diff --git a/vllm/models/kimi_k3/amd/kda.py b/vllm/models/kimi_k3/amd/kda.py
index 48ac78338571..7e193054c5c7 100644
--- a/vllm/models/kimi_k3/amd/kda.py
+++ b/vllm/models/kimi_k3/amd/kda.py
@@ -73,6 +73,35 @@
 logger = init_logger(__name__)
 
 
+class _PaddedKimiGDNMergedColumnParallelLinear(_KimiGDNMergedColumnParallelLinear):
+    """Merged projection with ``padding_size`` zero rows appended per rank."""
+
+    def __init__(self, *args, padding_size: int = 0, **kwargs) -> None:
+        super().__init__(*args, **kwargs)
+        self.padding_size = padding_size
+        if padding_size:
+            assert isinstance(self.weight, torch.Tensor)
+            # Exclude padding rows from the load-completion count.
+            set_weight_attrs(
+                self.weight,
+                {"weight_loader_numel": self.weight[:-padding_size].numel()},
+            )
+            self.weight.data[-padding_size:].zero_()
+
+    def weight_loader(self, param, loaded_weight, loaded_shard_id=None) -> None:
+        super().weight_loader(param, loaded_weight, loaded_shard_id)
+        self._zero_padding(param)
+
+    def weight_loader_v2(self, param, loaded_weight, loaded_shard_id=None) -> None:
+        super().weight_loader_v2(param, loaded_weight, loaded_shard_id)
+        self._zero_padding(param)
+
+    def _zero_padding(self, param: torch.Tensor) -> None:
+        # Online processing materializes the weight after __init__ zeroing.
+        if self.padding_size and param is self.weight:
+            param.data[-self.padding_size :].zero_()
+
+
 class KimiK3DeltaAttention(GatedDeltaNetAttention):
     def get_attn_backend(self) -> type[AttentionBackend]:
         return KimiK3ROCmKDABackend
@@ -159,17 +188,16 @@ def __init__(
         if self.in_proj_padding:
             in_proj_output_sizes.append(self.in_proj_padding * self.tp_size)
 
-        self.in_proj_qkvgfab = _KimiGDNMergedColumnParallelLinear(
+        self.in_proj_qkvgfab = _PaddedKimiGDNMergedColumnParallelLinear(
             self.hidden_size,
             in_proj_output_sizes,
             replicated_shard_id=4,
             tp_size=self.tp_size,
+            padding_size=self.in_proj_padding,
             bias=False,
             quant_config=self.quant_config,
             prefix=f"{prefix}.in_proj_qkvgfab",
         )
-        if self.in_proj_padding:
-            self.in_proj_qkvgfab.weight.data[-self.in_proj_padding :].zero_()
 
         self.f_b_proj = ColumnParallelLinear(
             self.head_dim,