[model] support hy_v3

README.md

@@ -90,6 +90,7 @@ The following is the list of models supported by MCore-Bridge:
 | Ovis     | ovis2_5                                                      |
 | Llama    | llama, llama4                                                |
 | GPT-OSS  | gpt_oss                                                      |
+| Hunyuan  | hy_v3                                                        |
 | ERNIE    | ernie4_5, ernie4_5_moe                                       |
 | MiMo     | mimo                                                         |
 | Dots     | dots1                                                        |

README_zh.md

@@ -89,6 +89,7 @@ uv pip install -e . --torch-backend=auto
 | Ovis     | ovis2_5                                                      |
 | Llama    | llama, llama4                                                |
 | GPT-OSS  | gpt_oss                                                      |
+| Hunyuan  | hy_v3                                                        |
 | ERNIE    | ernie4_5, ernie4_5_moe                                       |
 | MiMo     | mimo                                                         |
 | Dots     | dots1                                                        |

src/mcore_bridge/config/parser.py

@@ -39,7 +39,7 @@
     'qk_pos_emb_head_dim': ['qk_rope_head_dim'],
     'v_head_dim': ['v_head_dim'],
     'moe_router_topk_scaling_factor': ['routed_scaling_factor'],
-    'qk_layernorm': ['use_qk_norm'],
+    'qk_layernorm': ['use_qk_norm', 'qk_norm'],
     # qwen3_next/qwen3_5
     'linear_attention_freq': ['full_attention_interval'],
     'linear_num_key_heads': ['linear_num_key_heads'],
@@ -56,7 +56,7 @@
     'original_max_position_embeddings': ['original_max_position_embeddings'],
     'partial_rotary_factor': ['partial_rotary_factor'],
     'first_k_dense_replace': ['first_k_dense_replace', 'moe_layer_start_index'],
-    'n_shared_experts': ['n_shared_experts', 'num_shared_expert', 'moe_num_shared_experts'],
+    'n_shared_experts': ['n_shared_experts', 'num_shared_expert', 'moe_num_shared_experts', 'num_shared_experts'],
     'window_size': ['sliding_window'],
     'layer_types': ['layer_types'],
     'interleave_moe_layer_step': ['interleave_moe_layer_step'],

src/mcore_bridge/model/constant.py

@@ -7,6 +7,7 @@ class LLMModelType:
     olmoe = 'olmoe'
     glm4 = 'glm4'
     minimax_m2 = 'minimax_m2'
+    hy_v3 = 'hy_v3'
 
     qwen3_emb = 'qwen3_emb'
 

src/mcore_bridge/model/gpts/__init__.py

@@ -1,2 +1,2 @@
 # Copyright (c) ModelScope Contributors. All rights reserved.
-from . import glm4, llm, minimax_m2, olmoe, qwen3_emb, qwen3_next
+from . import glm4, hunyuan, llm, minimax_m2, olmoe, qwen3_emb, qwen3_next

src/mcore_bridge/model/gpts/hunyuan.py

@@ -0,0 +1,19 @@
+# Copyright (c) ModelScope Contributors. All rights reserved.
+
+from mcore_bridge.bridge import GPTBridge
+
+from ..constant import ModelType
+from ..register import ModelMeta, register_model
+
+
+class HyV3Bridge(GPTBridge):
+    hf_gate_key = 'router.gate.weight'
+    hf_expert_bias_key = 'expert_bias'
+    hf_shared_expert_key = 'shared_mlp'
+
+
+register_model(ModelMeta(
+    ModelType.hy_v3,
+    ['hy_v3'],
+    bridge_cls=HyV3Bridge,
+))

src/mcore_bridge/model/gpts/llm.py

@@ -1,6 +1,4 @@
 # Copyright (c) ModelScope Contributors. All rights reserved.
-from mcore_bridge.bridge import GPTBridge
-
 from ..constant import ModelType
 from ..register import ModelMeta, register_model