use crate::device::WgpuDevice;
use crate::error::Result;
use crate::tensor::Tensor;
use std::sync::Arc;
async fn create_test_tensor(
device: Arc<WgpuDevice>,
shape: Vec<usize>,
value: f32,
) -> Result<Tensor> {
let size: usize = shape.iter().product();
let data: Vec<f32> = vec![value; size];
Tensor::from_vec_on(data, shape, device).await
}
#[tokio::test]
async fn test_scaled_dot_product_attention_basic() {
let dev = crate::device::test_pool::get_test_device().await;
let query = create_test_tensor(dev.clone(), vec![1, 1, 2, 2], 0.5)
.await
.unwrap();
let key = create_test_tensor(dev.clone(), vec![1, 1, 2, 2], 0.5)
.await
.unwrap();
let value = create_test_tensor(dev.clone(), vec![1, 1, 2, 2], 1.0)
.await
.unwrap();
let output = query.scaled_dot_product_attention(key, value).unwrap();
assert_eq!(output.shape(), &[1, 1, 2, 2]);
}
#[tokio::test]
async fn test_scaled_dot_product_attention_multi_head() {
let dev = crate::device::test_pool::get_test_device().await;
let query = create_test_tensor(dev.clone(), vec![2, 4, 8, 16], 0.5)
.await
.unwrap();
let key = create_test_tensor(dev.clone(), vec![2, 4, 8, 16], 0.5)
.await
.unwrap();
let value = create_test_tensor(dev.clone(), vec![2, 4, 8, 16], 1.0)
.await
.unwrap();
let output = query.scaled_dot_product_attention(key, value).unwrap();
assert_eq!(output.shape(), &[2, 4, 8, 16]);
}
#[tokio::test]
async fn test_scaled_dot_product_attention_shape_validation() {
let dev = crate::device::test_pool::get_test_device().await;
let query = create_test_tensor(dev.clone(), vec![1, 1, 4, 4], 0.5)
.await
.unwrap();
let key = create_test_tensor(dev.clone(), vec![1, 1, 4, 4], 0.5)
.await
.unwrap();
let value = create_test_tensor(dev.clone(), vec![1, 1, 4, 5], 1.0)
.await
.unwrap();
let result = query.scaled_dot_product_attention(key, value);
assert!(result.is_err()); }