diff --git a/vllm/models/deepseek_v4/nvidia/dspark.py b/vllm/models/deepseek_v4/nvidia/dspark.py
index e4e258372a2..bde088edae1 100644
--- a/vllm/models/deepseek_v4/nvidia/dspark.py
+++ b/vllm/models/deepseek_v4/nvidia/dspark.py
@@ -43,6 +43,7 @@ from vllm.model_executor.models.utils import maybe_prefix
 
 from .model import (
     DeepseekV4DecoderLayer,
+    DeepseekV4Model,
     make_deepseek_v4_expert_params_mapping,
 )
 
@@ -277,6 +278,11 @@ class DSparkDeepseekV4ForCausalLM(nn.Module):
         assert vllm_config.speculative_config is not None
         self.draft_model_config = vllm_config.speculative_config.draft_model_config
         self.config = self.draft_model_config.hf_config
+        self.quant_config = vllm_config.quant_config
+        self.pad_shared_expert = (
+            getattr(self.quant_config, "weight_block_size", None) is not None
+            and not vllm_config.parallel_config.use_sequence_parallel_moe
+        )
         self.model = DSparkDeepseekV4Model(
             vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
         )
@@ -396,6 +402,12 @@ class DSparkDeepseekV4ForCausalLM(nn.Module):
                     else ".weight_scale_inv"
                 )
                 name = name.removesuffix(".scale") + suffix
+            if ".shared_experts.w2" in name:
+                name = name.replace(".shared_experts.w2", ".shared_experts.down_proj")
+            if self.pad_shared_expert and ".shared_experts." in name:
+                loaded_weight = DeepseekV4Model._pad_shared_expert_weight(
+                    self.quant_config, name, loaded_weight
+                )
 
             # E8M0 expert scales: keep raw exponent bytes.
             if ".experts." in name:
@@ -440,10 +452,6 @@ class DSparkDeepseekV4ForCausalLM(nn.Module):
                     params_dict[name][: narrow.shape[0]].copy_(narrow)
                     loaded_params.add(name)
                     continue
-                if ".shared_experts.w2" in name:
-                    name = name.replace(
-                        ".shared_experts.w2", ".shared_experts.down_proj"
-                    )
                 if name.endswith(".ffn.gate.bias"):
                     name = name.replace(
                         ".ffn.gate.bias", ".ffn.gate.e_score_correction_bias"
diff --git a/vllm/models/deepseek_v4/nvidia/model.py b/vllm/models/deepseek_v4/nvidia/model.py
index ddc2fe0f4bc..96a3c77074a 100644
--- a/vllm/models/deepseek_v4/nvidia/model.py
+++ b/vllm/models/deepseek_v4/nvidia/model.py
@@ -1181,7 +1181,9 @@ class DeepseekV4Model(nn.Module, EagleModelMixin):
 
         for name, loaded_weight in weights:
             if pad_shared_expert and ".shared_experts." in name:
-                loaded_weight = self._pad_shared_expert_weight(name, loaded_weight)
+                loaded_weight = self._pad_shared_expert_weight(
+                    self.quant_config, name, loaded_weight
+                )
             for param_name, weight_name, shard_id in stacked_params_mapping:
                 # Skip non-stacked layers and experts (experts handled below).
                 if ".experts." in name:
@@ -1256,15 +1258,18 @@ class DeepseekV4Model(nn.Module, EagleModelMixin):
 
         return loaded_params
 
+    @staticmethod
     def _pad_shared_expert_weight(
-        self, name: str, loaded_weight: torch.Tensor
+        quant_config: QuantizationConfig | None,
+        name: str,
+        loaded_weight: torch.Tensor,
     ) -> torch.Tensor:
         """Zero-pad a block-FP8 shared-expert weight/scale on its intermediate
         axis so the standard TP loaders split it into even, block-aligned shards
         (trailing ranks get the zero pad). gate (w1)/up (w3) [I, H] pad dim 0;
         down (w2 -> down_proj) [H, I] pads dim 1.
         """
-        block_size = getattr(self.quant_config, "weight_block_size", None)
+        block_size = getattr(quant_config, "weight_block_size", None)
         assert block_size is not None
         # Round the intermediate axis up to a whole number of TP shards. The axis
         # is in elements for weights (step = block) and in blocks for scales.
diff --git a/vllm/models/deepseek_v4/nvidia/mtp.py b/vllm/models/deepseek_v4/nvidia/mtp.py
index 8aa3d2c9a29..ee036967a7e 100644
--- a/vllm/models/deepseek_v4/nvidia/mtp.py
+++ b/vllm/models/deepseek_v4/nvidia/mtp.py
@@ -52,6 +52,7 @@ from vllm.sequence import IntermediateTensors
 
 from .model import (
     DeepseekV4DecoderLayer,
+    DeepseekV4Model,
     make_deepseek_v4_expert_params_mapping,
 )
 
@@ -265,6 +266,10 @@ class DeepSeekV4MTP(nn.Module):
         super().__init__()
         self.config = vllm_config.model_config.hf_config
         self.quant_config = vllm_config.quant_config
+        self.pad_shared_expert = (
+            getattr(self.quant_config, "weight_block_size", None) is not None
+            and not vllm_config.parallel_config.use_sequence_parallel_moe
+        )
         self.model = DeepSeekV4MultiTokenPredictor(
             vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
         )
@@ -387,6 +392,12 @@ class DeepSeekV4MTP(nn.Module):
                     else ".weight_scale_inv"
                 )
                 name = name.removesuffix(".scale") + suffix
+            if ".shared_experts.w2" in name:
+                name = name.replace(".shared_experts.w2", ".shared_experts.down_proj")
+            if self.pad_shared_expert and ".shared_experts." in name:
+                loaded_weight = DeepseekV4Model._pad_shared_expert_weight(
+                    self.quant_config, name, loaded_weight
+                )
             for param_name, weight_name, shard_id in stacked_params_mapping:
                 # Skip non-stacked layers and experts (experts handled below).
                 if ".experts." in name:
@@ -442,10 +453,6 @@ class DeepSeekV4MTP(nn.Module):
                     loaded_params.add(name)
                     continue
                 else:
-                    if ".shared_experts.w2" in name:
-                        name = name.replace(
-                            ".shared_experts.w2", ".shared_experts.down_proj"
-                        )
                     if name.endswith(".ffn.gate.bias"):
                         # ``e_score_correction_bias`` lives on the gate
                         # under a different attribute name.