diff --git a/fastvideo-kernel/attn_qat_infer/quantization/bench/bench_quant_k.py b/fastvideo-kernel/attn_qat_infer/quantization/bench/bench_quant_k.py
index 5fc97f5ee..e8360957c 100644
--- a/fastvideo-kernel/attn_qat_infer/quantization/bench/bench_quant_k.py
+++ b/fastvideo-kernel/attn_qat_infer/quantization/bench/bench_quant_k.py
@@ -14,7 +14,7 @@ See the License for the specific language governing permissions and
 limitations under the License.
 """
 import torch
-import fp4quant
+import fp4quant_cuda as fp4quant
 from triton.tools.mxfp import MXFP4Tensor
 
 from bench_utils import bench_kineto
diff --git a/fastvideo-kernel/attn_qat_infer/quantization/bench/bench_quant_q.py b/fastvideo-kernel/attn_qat_infer/quantization/bench/bench_quant_q.py
index 1d0576d59..f05c292e3 100644
--- a/fastvideo-kernel/attn_qat_infer/quantization/bench/bench_quant_q.py
+++ b/fastvideo-kernel/attn_qat_infer/quantization/bench/bench_quant_q.py
@@ -14,7 +14,7 @@ See the License for the specific language governing permissions and
 limitations under the License.
 """
 import torch
-import fp4quant
+import fp4quant_cuda as fp4quant
 from triton.tools.mxfp import MXFP4Tensor
 
 from bench_utils import bench_kineto
diff --git a/fastvideo-kernel/attn_qat_infer/quantization/bench/bench_quant_v.py b/fastvideo-kernel/attn_qat_infer/quantization/bench/bench_quant_v.py
index 0592aa888..a17885ec2 100644
--- a/fastvideo-kernel/attn_qat_infer/quantization/bench/bench_quant_v.py
+++ b/fastvideo-kernel/attn_qat_infer/quantization/bench/bench_quant_v.py
@@ -14,7 +14,7 @@ See the License for the specific language governing permissions and
 limitations under the License.
 """
 import torch
-import fp4quant
+import fp4quant_cuda as fp4quant
 from triton.tools.mxfp import MXFP4Tensor
 
 from bench_utils import bench_kineto
diff --git a/fastvideo-kernel/csrc/turbodiffusion/quant/quant.cu b/fastvideo-kernel/csrc/turbodiffusion/quant/quant.cu
index 9a56e3881..672013047 100644
--- a/fastvideo-kernel/csrc/turbodiffusion/quant/quant.cu
+++ b/fastvideo-kernel/csrc/turbodiffusion/quant/quant.cu
@@ -31,6 +31,11 @@ auto quant(
   std::optional<torch::Tensor>& Output_S
 ) {
 
+  TORCH_CHECK(
+    Input.scalar_type() == torch::kHalf || Input.scalar_type() == torch::kBFloat16,
+    "Unsupported input data type for quant_cuda (int8 quantization)."
+  );
+
   using ElementOut = int8_t;
   static constexpr int BlockSize = 128;
   static constexpr int NumThrPerCta = 256;
@@ -62,8 +67,7 @@ auto quant(
     }
 
     default: {
-      std::cerr << "Observing: " << Input.scalar_type() << " for the input datatype which is invalid";
-      throw std::runtime_error("Unsupported input data type for quantize_to_fp4.");
+      TORCH_CHECK(false, "Unsupported input data type for quant_cuda (int8 quantization).");
     }
   }
   
diff --git a/fastvideo-kernel/python/fastvideo_kernel/triton_kernels/attn_qat_train.py b/fastvideo-kernel/python/fastvideo_kernel/triton_kernels/attn_qat_train.py
index 52ef4294c..0fd0aedb9 100644
--- a/fastvideo-kernel/python/fastvideo_kernel/triton_kernels/attn_qat_train.py
+++ b/fastvideo-kernel/python/fastvideo_kernel/triton_kernels/attn_qat_train.py
@@ -1054,6 +1054,15 @@ class _attention(torch.autograd.Function):
         use_global_sf_P=True,
         use_global_sf_QKV=True,
     ):
+        if smooth_q:
+            # Not implemented, not merely disabled: only the backward kernels
+            # carry a SMOOTH_Q path (_attn_fwd has none), so restoring the
+            # triton_group_mean wiring would yield silently wrong gradients
+            # rather than a feature. The flag is kept for signature
+            # compatibility only.
+            raise NotImplementedError("smooth_q=True is not supported: the forward kernel has no "
+                                      "SMOOTH_Q path, so enabling it would produce incorrect gradients")
+
         # shape constraints
         HEAD_DIM_Q, HEAD_DIM_K = q.shape[-1], k.shape[-1]
         # when v is in float8_e5m2 it is transposed.
@@ -1364,10 +1373,8 @@ class _attention(torch.autograd.Function):
                                        BLOCK_M=PRE_BLOCK,
                                        HEAD_DIM=ctx.HEAD_DIM)
 
+        # smooth_q is rejected in forward(); ctx.smooth_q is always False here.
         q_m = None
-        if ctx.smooth_q:
-            # _, q_m = triton_group_mean(q)
-            q_m = q_m.repeat_interleave(q.shape[2] // q_m.shape[2], dim=2)  # B,H,L,D
 
         sm100_optimized_backward = (getattr(ctx, "sm100_optimized", False) and ctx.use_qat_qkv_backward
                                     and not ctx.smooth_k and not ctx.smooth_q and N_CTX_KV % 16 == 0)
diff --git a/fastvideo-kernel/tests/test_attn_qat_train.py b/fastvideo-kernel/tests/test_attn_qat_train.py
index 22755e832..e221b50aa 100644
--- a/fastvideo-kernel/tests/test_attn_qat_train.py
+++ b/fastvideo-kernel/tests/test_attn_qat_train.py
@@ -4,6 +4,7 @@ Simple test for attn_qat_train.
 Tests forward and backward passes with and without QAT enabled.
 """
 
+import pytest
 import torch
 from fastvideo_kernel.triton_kernels.attn_qat_train import _attention
 from fastvideo_kernel.triton_kernels.fused_attention import attention as fused_attention
@@ -13,6 +14,27 @@ attention = _attention.apply
 DEVICE = torch.device("cuda")
 
 
+def test_smooth_q_fails_before_tensor_or_kernel_work():
+    dummy = torch.empty((1, 1, 1, 16))
+
+    with pytest.raises(NotImplementedError, match="smooth_q=True is not supported"):
+        attention(
+            dummy,
+            dummy,
+            dummy,
+            False,
+            1.0,
+            True,
+            True,
+            True,
+            True,
+            True,
+            True,
+            False,
+            True,
+        )
+
+
 def attn_qat_train_wrapper(q_BLHD, k_BLHD, v_BLHD, is_causal=False, sm_scale=None):
     """
     Wrapper function that mimics attn_qat_train from the backend wrapper.
diff --git a/fastvideo-kernel/tests/test_nvfp4_benchmark_imports.py b/fastvideo-kernel/tests/test_nvfp4_benchmark_imports.py
new file mode 100644
index 000000000..44f932ade
--- /dev/null
+++ b/fastvideo-kernel/tests/test_nvfp4_benchmark_imports.py
@@ -0,0 +1,21 @@
+import ast
+from pathlib import Path
+
+import pytest
+
+
+BENCH_DIR = Path(__file__).resolve().parents[1] / "attn_qat_infer" / "quantization" / "bench"
+
+
+@pytest.mark.parametrize("script_name", ["bench_quant_q.py", "bench_quant_k.py", "bench_quant_v.py"])
+def test_quant_benchmark_imports_built_extension(script_name: str) -> None:
+    tree = ast.parse((BENCH_DIR / script_name).read_text())
+    imports = {
+        (alias.name, alias.asname)
+        for node in ast.walk(tree)
+        if isinstance(node, ast.Import)
+        for alias in node.names
+    }
+
+    assert ("fp4quant_cuda", "fp4quant") in imports
+    assert ("fp4quant", None) not in imports
diff --git a/fastvideo-kernel/tests/test_turbodiffusion.py b/fastvideo-kernel/tests/test_turbodiffusion.py
index 771ab7ffa..02ae43bdf 100644
--- a/fastvideo-kernel/tests/test_turbodiffusion.py
+++ b/fastvideo-kernel/tests/test_turbodiffusion.py
@@ -26,6 +26,14 @@ def rms_norm_ref(x, w, eps=1e-6):
 @pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA not available")
 class TestTurboDiffusion:
 
+    def test_quant_invalid_dtype_names_int8_operation(self):
+        if turbodiffusion_ops.quant_cuda is None:
+            pytest.skip("quant_cuda not available")
+
+        x = torch.ones((16, 128), dtype=torch.float32, device="cuda")
+        with pytest.raises(RuntimeError, match=r"Unsupported input data type for quant_cuda \(int8 quantization\)\."):
+            turbodiffusion_ops.quant_cuda(x, None, None)
+
     @pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16])
     @pytest.mark.parametrize("shape", [(16, 128), (32, 256), (1, 1024)])
     def test_quant_correctness(self, dtype, shape):