Enable `NemotronHPuzzle` + `NemotronHMTP` (#37803)

vllm/config/speculative.py

@@ -280,7 +280,7 @@ class SpeculativeConfig:
             )
 
         if (
-            hf_config.model_type == "nemotron_h"
+            hf_config.model_type in {"nemotron_h", "nemotron_h_puzzle"}
             and hasattr(hf_config, "num_nextn_predict_layers")
             and hf_config.num_nextn_predict_layers > 0
         ):

vllm/model_executor/models/nemotron_h_mtp.py

@@ -395,13 +395,16 @@ class NemotronHMTP(nn.Module, SupportsPP):
         ]
 
         expert_params_mapping = []
-        if hasattr(self.config, "n_routed_experts") and self.config.n_routed_experts:
+        num_experts = getattr(self.config, "n_routed_experts", None)
+        if getattr(self.config, "model_type", None) == "nemotron_h_puzzle":
+            num_experts = self.config.mtp_n_routed_experts
+        if num_experts is not None:
             expert_params_mapping = FusedMoE.make_expert_params_mapping(
                 self,
                 ckpt_gate_proj_name="up_proj",
                 ckpt_down_proj_name="down_proj",
                 ckpt_up_proj_name="",  # Empty - non-gated MoE
-                num_experts=self.config.n_routed_experts,
+                num_experts=num_experts,
                 num_redundant_experts=self.num_redundant_experts,
             )