[https://nvbugs/6264844][fix] Fix wrong NCCL fallback in nemotron-h (#15294)

Signed-off-by: Wanli Jiang <35160485+Wanli-Jiang@users.noreply.github.com>

Author: Wanli-Jiang

tensorrt_llm/_torch/models/modeling_nemotron_h.py

@@ -277,9 +277,11 @@ def _moe(name):
         )
 
         if reduce_output:
+            # AllReduce needs dtype at construction to build fused MNNVL paths.
             self.allreduce = AllReduce(
                 mapping=model_config.mapping,
                 strategy=model_config.allreduce_strategy,
+                dtype=config.torch_dtype,
             )
         else:
             self.allreduce = None
@@ -488,9 +490,11 @@ def __init__(
         )
 
         if fuse_allreduce_norm and layer_idx > 0:
+            # AllReduce needs dtype at construction to build fused MNNVL paths.
             self.pre_allreduce = AllReduce(
                 mapping=model_config.mapping,
                 strategy=model_config.allreduce_strategy,
+                dtype=config.torch_dtype,
             )
 
         # Mixer creation.  The fuse_allreduce_norm optimization is orthogonal
@@ -717,9 +721,11 @@ def __init__(self, model_config: NemotronHModelConfig):
 
         # AllReduce for fusing with final norm (after last layer's mixer)
         if self.fuse_allreduce_norm:
+            # AllReduce needs dtype at construction to build fused MNNVL paths.
             self.final_allreduce = AllReduce(
                 mapping=model_config.mapping,
                 strategy=model_config.allreduce_strategy,
+                dtype=config.torch_dtype,
             )
 
     def forward(