diff --git a/setup.cfg b/setup.cfg
index bab3eebff..2a51a439f 100644
--- a/setup.cfg
+++ b/setup.cfg
@@ -68,7 +68,9 @@ STREAMING =
 VISION =
     # Vision Tools
     webp>=0.4.0
-    pillow-simd>=9.5.0
+    # pillow-simd is an x86-only (SSE4) fork; fall back to stock Pillow on other architectures (e.g. aarch64).
+    pillow-simd>=9.5.0; platform_machine == "x86_64"
+    Pillow>=9.5.0; platform_machine != "x86_64"
     torchvision>=0.24.0
 
 DEV =
diff --git a/tests/functional/test_functional.py b/tests/functional/test_functional.py
index 07b8768da..bd0414d11 100644
--- a/tests/functional/test_functional.py
+++ b/tests/functional/test_functional.py
@@ -87,7 +87,6 @@ def test_mlp_recomputation(gated, activation, testing_device):
 
 # Takes ~6s, much more if it needs to compile, reducing the hidden size doesn't help.
 @pytest.mark.slow
-@pytest.mark.skip("Dropless MoE is broken")
 def test_dropless_mlp(testing_device):
     device = torch.device(testing_device)
     num_experts = 4