[Misc] Remove duplicated DeepSeek V2/V3 model definition (#12793)

vllm/config.py

@@ -754,7 +754,6 @@ class ModelConfig:
 
     @property
     def is_deepseek_mla(self) -> bool:
-        # TODO add deepseek_v3
         return (hasattr(self.hf_text_config, "model_type")) \
                 and (self.hf_text_config.model_type in \
                     ('deepseek_v2', 'deepseek_v3'))\

vllm/model_executor/models/deepseek_v2.py

@@ -21,7 +21,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-"""Inference-only DeepseekV2 model."""
+"""Inference-only DeepseekV2/DeepseekV3 model."""
 from typing import Any, Dict, Iterable, List, Optional, Set, Tuple, Union
 
 import torch
@@ -115,7 +115,19 @@ class DeepseekV2MoE(nn.Module):
             raise ValueError(f"Unsupported activation: {config.hidden_act}. "
                              "Only silu is supported for now.")
 
-        self.experts = FusedMoE(num_experts=config.n_routed_experts,
+        self.gate = ReplicatedLinear(config.hidden_size,
+                                     config.n_routed_experts,
+                                     bias=False,
+                                     quant_config=None,
+                                     prefix=f"{prefix}.gate")
+        if config.topk_method == "noaux_tc":
+            self.gate.e_score_correction_bias = nn.Parameter(
+                torch.empty(config.n_routed_experts))
+        else:
+            self.gate.e_score_correction_bias = None
+
+        self.experts = FusedMoE(
+            num_experts=config.n_routed_experts,
             top_k=config.num_experts_per_tok,
             hidden_size=config.hidden_size,
             intermediate_size=config.moe_intermediate_size,
@@ -125,13 +137,10 @@ class DeepseekV2MoE(nn.Module):
             use_grouped_topk=True,
             num_expert_group=config.n_group,
             topk_group=config.topk_group,
-                                prefix=f"{prefix}.experts")
+            prefix=f"{prefix}.experts",
+            scoring_func=config.scoring_func,
+            e_score_correction_bias=self.gate.e_score_correction_bias)
 
-        self.gate = ReplicatedLinear(config.hidden_size,
-                                     config.n_routed_experts,
-                                     bias=False,
-                                     quant_config=None,
-                                     prefix=f"{prefix}.gate")
         if config.n_shared_experts is not None:
             intermediate_size = (config.moe_intermediate_size *
                                  config.n_shared_experts)
@@ -732,6 +741,15 @@ class DeepseekV2ForCausalLM(nn.Module, SupportsPP):
         for name, loaded_weight in weights:
             if "rotary_emb.inv_freq" in name:
                 continue
+
+            # TODO(simon): support nextn predict layers
+            if hasattr(self.config, "num_nextn_predict_layers"
+                       ) and self.config.num_nextn_predict_layers > 0:
+                assert self.config.num_nextn_predict_layers == 1
+                layer_idx = self.config.num_hidden_layers
+                if name.startswith(f"model.layers.{layer_idx}"):
+                    continue
+
             for (param_name, weight_name, shard_id) in stacked_params_mapping:
                 # Skip non-stacked layers and experts (experts handled below).
                 if weight_name not in name:
@@ -793,3 +811,7 @@ class DeepseekV2ForCausalLM(nn.Module, SupportsPP):
                     weight_loader(param, loaded_weight)
             loaded_params.add(name)
         return loaded_params
+
+
+class DeepseekV3ForCausalLM(DeepseekV2ForCausalLM):
+    pass

vllm/model_executor/models/registry.py

@@ -45,7 +45,7 @@ _TEXT_GENERATION_MODELS = {
     "DeciLMForCausalLM": ("decilm", "DeciLMForCausalLM"),
     "DeepseekForCausalLM": ("deepseek", "DeepseekForCausalLM"),
     "DeepseekV2ForCausalLM": ("deepseek_v2", "DeepseekV2ForCausalLM"),
-    "DeepseekV3ForCausalLM": ("deepseek_v3", "DeepseekV3ForCausalLM"),
+    "DeepseekV3ForCausalLM": ("deepseek_v2", "DeepseekV3ForCausalLM"),
     "ExaoneForCausalLM": ("exaone", "ExaoneForCausalLM"),
     "FalconForCausalLM": ("falcon", "FalconForCausalLM"),
     "Fairseq2LlamaForCausalLM": ("fairseq2_llama", "Fairseq2LlamaForCausalLM"),