test: isolate fused SwiGLU — test no-clamp first

tests/unit/test_fused_swiglu_kernel.py

@@ -92,8 +92,30 @@ def test_fused_swiglu_compilation():
     x_padded = torch.zeros(128, K_packed, dtype=torch.uint8, device=device).view(torch.float4_e2m1fn_x2)
     x_padded.view(torch.uint8)[:tokens] = x_fp4.view(torch.uint8)
 
-    # Run FUSED path
-    print("  Running fused SwiGLU GEMM...")
+    # Test WITHOUT clamp first (swiglu_limit=0) to isolate SiLU-only path
+    print("\n--- Smoke test: fused kernel (no clamp) ---")
+    swiglu_limit_test = 0.0
+    warmup_fused_swiglu_compilation(
+        num_experts, K_packed, N_packed, device,
+        swiglu_limit=swiglu_limit_test,
+    )
+    l1_fused_noclamp = run_fused_swiglu_grouped_gemm(
+        mat_a=x_padded, mat_b=mat_b,
+        scale_a=scale_a, scale_b=scale_b,
+        expert_offsets=padded_offsets,
+        global_scale_a=gsa, global_scale_b=gsb,
+        swiglu_limit=swiglu_limit_test,
+    )[:tokens]
+    has_nan_nc = torch.isnan(l1_fused_noclamp).any().item()
+    max_val_nc = l1_fused_noclamp.abs().max().item()
+    print(f"  No-clamp output: NaN={has_nan_nc} max={max_val_nc:.4f}")
+    if has_nan_nc or max_val_nc > 1e6:
+        print("  ❌ No-clamp path has issues")
+        return False
+    print("  ✅ No-clamp path works")
+
+    # Test WITH clamp
+    print("\n--- Smoke test: fused kernel (with clamp) ---")
     l1_fused = run_fused_swiglu_grouped_gemm(
         mat_a=x_padded, mat_b=mat_b,
         scale_a=scale_a, scale_b=scale_b,