diff --git a/vllm/models/deepseek_v4/nvidia/dspark.py b/vllm/models/deepseek_v4/nvidia/dspark.py
index 6ddf70e1c16a..69bb42c791cd 100644
--- a/vllm/models/deepseek_v4/nvidia/dspark.py
+++ b/vllm/models/deepseek_v4/nvidia/dspark.py
@@ -700,13 +700,17 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         self.max_batch = max(1, int(vllm_config.scheduler_config.max_num_seqs))
         hidden_size = int(config.hidden_size)
         dtype = vllm_config.model_config.dtype
-        quant_config = get_draft_quant_config(vllm_config)
+        self.quant_config = get_draft_quant_config(vllm_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.dspark_aux_hidden_size = hidden_size * len(self.target_layer_ids)
 
         self.embed_tokens = VocabParallelEmbedding(
             config.vocab_size,
             hidden_size,
-            quant_config=quant_config,
+            quant_config=self.quant_config,
             prefix=f"{prefix}embed_tokens",
         )
         self.main_proj = ReplicatedLinear(
@@ -714,7 +718,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
             hidden_size,
             bias=False,
             params_dtype=dtype,
-            quant_config=quant_config,
+            quant_config=self.quant_config,
             prefix=f"{prefix}mtp.0.main_proj",
             return_bias=False,
         )
@@ -1044,6 +1048,12 @@ def map_name(name: str) -> str | None:
                     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:
                 if not name.startswith("layers."):
                     continue