diff --git a/tensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.py b/tensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.py
index 79c2f064e1ac..ba01e7ad536a 100644
--- a/tensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.py
+++ b/tensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.py
@@ -339,6 +339,11 @@ def is_situ_activation(self) -> bool:
         return self.trtllm_gen_activation_type == ActType_TrtllmGen.SiTu
 
     def _validate_backend_local_activation(self) -> None:
+        # Runs from __init__, before create_weights, so the swiglu_* attributes
+        # checked below are still the constructor-provided values. For SiTu,
+        # create_weights later reuses the swiglu_alpha/swiglu_beta storage for
+        # the backend-local activation parameters (SiTu and SwiGLU are mutually
+        # exclusive and feed the same gemm1_alpha/gemm1_beta op slots).
         if self.trtllm_gen_activation_type is None:
             if (self.trtllm_gen_activation_alpha is not None
                     or self.trtllm_gen_activation_beta is not None):
@@ -549,7 +554,9 @@ def _check_configs(self):
                 raise ValueError(
                     "TRTLLM-Gen SiTu requires MXFP4 scaling vector size 32, "
                     f"got {self.scaling_vector_size}.")
-            for name in ("situ_alpha", "situ_beta"):
+            # For SiTu these hold the backend-local activation parameters
+            # (populated by create_weights, which runs before this check).
+            for name in ("swiglu_alpha", "swiglu_beta"):
                 value = getattr(self, name)
                 if (value.dtype != torch.float32
                         or value.shape != (self.expert_size_per_partition, )
@@ -594,19 +601,26 @@ def create_weights(self):
         else:
             self.quant_method.create_weights(self)
 
+        # SiTu reuses the swiglu_alpha/swiglu_beta storage: SiTu and SwiGLU are
+        # mutually exclusive (constructor-provided SwiGLU parameters are
+        # rejected by _validate_backend_local_activation) and feed the same
+        # gemm1_alpha/gemm1_beta op slots. Safe with respect to the
+        # `swiglu_alpha is not None` gates: create_moe.py checks the
+        # constructor kwargs (None for SiTu); _get_quant_method consults
+        # swiglu_alpha only on the nvfp4 branch (SiTu requires
+        # W4A8_MXFP4_MXFP8) and has already run above; _check_configs runs
+        # after this point and its swiglu gate admits w4a8_mxfp4_mxfp8.
         if self.is_situ_activation:
-            situ_alpha = nn.Parameter(torch.full(
+            self.swiglu_alpha = nn.Parameter(torch.full(
                 (self.expert_size_per_partition, ),
                 float(self.trtllm_gen_activation_alpha),
                 dtype=torch.float32),
-                                      requires_grad=False)
-            situ_beta = nn.Parameter(torch.full(
+                                             requires_grad=False)
+            self.swiglu_beta = nn.Parameter(torch.full(
                 (self.expert_size_per_partition, ),
                 float(self.trtllm_gen_activation_beta),
                 dtype=torch.float32),
-                                     requires_grad=False)
-            self.register_parameter("situ_alpha", situ_alpha)
-            self.register_parameter("situ_beta", situ_beta)
+                                            requires_grad=False)
 
         self._weights_created = True
         self._check_configs()
@@ -630,8 +644,9 @@ def cache_derived_state(self) -> None:
         if self.is_situ_activation:
             # Reinitialize constants after meta-device materialization. These
             # are backend configuration, not checkpoint weights.
-            self.situ_alpha.data.fill_(float(self.trtllm_gen_activation_alpha))
-            self.situ_beta.data.fill_(float(self.trtllm_gen_activation_beta))
+            self.swiglu_alpha.data.fill_(float(
+                self.trtllm_gen_activation_alpha))
+            self.swiglu_beta.data.fill_(float(self.trtllm_gen_activation_beta))
 
     def load_weights(self,
                      weights: List[Dict],
@@ -901,10 +916,10 @@ def run_moe(
             ] else 2
             intermediate_size_per_partition_padded = self.w3_w1_weight.shape[
                 -2] // factor
-            gemm1_alpha = (self.situ_alpha
-                           if self.is_situ_activation else self.swiglu_alpha)
-            gemm1_beta = (self.situ_beta
-                          if self.is_situ_activation else self.swiglu_beta)
+            # Holds SwiGLU's per-expert alpha/beta, or SiTu's backend-local
+            # activation parameters (which reuse this storage; see
+            # create_weights).
+            gemm1_alpha, gemm1_beta = self.swiglu_alpha, self.swiglu_beta
 
             output1_scale_scalar = self._get_data_or_none("fc31_scale_c")
             output1_scale_gate_scalar = self._get_data_or_none("fc31_alpha")
diff --git a/tensorrt_llm/_torch/utils.py b/tensorrt_llm/_torch/utils.py
index cb62ec99f76b..3899c560eb69 100644
--- a/tensorrt_llm/_torch/utils.py
+++ b/tensorrt_llm/_torch/utils.py
@@ -63,6 +63,12 @@ class ActivationType(IntEnum):
     Relu2 = 8
 
 
+# TRTLLM-Gen-local activation encoding, kept separate from the shared
+# ActivationType above ON PURPOSE: ActivationType mirrors the cutlass enum in
+# common.h and drives cutlass MoE kernels, whereas SiTu exists only in the
+# trtllm-gen batched-GEMM kernels. Adding SiTu to the shared ActivationType
+# would force a matching cutlass enum member that no cutlass kernel implements.
+# So SiTu stays here (TRTLLM-15177 item 1.2(a): decided keep-backend-local).
 # Keep this in sync with the ActType enum in
 # cpp/tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/KernelRunner.h
 class ActType_TrtllmGen(IntEnum):
diff --git a/tests/integration/test_lists/waives.txt b/tests/integration/test_lists/waives.txt
index 82bbc5e7f0db..bd7595f02b6f 100644
--- a/tests/integration/test_lists/waives.txt
+++ b/tests/integration/test_lists/waives.txt
@@ -345,6 +345,7 @@ perf/test_perf_sanity.py::test_e2e[disagg_upload-gen_only-gb300_glm-5-fp4_8k1k_c
 perf/test_perf_sanity.py::test_e2e[disagg_upload-gen_only-gb300_kimi-k25-thinking-fp4_8k1k_con1024_ctx1_dep4_gen1_dep32_eplb416_mtp3_ccb-NIXL] SKIP (https://nvbugs/6490049)
 test_e2e.py::test_multi_nodes_eval[MiniMax-M3-tp16-mmlu] SKIP (https://nvbugs/6373561)
 test_e2e.py::test_ptp_quickstart_advanced_deepseek_r1_w4afp8_8gpus[DeepSeek-R1-W4AFP8-DeepSeek-R1/DeepSeek-R1-W4AFP8] SKIP (https://nvbugs/5836830)
+test_e2e.py::test_ptp_quickstart_bert[TRTLLM-BertForSequenceClassification-bert/bert-base-uncased-yelp-polarity] SKIP (bug pending, tracked in PR 17414)
 test_e2e.py::test_trtllm_bench_llmapi_launch[pytorch_backend-llama-v3-llama3-8b] SKIP (https://nvbugs/6568058)
 unittest/_torch/attention/sparse/dsa/test_req_idx_per_token.py::test_on_update_kv_lens_rebuilds_stale_map SKIP (https://nvbugs/6574939)
 unittest/_torch/attention/sparse/rocketkv/test_rocketkv.py::test_model[TRTLLM-llama-3.1-model/Llama-3.1-8B-Instruct-pytorch] SKIP (https://nvbugs/6602094)
@@ -391,6 +392,7 @@ unittest/_torch/visual_gen/test_cosmos3_distilled.py::TestDistilledDenoiseLoop::
 unittest/auto_deploy/multigpu/custom_ops SKIP (https://nvbugs/6403920)
 unittest/auto_deploy/singlegpu/models/test_glm4_moe_modeling.py::test_glm4_moe_decoder_layer_equivalence[1-dtype0-1-8] SKIP (https://nvbugs/6566730)
 unittest/auto_deploy/singlegpu/models/test_glm4_moe_modeling.py::test_glm4_moe_moe_equivalence[dtype0-1-8] SKIP (https://nvbugs/6566730)
+unittest/auto_deploy/singlegpu/shim SKIP (bug pending, tracked in PR 17414)
 unittest/bindings/test_transfer_agent_bindings.py::TestMooncakeFunctionalTransfer::test_mooncake_wait_in_progress_on_zero_timeout SKIP (https://nvbugs/6517836)
 unittest/bindings/test_transfer_agent_bindings.py::TestMooncakeFunctionalTransfer::test_mooncake_write_transfer_gpu_tensor SKIP (https://nvbugs/6517836)
 unittest/bindings/test_transfer_agent_bindings.py::TestMooncakeFunctionalTransfer::test_mooncake_write_transfer_multiple_chunks SKIP (https://nvbugs/6517836)