feat: improve get local kernel importing

[drbh]

This PR update the get_local_kernel function to handle more paths and infer the variant if needed.

These changes handle specifying the top level of the kernel repo or a specific variant directory as shown in the example below

git clone https://huggingface.co/kernels-community/activation
uv run get-local.py

# /// script
# requires-python = ">=3.10"
# dependencies = [
#     "numpy",
#     "torch==2.7.0",
#     "kernels",
# ]
#
# [tool.uv.sources]
# kernels = { git = "https://github.com/huggingface/kernels.git", branch = "improve-get-local-kernel" }
# ///
import torch
from kernels import get_local_kernel
from pathlib import Path

# set seed for reproducibility
torch.manual_seed(0)

activation = get_local_kernel(
    repo_path=Path("./activation"),
    # repo_path=Path("/home/ubuntu/.cache/huggingface/hub/models--kernels-community--activation/snapshots/2fafa6a3a38ccb57a1a98419047cf7816ecbc071/"),
    # repo_path=Path("activation/build/torch27-cxx11-cu126-x86_64-linux"),
    package_name="activation",
)

print(activation)

# Random tensor
x = torch.randn((10, 10), dtype=torch.float16, device="cuda")

# Run the kernel
y = torch.empty_like(x)
activation.gelu_fast(y, x)

# print(y)
sum_of_y = y.sum().item()
print(f"Sum of y: {sum_of_y}")

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[pcuenca]

Nice, haven't checked the details but looks good conceptually (and the example scripts looks quite complete).

[MekkCyber]

LGTM thanks for adding this 🤗

[danieldk]

Could you add test cases for these? Not for universal vs. non-universal, but for the three different types of paths that are supported?

Probably best to change the fixture here:

kernels/tests/test_basic.py

Line 13 in 1caa4c1

def local_kernel():

to return the path and not the loaded kernel. Then the existing test could be extended to test these.

[drbh]

yea no problem, totally agree that we should have tests for this. Just opened a PR here #134