Fix Qwen3 when --dtype float16 on CPU / MPS

[alvarobartt]

What does this PR do?

This PR is a follow up fix of #648, since the float16 inference for Qwen3 models was broken due to the attention_bias dtype, which was defaulting to float32, so this PR adds the dtype field for Qwen3Model so that the dtype is inherited.

As a side-fix, this PR also removes the to_device method from Tensor::zeros_like as it's not required, and by default placed in the same device as the tensor used.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline, Pull Request section?
• Was this discussed/approved via a GitHub issue or the forum? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the documentation guidelines, and here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

@Narsil or @kozistr

[Narsil]

index 989968d..cfe056c 100644
--- a/backends/candle/src/models/qwen3.rs
+++ b/backends/candle/src/models/qwen3.rs
@@ -454,13 +454,14 @@ impl Qwen3Model {
         let causal_mask = Tensor::from_slice(&mask, (seq_len, seq_len), &Device::Cpu)?;
         let causal_mask = causal_mask.expand(&[bs, dim, seq_len, seq_len])?;+        let device = attention_bias.device();
         let negatives =
             Tensor::full(f32::MIN, attention_bias.shape(), &Device::Cpu)?.to_dtype(self.dtype)?;
-        let zeros = Tensor::zeros_like(&attention_bias)?.to_dtype(self.dtype)?;
+        let zeros = Tensor::zeros_like(&attention_bias)?.to_device(&Device::Cpu)?.to_dtype(self.dtype)?;         let causal_mask = causal_mask
             .where_cond(&negatives, &zeros)?
-            .to_device(&self.device)?;
+            .to_device(device)?;         attention_bias.broadcast_add(&causal_mask)
     }

This one is working for me. Here there is an issue because the zeros are on the CPU, while the negatives are on CPU.

[kozistr]

thanks for catching this! I missed that part, sorry for that.

[alvarobartt]

Closed in favor of #654