debug: add L2, SiLU, and scatter pipeline prints

src/nvfp4_megamoe_kernel/nvfp4_mega_moe.py

@@ -195,6 +195,7 @@ def nvfp4_mega_moe_l2(
         slot_expert_ids,
         per_expert_alpha=per_expert_alpha,
     )
+    print(f"[L2-GEMM-OUT] slots={slot_out.shape[0]} N={N} amax={slot_out.abs().max().item():.4e} mean={slot_out.float().mean().item():.4e} nan={torch.isnan(slot_out).any().item()}")
     return slot_out  # (num_slots, HIDDEN) bfloat16
 
 
@@ -385,14 +386,12 @@ def nvfp4_mega_moe_full(
 
     # Step 3: SiLU + Mul PER SLOT — nonlinearity before combining paths
     gate, up = l1_slots.chunk(2, dim=-1)
+    print(f"[L1-SPLIT] gate amax={gate.abs().max().item():.4e} mean={gate.float().mean().item():.4e} | up amax={up.abs().max().item():.4e} mean={up.float().mean().item():.4e}")
     activated = torch.nn.functional.silu(gate) * up
+    print(f"[SILU-ACT] amax={activated.abs().max().item():.4e} mean={activated.float().mean().item():.4e} nan={torch.isnan(activated).any().item()}")
     if activation_clamp is not None:
         activated = activated.clamp(max=activation_clamp)
 
-    if MEGA_MOE_DEBUG:
-        print(f"[silu] nan={torch.isnan(activated).any().item()} "
-              f"abs_max={activated.abs().max().item():.4e}")
-
     # Step 4: Quantize activated slots → FP4
     l1_fp4, l1_sf_out, l2_global_scale = stage_activation(activated)
 
@@ -427,3 +426,4 @@ def nvfp4_mega_moe_full(
         slot_token,
         l2_slots * slot_weight.to(l2_slots.dtype).unsqueeze(1),
     )
+    print(f"[SCATTER] y amax={y.abs().max().item():.4e} mean={y.float().mean().item():.4e} nan={torch.isnan(y).any().item()} slots={num_slots}")