[Fix] Keep mHC context out of non-V4 compiled MoE forwards (#40353)

python/sglang/srt/models/deepseek_v2.py

@@ -608,6 +608,7 @@ class DeepseekV2MoE(nn.Module):
         self.alt_stream = alt_stream
         self.routed_quant_stream = routed_quant_stream
         self.is_nextn = is_nextn
+        self.is_deepseek_v4 = is_deepseek_v4
         self._fuse_finalize_all_reduce = (
             is_deepseek_v4
             and getattr(config, "hc_pre_from_prev_sublayer", False)
@@ -1185,8 +1186,10 @@ class DeepseekV2MoE(nn.Module):
             )
 
         if not all_reduce_done:
-            if self.tp_size > 1 and not should_skip_post_experts_all_reduce(
-                is_tp_path=True
+            if (
+                self.is_deepseek_v4
+                and self.tp_size > 1
+                and not should_skip_post_experts_all_reduce(is_tp_path=True)
             ):
                 from sglang.srt.layers.moe.mhc_post_fusion import (
                     current_mhc_post_fusion,
@@ -1341,8 +1344,10 @@ class DeepseekV2MoE(nn.Module):
             self.routed_scaling_factor,
         )
 
-        if self.tp_size > 1 and not should_skip_post_experts_all_reduce(
-            is_tp_path=True,
+        if (
+            self.is_deepseek_v4
+            and self.tp_size > 1
+            and not should_skip_post_experts_all_reduce(is_tp_path=True)
         ):
             from sglang.srt.layers.moe.mhc_post_fusion import current_mhc_post_fusion