Studio already falls back to the torch2.10 flash-attn / causal-conv1d /
mamba-ssm wheels when it finds torch 2.11, because upstream publishes no
2.11-tagged builds. torch 2.12 is in exactly the same position, and the same
wheels work there, so a 2.12 install currently drops to a source build for no
reason.
Measured on a B200, python 3.12, fresh uv venv on torch 2.12.1+cu130, wheels
installed with --no-deps and torch verified unmoved afterwards, importing the
compiled .so directly rather than only the Python package:
causal-conv1d 1.6.1 9412 passed / 3888 skipped / 0 failed
mamba-ssm 2.3.1 tests/ops, 20 passed
flash-attn 2.8.1 splitkv + qkvpacked subset, 848 passed
Against a torch 2.10 control the pass/fail/skip counts match and the failing
test-ID sets are byte identical.
The reuse window is bounded rather than open ended, so the comment now records
that. flash-attn v2.8.3.post1's torch2.9 wheel fails to import on torch 2.10
and on torch 2.12 alike, with an undefined symbol out of flash_attn_2_cuda:
torch broke extension ABI between 2.9 and 2.10 and has held it from 2.10
through 2.12. A wheel cannot skip a torch minor backwards, so torch 2.13 is
deliberately left out of the table until it is measured.
The torch2.10 flash-attn pin stays at 2.8.1. v2.8.3 looks like a free upgrade
but publishes only 2 of the 8 torch2.10 assets that v2.8.1 does, keeping just
cu13/cp312 for x86_64 and aarch64 and dropping every cu12 and every cp313
torch2.10 wheel, while v2.8.3.post1 dropped the torch2.10 assets entirely.
Bumping the pin would silently 404 most users back to a source build, so the
constant now carries that warning.
Tests cover the 2.12 mapping through both direct_wheel_url and the flash-attn
URL builder, that reuse only ever targets torch2.10, and that the selected
flash-attn version is never a .post release.
741bebb36a