feature request: Add spacy support cupy>13 (CUDA 13)

[KennethEnevoldsen]

Currently, spacy does not (officially) support cupy>=13, which also means that I can't use spacy with cuda>13 which e.g. also prevents usage on blackwell GPUs (B200). At least it is not clear in the docs.

However I can circumvent it with the following fix:

# /// script
# requires-python = ">=3.10"
# dependencies = [
#     "spacy",
#     "cupy-cuda13x[ctk]",
#     "en-core-web-sm",
# ]
#
# [tool.uv.sources]
# en-core-web-sm = { url = "https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-3.8.0/en_core_web_sm-3.8.0-py3-none-any.whl" }
# ///
"""Minimal repro: spaCy on a CUDA 13 / Blackwell GPU (e.g. B200, sm_100).

Run with:  uv run test_spacy.py
"""

import cupy

# Sanity: CuPy sees the GPU
print("CuPy:", cupy.__version__, "| device:",
      cupy.cuda.runtime.getDeviceProperties(0)["name"].decode())

import spacy

print("Require gpu:", spacy.require_gpu())

# Run a real pipeline on the GPU
nlp = spacy.load("en_core_web_sm")
doc = nlp("Apple is looking at buying a U.K. startup for $1 billion.")
print("Entities:", [(ent.text, ent.label_) for ent in doc.ents])

Which reproduces the expected:

CuPy: 14.1.1 | device: NVIDIA B200 MIG 1g.23gb
Require gpu: True
Entities: [('Apple', 'ORG'), ('U.K.', 'GPE'), ('$1 billion', 'MONEY')]

This is not entirely clear from the docs that this is possible, where the most recent version seems to be pip install 'spacy[cuda12x]' (similar also the case if you look at the setup.cfg). I suspect the fix is to add a cuda13x extra and raise the cupy<13 cap.

There might be issues that I have not discovered with this approach.

[rehan243]

Oh yeah, we've hit this too with newer CUDA versions. Turns out spaCy's CuPy dep is pretty version-sensitive, but you can override it without too much hassle. Your workaround looks solid, but just FYI, you can also specify the CuPy version directly when installing spaCy to avoid conflicts. Something like:

pip install spacy[cuda113] cupy-cuda113

We’ve used this pattern to lock CuPy versions for spaCy on A100s (sm_80) before. For B200 GPUs (sm_100), you’d just swap cuda113 with cuda13x based on your setup.

Also, you’re right that the docs don’t mention this edge case—super confusing if you’re trying to stay up-to-date with CUDA. Might be worth flagging that in a PR or opening an issue for clearer guidance?