diff --git a/single_shot_inference.py b/single_shot_inference.py
index 1f47d0c1..9c011392 100644
--- a/single_shot_inference.py
+++ b/single_shot_inference.py
@@ -414,8 +414,14 @@ def moe_forward(x, li, moe_runner, se_runner, router, token_id):
     torch.cuda.synchronize(x.device)
     if topk_ids.max().item() >= 384 or topk_ids.min().item() < 0:
         print(f"  L{li} BAD topk_ids: min={topk_ids.min().item()} max={topk_ids.max().item()}", flush=True)
+    if li < 3:
+        print(f"  L{li} MoE input: |x|={x.abs().max().item():.4f} has_nan={torch.isnan(x).any().item()}", flush=True)
     routed_out = moe_runner.run(x, topk_w, topk_ids)
+    if li < 3:
+        print(f"  L{li} MoE routed: |out|={routed_out.abs().max().item():.4f} has_nan={torch.isnan(routed_out).any().item()}", flush=True)
     shared_out = se_runner.run(x)
+    if li < 3:
+        print(f"  L{li} MoE shared: |out|={shared_out.abs().max().item():.4f} has_nan={torch.isnan(shared_out).any().item()}", flush=True)
     return routed_out + shared_out
 
 # =====================================================================