update to 0.19.1 by benmoss · Pull Request #30 · conda-forge/vllm-feedstock

benmoss · 2026-04-20T17:44:48Z

Checklist

Used a personal fork of the feedstock to propose changes
Bumped the build number (if the version is unchanged)
Reset the build number to 0 (if the version changed)
Re-rendered with the latest conda-smithy (Use the phrase @conda-forge-admin, please rerender in a comment in this PR for automated rerendering)
Ensured the license file is being packaged.

- Bump version to 0.19.1, build_number to 0 - Update pytorch_version to 2.10.0, torchvision to 0.25.0 - Update flash-attention vendored source to new commit - Drop patches 0003/0004/0006 (gettid, CMAKE_ARGS, cpu.txt torch pin) upstreamed or no longer needed in 0.19.1 - Update patches 0002/0005 for changed context in 0.19.1 - Remove cutlass <4 host dep (vllm now uses cutlass 4.2.1 via FetchContent) - Remove ray-cgraph, xformers, scipy, triton from run deps - Add anthropic, ijson, mcp, opentelemetry-{api,sdk,exporter-otlp} to run deps - Update many dep version constraints per requirements/common.txt

…6.04.18.15.28.11

pytorch 2.10.0 requires libcusparse >=12.5.10.65, but the cuda 12.9 meta-package pins cuda-libraries 12.9.0.* which provides libcusparse 12.5.9.5.*. Removing the cuda meta-package lets the solver pick a compatible libcusparse. The specific CUDA packages already listed in host (cuda-cudart-dev, cuda-nvrtc-dev, etc.) are sufficient.

conda-forge-admin · 2026-04-20T17:46:14Z

Hi! This is the friendly automated conda-forge-linting service.

I wanted to let you know that I linted all conda-recipes in your PR (recipe/recipe.yaml) and found some lint.

Here's what I've got...

For recipe/recipe.yaml:

❌ In conda-forge.yml: $.github_actions = {'self_hosted': True, 'timeout_minutes': 1200, 'triggers': ['push', 'pull_request']}.

{'self_hosted': True, 'timeout_minutes': 1200, 'triggers': ['push', 'pull_request']} is not valid under any of the given schemas
Schema
```
{
  "anyOf": [
    {
      "$ref": "#/$defs/GithubActionsConfig"
    },
    {
      "type": "null"
    }
  ]
}
```

_{This message was generated by GitHub Actions workflow run https://github.com/conda-forge/conda-forge-webservices/actions/runs/24685112523. Examine the logs at this URL for more detail.}

The package is not in conda-forge and is only needed for the AWS SageMaker entrypoint. vllm imports it at module load time via api_server.py -> sagemaker/api_router.py, which breaks the build. Add a no-op stub so the module imports cleanly; runtime SageMaker functionality will still require the real package.

conda-forge ships libgomp as a separate package rather than bundling it inside the torch wheel, so vllm's shim probe finds nothing and find_library fails. Add a fallback to PREFIX/lib before the find_library call.

conda-forge-admin · 2026-04-20T20:17:29Z

Hi! This is the friendly automated conda-forge-linting service.

I just wanted to let you know that I linted all conda-recipes in your PR (recipe/recipe.yaml) and found it was in an excellent condition.

…6.04.18.15.28.11

MNT: Re-rendered with conda-smithy 3.61.1 and conda-forge-pinning 2026.04.21.11.53.26 Other tools: - conda-build 26.3.0 - rattler-build 0.62.2 - rattler-build-conda-compat 1.4.13

…tHub tarballs)

ACL (Arm Compute Library) is downloaded from GitHub and built from source during cmake configure. This fails on conda-forge because oneDNN's FindACL.cmake cannot locate the built artifacts. Add a VLLM_SKIP_ACL env var guard around the entire ACL section and set it for aarch64 builds. oneDNN still works without ACL, just without the ACL-accelerated kernels.

benmoss · 2026-04-24T13:31:49Z

Not sure if renumbering the patches is high priority or not :)

shermansiu · 2026-04-27T14:08:06Z

And 0.20.0 is also out, wow

shermansiu · 2026-04-27T14:05:15Z

    - aarch64 and use_cuda  # Still have issues locating CUDA for the aarch64 build
    # - not use_cuda  # Just build CUDA for now
    # - match(python, "!=3.10")  # Until all the builds succeed
+    # See XXX on why we need this skip to distinguish runners for GPU builds vs CPU builds


Could we please fill in the actual link here?

Wait, that's from jaimergp's commit...

Yeah.... it's supposed to be a reference to conda-forge/conda-forge-pinning-feedstock#6910

Thanks! I suppose we can update it for the 0.20.0 build!

shermansiu · 2026-04-27T14:11:29Z

Thanks so much, @benmoss! 😄

h-vetinari · 2026-04-28T01:52:53Z

    script:
      # Pick an arbitrary test to run: some of the other ones rely on a bunch of external packages
-      - pytest ./tests/core/test_scheduler.py
+      - pytest ./tests/config/test_config_utils.py


this is a test for non-core functionality. Why was this changed?

Probably because the tests/core directory was removed in 0.19.1, so he picked another random test.

https://github.com/vllm-project/vllm/tree/v0.19.1/tests/

Though I suppose there's this? https://github.com/vllm-project/vllm/blob/v0.19.1/tests/v1/core/test_scheduler.py

tests/core → tests/v1/core

benmoss added 5 commits April 20, 2026 10:14

MNT: Re-rendered with conda-smithy 3.60.0 and conda-forge-pinning 202…

3156f8f

…6.04.18.15.28.11

add libcusparse-dev

32f5553

add libcusolver-dev

ec9b43a

benmoss added 6 commits April 20, 2026 14:07

run arm64/aarch64 natively on azure

8b05e75

fix: add libgomp to linux host deps for OpenMP discovery on aarch64

0d7f881

fix: update test path, tests/core/test_scheduler.py removed in 0.19.1

aa28c8d

fix: fall back to PREFIX/lib for libgomp when PyTorch does not bundle it

449ffa1

conda-forge ships libgomp as a separate package rather than bundling it inside the torch wheel, so vllm's shim probe finds nothing and find_library fails. Add a fallback to PREFIX/lib before the find_library call.

fix deprecated self_hosted, switch to blacksmith runner

0150654

benmoss added 2 commits April 20, 2026 16:17

MNT: Re-rendered with conda-smithy 3.60.0 and conda-forge-pinning 202…

017c2d4

…6.04.18.15.28.11

chore: remove spurious git version footer from patch 0005

25ba4e5

jaimergp reviewed Apr 20, 2026

View reviewed changes

Comment thread recipe/conda_build_config.yaml Outdated

jaimergp mentioned this pull request May 18, 2026

Migrate to namespace / blacksmith conda-forge/admin-requests#2089

Open

22 tasks

benmoss added 2 commits April 21, 2026 09:17

only use blacksmith vms for cuda builds

9868633

fix: add libcurand-dev to host for curand_kernel.h (needed by CUTLASS)

0c1c6a4

benmoss force-pushed the 0.19.1 branch from dbe8ead to 0c1c6a4 Compare April 21, 2026 13:24

jaimergp and others added 3 commits April 23, 2026 12:24

Split CPU vs GPU jobs

43fa6ed

MNT: Re-rendered with conda-smithy 3.61.1 and conda-forge-pinning 2026.04.21.11.53.26 Other tools: - conda-build 26.3.0 - rattler-build 0.62.2 - rattler-build-conda-compat 1.4.13

fix: vendor cutlass submodule for flash-attention (not included in Gi…

2cc35d7

…tHub tarballs)

benmoss marked this pull request as ready for review April 24, 2026 13:01

benmoss requested review from h-vetinari, maresb and shermansiu as code owners April 24, 2026 13:01

shermansiu reviewed Apr 27, 2026

View reviewed changes

shermansiu merged commit 9e256e0 into conda-forge:main Apr 27, 2026
17 checks passed

benmoss deleted the 0.19.1 branch April 27, 2026 14:11

h-vetinari reviewed Apr 28, 2026

View reviewed changes

Uh oh!

Conversation

benmoss commented Apr 20, 2026

Uh oh!

conda-forge-admin commented Apr 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

conda-forge-admin commented Apr 20, 2026

Uh oh!

Uh oh!

benmoss commented Apr 24, 2026

Uh oh!

shermansiu commented Apr 27, 2026

Uh oh!

shermansiu Apr 27, 2026

Choose a reason for hiding this comment

Uh oh!

shermansiu Apr 27, 2026

Choose a reason for hiding this comment

Uh oh!

benmoss Apr 27, 2026

Choose a reason for hiding this comment

Uh oh!

shermansiu Apr 27, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

shermansiu commented Apr 27, 2026

Uh oh!

h-vetinari Apr 28, 2026

Choose a reason for hiding this comment

Uh oh!

shermansiu Apr 28, 2026

Choose a reason for hiding this comment

Uh oh!

shermansiu Apr 28, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

conda-forge-admin commented Apr 20, 2026 •

edited

Loading