Apply the cast before the scaling. (huggingface#2135)

candle-nn/src/ops.rs

@@ -70,7 +70,7 @@ pub fn dropout(xs: &Tensor, drop_p: f32) -> Result<Tensor> {
     let rand = Tensor::rand(0f32, 1f32, xs.shape(), xs.device())?;
     let scale = 1.0 / (1.0 - drop_p as f64);
     let drop_p = Tensor::new(drop_p, xs.device())?.broadcast_as(xs.shape())?;
-    let mask = (rand.ge(&drop_p)? * scale)?.to_dtype(xs.dtype())?;
+    let mask = (rand.ge(&drop_p)?.to_dtype(xs.dtype())? * scale)?;
     xs * mask
 }