diff --git a/Source/Cmlx/mlx/mlx/compile.cpp b/Source/Cmlx/mlx/mlx/compile.cpp
index ca5f0699..48708cd6 100644
--- a/Source/Cmlx/mlx/mlx/compile.cpp
+++ b/Source/Cmlx/mlx/mlx/compile.cpp
@@ -135,8 +135,9 @@ Compiled::Compiled(
     }
   }
   os << "_";
-  for (const auto& x : inputs) {
-    if (constant_ids.find(x.id()) != constant_ids.end()) {
+  // Iterate the moved-into members; the parameters are moved-from above.
+  for (const auto& x : inputs_) {
+    if (constant_ids_.find(x.id()) != constant_ids_.end()) {
       continue;
     }
     os << kindof(x.dtype()) << x.itemsize();
diff --git a/Tests/MLXTests/TransformTests.swift b/Tests/MLXTests/TransformTests.swift
index 6c74f6de..5c7bbad6 100644
--- a/Tests/MLXTests/TransformTests.swift
+++ b/Tests/MLXTests/TransformTests.swift
@@ -163,6 +163,22 @@ class TransformTests: XCTestCase {
         XCTAssertEqual(traces, 2)
     }
 
+    func testCompiledKernelsDifferByInputDtype() {
+        // bf16 and f16 inputs upcast to float32 give the same fused graph, so
+        // only the input dtypes tell the two kernels apart.
+        let compiled = compile { (inputs: [MLXArray]) -> [MLXArray] in
+            let x = inputs[0].asType(.float32)
+            return [exp(x) * x]
+        }
+
+        let values = MLXArray(Array(stride(from: Float(-2), to: 2, by: 0.25)))
+        for dtype in [DType.bfloat16, .float16] {
+            let x = values.asType(dtype)
+            let expected = exp(values) * values
+            assertEqual(compiled([x])[0], expected)
+        }
+    }
+
     class CompileTestState: CustomStringConvertible, Updatable {
         var y: MLXArray
         var o: MLXArray?