Upgrade to numpy >= 2.0.0

[fgebhart]

Hi 👋

Would it be possible to upgrade the dependency of vLLM to numpy to remove the pinning here?

We are running into a dependency conflict issue in our dependency chain and would appreciate the move to support more recent numpy versions.

Thanks!

[youkaichao]

vLLM does not have a direct dependency on numpy, this is required for dependencies of vLLM, i.e. an indirect dependency. We need to wait for the open-source community to adopt numpy 2.0 first, only after that can we safely remove the constraint.

[chris-aeviator]

this is entirely confusing, e.g. pixi complains about vllm being the reason

  × failed to solve the pypi requirements of 'ray' 'linux-64'
  ├─▶ failed to resolve pypi dependencies
  ╰─▶ Because vllm>=0.6.0 depends on numpy<2.0.0 and numpy==2.1.2, we can conclude that vllm>=0.6.0 cannot be used.
      And because only the following versions of vllm are available:
          vllm<=0.6.0
          vllm==0.6.1
          vllm==0.6.1.post1
          vllm==0.6.1.post2
          vllm==0.6.2
          vllm==0.6.3
          vllm==0.6.3.post1
      and you require vllm>=0.6.0, we can conclude that your requirements are unsatisfiable.

[chris-aeviator]

vLLM does not have a direct dependency on numpy, this is required for dependencies of vLLM, i.e. an indirect dependency. We need to wait for the open-source community to adopt numpy 2.0 first, only after that can we safely remove the constraint.

@youkaichao this is false (or I misunderstand you) - see

vllm/requirements-common.txt

Line 3 in ea7bd68

numpy < 2.0.0

for the direct dependency

[youkaichao]

@chris-aeviator I added this line to prevent numpy 2.0 . see #5582 (comment)

[rgommers]

Hi @youkaichao, the issue with Outlines was fixed in its 0.1.6 release ~1.5 months ago, see dottxt-ai/outlines#1265. According to the discussion in gh-5582, I think you can remove the pin now.

[rgommers]

It looks like this is fixed now since vllm 0.8.3. The last remaining fix was gh-15692, which upgraded the numba pin from 0.60 to 0.61 to allow using numpy 2.x.

Right now, running pip install vllm in a clean python 3.12 venv results in numpy 2.1.3 being installed.

---

gh-12759 is still helpful to test with recent versions of numpy, but isn't required to unblock users of an installed vllm.

[fgebhart]

Great, thanks! 👏