fix(multimodal): gate the Qwen2-VL prefill export on the feature that calls it

src/multimodal/host_preprocessor.rs

@@ -44,10 +44,14 @@ mod export;
 use super::qwen_vl::insert_qwen_vl_image_tokens;
 use export::export_mlx_tensor;
 use export::{
-    build_prepared_prefill, export_llava_prefill, export_qwen2_vl_prefill, usize_to_i32,
-    validate_embedding_shape, validate_processor_shape, validate_projected_shape,
-    validate_sequence_capacity,
+    build_prepared_prefill, export_llava_prefill, usize_to_i32, validate_embedding_shape,
+    validate_processor_shape, validate_projected_shape, validate_sequence_capacity,
 };
+// Only the `xla-iree` Qwen2-VL prefill path calls this, so importing it
+// unconditionally leaves it unused in a default-feature build and `-D warnings`
+// turns that into a hard error. The gate has to match the single call site.
+#[cfg(any(feature = "xla-iree", test))]
+use export::export_qwen2_vl_prefill;
 
 /// Vision implementation selected for OpenXLA multimodal preprocessing.
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]

src/multimodal/host_preprocessor_export.rs

@@ -151,6 +151,11 @@ pub(super) fn export_llava_prefill(
     )
 }
 
+// Reachable in production only from the `xla-iree` Qwen2-VL prefill path, and
+// exercised by the module tests regardless, so the gate carries `test` too.
+// Without it a default-feature `cargo clippy --lib -- -D warnings` fails the
+// build on `function is never used`.
+#[cfg(any(feature = "xla-iree", test))]
 pub(super) fn export_qwen2_vl_prefill(
     logical_tokens: Vec<i32>,
     merged: InputEmbeddings,