The Studio installer skipped Flash Attention on any Blackwell GPU
(compute capability >= 10.0: sm_100 B200/B100, sm_120 RTX 50 series),
on the assumption that no prebuilt wheels existed for those archs. That
assumption is stale: Dao-AILab has shipped wheels with sm_100 / sm_120
kernels since flash-attn v2.7.3, built for CUDA >= 12.8.
Replace the blanket has_blackwell_gpu() skip in _ensure_flash_attn and
_ensure_flash_attn_for_long_context with a CUDA-version-aware gate. The
normal prebuilt-wheel path already builds the exact wheel URL, HEAD-checks
it, and falls back gracefully, so it now runs on Blackwell unless the
installed torch's CUDA build is older than 12.8 (in which case the
matching wheel would lack Blackwell kernels and crash at kernel launch).
- wheel_utils: probe now reports full cuda_version; add
cuda_supports_blackwell_flash_attn and should_skip_flash_attn_for_blackwell.
- install_python_stack / training worker: use the new gate.
- Update the corresponding tests.
504e253674