Update unit test of fused moe w4afp8

Signed-off-by: Min Yu <[email protected]>

Author: yumin066

tests/unittest/_torch/modules/test_fused_moe.py

@@ -1328,15 +1328,12 @@ def test_fused_moe_w4afp8(dtype, weight_loading_mode):
                                       dtype=torch.int8).cuda()
 
             # The pre-quant scale to be multiplied with the input activation.
-            w1_pre_quant_scale = torch.ones(HIDDEN_SIZE,
-                                            dtype=dtype,
-                                            device="cuda")
-            w2_pre_quant_scale = torch.ones(INTERMEDIATE_SIZE,
-                                            dtype=dtype,
-                                            device="cuda")
-            w3_pre_quant_scale = torch.ones(HIDDEN_SIZE,
-                                            dtype=dtype,
-                                            device="cuda")
+            w1_pre_quant_scale = torch.rand(
+                HIDDEN_SIZE, dtype=dtype, device="cuda") * 0.1 + 0.95
+            w2_pre_quant_scale = torch.rand(
+                INTERMEDIATE_SIZE, dtype=dtype, device="cuda") * 0.1 + 0.95
+            w3_pre_quant_scale = torch.rand(
+                HIDDEN_SIZE, dtype=dtype, device="cuda") * 0.1 + 0.95
 
             # The weight scale to dequantize int4 weights (by multiplication).
             w1_scale = torch.randn(