fix: use NVFP4 SymmBuffer (2x SF size for group_size=16)

The NVFP4 mega_moe kernel needs a larger symmetric buffer because
group_size=16 produces 2x more scale factor entries than MXFP4's 32.
Switch from deep_gemm.get_symm_buffer_for_mega_moe to
deep_gemm.mega.nvfp4.get_symm_buffer_for_nvfp4_mega_moe.

patches/deepseek_v4.py

@@ -761,7 +761,7 @@ class DeepseekV4MegaMoEExperts(nn.Module):
         return w_packed, scale_exp
 
     def get_symm_buffer(self):
-        import vllm.third_party.deep_gemm as deep_gemm
+        from deep_gemm.mega import nvfp4 as nvfp4_mega
 
         group = get_ep_group().device_group
         device = torch.accelerator.current_device_index()
@@ -776,7 +776,8 @@ class DeepseekV4MegaMoEExperts(nn.Module):
         )
         symm_buffer = self._symm_buffer_cache.get(key)
         if symm_buffer is None:
-            symm_buffer = deep_gemm.get_symm_buffer_for_mega_moe(
+            # NVFP4 SymmBuffer: 2x SF size due to group_size=16 (vs MXFP4's 32)
+            symm_buffer = nvfp4_mega.get_symm_buffer_for_nvfp4_mega_moe(
                 group,
                 self.num_experts,
                 self.max_num_tokens,