[CUDA] Fix half type matmul in cutlass kernels (#3469)

2026-05-06 08:35:53 +09:00
parent 1fdd4e23c2
commit 80bcd1c658
8 changed files with 28 additions and 34 deletions
@@ -1046,7 +1046,7 @@ class TestQuantized(mlx_tests.MLXTestCase):
                y3 = scatter_unsort(y3, inv_order, indices.shape)
                y4 = scatter_unsort(y4, inv_order, indices.shape)

-                tol = 1.5e-5 if (dtype == mx.float32) else 2.5e-4
+                tol = 1.5e-5 if (dtype == mx.float32) else 1e-3

                self.assertLess((y1 - y2).abs().max(), tol)
                self.assertLess((y1 - y3).abs().max(), tol)