diff --git a/tests/kernels/quantization/test_triton_w4a16.py b/tests/kernels/quantization/test_triton_w4a16.py
index 42f163dea4..3c46829116 100644
--- a/tests/kernels/quantization/test_triton_w4a16.py
+++ b/tests/kernels/quantization/test_triton_w4a16.py
@@ -306,6 +306,7 @@ def test_triton_w4a16_process_weights_after_loading_repacks_layout():
 
 @pytest.mark.skipif(not current_platform.is_rocm(), reason="ROCm only")
 def test_triton_w4a16_process_weights_after_loading_keeps_gptq_qzeros_layout():
+    """AutoGPTQ qzeros are already [K//G, N//8] (output_dim=1): no transpose."""
     if not torch.cuda.is_available():
         pytest.skip("CUDA/HIP device not available")
 
@@ -413,6 +414,7 @@ def test_triton_w4a16_process_weights_after_loading_keeps_gptq_qzeros_layout():
 
 @pytest.mark.skipif(not current_platform.is_rocm(), reason="ROCm only")
 def test_triton_w4a16_symmetric_apply_ignores_qzeros(monkeypatch):
+    """For symmetric (uint4b8) layers, apply_weights must pass qzeros=None."""
     from vllm.model_executor.kernels.linear.mixed_precision.MPLinearKernel import (
         MPLinearLayerConfig,
     )
diff --git a/vllm/model_executor/kernels/linear/mixed_precision/triton_w4a16.py b/vllm/model_executor/kernels/linear/mixed_precision/triton_w4a16.py
index bc0a587b67..8a4a493c6c 100644
--- a/vllm/model_executor/kernels/linear/mixed_precision/triton_w4a16.py
+++ b/vllm/model_executor/kernels/linear/mixed_precision/triton_w4a16.py
@@ -400,29 +400,40 @@ class TritonW4A16LinearKernel(MPLinearKernel):
         if self.w_zp_name is not None:
             zp = getattr(layer, self.w_zp_name, None)
             if zp is not None:
-                c = self.config
-                K, N = c.partition_weight_shape
-                group_size = c.group_size if c.group_size != -1 else K
-                expected_shape = (K // group_size, N // 8)
-                transposed_shape = (N // 8, K // group_size)
-
-                if tuple(zp.data.shape) == expected_shape:
-                    # GPTQ/AutoGPTQ already stores qzeros in the kernel layout.
-                    qzeros = zp.data.contiguous()
-                elif tuple(zp.data.shape) == transposed_shape:
-                    # Compressed-tensors stores qzeros transposed from what the
-                    # kernel needs.
-                    qzeros = zp.data.t().contiguous()
+                # Kernel needs [K//G, N//8]:
+                #        input(K) at dim 0, output(N) packed at dim 1.
+                # AutoGPTQ:
+                #        output_dim=1 -> already [K//G, N//8], no transpose.
+                # compressed-tensors:
+                #        output_dim=0 -> [N//8, K//G], needs transpose.
+                # None (unknown):
+                #        infer from shape; if square (ambiguous), default to transpose.
+                zp_output_dim = getattr(zp, "output_dim", None)
+                if zp_output_dim is not None:
+                    needs_transpose = zp_output_dim != 1
                 else:
-                    raise AssertionError(
-                        f"{self.w_zp_name} shape mismatch: {zp.data.shape}; "
-                        f"expected {expected_shape} or {transposed_shape}"
-                    )
-
+                    # in case output_dim is None
+                    c = self.config
+                    K, N = c.partition_weight_shape
+                    group_size = c.group_size if c.group_size != -1 else K
+                    expected_shape = (K // group_size, N // 8)
+                    transposed_shape = (N // 8, K // group_size)
+                    if (
+                        tuple(zp.data.shape) == expected_shape
+                        and expected_shape != transposed_shape
+                    ):
+                        needs_transpose = False
+                    else:
+                        needs_transpose = True
+                zp_data = (
+                    zp.data.t().contiguous()
+                    if needs_transpose
+                    else zp.data.contiguous()
+                )
                 replace_parameter(
                     layer,
                     self.w_zp_name,
-                    torch.nn.Parameter(qzeros, requires_grad=False),
+                    torch.nn.Parameter(zp_data, requires_grad=False),
                 )
 
     def apply_weights(
@@ -437,17 +448,18 @@ class TritonW4A16LinearKernel(MPLinearKernel):
         K = c.partition_weight_shape[0]
         group_size = c.group_size if c.group_size != -1 else K
 
-        # For symmetric types (uint4b8), use the scalar bias; no zeros tensor.
-        # Some checkpoint loaders still register qzeros parameters for GPTQ
-        # layers, but they are not part of the symmetric kernel contract.
-        zp_bias = c.weight_type.bias if c.weight_type.has_bias() else 0
-        qzeros = None if c.weight_type.has_bias() else w_zp
+        # For symmetric types (uint4b8), use the scalar bias; no zeros tensor
+        if c.weight_type.has_bias():
+            zp_bias = c.weight_type.bias
+            w_zp = None  # symmetric: ignore qzeros, use scalar bias instead
+        else:
+            zp_bias = 0
 
         output = triton_w4a16_gemm(
             a=x_2d,
             b_q=w_q,
             scales=w_s,
-            qzeros=qzeros,
+            qzeros=w_zp,
             group_size=group_size,
             zp_bias=zp_bias,
         )