feat(mlx-lm): add Qwen3.5 hybrid attention model support

mlx-lm/src/lib.rs

@@ -8,6 +8,7 @@ pub mod utils;
 use mlx_rs::Array;
 
 use crate::models::qwen3;
+use crate::models::qwen3_5;
 
 pub struct ModelInputBuilder<'a, C, T> {
     pub y: &'a Array,
@@ -31,6 +32,18 @@ impl<'a, C> ModelInput<'a, C, Option<Array>> for qwen3::ModelInput<'a, C> {
     }
 }
 
+impl<'a, C> ModelInput<'a, C, Option<Array>> for qwen3_5::ModelInput<'a, C> {
+    fn from_model_input_builder(builder: ModelInputBuilder<'a, C, Option<Array>>) -> Self {
+        let ModelInputBuilder { y, cache, state } = builder;
+
+        Self {
+            inputs: y,
+            mask: state.as_ref(),
+            cache,
+        }
+    }
+}
+
 pub trait ModelOutput {
     fn logits(&self) -> &Array;
 }

mlx-lm/src/models/mod.rs

@@ -1,2 +1,3 @@
 pub mod llama;
 pub mod qwen3;
+pub mod qwen3_5;