[Fix] Fix Nemotron-H Mamba illegal memory access under DP attention with CUDA graph (#34561)

Co-authored-by: Brayden Zhong <b8zhong@uwaterloo.ca>

python/sglang/srt/models/nemotron_h.py

@@ -563,7 +563,16 @@ class NemotronHMambaDecoderLayer(NemotronHAttnLikeDecoderLayer):
             if get_real_num_tokens(hidden_states, forward_batch) == 0:
                 return torch.zeros_like(hidden_states), residual
 
-            output = self._forward_mamba(hidden_states, forward_batch)
+            if is_in_breakable_cuda_graph():
+                output = torch.empty_like(hidden_states)
+                breakable_nemotron_mamba2_with_output(
+                    hidden_states, output, self.layer_id, False
+                )
+            elif is_in_tc_piecewise_cuda_graph():
+                output = torch.empty_like(hidden_states)
+                nemotron_mamba2_with_output(hidden_states, output, self.layer_id, False)
+            else:
+                output = self._forward_mamba(hidden_states, forward_batch)
             return output, residual
 
         hidden_states, residual = input_norm_maybe_fuse_allreduce(

test/registered/4-gpu-models/test_nvidia_nemotron_3_super_nvfp4.py

@@ -52,8 +52,6 @@ DP_ATTENTION_EP_ARGS = [
     "0.93",
     "--max-prefill-tokens",
     "8192",
-    "--cuda-graph-backend-prefill",
-    "disabled",
 ]
 
 MTP_ARGS = [