Fix swiglu_decode intermediate comparison to use chained reference

iron/operators/swiglu_decode/test.py

@@ -13,7 +13,14 @@
 
 
 def get_params():
-    params_list = [(2048, 2048)]
+    # (embedding_dim, hidden_dim)
+    # Square shape is the historical smoke-test config; the rectangular
+    # shape reflects real decoder-model FFN dims (e.g. Qwen3.5-0.8B
+    # embedding=1024, hidden=3584) that downstream runtimes actually hit.
+    params_list = [
+        (2048, 2048),
+        (1024, 3584),
+    ]
 
     params = []
     for p in params_list:
@@ -55,25 +62,43 @@ def test_swiglu_decode(embedding_dim, hidden_dim, aie_context):
     print(f"Effective Bandwidth: {bandwidth_gbps:.4f} GB/s")
 
     errors = {}
-    # Verify intermediate result
-    # Reshape to (1, hidden_dim) using the unpadded dimension to match the golden reference shape.
-    # Note: op.hidden_dim_padded may differ if padding was applied; we use hidden_dim here
-    # because the golden reference was generated with the unpadded hidden_dim.
+
+    # Verify intermediate result (left_swished * right) against a chained
+    # reference built from the observed AIE left_swished and right buffers.
+    # This isolates eltwise_mul from any sub-tolerance drift accumulated in
+    # the upstream gemv_1 / silu stages that would otherwise be amplified by
+    # multiplication against a large-magnitude right operand (e.g. silu
+    # outputs that land near zero for very-negative inputs, where bf16
+    # rounding asymmetrically flushes NPU vs fp32-CPU). This mirrors the
+    # approach used by swiglu_prefill/test.py.
+    # Reshape to (1, hidden_dim) using the unpadded dimension to match the
+    # reference shape. Note: op.hidden_dim_padded may differ if padding was
+    # applied; we use hidden_dim here because the golden reference was
+    # generated with the unpadded hidden_dim.
+    left_swished = op_func.left_swished.to_torch().reshape((1, hidden_dim))
+    right = op_func.right.to_torch().reshape((1, hidden_dim))
+    ref_intermediate = left_swished * right
+
     intermediate = op_func.intermediate.to_torch().reshape((1, hidden_dim))
     errors_intermediate = verify_buffer(
         intermediate,
         "intermediate",
-        golden_ref["intermediate"],
-        rel_tol=0.07,
-        abs_tol=0.7,
+        ref_intermediate,
+        rel_tol=0.04,
+        abs_tol=0.4,
     )
     if errors_intermediate:
         errors["intermediate"] = errors_intermediate
 
-    # Verify output using intermediate result
-    ref_2 = intermediate @ golden_ref["w_down"]
+    # Verify output using intermediate result.
+    # Note: we use the AIE intermediate buffer as reference (rather than
+    # golden_ref["output"]) because this better matches the bfloat16 precision
+    # path and isolates errors to gemv_2.
+    ref_output = intermediate @ golden_ref["w_down"]
     output = output_buf.to_torch().reshape((1, embedding_dim))
-    errors_output = verify_buffer(output, "output", ref_2, rel_tol=0.04, abs_tol=0.4)
+    errors_output = verify_buffer(
+        output, "output", ref_output, rel_tol=0.04, abs_tol=0.4
+    )
     if errors_output:
         errors["output"] = errors_output