python3Packages.vllm: 0.10.1.1 -> 0.10.2#447722
Merged
GaetanLepage merged 2 commits intoNixOS:masterfrom Oct 4, 2025
Merged
Conversation
nixpkgs-ci
bot
added
10.rebuild-linux: 1-10
Contributor
Author
|
happysalada
approved these changes
Oct 2, 2025
nixpkgs-ci
bot
added
12.approvals: 1
Contributor
Author
|
I'm actually getting a build failure when building the package with CUDA support. See this link for more details and logs. Relevant lines from failed build log: -- FlashMLA is available at /nix/store/ab1x9ra2sc9k9r2nxsrhhcm0izncxbsz-flashmla-1.0.0
-- Build type: RelWithDebInfo
-- Target device: cuda
-- Found Python: /nix/store/829wb290i87wngxlh404klwxql5v18p4-python3-3.13.7/bin/python3.13 (found version "3.13.7") found components: Interpreter Development.Module Development.SABIModule
CMake Warning at /nix/store/s9xqh0qzpk19hx5n0i52r2rf2fpxhw9a-vllm-flash-attn-2.7.4.post1/CMakeLists.txt:75 (message):
Pytorch version 2.4.0 expected for CUDA build, saw 2.8.0 instead.
-- CUDA target architectures: 7.5;8.0;8.6;8.9;9.0;10.0;12.0
-- CUDA supported target architectures: 8.0;8.6;8.9;9.0;10.0;12.0
-- FA2_ARCHS: 8.0+PTX
-- FA3_ARCHS: 9.0a;8.0
-- vllm-flash-attn is available at /nix/store/s9xqh0qzpk19hx5n0i52r2rf2fpxhw9a-vllm-flash-attn-2.7.4.post1
-- Configuring done (28.7s)
CMake Error at /nix/store/a94d5zmalqava26y3hqsnj5l11l5kl5y-cmake-3.31.7/share/cmake-3.31/Modules/FindPython/Support.cmake:4240 (add_library):
Cannot find source file:
/nix/store/ab1x9ra2sc9k9r2nxsrhhcm0izncxbsz-flashmla-1.0.0/csrc/kernels_fp8/flash_fwd_mla_fp8_sm90.cu
Call Stack (most recent call first):
/nix/store/a94d5zmalqava26y3hqsnj5l11l5kl5y-cmake-3.31.7/share/cmake-3.31/Modules/FindPython.cmake:692 (__Python_add_library)
cmake/utils.cmake:462 (Python_add_library)
cmake/external_projects/flashmla.cmake:53 (define_gpu_extension_target)
CMakeLists.txt:942 (include)
CMake Error at /nix/store/a94d5zmalqava26y3hqsnj5l11l5kl5y-cmake-3.31.7/share/cmake-3.31/Modules/FindPython/Support.cmake:4240 (add_library):
No SOURCES given to target: _flashmla_C
Call Stack (most recent call first):
/nix/store/a94d5zmalqava26y3hqsnj5l11l5kl5y-cmake-3.31.7/share/cmake-3.31/Modules/FindPython.cmake:692 (__Python_add_library)
cmake/utils.cmake:462 (Python_add_library)
cmake/external_projects/flashmla.cmake:53 (define_gpu_extension_target)
CMakeLists.txt:942 (include)
CMake Generate step failed. Build files cannot be regenerated correctly. |
Contributor
Author
|
Command to reproduce building with CUDA support from my fork, for reference: nix-build \
-I nixpkgs=https://github.com/daniel-fahey/nixpkgs/archive/fbc1629b8775a9eeb932f1b79d14fa973adb67ec.tar.gz \
--expr 'with import <nixpkgs> { config = { allowUnfree = true; cudaSupport = true; }; }; python313Packages.vllm'Edit: see nixbuild.net failed build #4558385 for derivation |
Contributor
|
Contributor
|
Thanks for the PR @daniel-fahey ! |
Contributor
Author
Not yet, good suggestion! I'm busy with other stuff so am just showing my working as I go along in drips-and-drabs in-case someone else needs to pick this up. |
GaetanLepage
force-pushed
the
update-vllm
branch
3 times, most recently
from
3c5d6cc to
9058b43
Compare
daniel-fahey
force-pushed
the
update-vllm
branch
2 times, most recently
from
142c7e5 to
e7fb896
Compare
Contributor
|
Contributor
|
Now the CPU build fails with: |
GaetanLepage
changed the title
Contributor
|
How was it building on this branch earlier today??? |
daniel-fahey
force-pushed
the
update-vllm
branch
from
44bd507 to
e7fb896
Compare
Contributor
Author
|
Contributor
Author
This builds the working Python 3.12 CPU version: [daniel@laptop:~/Source/nixpkgs]$ nix-build -I nixpkgs=https://github.com/daniel-fahey/nixpkgs/archive/fbc1629b8775a9eeb932f1b79d14fa973adb67ec.tar.gz --expr 'with import <nixpkgs> { }; python312Packages.vllm'
/nix/store/zdm5svz7ygij3dj13djsizm1zggbnapw-python3.12-vllm-0.10.2And this is the broken one: [daniel@laptop:~/Source/nixpkgs]$ nix-build -I nixpkgs=https://github.com/daniel-fahey/nixpkgs/archive/e7fb8960241ab2211c471004bcd2b04683767c33.tar.gz --expr 'with import <nixpkgs> { }; python312Packages.vllm'
this derivation will be built:
/nix/store/j98dgdx87kjahfj73cd0wvzp3c0px4dh-python3.12-vllm-0.10.2.drv
building '/nix/store/j98dgdx87kjahfj73cd0wvzp3c0px4dh-python3.12-vllm-0.10.2.drv' on 'ssh://eu.nixbuild.net'...
copying 0 paths...
[nixbuild.net] Cached build failure. The id of the previously failed build is 4565380. See this link for the failed build details and logs: https://nixbuild.net/builds/4565380?t=EtkBCm8KBWJ1aWxkCgpidWlsZDpyZWFkGAMiCQoHCAcSAxD5BSINCgsIBBIHOgUKAxiBCDImCiQKAggbEgYIBRICCAUaFgoECgIIBQoICgYg9eyp0wYKBBoCCAAyFgoUCgIIGxIOCAISAxiACBIFEITTlgISJAgAEiAGRu1AM51kzPYGl4nShP2nj4SwLZdIQEZhBaIO6KS15xpATQzlvJ7Y-SLutaWbRFkbSrm3qOrarrHr60v_1kS5vx08FGnOtkZSZjwNvMDFTJwiO1O-MfmWac0BNAXtrNz8ASIiCiAuWrCOa6uQpEanWP7s1LJ8iKWZhtugtJoeboULO7JD1w==
[nixbuild.net] To turn off build failure caching, see https://docs.nixbuild.net/settings/#reuse-build-failures
error: build of '/nix/store/j98dgdx87kjahfj73cd0wvzp3c0px4dh-python3.12-vllm-0.10.2.drv' on 'ssh://eu.nixbuild.net' failed: Cached build failure (build 4565380): Cannot build '/nix/store/j98dgdx87kjahfj73cd0wvzp3c0px4dh-python3.12-vllm-0.10.2.drv'.
Reason: builder failed with exit code 1.
Output paths:
/nix/store/cfj51613p4smfza7150mzsk21smfrzxl-python3.12-vllm-0.10.2
/nix/store/l2gjwc39xss9hg63rbpgh66is87ay4iw-python3.12-vllm-0.10.2-dist
error: Cannot build '/nix/store/j98dgdx87kjahfj73cd0wvzp3c0px4dh-python3.12-vllm-0.10.2.drv'.
Reason: builder failed with exit code 1.
Output paths:
/nix/store/cfj51613p4smfza7150mzsk21smfrzxl-python3.12-vllm-0.10.2
/nix/store/l2gjwc39xss9hg63rbpgh66is87ay4iw-python3.12-vllm-0.10.2-dist
Last 2 log lines:
> [nixbuild.net] Cached build failure. The id of the previously failed build is 4565380. See this link for the failed build details and logs: https://nixbuild.net/builds/4565380?t=EtkBCm8KBWJ1aWxkCgpidWlsZDpyZWFkGAMiCQoHCAcSAxD5BSINCgsIBBIHOgUKAxiBCDImCiQKAggbEgYIBRICCAUaFgoECgIIBQoICgYg9eyp0wYKBBoCCAAyFgoUCgIIGxIOCAISAxiACBIFEITTlgISJAgAEiAGRu1AM51kzPYGl4nShP2nj4SwLZdIQEZhBaIO6KS15xpATQzlvJ7Y-SLutaWbRFkbSrm3qOrarrHr60v_1kS5vx08FGnOtkZSZjwNvMDFTJwiO1O-MfmWac0BNAXtrNz8ASIiCiAuWrCOa6uQpEanWP7s1LJ8iKWZhtugtJoeboULO7JD1w==
> [nixbuild.net] To turn off build failure caching, see https://docs.nixbuild.net/settings/#reuse-build-failures
For full logs, run:
nix log /nix/store/j98dgdx87kjahfj73cd0wvzp3c0px4dh-python3.12-vllm-0.10.2.drvComparing the build inputs: [daniel@laptop:~/Source/nixpkgs]$ diff \
> <(nix derivation show --impure -I nixpkgs=https://github.com/daniel-fahey/nixpkgs/archive/fbc1629b8775a9eeb932f1b79d14fa973adb67ec.tar.gz --expr 'with import <nixpkgs> { }; python312Packages.vllm' | jq -r '.[] | .inputDrvs | keys[]') \
> <(nix derivation show --impure -I nixpkgs=https://github.com/daniel-fahey/nixpkgs/archive/e7fb8960241ab2211c471004bcd2b04683767c33.tar.gz --expr 'with import <nixpkgs> { }; python312Packages.vllm' | jq -r '.[] | .inputDrvs | keys[]')
57a58
> /nix/store/lnybvwfry02293c564dm33kv7ash7bck-python3.12-blake3-1.0.7.drv
71d71
< /nix/store/rc213bx4f77s732gj49y51r87z5mcjzy-python3.12-blake3-1.0.6.drvBlake3-py was updated in 29c1e9e, maybe somehow relevant? It is a dependency. |
daniel-fahey
force-pushed
the
update-vllm
branch
from
e7fb896 to
5d92454
Compare
Contributor
Author
|
daniel-fahey
force-pushed
the
update-vllm
branch
from
a4e10ee to
1486864
Compare
Contributor
Author
|
Anyone got a better alternative than reverting 29c1e9e? I think it's valid to do so, and I'll report the blake3 v1.0.7 compatibility issue with Python 3.12 vLLM v0.10.2+CPU upstream? |
Contributor
Author
|
Right, options: Option 1: blake3 pin for Python 3.12 CPU builds:# something like this (I've not tested it yet)
dependencies = [
aioprometheus
# Pin blake3 to 1.0.6 for Python 3.12 CPU builds
# blake3 1.0.7 causes build failures - see https://github.com/NixOS/nixpkgs/pull/447722
(if (cpuSupport && pythonOlder "3.13") then
blake3.overrideAttrs (old: {
version = "1.0.6";
src = fetchPypi {
pname = "blake3";
version = "1.0.6";
hash = "sha256-/Jhu3+7XSW024FMuUvwid3B95wSRfOH3T9yN3wMy0tY=";
};
})
else
blake3
)
cachetools
cbor2
# ... rest of deps ...
];Option 2: just mark Python 3.12 CPU builds as broken:meta = {
description = "High-throughput and memory-efficient inference and serving engine for LLMs";
# ... existing meta ...
broken = (cpuSupport && pythonOlder "3.13"); # blake3 1.0.7 incompatibility, see https://github.com/NixOS/nixpkgs/pull/447722
badPlatforms = [
"x86_64-darwin"
];
};I'm leaning toward Option 2 (marking broken) since:
Reverting the blake3 update globally would be wrong - that affects the entire ecosystem and blake3 1.0.7 is already in master. The pin would work but adds complexity for a configuration that might not have many users. Will mark as broken for now and file an upstream issue with vLLM about the blake3 1.0.7 incompatibility with Python 3.12 CPU builds? |
Contributor
|
blake3-py 1.0.7 basically only updates dependencies, so I didn't expect it to break anything. If Python 3.12 support for vllm is not important then I also think Option 2 (marking broken) is the better option til we figure out what the root cause is. Please file an issue with vllm, perhaps its authors have more insight. |
daniel-fahey
force-pushed
the
update-vllm
branch
from
1486864 to
470418e
Compare
daniel-fahey
force-pushed
the
update-vllm
branch
from
470418e to
ded72cb
Compare
Contributor
Author
|
GaetanLepage
approved these changes
Oct 3, 2025
Contributor
Could you do so @daniel-fahey please? You may then add a link to the issue in the derivation. |
nixpkgs-ci
bot
added
12.approvals: 2
Contributor
Author
Will do tomorrow, I'm out 🪩🕺 |
Closed
1 task
Notably adds Python 3.13 support (see vllm-project/vllm#13164) among many other things (https://github.com/vllm-project/vllm/releases/tag/v0.10.2). Co-authored-by: Gaétan Lepage <gaetan@glepage.com>
daniel-fahey
force-pushed
the
update-vllm
branch
from
ded72cb to
cc60981
Compare
GaetanLepage
approved these changes
Oct 4, 2025
Contributor
GaetanLepage
left a comment
GaetanLepage
left a comment
There was a problem hiding this comment.
Thanks @daniel-fahey !
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Notably adds Python 3.13 support (see vllm-project/vllm#13164) among many other things (https://github.com/vllm-project/vllm/releases/tag/v0.10.2).
Supersedes existing PR #442802 from @r-ryantm.
Python 3.13 support is also needed for vLLM to appear in the builds on the new (and fantastic initiative, btw) nixos-cuda Hydra, configured by @GaetanLepage.
For reference I share part of my engineer's log re: the nixos-cuda builder herein:
The new hydra.nixos-cuda.org instance is building the cuda-packages jobset, which is defined by
release-cuda.nix. Whilepython3Packages.vllmis explicitly listed in that file, it doesn't appear in any evaluation results.This is because vLLM is being filtered out during the evaluation phase:
b6f6c61, see below)disabled = pythonAtLeast "3.13";in its package definitionerror: vllm-0.10.1.1 not supported for interpreter python3.13To check which nixpkgs revision the new nixos-cuda Hydra is currently using:
To reproduce the evaluation failure:
And to verify vLLM is missing from the evaluation:
Things done
passthru.tests.nixpkgs-reviewon this PR. See nixpkgs-review usage../result/bin/.Add a 👍 reaction to pull requests you find important.