fix: use randint for float4 dummy weights in cudagraph test

tests/cudagraph_test.py

@@ -106,15 +106,18 @@ def make_dummy_runner(num_experts=32, hidden_size=7168, intermediate_size=3072,
     runner = CuTeDSLMoERunner(num_experts, hidden_size, intermediate_size, device=device)
     
     # Create minimal dummy weights
-    l1_fp4 = [torch.randn(3584, intermediate_size * 2 // 2, dtype=torch.float4_e2m1fn_x2, device=device)
-              for _ in range(num_experts)]
-    l1_sf = [torch.randn(3584 // 16, intermediate_size * 2, dtype=torch.float8_e4m3fn, device=device)
-             for _ in range(num_experts)]
+    # Create minimal dummy weights (uint8 → view as float4)
+    def rand_fp4(*shape, device="cuda"):
+        return torch.randint(0, 256, shape, dtype=torch.uint8, device=device).view(torch.float4_e2m1fn_x2)
+    
+    def rand_sf(*shape, device="cuda"):
+        return torch.rand(shape, dtype=torch.float8_e4m3fn, device=device)
+    
+    l1_fp4 = [rand_fp4(3584, intermediate_size, device=device) for _ in range(num_experts)]
+    l1_sf = [rand_sf(3584 // 16, intermediate_size * 2, device=device) for _ in range(num_experts)]
     l1_gs = [0.1] * num_experts
-    l2_fp4 = [torch.randn(1536, hidden_size // 2, dtype=torch.float4_e2m1fn_x2, device=device)
-              for _ in range(num_experts)]
-    l2_sf = [torch.randn(1536 // 16, hidden_size, dtype=torch.float8_e4m3fn, device=device)
-             for _ in range(num_experts)]
+    l2_fp4 = [rand_fp4(1536, hidden_size // 2, device=device) for _ in range(num_experts)]
+    l2_sf = [rand_sf(1536 // 16, hidden_size, device=device) for _ in range(num_experts)]
     l2_gs = [0.1] * num_experts
     
     runner.prepare_weights_direct(l1_fp4, l1_sf, l1_gs, l2_fp4, l2_sf, l2_gs)