diff --git a/tests/kernels/test_scaled_mm_triton.py b/tests/kernels/test_scaled_mm_triton.py
index ed5ca9de2971d..7db770452d1f5 100644
--- a/tests/kernels/test_scaled_mm_triton.py
+++ b/tests/kernels/test_scaled_mm_triton.py
@@ -8,7 +8,7 @@
 import pytest
 import torch
 
-from vllm.utils import seed_everything
+from vllm.platforms import current_platform
 
 device = "cuda"
 
@@ -42,7 +42,7 @@ def test_scaled_mm(M, N, K, in_dtype, out_dtype, use_scalar_scale_a,
     is_floating_point_type = lambda t: torch.tensor([1, 1], dtype=t
                                                     ).is_floating_point()
 
-    seed_everything(0)
+    current_platform.seed_everything(0)
 
     # NOTE: There are cases, where if the matrix is large enough, an output
     # like 65504.4 can be produced, and can easily turn into inf when
@@ -70,7 +70,7 @@ def test_scaled_mm(M, N, K, in_dtype, out_dtype, use_scalar_scale_a,
     if use_scalar_scale_b:
         scale_b = torch.rand((1, 1), device=device)
     else:
-        scale_b = 0.25 * torch.rand((1, 1), device=device)
+        scale_b = 0.25 * torch.rand((N, 1), device=device)
 
     bias = None
     if use_bias: