diff --git a/cutedsl/bridge.py b/cutedsl/bridge.py
index b96211b7..8a58edb3 100644
--- a/cutedsl/bridge.py
+++ b/cutedsl/bridge.py
@@ -85,7 +85,7 @@ def quantize_to_nvfp4(x_bf16, block_size=SF_VEC_SIZE):
     abs_scaled = x_scaled.abs().clamp(max=6.0)
     
     half_steps = (abs_scaled * 2.0).round().clamp(0, 12).to(torch.int8)
-    step_to_idx = torch.tensor([0,1,2,3,4,4,5,5,6,6,6,7,7], dtype=torch.int8, device=x_bf16.device)
+    step_to_idx = x_bf16.new_tensor([0,1,2,3,4,4,5,5,6,6,6,7,7], dtype=torch.int8)
     indices = step_to_idx[half_steps.long()]
 
     nibbles = torch.where(signs < 0, indices + 8, indices).to(torch.uint8)
@@ -141,7 +141,7 @@ def quantize_activation_nvfp4(x_bf16, global_scale, block_size=SF_VEC_SIZE):
     abs_scaled = x_scaled.abs().clamp(max=6.0)
     
     half_steps = (abs_scaled * 2.0).round().clamp(0, 12).to(torch.int8)
-    step_to_idx = torch.tensor([0,1,2,3,4,4,5,5,6,6,6,7,7], dtype=torch.int8, device=x_bf16.device)
+    step_to_idx = x_bf16.new_tensor([0,1,2,3,4,4,5,5,6,6,6,7,7], dtype=torch.int8)
     indices = step_to_idx[half_steps.long()]
 
     nibbles = torch.where(signs < 0, indices + 8, indices).to(torch.uint8)