//! Minimal row-major f32 CPU tensor: what `WeightStore::load_f32` returns (load-time conversions only; the data path is CUDA).
#[derive(Debug, Clone, PartialEq)]
pub struct Tensor {
    pub shape: Vec<usize>,
    pub data: Vec<f32>,
}

impl Tensor {
    pub fn new(shape: Vec<usize>, data: Vec<f32>) -> Self {
        assert_eq!(shape.iter().product::<usize>(), data.len(), "shape {:?} vs len {}", shape, data.len());
        Tensor { shape, data }
    }
    pub fn numel(&self) -> usize {
        self.data.len()
    }
    /// rows of a matrix `[rows, cols]` (all leading dims flattened)
    pub fn rows(&self) -> usize {
        self.data.len() / self.cols()
    }
    pub fn cols(&self) -> usize {
        *self.shape.last().unwrap_or(&1)
    }
    pub fn row(&self, i: usize) -> &[f32] {
        let c = self.cols();
        &self.data[i * c..(i + 1) * c]
    }
}