add qwen3vl 4b, 8b, 32b

src/utils/mod.rs

@@ -21,7 +21,7 @@ pub fn get_device(device: Option<&Device>) -> Device {
         None => {
             #[cfg(feature = "cuda")]
             {
-                Device::new_cuda(0).unwrap_or(Device::Cpu)
+                Device::new_cuda(6).unwrap_or(Device::Cpu)
             }
             #[cfg(not(feature = "cuda"))]
             {

src/utils/tensor_utils.rs

@@ -8,10 +8,15 @@ pub fn prepare_causal_attention_mask(
     device: &Device,
 ) -> Result<Tensor> {
     // Sliding window mask?
-    let mask: Vec<_> = (0..tgt_len)
-        .flat_map(|i| (0..tgt_len).map(move |j| if i < j { f32::NEG_INFINITY } else { 0. }))
-        .collect();
-    let mask = Tensor::from_slice(&mask, (tgt_len, tgt_len), device)?;
+    // let mask: Vec<f32> = (0..tgt_len)
+    //     .flat_map(|i| (0..tgt_len).map(move |j| if i < j { f32::NEG_INFINITY } else { 0. }))
+    //     .collect();
+    // let mask = Tensor::from_vec(mask, (tgt_len, tgt_len), device)?;
+    let arange = Tensor::arange(0u32, tgt_len as u32, device)?;
+    let arange = arange.unsqueeze(1)?.broadcast_as((tgt_len, tgt_len))?;
+    let upper_triangle = arange.t()?.gt(&arange)?;
+    let mask = upper_triangle.where_cond(&Tensor::new(f32::NEG_INFINITY, device)?.broadcast_as(arange.shape())?, &Tensor::new(0f32, device)?.broadcast_as(arange.shape())?)?;
+
     let mask = if seqlen_offset > 0 {
         let mask0 = Tensor::zeros((tgt_len, seqlen_offset), DType::F32, device)?;
         Tensor::cat(&[&mask0, &mask], D::Minus1)?