restore v32 support

vllm/model_executor/models/config.py

@@ -410,8 +410,8 @@ class DeepseekV32ForCausalLM(VerifyAndUpdateConfig):
         hf_config = vllm_config.model_config.hf_config
 
         # Mirror the check in vllm/model_executor/models/deepseek_v2.py
-        # is_v32 = hasattr(hf_config, "index_topk")
-        # assert is_v32
+        is_v32 = hasattr(hf_config, "index_topk")
+        assert is_v32
 
         # For DeepSeekV3.2, we use a custom fp8 format as default (i.e.
         #   "auto")

vllm/model_executor/models/deepseek_mtp.py

@@ -65,8 +65,7 @@ class DeepSeekMultiTokenPredictorLayer(nn.Module):
                                  config.hidden_size,
                                  bias=False)
 
-        # self.is_v32 = hasattr(config, "index_topk")
-        self.is_v32 = False
+        self.is_v32 = hasattr(config, "index_topk")
         if self.is_v32:
             topk_tokens = config.index_topk
             topk_indices_buffer = torch.empty(

vllm/model_executor/models/deepseek_v2.py

@@ -954,8 +954,7 @@ class DeepseekV2MLAAttention(nn.Module):
             mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
             self.scaling = self.scaling * mscale * mscale
 
-        # self.is_v32 = hasattr(config, "index_topk")
-        self.is_v32 = False
+        self.is_v32 = hasattr(config, "index_topk")
 
         if self.is_v32:
             self.indexer = Indexer(vllm_config, config, hidden_size,
@@ -1176,8 +1175,7 @@ class DeepseekV2Model(nn.Module):
         self.config = config
 
         self.vocab_size = config.vocab_size
-        # self.is_v32 = hasattr(config, "index_topk")
-        self.is_v32 = False
+        self.is_v32 = hasattr(config, "index_topk")
         if self.is_v32:
             topk_tokens = config.index_topk
             topk_indices_buffer = torch.empty(