fix

Author: sufubao

lightllm/common/basemodel/basemodel.py

@@ -110,7 +110,7 @@ def __init__(self, kvargs):
         self._init_some_value()
         self._init_custom()
         self._init_inferstate_cls()
-        # self._autotune_warmup()
+        self._autotune_warmup()
         self._init_padded_req()
         # wait必须在init cudagraph 之前，避免错误捕获
         self._wait_other_modules_ready()

lightllm/models/deepseek2/model.py

@@ -50,7 +50,7 @@ def __init__(self, model):
                 self.softmax_scale = self.softmax_scale * mscale * mscale
 
 
-@ModelRegistry(["deepseek_v2", "deepseek_v3", "deepseek_v32"])
+@ModelRegistry(["deepseek_v2", "deepseek_v3"])
 class Deepseek2TpPartModel(LlamaTpPartModel):
     # weight class
     transformer_weight_class = Deepseek2TransformerLayerWeight

lightllm/models/deepseek3_2/model.py

@@ -5,7 +5,7 @@
 from lightllm.utils.envs_utils import get_env_start_args
 from lightllm.models.deepseek3_2.infer_struct import Deepseek3_2FlashAttentionStateInfo
 from lightllm.models.deepseek3_2.mem_manager import Deepseek3_2MemoryManager, Deepseek3_2FP8KVMemoryManager
-# @ModelRegistry(["deepseek_v32"])
+@ModelRegistry(["deepseek_v32"])
 class Deepseek3_2TpPartModel(Deepseek2TpPartModel):
     # weight class
     transformer_weight_class = Deepseek3_2TransformerLayerWeight