monkeypatch.attention.flash_attn_4
monkeypatch.attention.flash_attn_4
Route flash attention through native FA4 when it is available on SM90+ hardware.
Functions
| Name | Description |
|---|---|
| configure_fa4 | Prepare the process to run native FA4. |
| fa4_usable | Whether native FA4 can serve attention for this model on this GPU. |
configure_fa4
monkeypatch.attention.flash_attn_4.configure_fa4()Prepare the process to run native FA4.
Silences the harmless first-compile AuxData warning and, for an explicitly requested
flash_attention_4, surfaces the stale-quack and Blackwell nested-election warnings
(the auto-upgrade path checks both in fa4_usable before reaching here).
fa4_usable
monkeypatch.attention.flash_attn_4.fa4_usable(model_config=None)Whether native FA4 can serve attention for this model on this GPU.
Checks GPU arch (SM90/100/110), flash_attn.cute import, FA4 head-dim limits, the
quack-kernels floor, and the Blackwell nested-election hang. Warns (with the fix) when
head dims, quack, or the Blackwell backward are the blocker.