[bugfix] fix bug

README.md

@@ -52,7 +52,7 @@ You can contact us and communicate with us by adding our group:
 
 ## 📝 Introduction
 
-**mcore-bridge** is a large language model and multimodal large model definition library built on the Megatron-Core ecosystem, developed by the ModelScope community. It currently supports 300+ text-only models and 200+ multimodal models, including large language models such as Qwen3-Next, GLM5.1, DeepSeek-V3.2, Minimax2.7, Kimi K2.5, and GPT-OSS, as well as multimodal large models such as Qwen3.5, Qwen3-Omni, Gemma4, GLM4.6-V, InternVL3.5, and Ovis2.5.
+**mcore-bridge** is a large language model and multimodal large model definition library built on the Megatron-Core ecosystem, developed by the ModelScope community. It currently supports 300+ text-only models and 200+ multimodal models, including large language models such as Qwen3-Next, GLM-5.1, DeepSeek-V4, Minimax-2.7, Kimi-K2.5, and GPT-OSS, as well as multimodal large models such as Qwen3.5, Qwen3-Omni, Gemma4, GLM4.6-V, InternVL3.5, and Ovis2.5.
 
 ------
 

README_zh.md

@@ -51,7 +51,7 @@
 
 ## 📝 简介
 
-**mcore-bridge** 是由魔搭社区推出的、基于 Megatron-Core 生态构建的大模型与多模态大模型定义库。目前已支持 300+ 纯文本模型与 200+ 多模态模型。其中大语言模型包括 Qwen3-Next、GLM5.1、DeepSeek-V3.2、Minimax2.7、Kimi K2.5、GPT-OSS 等；多模态大模型包括 Qwen3.5、Qwen3-Omni、Gemma4、GLM4.6-V、InternVL3.5、Ovis2.5 等。
+**mcore-bridge** 是由魔搭社区推出的、基于 Megatron-Core 生态构建的大模型与多模态大模型定义库。目前已支持 300+ 纯文本模型与 200+ 多模态模型。其中大语言模型包括 Qwen3-Next、GLM-5.1、DeepSeek-V4、Minimax-2.7、Kimi-K2.5、GPT-OSS 等；多模态大模型包括 Qwen3.5、Qwen3-Omni、Gemma4、GLM4.6-V、InternVL3.5、Ovis2.5 等。
 
 ------
 

src/mcore_bridge/config/model_config.py

@@ -330,6 +330,11 @@ def __post_init__(self):
             self.mtp_unroll_steps = self.mtp_num_layers
         if self.csa_compress_ratios is not None and self.mtp_num_layers is not None:
             self.csa_compress_ratios += [0] * self.mtp_num_layers
+        if self.multi_latent_attention:
+            # multi_latent_attention always uses interleave, which corresponds to is_neox_style=False.
+            # deepseek-v32: interleave + non-interleave (indexer)
+            # glm-5: interleave + interleave (indexer)
+            self.rotary_interleaved = False
         super().__post_init__()
 
         self._check_npu()

src/mcore_bridge/config/parser.py

@@ -133,6 +133,7 @@ def hf_to_mcore_config(hf_config: PretrainedConfig) -> Dict[str, Any]:
     mlp_ffn_hidden_size = res.pop('mlp_ffn_hidden_size', None)
     interleave_moe_layer_step = res.pop('interleave_moe_layer_step', None)
     window_size = res.pop('window_size', None)
+    moe_n_hash_layers = res.pop('moe_n_hash_layers', None)
     rope_scaling = res.get('rope_scaling') or {}
     if llm_model_type in {'qwen3', 'qwen3_moe', 'qwen3_next'} or hf_model_type in {
             'qwen3_omni_moe', 'qwen3_omni', 'qwen3_vl', 'qwen3_vl_moe', 'qwen3_5', 'qwen3_5_moe', 'llavaonevision1_5'
@@ -161,7 +162,6 @@ def hf_to_mcore_config(hf_config: PretrainedConfig) -> Dict[str, Any]:
             res['enable_hyper_connections'] = True
             csa_compress_ratios = res.pop('csa_compress_ratios', None)
             res['csa_compress_ratios'] = [csa_compress_ratios.get(layer_type, 0) for layer_type in layer_types]
-            moe_n_hash_layers = res.pop('moe_n_hash_layers', None)
             res['moe_n_hash_layers'] = len([layer for layer in moe_n_hash_layers if layer == 'hash_moe'])
     elif llm_model_type == 'hunyuan':
         # Since HunYuan’s attention applies RoPE before using q/k_layernorm,