dspark-nvfp4-draft-routing.patch

diff --git a/vllm/models/deepseek_v4/quant_config.py b/vllm/models/deepseek_v4/quant_config.py
index 89cf695..5918fa0 100644
--- a/vllm/models/deepseek_v4/quant_config.py
+++ b/vllm/models/deepseek_v4/quant_config.py
@@ -4,6 +4,7 @@
 
 from __future__ import annotations
 
+import re
 from typing import TYPE_CHECKING, cast
 
 from vllm.config import get_current_vllm_config
@@ -175,7 +176,7 @@ class DeepseekV4FP8Config(Fp8Config):
             ):
                 return UnquantizedFusedMoEMethod(layer.moe_config)
             if self.expert_dtype == "fp4":
-                if self.moe_quant_algo == "NVFP4":
+                if self.moe_quant_algo == "NVFP4" and not self._is_draft_layer(prefix):
                     from vllm.model_executor.layers.quantization.modelopt import (
                         ModelOptNvFp4FusedMoE,
                     )
@@ -189,7 +190,20 @@ class DeepseekV4FP8Config(Fp8Config):
             # returns Fp8MoEMethod with block-wise float32 scales.
         return super().get_quant_method(layer, prefix)
 
+    def _is_draft_layer(self, prefix: str) -> bool:
+        # NVFP4 conversions quantize only the main stack; the speculative
+        # module (runtime layer index >= num_hidden_layers) keeps its
+        # original MXFP4 experts and must load through the MXFP4 path.
+        m = re.search(r"layers\.(\d+)\.", prefix)
+        if m is None:
+            return False
+        try:
+            hf_config = get_current_vllm_config().model_config.hf_config
+        except Exception:
+            return False
+        return int(m.group(1)) >= hf_config.num_hidden_layers
+
     def is_mxfp4_quant(self, prefix, layer):
         if not isinstance(layer, RoutedExperts) or self.expert_dtype != "fp4":
             return False
-        return self.moe_quant_algo != "NVFP4"
+        return self.moe_quant_algo != "NVFP4" or self._is_draft_layer(prefix)