[bugfix] fix qwen3_next norm sp

src/mcore_bridge/model/gpts/glm4.py

@@ -1,7 +1,6 @@
 # Copyright (c) ModelScope Contributors. All rights reserved.
 from megatron.core.dist_checkpointing.mapping import ShardedStateDict
 from megatron.core.extensions.transformer_engine import TENorm
-from megatron.core.transformer import transformer_layer
 from megatron.core.transformer.attention import SelfAttention
 from megatron.core.transformer.mlp import MLP, apply_swiglu_sharded_factory
 from megatron.core.transformer.spec_utils import build_module

src/mcore_bridge/model/gpts/qwen3_next.py

@@ -81,6 +81,8 @@ def __init__(self, config: ModelConfig, hidden_size: int, eps: float = 1e-5):
         self.eps = eps
         # Initialize weight to zeros (Zero-Centered), matching HuggingFace Qwen3NextRMSNorm
         self.weight = torch.nn.Parameter(torch.zeros(hidden_size))
+        # Mark weight for SP gradient AllReduce across TP domain (consistent with TENorm/MCoreRMSNorm)
+        setattr(self.weight, 'sequence_parallel', config.sequence_parallel)
 
     def _norm(self, x):
         return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)