fix xpu

Author: YuanRisheng

fastdeploy/engine/engine.py

@@ -1194,10 +1194,10 @@ def detect_thread():
                                              r'set state for layer (\d+)',
                                              line)):
                     progress = eval(match.group(
-                        1)) * 1.0 / self.cfg.model_config.num_layers
+                        1)) * 1.0 / self.cfg.model_config.num_hidden_layers
                     self.worker_init_status["layer_loadding"] = progress
                     if self.worker_init_status[
-                            "layer_loadding"] == self.cfg.model_config.num_layers - 1:
+                            "layer_loadding"] == self.cfg.model_config.num_hidden_layers - 1:
                         self.worker_init_status["finished"] = True
 
         self.checking_worker_status_thread = threading.Thread(

fastdeploy/model_executor/layers/attention/iluvatar_attn_backend.py

@@ -101,7 +101,7 @@ def __init__(self, llm_config: FDConfig, kv_num_heads: int, num_heads: int,
         self.head_dim = head_dim
         # note: scale need to change if using MLA
         self.attention_metadata.scale = 1.0 / sqrt(head_dim)
-        self.num_layers = llm_config.model_config.num_layers
+        self.num_layers = llm_config.model_config.num_hidden_layers
         self.record_block_table_metadata = {}
         self.only_use_flash_attn = int(
             os.getenv("FD_ILUVATAR_ONLY_USE_FLASH_ATTN", 0)) == 1

fastdeploy/model_executor/layers/backends/gcu/attention/flash_attn_backend.py

@@ -91,7 +91,7 @@ def __init__(self, fd_config: FDConfig, kv_num_heads: int, num_heads: int,
         self.num_heads = num_heads
         self.head_dim = head_dim
         self.scaling = 1.0 / (self.head_dim**0.5)
-        self.num_layers = fd_config.model_config.num_layers
+        self.num_layers = fd_config.model_config.num_hidden_layers
         self.position_ids_base = paddle.arange(self.max_seq_len)
 
         # TODO(zhengjun): Need to adapt the allocation logic and

fastdeploy/model_executor/layers/backends/gcu/attention/mem_efficient_attn_backend.py

@@ -90,7 +90,7 @@ def __init__(self, fd_config: FDConfig, kv_num_heads: int, num_heads: int,
         self.num_heads = num_heads
         self.head_dim = head_dim
         self.scaling = 1.0 / (self.head_dim**0.5)
-        self.num_layers = fd_config.model_config.num_layers
+        self.num_layers = fd_config.model_config.num_hidden_layers
         self.position_ids_base = paddle.arange(self.max_seq_len)
 
         # TODO(zhengjun): Need to adapt the allocation logic and

fastdeploy/model_executor/load_weight_utils.py

@@ -43,7 +43,7 @@ def load_ep_checkpoint(model_path: str,
     filtered_map = {k: v for k, v in weight_list.items() if "experts" not in k}
     num_local_ffn_keys = []
 
-    for i in range(config.moe_layer_start_index, config.num_layers):
+    for i in range(config.moe_layer_start_index, config.num_hidden_layers):
         for j in range(
                 config.num_experts_start_offset,
                 config.num_experts_start_offset + config.num_experts_per_rank,

fastdeploy/model_executor/models/deepseek_v3.py

@@ -758,5 +758,5 @@ def get_tensor_parallel_split_mappings(num_layers):
 
             return final_actions
 
-        mappings = get_tensor_parallel_split_mappings(config.num_layers)
+        mappings = get_tensor_parallel_split_mappings(config.num_hidden_layers)
         return mappings

fastdeploy/model_executor/models/ernie4_5_moe.py

@@ -618,7 +618,7 @@ def get_tensor_parallel_split_mappings(num_layers, moe_num_experts,
         elif isinstance(config.moe_layer_start_index, int):
             moe_layer_start_index = config.moe_layer_start_index
 
-        mappings = get_tensor_parallel_split_mappings(config.num_layers,
+        mappings = get_tensor_parallel_split_mappings(config.num_hidden_layers,
                                                       moe_num_experts,
                                                       moe_layer_start_index,
                                                       config.prefix_name)

fastdeploy/model_executor/models/ernie4_5_mtp.py

@@ -237,7 +237,7 @@ def get_tensor_parallel_split_mappings(num_layers, moe_num_experts,
 
         moe_num_experts = 0
         mappings = get_tensor_parallel_split_mappings(
-            config.num_layers,
+            config.num_hidden_layers,
             moe_num_experts,
             config.moe_layer_start_index,
         )

fastdeploy/model_executor/models/qwen2.py

@@ -427,6 +427,6 @@ def get_tensor_parallel_split_mappings(num_layers):
 
             return final_actions
 
-        mappings = get_tensor_parallel_split_mappings(config.num_layers)
+        mappings = get_tensor_parallel_split_mappings(config.num_hidden_layers)
 
         return mappings

fastdeploy/model_executor/models/qwen3.py

@@ -357,5 +357,5 @@ def get_tensor_parallel_split_mappings(num_layers):
 
             return final_actions
 
-        mappings = get_tensor_parallel_split_mappings(config.num_layers)
+        mappings = get_tensor_parallel_split_mappings(config.num_hidden_layers)
         return mappings