torchao.prototype.attention (prototype)

Created On: Mar 24, 2026 | Last Updated On: Mar 24, 2026

High-Level API

apply_low_precision_attention	Apply low-precision attention to a model.
AttentionBackend	Backend kernel for computing attention.

Direct Usage (FA3)

fp8_fa3_sdpa	FP8 SDPA shared by all backends.
fp8_fa3_rope_sdpa	Fused RoPE + FP8 SDPA shared by all backends.