You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
compute/xla: cast bias to bf16 before custom-call in fwd and bwd
Forward and backward flash paths appended options.Bias to operands
without casting, producing mixed dtypes (bf16 q/k/v vs f16/f32 bias)
that cuDNN rejects. Cast via f.bf16 after validateBias (validation
still accepts half/float32 input; the kernel always gets bf16).
Also tighten the validateBias doc comment from "broadcastable to
[B,H,S,Skv]" to exact shape [B,H,S,S] required (no per-axis
broadcast), and trim the dtype-gate comment at line 49.
0 commit comments