fix: use F.silu in test script for old corex torch

ex_engine/csrc/build_test_moe_tcu.sh

@@ -66,7 +66,7 @@ for k in range(top_k):
     eid = expert_ids[k].item()
     w = expert_weights[k].item()
     gate_up = F.linear(hidden, w13[eid])
-    gate = torch.silu(gate_up[:, :I])
+    gate = F.silu(gate_up[:, :I])
     up = gate_up[:, I:]
     act = gate * up
     expert_out = F.linear(act, w2[eid])
@@ -131,7 +131,7 @@ for _ in range(3):
         eid = expert_ids[k].item()
         w = expert_weights[k].item()
         gate_up = F.linear(hidden, w13[eid])
-        gate = torch.silu(gate_up[:, :I])
+        gate = F.silu(gate_up[:, :I])
         up = gate_up[:, I:]
         act = gate * up
         out_py += w * F.linear(act, w2[eid])
@@ -144,7 +144,7 @@ for _ in range(100):
         eid = expert_ids[k].item()
         w = expert_weights[k].item()
         gate_up = F.linear(hidden, w13[eid])
-        gate = torch.silu(gate_up[:, :I])
+        gate = F.silu(gate_up[:, :I])
         up = gate_up[:, I:]
         act = gate * up
         out_py += w * F.linear(act, w2[eid])

ex_engine/csrc/moe_tcu_dispatch.cpp

@@ -49,7 +49,7 @@ torch::Tensor moe_decode(
         auto gate_up = torch::mm(hidden, gate_up_weights[eid].t());
 
         // SiLU and mul
-        auto gate = torch::silu(gate_up.slice(1, 0, inter));
+        auto gate_slice = gate_up.slice(1, 0, inter); auto gate = gate_slice * torch::sigmoid(gate_slice);
         auto up = gate_up.slice(1, inter, inter2);
         auto act = gate * up;  // (1, I)
 
@@ -117,7 +117,7 @@ torch::Tensor moe_prefill(
         auto gate_up = torch::mm(tokens, gate_up_weights[eid].t());
 
         // SiLU and mul
-        auto gate = torch::silu(gate_up.slice(1, 0, inter));
+        auto gate_slice = gate_up.slice(1, 0, inter); auto gate = gate_slice * torch::sigmoid(gate_slice);
         auto up = gate_up.slice(1, inter, inter2);
         auto act = gate * up;  // (count, I)