Preserve image-aware sparse QSA for Qwen vision

src/engine/metal/qwen_mtplx/indexing.rs

@@ -1,5 +1,5 @@
 //! Select, take/Gather, GatherAxis, Arange and Scan from the pinned runtime.
-//! take exports BF16/F16/F32/U32 with I64 and I32 indices.
+//! take exports BF16/F16/F32/I32/U32 with I64 and I32 indices.
 use super::super::gpu::Buffer;
 use super::array::{Array, Dtype, Layout};
 use super::stream::Stream;
@@ -503,7 +503,7 @@ pub(super) fn take(table: &Array, ids: &Array, axis: i32, stream: Stream) -> Res
         || !(il.dtype() == Dtype::I64 || il.dtype() == Dtype::U32 || il.dtype() == Dtype::I32)
         || !matches!(
             sl.dtype(),
-            Dtype::BF16 | Dtype::F16 | Dtype::F32 | Dtype::U32
+            Dtype::BF16 | Dtype::F16 | Dtype::F32 | Dtype::I32 | Dtype::U32
         )
         || (sl.shape()[axis] == 0 && il.size() != 0)
     {
@@ -551,7 +551,8 @@ fn gather_take_evaluate(
         Dtype::BF16 => "bf16",
         Dtype::F16 => "f16",
         Dtype::F32 => "f32",
-        Dtype::U32 => "u32",
+        // Gather copies 32-bit payloads without value arithmetic.
+        Dtype::I32 | Dtype::U32 => "u32",
         _ => unreachable!(),
     };
     let large = sl.size() > i32::MAX as usize