Reenable breakable CUDA graph for NemotronH (#34538)

python/sglang/srt/configs/model_config.py

@@ -133,14 +133,6 @@ def is_deepseek_v4(config) -> bool:
     )
 
 
-def is_nemotron_h(config) -> bool:
-    return _hf_arch(config) in (
-        "NemotronHForCausalLM",
-        "NemotronHPuzzleForCausalLM",
-        "NemotronHForCausalLMMTP",
-    )
-
-
 def get_dsa_index_head_dim(config: PretrainedConfig) -> int:
     assert is_deepseek_dsa(config) or is_deepseek_v4(config)
     return config.index_head_dim

python/sglang/srt/server_args.py

@@ -4568,17 +4568,10 @@ class ServerArgs:
         memory-saver rejection in its own __init__; config-time rules can be
         added here as they're discovered.
         """
-        from sglang.srt.configs.model_config import is_deepseek_v4, is_nemotron_h
+        from sglang.srt.configs.model_config import is_deepseek_v4
         from sglang.srt.layers.cp.bcg import supports_prefill_cp_bcg
 
         rules = [
-            # NemotronH's hybrid Mamba2 prefill is not BCG-safe: the mamba
-            # state-track write is not wired into the captured buffers, so a
-            # replay can commit a cache slot it never wrote.
-            (
-                "NemotronH (hybrid Mamba2 prefill)",
-                lambda: is_nemotron_h(self.get_model_config().hf_config),
-            ),
             # DSV4 is BCG-compatible but introduces heavy memory pressure: the
             # c4 indexer scratch is pinned in the capture pool and OOMs. Disable.
             (

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

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