// SPDX-License-Identifier: AGPL-3.0-or-later
// Attention Apply: Apply attention weights to values
// Pass 3 of multi-pass attention implementation
//
// Computes: output[i,d] = sum_j (weights[i,j] * V[j,d])
// weights: [batch, heads, q_seq_len, kv_seq_len]
// V: [batch, heads, kv_seq_len, head_dim]
// output: [batch, heads, q_seq_len, head_dim]
struct AttentionParams {
batch_size: u32,
num_heads: u32,
q_seq_len: u32,
kv_seq_len: u32,
head_dim: u32,
_pad0: u32,
_pad1: u32,
_pad2: u32,
}
@group(0) @binding(0) var<storage, read> weights: array<f64>;
@group(0) @binding(1) var<storage, read> value: array<f64>;
@group(0) @binding(2) var<storage, read_write> output: array<f64>;
@group(0) @binding(3) var<uniform> params: AttentionParams;
@compute @workgroup_size(16, 16, 1)
fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {
let batch_head = global_id.z;
let i = global_id.y; // query position (output row)
let d = global_id.x; // head dimension
if (i >= params.q_seq_len || d >= params.head_dim) {
return;
}
let batch = batch_head / params.num_heads;
let head = batch_head % params.num_heads;
if (batch >= params.batch_size) {
return;
}
var weighted_sum = 0.0;
for (var j = 0u; j < params.kv_seq_len; j = j + 1u) {
let weight_idx = batch * params.num_heads * params.q_seq_len * params.kv_seq_len
+ head * params.q_seq_len * params.kv_seq_len
+ i * params.kv_seq_len
+ j;
let value_idx = batch * params.num_heads * params.kv_seq_len * params.head_dim
+ head * params.kv_seq_len * params.head_dim
+ j * params.head_dim
+ d;
weighted_sum += weights[weight_idx] * value[value_idx];
}
let out_idx = batch * params.num_heads * params.q_seq_len * params.head_dim
+ head * params.q_seq_len * params.head_dim
+ i * params.head_dim
+ d;
output[out_idx] = weighted_sum;
}