Skip to main content

psa_pack_qkv_backward

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

Function psa_pack_qkv_backward 

Source 
pub fn psa_pack_qkv_backward<T: Triton, D: Float, const KEY_DIM: i32>(
    d_packed_ptr: T::Pointer<D>,
    d_qkv_ptr: T::Pointer<D>,
    qkv_h: i32,
    H: i32,
    W: i32,
    B: 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_pack_qkv: scatters d_packed back to d_qkv.

Grid / block matches the forward pass: [4 * BH * N, 1, 1], block [KEY_DIM, 1, 1].

Uses atomic_add because both psa_pack_qkv_backward (all four sections) and psa_extract_v_backward (V_lo/V_hi sections) write to the same d_qkv buffer.