Disable breakable CUDA graph for NemotronH (#33130)

python/sglang/srt/configs/model_config.py

@@ -129,6 +129,14 @@ 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

@@ -4426,7 +4426,11 @@ 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_dsa, is_deepseek_v4
+        from sglang.srt.configs.model_config import (
+            is_deepseek_dsa,
+            is_deepseek_v4,
+            is_nemotron_h,
+        )
 
         rules = [
             # MLA prefill under BCG takes forward_mha, which has no eager
@@ -4437,6 +4441,13 @@ class ServerArgs:
                 lambda: self.use_mla_backend()
                 and not is_deepseek_dsa(self.get_model_config().hf_config),
             ),
+            # 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.
             (