederic-Emmanuel wrote:


    some news ?

    friendly ping :))


    I was waiting to see if I could get some advice from upstream.

    I added a bit more to the upstream bug about what's different between
    llvm 15 and 19 for one of the failing test.

    https://github.com/numba/llvmlite/pull/1092

    I pushed my backport of the llvm-19 patch and general package updates
    to llvm-19 to https://salsa.debian.org/diane/llvmlite

    This currently has a patch to mark the two failing tests as
    "expectedFailures" because I thought it'd be useful to take this build
    and see how well numba works with it.

    Diane

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)
  • From Diane Trout@21:1/5 to All on Wed Apr 9 08:10:01 2025
    Hi,

    I have a version of llvmlite that builds against llvm-19 with 2 test
    failures that don't look to important. One of them is definitely just
    the layout of the object file changed from what was expected.

    I got most of the help in the comments in here: https://github.com/numba/llvmlite/pull/1092

    Additionally with some other patches I was able to get numba to build
    with this version of llvmlite and with numpy 2.2 and all it's tests
    passed on x86_64. It also got through some tests on arm64 before it
    timed out.

    I pushed all my changes to llvmlite to
    https://salsa.debian.org/diane/llvmlite
    at commit 17e84a9243ba92a9248e7415c84a980e0ff96886

    The important bit is this patch from upstream https://salsa.debian.org/diane/llvmlite/-/blob/master/debian/patches/llvm-19-support-pr1182.patch?ref_type=heads
    And this patch skips/expects the two failed tests https://salsa.debian.org/diane/llvmlite/-/blob/master/debian/patches/skip-llvm-19-broken-tests.patch?ref_type=heads

    I also pushed all my changes to numba for it to build with the above
    version of llvmlite and numpy 2.2 https://salsa.debian.org/science-team/numba.git

    (numba should work, though I had to merge a small change to the tests
    that I forgot to push to salsa a while ago, and it's building again,
    but I'm tired and it'll take a while to run through all the tests
    again)

    I hope it's still not too late to get numba back in to testing.

    Diane

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)
  • From Nicholas D Steeves@21:1/5 to Diane Trout on Thu Apr 10 03:00:01 2025
    Hi Diane!

    Diane Trout <diane@ghic.org> writes:

    Hi,

    I have a version of llvmlite that builds against llvm-19 with 2 test
    failures that don't look to important. One of them is definitely just
    the layout of the object file changed from what was expected.

    I got most of the help in the comments in here: https://github.com/numba/llvmlite/pull/1092
    [snip]
    I also pushed all my changes to numba for it to build with the above
    version of llvmlite and numpy 2.2 https://salsa.debian.org/science-team/numba.git

    (numba should work, though I had to merge a small change to the tests
    that I forgot to push to salsa a while ago, and it's building again,
    but I'm tired and it'll take a while to run through all the tests
    again)

    I hope it's still not too late to get numba back in to testing.

    Has a bug been filed against llvmlite? If not, shouldn't there be one?
    It seems to be that the worst case scenario would be a friendly NMU, and
    a bug is required for that.

    If llvmlite counts as a toolchain package then release team approval for
    an unblock will also be needed. The release team tends to get
    progressively busier as the release approaches, so filing an unblock authorisation bug sooner rather than later demonstrates courtesy and
    goodwill imho. I believe that you will find support for a cherry-picked llvmlite fix, because you have been working towards supporting Trixie's
    release goals since February :)

    I'm not sure if this link:

    https://lists.debian.org/msgid-search/128a5dff67746a5120e3ab36650282b267a3652d.camel@ghic.org

    should go at the llvmlite bug, or the release team bug, or both.


    Please feel free to keep me in CC, because I'd appreciate the
    encouragement of seeing that our processes still work.

    Kind regards,
    Nicholas

    -----BEGIN PGP SIGNATURE-----

    iQJEBAEBCgAuFiEE4qYmHjkArtfNxmcIWogwR199EGEFAmf3F7cQHHN0ZW5AZGVi aWFuLm9yZwAKCRBaiDBHX30QYb2dD/9TF8o8PmqENwL5c2Mqkqsv9QeBrFnUULuG YaYkzgkarAE/5/2fgTy7CrWNqqrin3mFsfS2ONIVxCPlQ0Z6IVa9Ft9IFQFXgDex yZJ0KNfos8PNs+s8zSDPAoVR+VJSRRWtbS/lPzGxj0eHTjsSELheFF/UhckKedEr blenkWzAXs+YRALB3gqmHVh8fscM05V0Wnlw1YQaJF8MYMVlmKa3VxH6FSYzo6gU ktQfvhF20XgNQ7SpsjZdKwJ9p1yJB7svbmxNZSbaMkDjjY9BiP8C4CtcZPfmUI4g BM3IANYcqgKzveKcWDozBagGH9/tJ57bM4WnMpmtDvnfYO4t6VtHlkt1Loune83Y n1X3Z8avTqFaxIBqj67cG6TemFhaQJyfMxL4lNGNjbvMwczlSWiptMJRmoEGceFh YcVzjYwIerrWkrggxfq92BSCQR4woBYOmDY+z16k4H0AL+7n1d4Tx6nw17QBbQea nS91dtpFRgJJ5ALgmc0ZXYWM8pXSnWdHKJF3qYFJK82SUbwtdDb4ki02dP4tRz4L 2vEsJnMwIZF2xT8lyBEyQQ9ZMCtOuBfYiGP3mnJkMH2n1r0ripsX7+LTr/K6LoXJ aLFGPXswNWS3lqmNQCMs/OWnd1vcPeAQCf0RprNex0uJirDHDHGhR0AtrMqz3DQY
    Bh8yY5HuKg==
    =av3m
    -----END PGP SIGNATURE-----

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)
  • From M. Zhou@21:1/5 to Diane Trout on Thu Apr 10 03:20:01 2025
    Hi Diane,

    Thank you for working on this.

    Do you have LLVM team write access? If so please help your self and
    directly push the commits there. If not, please open a merge request
    and I'll process it quickly.

    On Tue, 2025-04-08 at 22:59 -0700, Diane Trout wrote:
    Hi,

    I have a version of llvmlite that builds against llvm-19 with 2 test
    failures that don't look to important. One of them is definitely just
    the layout of the object file changed from what was expected.

    I got most of the help in the comments in here: https://github.com/numba/llvmlite/pull/1092

    Additionally with some other patches I was able to get numba to build
    with this version of llvmlite and with numpy 2.2 and all it's tests
    passed on x86_64. It also got through some tests on arm64 before it
    timed out.

    I pushed all my changes to llvmlite to https://salsa.debian.org/diane/llvmlite
    at commit 17e84a9243ba92a9248e7415c84a980e0ff96886

    The important bit is this patch from upstream https://salsa.debian.org/diane/llvmlite/-/blob/master/debian/patches/llvm-19-support-pr1182.patch?ref_type=heads
    And this patch skips/expects the two failed tests https://salsa.debian.org/diane/llvmlite/-/blob/master/debian/patches/skip-llvm-19-broken-tests.patch?ref_type=heads

    I also pushed all my changes to numba for it to build with the above
    version of llvmlite and numpy 2.2 https://salsa.debian.org/science-team/numba.git

    (numba should work, though I had to merge a small change to the tests
    that I forgot to push to salsa a while ago, and it's building again,
    but I'm tired and it'll take a while to run through all the tests
    again)

    I hope it's still not too late to get numba back in to testing.

    Diane

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)
  • From Diane Trout@21:1/5 to Nicholas D Steeves on Thu Apr 10 06:10:01 2025
    On Wed, 2025-04-09 at 20:58 -0400, Nicholas D Steeves wrote:

    Hello,


    Has a bug been filed against llvmlite?  If not, shouldn't there be
    one?

    So it's not clear if there is a bug here.

    The problem is the llvmlite upstream is currently release is targeting
    llvm-15, which debian removed for being.

    Picking up a fix someone else made and updating it I got llvmlite to
    build against llvm-19 which is in debian, they helped clean up my fixes
    a bit more and suggested a new PR to use.

    Updating to 19 is just listed in their
    "Continuous Discussin: LLVM update plan" https://github.com/numba/llvmlite/issues/1048


    It seems to be that the worst case scenario would be a friendly NMU,
    and
    a bug is required for that.

    Ah ok. the regular maintainer for llvmlite M. Zhou responded a bit
    after you and it looks like they're willing to update it shortly.


    If llvmlite counts as a toolchain package then release team approval
    for
    an unblock will also be needed.  The release team tends to get

    I don't know if llvmlite is a toolchain package, though it also fell
    out of testing because the version of llvm it used was removed, so this
    package looks to be an improvement over no package.


    llvmlite fix, because you have been working towards supporting
    Trixie's
    release goals since February :)

    I don't lot of free time, and was hoping upstream would fix it but
    realized we were almost out of time and managed to get the last bits
    pushed over the finish line.


    I'm not sure if this link:

      https://lists.debian.org/msgid-search/128a5dff67746a5120e3ab36650282b267a3652d.camel@ghic.org

    should go at the llvmlite bug, or the release team bug, or both.

    Ok.

    I guess my plan is pass the fixes to llvmlite to lumin, then push
    numba, see how things build, and then go talk to the release team?

    Diane

    -----BEGIN PGP SIGNATURE-----

    iQIzBAABCgAdFiEETQVcMeSBIEX5AQ11mQ04NnM013AFAmf3RBoACgkQmQ04NnM0 13CobA//aSLR6ia75+d3TaOokSzb29sbVKxTImJatXNaBglRZMDi8mXYdeYsAMNE yv1pqIEyx2F7bqerChD1LOlR2Us5ygRP+md8l47wfdkZhkblt3kMFUvQ8mY3Doet /L+zOkdz5xOyBJb92AXkNdGCwSGN+e414c90Fq6YTsSprb8CpyG/yBODj1IfvSIE aeMzFhq3VvybLGjP/OpkvfTx7k1MdSKVky2B0moPRaxzWlRXp87aAYKdyU4ii+uy evzPSaQoLtc+6I9MYfgp2tr4VMBLPetegrEI8te+KxiD7Nfv4Hojm0/WcwuyoipK orfWm/JI4vr58JEhaqq0fqU8wTnrM/TEBYH1i4X2bMwqldFKuFtgPBIiT1fxnl6x Zk7ydWY/ld2aSUUQcV8ziwE8aHYX5vJAiZrH1QovD2WmZ9/LIQjadiaKdIdnqVYA rhP19pfXRJXtu+fbhRx+qY1LnA6P2xGytazW8GeIH3JVtABFPbKgdXgNs7qEWFOU 9bsTzDbvI/fOGaK933u2wJqmmopg/xTS5ECzj46C89e9fthermTrQnp9i+Vz1FaV eK6bQ+nFZIWJu2kmIffcbpfJtxccynye4yEit4asXI+fxnMmBAjp/H5aaP+2Lc6l d0Ccql/bYf6Zo0sEopOLv7Vl6hUV0UtFZ52HvEgaGNJBdOWQyvU=
    =RJCe
    -----END PGP SIGNATURE-----

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)
  • From Diane Trout@21:1/5 to M. Zhou on Thu Apr 10 06:20:01 2025
    On Wed, 2025-04-09 at 21:00 -0400, M. Zhou wrote:
    Hi Diane,

    Thank you for working on this.

    Do you have LLVM team write access? If so please help your self and
    directly push the commits there. If not, please open a merge request
    and I'll process it quickly.


    I don't have llvm team write access, which is good because I almost accidentally pushed a change to your repository instead of mine.

    I don't know if you need the pristine-tar and upstream branches as
    merge requests, as I haven't done a salsa based pull request before.

    but here's the important one.

    https://salsa.debian.org/pkg-llvm-team/llvmlite/-/merge_requests/5

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)