Skip to main content

psa_merge_attn_backward

vision_rs::models::yolo::kernels::attention::psa

Function psa_merge_attn_backward 

Source 
pub fn psa_merge_attn_backward<T: Triton, D: Float, const KEY_DIM: i32>(
    d_merged_ptr: T::Pointer<D>,
    d_lo_ptr: T::Pointer<D>,
    d_hi_ptr: T::Pointer<D>,
    c: i32,
    H: i32,
    W: i32,
    num_heads: i32,
)
where
    T::I32Tensor: Tensor<i32, 1> + Comparison<i32, BoolTensor = T::BoolTensor>,
    T::Pointer<D>: AddOffsets<i32, 1, T::I32Tensor, Output = T::Tensor<T::Pointer<D>>>,
Expand description

Backward of psa_merge_attn_nchw: scatters d_merged back to d_lo and d_hi flat buffers.

Grid / block matches the forward: [BH * N, 1, 1], block [KEY_DIM, 1, 1]. Regular stores are safe: each (bh, n, d) position maps to a unique merged channel, so d_lo and d_hi receive no overlapping writes.