[cuda] int4: stabilize two-layer decode test via CUDA-seeded init

_make_int4_linear built the throwaway nn.Linear on CPU, so reset_parameters()
drew from the CPU RNG between the two layer constructions and shifted the stream
that seeds the quantized weights. That pushed test_two_layer_mlp's genuine INT4
error from 0.1405 to 0.1556, crossing the 0.15 bound. Build the module with
device=cuda so init draws from the CUDA RNG, leaving the CPU stream (and the
measured error) deterministic. Test-only; dequant math is unchanged.

Author: Gasoonjia

backends/cuda/tests/test_int4_dispatch.py

@@ -59,7 +59,10 @@ def _make_int4_linear(N, K, group_size=128, symmetric=False, bias=False):
     )
     int4_w = quantize_weight(w_bf16, config)
 
-    module = nn.Linear(K, N, bias=bias, dtype=torch.bfloat16)
+    # device="cuda" so the random init draws from the CUDA RNG to match the
+    # same random weight as regular int4 dispatch and fit the same numerical
+    # error tolerance.
+    module = nn.Linear(K, N, bias=bias, dtype=torch.bfloat16, device="cuda")
     pack_linear_for_cuda(module, {"weight": int4_w})
     module.cuda()
     return module, w_bf16.cuda()