Cannot load the FA2 kernel

[remi-or]

Description

kernels fails to load the flash attention 2 kernel from the hub, seemingly because some metadata is missing.

Steps to reproduce

import kernels

print(f"{kernels.__version__ = }")


kwargs = {
    'repo_id': 'kernels-community/flash-attn2',
    'revision': None,
    'version': 1,
    'trust_remote_code': False,
}
print(f"{kwargs = }")

k = kernels.get_kernel(**kwargs)
print(f"{k = }")

The output is:

kernels.__version__ = '0.15.2'
kwargs = {'repo_id': 'kernels-community/flash-attn2', 'revision': None, 'version': 1, 'trust_remote_code': False}
Fetching ... files: 22it [00:00, 3862.16it/s]00, ?B/s]
Download complete: : 0.00B [00:00, ?B/s]              Traceback (most recent call last):
  File "/fsx/remi_ouazan/transformers/wip.py", line 14, in <module>
    k = kernels.get_kernel(**kwargs)
  File "/fsx/remi_ouazan/transformers/tcb/lib/python3.13/site-packages/kernels/utils.py", line 480, in get_kernel
    return _import_from_path(variant_path, repo_info=repo_info)
  File "/fsx/remi_ouazan/transformers/tcb/lib/python3.13/site-packages/kernels/utils.py", line 199, in _import_from_path
    metadata = Metadata.read_from_file(variant_path / "metadata.json")
ValueError: Cannot parse metadata from `"/fsx/remi_ouazan/.cache/kernels--kernels-community--flash-attn2/snapshots/fbb79e60c84ae9db159e5d0ea74e9724558361ae/build/torch29-cxx11-cu128-x86_64-linux/metadata.json"`: missing field `name` at line 14 column 1

Expected behavior

I would expect the kernel to load, maybe with a warning about the metadata.

Environment

• kernels 0.15.2
• Python 3.13.5
• torch / cuda: '2.9.0+cu128'
• os: linux

Additional context

No response

[sayakpaul]

Cc: @drbh

[danieldk]

kernels > 0.14 requires certain metadata to exist (0.14 was a hard cut for many things, including deprecating model repos) and these builds for old Torch versions are pre-0.14 and do not have the required metadata. We should probably yank these old builds.

Moving to kernels-community, since it's an issue with old files lingering around.

[danieldk]

@drbh would it be possible to switch FA2 over to the stable ABI as well, so that we can run on >= Torch 2.9?

[sayakpaul]

@drbh a gentle ping on this.

[drbh]

stable support for fa2 is in progress here

• feat: prefer torch stable abi for cuda fa2 #997

note that we needed to upstream the ability to target stable abi for a subset of the kernel backends which was already merged here

• feat: scope stable abi to backend kernels#682

fa2 stable support should be merged soon!

update:
a preview build of stable fa2 is accessible here https://huggingface.co/kernels/kernels-staging/flash-attn2/tree/pr-997

[sayakpaul]

@remi-or can we close this?

[remi-or]

Works for me! So I would say yes. Thank you!