diff --git a/Sources/FluidAudio/Diarizer/Offline/Core/OfflineDiarizerModels.swift b/Sources/FluidAudio/Diarizer/Offline/Core/OfflineDiarizerModels.swift
index 7f882f028..51f98e6f5 100644
--- a/Sources/FluidAudio/Diarizer/Offline/Core/OfflineDiarizerModels.swift
+++ b/Sources/FluidAudio/Diarizer/Offline/Core/OfflineDiarizerModels.swift
@@ -85,7 +85,12 @@ public struct OfflineDiarizerModels: Sendable {
         logger.info("Loading offline diarization models from \(modelsDirectory.path)")
 
         let loadStart = Date()
-        let inferenceComputeUnits: MLComputeUnits = .all
+        // Honor a caller-supplied compute-units choice (the `configuration:`
+        // parameter was previously accepted but ignored here). Defaults to `.all`,
+        // so existing callers are unaffected; callers that need CPU/GPU-only
+        // inference (e.g. to avoid cross-device ANE numeric variance in the
+        // embeddings) can now pass `MLModelConfiguration(computeUnits:)`.
+        let inferenceComputeUnits: MLComputeUnits = configuration?.computeUnits ?? .all
 
         let segmentationAndEmbeddingNames: [String] = [
             ModelNames.OfflineDiarizer.segmentationPath,
diff --git a/Tests/FluidAudioTests/CI/BasicInitializationTests.swift b/Tests/FluidAudioTests/CI/BasicInitializationTests.swift
index fe05d44c1..bfa0ab25e 100644
--- a/Tests/FluidAudioTests/CI/BasicInitializationTests.swift
+++ b/Tests/FluidAudioTests/CI/BasicInitializationTests.swift
@@ -348,6 +348,30 @@ extension CoreMLDiarizerTests {
 
         XCTAssertEqual(models.embeddingModel.configuration.computeUnits, customConfig.computeUnits)
     }
+
+    /// Tests that the OFFLINE diarizer model loader honors a user-specified
+    /// configuration's compute units. This is parity with
+    /// `testModelLoadingCustomConfig` above (the streaming `DiarizerModels`
+    /// loader already honored it); `OfflineDiarizerModels.load(configuration:)`
+    /// previously accepted the parameter but ignored it and always loaded `.all`.
+    func testOfflineModelLoadingCustomConfig() async throws {
+
+        XCTExpectFailure("Download might fail in CI environment", strict: false)
+
+        let customConfig = MLModelConfiguration()
+        customConfig.computeUnits = .cpuOnly
+
+        let models = try await OfflineDiarizerModels.load(configuration: customConfig)
+
+        // Segmentation, embedding, and PLDA-rho models load with the requested
+        // compute units; the fbank front-end intentionally stays on `.cpuOnly`.
+        XCTAssertEqual(
+            models.segmentationModel.configuration.computeUnits, customConfig.computeUnits)
+        XCTAssertEqual(
+            models.embeddingModel.configuration.computeUnits, customConfig.computeUnits)
+        XCTAssertEqual(
+            models.pldaRhoModel.configuration.computeUnits, customConfig.computeUnits)
+    }
 }
 
 // MARK: - CoreML Backend Specific Test