kernels-community

/

flash-attn2

like 32

Follow

kernels-community 427

Under kernels>=0.14, loading this kernel fails on older torch build variants because their build/<variant>/metadata.json is missing the now-required name / id fields (and in some cases the file is missing entirely). Newer variants (torch 2.10+) have already been migrated correctly, so the breakage is silent and depends on the host's torch version.

Repro

On any environment that resolves to a torch 2.8 / cu128 build (e.g. the pytorch/pytorch:2.8.0-cuda12.8-cudnn9-runtime image):

from kernels import get_kernel
get_kernel("kernels-community/flash-attn2")

ValueError: Cannot parse metadata from
`.../build/torch28-cxx11-cu128-x86_64-linux/metadata.json`:
missing field `name` at line 4 column 1

Current state of metadata.json across variants (main)

For reference, a correct file (from torch210-cxx11-cu128-x86_64-linux):

{
  "name": "flash-attn2",
  "id": "_flash_attn2_cuda_042c80b",
  "version": 1,
  "license": "BSD-3-Clause",
  "python-depends": [],
  "backend": {"type": "cuda", "archs": ["10.0", "12.0", "8.0", "9.0"]}
}

Impact

Anyone on torch 2.7 / 2.8 / 2.9 — including the official pytorch/pytorch:2.8.0-cuda12.8 base image, which is widely used in CI and HF Jobs — can no longer load this kernel under kernels==0.14. transformers --attn-implementation kernels-community/flash-attn2 breaks on those hosts. Workarounds (pin kernels<0.14, or bump torch to 2.10) both work but require user-side changes.

Ask

Backfill metadata.json (with name, id, license, backend) on the older torch build variants so the kernel loads under kernels>=0.14 regardless of torch version. The torch210+ entries can serve as the template.

Environment

• kernels==0.14.0
• torch==2.8.0, CUDA 12.8, Linux x86_64
• kernels-community/flash-attn2 @ main (commit 46e3484 at time of writing)