[Chore] Clean up deepseek v2/v3 config copy (#28055)

Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>

vllm/model_executor/models/deepseek.py

@@ -292,6 +292,7 @@ class DeepseekDecoderLayer(nn.Module):
         rope_theta = getattr(config, "rope_theta", 10000)
         rope_scaling = getattr(config, "rope_scaling", None)
         max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
+        moe_layer_freq = getattr(config, "moe_layer_freq", 1)
         self.self_attn = DeepseekAttention(
             hidden_size=self.hidden_size,
             num_heads=config.num_attention_heads,
@@ -306,7 +307,7 @@ class DeepseekDecoderLayer(nn.Module):
         if (
             config.n_routed_experts is not None
             and layer_idx >= config.first_k_dense_replace
-            and layer_idx % config.moe_layer_freq == 0
+            and layer_idx % moe_layer_freq == 0
         ):
             self.mlp = DeepseekMoE(
                 config=config, quant_config=quant_config, prefix=f"{prefix}.mlp"

vllm/model_executor/models/deepseek_v2.py

@@ -994,6 +994,7 @@ class DeepseekV2DecoderLayer(nn.Module):
         rope_theta = getattr(config, "rope_theta", 10000)
         rope_scaling = getattr(config, "rope_scaling", None)
         max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
+        moe_layer_freq = getattr(config, "moe_layer_freq", 1)
         # DecoderLayers are created with `make_layers` which passes the prefix
         # with the layer's index.
         layer_idx = int(prefix.split(sep=".")[-1])
@@ -1024,7 +1025,7 @@ class DeepseekV2DecoderLayer(nn.Module):
         if (
             config.n_routed_experts is not None
             and layer_idx >= config.first_k_dense_replace
-            and layer_idx % config.moe_layer_freq == 0
+            and layer_idx % moe_layer_freq == 0
         ):
             self.mlp = DeepseekV2MoE(
                 config=config,

vllm/model_executor/models/kimi_vl.py

@@ -50,7 +50,7 @@ from typing import Annotated, Any, Literal
 
 import torch
 from torch import nn
-from transformers import BatchFeature
+from transformers import BatchFeature, DeepseekV2Config
 from transformers.activations import GELUActivation
 
 from vllm.config import VllmConfig
@@ -91,7 +91,6 @@ from vllm.multimodal.processing import (
 from vllm.multimodal.profiling import BaseDummyInputsBuilder
 from vllm.sequence import IntermediateTensors
 from vllm.transformers_utils.configs import KimiVLConfig, MoonViTConfig
-from vllm.transformers_utils.configs.deepseek_vl2 import DeepseekV2Config
 from vllm.utils.tensor_schema import TensorSchema, TensorShape
 
 from .utils import PPMissingLayer, is_pp_missing_parameter, maybe_prefix

vllm/transformers_utils/config.py

@@ -24,7 +24,7 @@ from huggingface_hub.utils import (
     RepositoryNotFoundError,
     RevisionNotFoundError,
 )
-from transformers import GenerationConfig, PretrainedConfig
+from transformers import DeepseekV3Config, GenerationConfig, PretrainedConfig
 from transformers.models.auto.image_processing_auto import get_image_processor_config
 from transformers.models.auto.modeling_auto import (
     MODEL_FOR_CAUSAL_LM_MAPPING_NAMES,
@@ -68,16 +68,18 @@ def _get_hf_token() -> str | None:
 
 class LazyConfigDict(dict):
     def __getitem__(self, key):
+        if isinstance(value := super().__getitem__(key), type):
+            return value
+
         import vllm.transformers_utils.configs as configs
 
-        return getattr(configs, super().__getitem__(key))
+        return getattr(configs, value)
 
 
 _CONFIG_REGISTRY: dict[str, type[PretrainedConfig]] = LazyConfigDict(
     chatglm="ChatGLMConfig",
     deepseek_vl_v2="DeepseekVLV2Config",
-    deepseek_v3="DeepseekV3Config",
+    deepseek_v32=DeepseekV3Config,
-    deepseek_v32="DeepseekV3Config",
     flex_olmo="FlexOlmoConfig",
     kimi_linear="KimiLinearConfig",
     kimi_vl="KimiVLConfig",

vllm/transformers_utils/configs/__init__.py

@@ -8,7 +8,6 @@ Model configs may be defined in this directory for the following reasons:
 """
 
 from vllm.transformers_utils.configs.chatglm import ChatGLMConfig
-from vllm.transformers_utils.configs.deepseek_v3 import DeepseekV3Config
 from vllm.transformers_utils.configs.deepseek_vl2 import DeepseekVLV2Config
 from vllm.transformers_utils.configs.dotsocr import DotsOCRConfig
 from vllm.transformers_utils.configs.eagle import EAGLEConfig
@@ -43,7 +42,6 @@ from vllm.transformers_utils.configs.ultravox import UltravoxConfig
 __all__ = [
     "ChatGLMConfig",
     "DeepseekVLV2Config",
-    "DeepseekV3Config",
     "DotsOCRConfig",
     "EAGLEConfig",
     "FlexOlmoConfig",

vllm/transformers_utils/configs/deepseek_v3.py

@@ -1,100 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-from transformers.configuration_utils import PretrainedConfig
-from transformers.utils import logging
-
-logger = logging.get_logger(__name__)
-
-
-class DeepseekV3Config(PretrainedConfig):
-    model_type = "deepseek_v3"
-    keys_to_ignore_at_inference = ["past_key_values"]
-
-    def __init__(
-        self,
-        vocab_size=129280,
-        hidden_size=7168,
-        intermediate_size=18432,
-        moe_intermediate_size=2048,
-        num_hidden_layers=61,
-        num_nextn_predict_layers=1,
-        num_attention_heads=128,
-        num_key_value_heads=128,
-        n_shared_experts=1,
-        n_routed_experts=256,
-        ep_size=1,
-        routed_scaling_factor=2.5,
-        kv_lora_rank=512,
-        q_lora_rank=1536,
-        qk_rope_head_dim=64,
-        v_head_dim=128,
-        qk_nope_head_dim=128,
-        topk_method="noaux_tc",
-        n_group=8,
-        topk_group=4,
-        num_experts_per_tok=8,
-        moe_layer_freq=1,
-        first_k_dense_replace=3,
-        norm_topk_prob=True,
-        scoring_func="sigmoid",
-        hidden_act="silu",
-        max_position_embeddings=4096,
-        initializer_range=0.02,
-        rms_norm_eps=1e-6,
-        use_cache=True,
-        pad_token_id=None,
-        bos_token_id=0,
-        eos_token_id=1,
-        tie_word_embeddings=False,
-        rope_theta=10000.0,
-        rope_scaling=None,
-        attention_bias=False,
-        attention_dropout=0.0,
-        **kwargs,
-    ):
-        self.vocab_size = vocab_size
-        self.max_position_embeddings = max_position_embeddings
-        self.hidden_size = hidden_size
-        self.intermediate_size = intermediate_size
-        self.moe_intermediate_size = moe_intermediate_size
-        self.num_hidden_layers = num_hidden_layers
-        self.num_nextn_predict_layers = num_nextn_predict_layers
-        self.num_attention_heads = num_attention_heads
-        self.n_shared_experts = n_shared_experts
-        self.n_routed_experts = n_routed_experts
-        self.ep_size = ep_size
-        self.routed_scaling_factor = routed_scaling_factor
-        self.kv_lora_rank = kv_lora_rank
-        self.q_lora_rank = q_lora_rank
-        self.qk_rope_head_dim = qk_rope_head_dim
-        self.v_head_dim = v_head_dim
-        self.qk_nope_head_dim = qk_nope_head_dim
-        self.topk_method = topk_method
-        self.n_group = n_group
-        self.topk_group = topk_group
-        self.num_experts_per_tok = num_experts_per_tok
-        self.moe_layer_freq = moe_layer_freq
-        self.first_k_dense_replace = first_k_dense_replace
-        self.norm_topk_prob = norm_topk_prob
-        self.scoring_func = scoring_func
-        # for backward compatibility
-        if num_key_value_heads is None:
-            num_key_value_heads = num_attention_heads
-
-        self.num_key_value_heads = num_key_value_heads
-        self.hidden_act = hidden_act
-        self.initializer_range = initializer_range
-        self.rms_norm_eps = rms_norm_eps
-        self.use_cache = use_cache
-        self.rope_theta = rope_theta
-        self.rope_scaling = rope_scaling
-        self.attention_bias = attention_bias
-        self.attention_dropout = attention_dropout
-
-        super().__init__(
-            pad_token_id=pad_token_id,
-            bos_token_id=bos_token_id,
-            eos_token_id=eos_token_id,
-            tie_word_embeddings=tie_word_embeddings,
-            **kwargs,
-        )

vllm/transformers_utils/configs/deepseek_vl2.py

@@ -3,7 +3,7 @@
 
 # adapted from https://github.com/deepseek-ai/DeepSeek-VL2/blob/faf18023f24b962b32d9f0a2d89e402a8d383a78/deepseek_vl2/models/modeling_deepseek_vl_v2.py#L115-L268
 
-from transformers.configuration_utils import PretrainedConfig
+from transformers import DeepseekV2Config, PretrainedConfig
 
 
 class VisionEncoderConfig(PretrainedConfig):
@@ -87,106 +87,6 @@ class MlpProjectorConfig(PretrainedConfig):
         super().__init__(**kwargs)
 
 
-class DeepseekV2Config(PretrainedConfig):
-    model_type = "deepseek_v2"
-    keys_to_ignore_at_inference = ["past_key_values"]
-
-    def __init__(
-        self,
-        vocab_size=102400,
-        hidden_size=4096,
-        intermediate_size=11008,
-        moe_intermediate_size=1407,
-        num_hidden_layers=30,
-        num_attention_heads=32,
-        num_key_value_heads=32,
-        n_shared_experts=None,
-        n_routed_experts=None,
-        ep_size=1,
-        routed_scaling_factor=1.0,
-        kv_lora_rank=512,
-        q_lora_rank=1536,
-        qk_rope_head_dim=64,
-        v_head_dim=128,
-        qk_nope_head_dim=128,
-        topk_method="gready",
-        n_group=None,
-        topk_group=None,
-        num_experts_per_tok=None,
-        moe_layer_freq=1,
-        first_k_dense_replace=0,
-        norm_topk_prob=False,
-        scoring_func="softmax",
-        aux_loss_alpha=0.001,
-        seq_aux=True,
-        hidden_act="silu",
-        max_position_embeddings=2048,
-        initializer_range=0.02,
-        rms_norm_eps=1e-6,
-        use_cache=True,
-        pad_token_id=None,
-        bos_token_id=100000,
-        eos_token_id=100001,
-        pretraining_tp=1,
-        tie_word_embeddings=False,
-        rope_theta=10000.0,
-        rope_scaling=None,
-        attention_bias=False,
-        attention_dropout=0.0,
-        use_mla=True,
-        **kwargs,
-    ):
-        self.vocab_size = vocab_size
-        self.max_position_embeddings = max_position_embeddings
-        self.hidden_size = hidden_size
-        self.intermediate_size = intermediate_size
-        self.moe_intermediate_size = moe_intermediate_size
-        self.num_hidden_layers = num_hidden_layers
-        self.num_attention_heads = num_attention_heads
-        self.n_shared_experts = n_shared_experts
-        self.n_routed_experts = n_routed_experts
-        self.ep_size = ep_size
-        self.routed_scaling_factor = routed_scaling_factor
-        self.kv_lora_rank = kv_lora_rank
-        self.q_lora_rank = q_lora_rank
-        self.qk_rope_head_dim = qk_rope_head_dim
-        self.v_head_dim = v_head_dim
-        self.qk_nope_head_dim = qk_nope_head_dim
-        self.topk_method = topk_method
-        self.n_group = n_group
-        self.topk_group = topk_group
-        self.num_experts_per_tok = num_experts_per_tok
-        self.moe_layer_freq = moe_layer_freq
-        self.first_k_dense_replace = first_k_dense_replace
-        self.norm_topk_prob = norm_topk_prob
-        self.scoring_func = scoring_func
-        self.aux_loss_alpha = aux_loss_alpha
-        self.seq_aux = seq_aux
-        # for backward compatibility
-        if num_key_value_heads is None:
-            num_key_value_heads = num_attention_heads
-
-        self.num_key_value_heads = num_key_value_heads
-        self.hidden_act = hidden_act
-        self.initializer_range = initializer_range
-        self.rms_norm_eps = float(rms_norm_eps)
-        self.pretraining_tp = pretraining_tp
-        self.use_cache = use_cache
-        self.rope_theta = rope_theta
-        self.rope_scaling = rope_scaling
-        self.attention_bias = attention_bias
-        self.attention_dropout = attention_dropout
-        self.use_mla = use_mla
-
-        super().__init__(
-            pad_token_id=pad_token_id,
-            bos_token_id=bos_token_id,
-            eos_token_id=eos_token_id,
-            tie_word_embeddings=tie_word_embeddings,
-            **kwargs,
-        )
-
-
 class DeepseekVLV2Config(PretrainedConfig):
     model_type = "deepseek_vl_v2"
     vision_config: VisionEncoderConfig

vllm/transformers_utils/configs/eagle.py

@@ -3,9 +3,7 @@
 
 import os
 
-from transformers import AutoConfig, PretrainedConfig
+from transformers import AutoConfig, DeepseekV2Config, PretrainedConfig
-
-from vllm.transformers_utils.configs.deepseek_vl2 import DeepseekV2Config
 
 
 class EAGLEConfig(PretrainedConfig):
@@ -20,13 +18,7 @@ class EAGLEConfig(PretrainedConfig):
     ):
         model_config: PretrainedConfig | DeepseekV2Config | None
         if isinstance(model, dict):
-            archs = model.get("architectures", [])
+            model_config = AutoConfig.for_model(**model)
-            target_archs = ["DeepseekV2ForCausalLM", "DeepseekV3ForCausalLM"]
-            if any(target_arch in archs for target_arch in target_archs):
-                # AutoConfig does not support DeepSeek MoE models yet
-                model_config = DeepseekV2Config(**model)
-            else:
-                model_config = AutoConfig.for_model(**model)
         else:
             model_config = model
 

vllm/transformers_utils/configs/kimi_vl.py

@@ -2,9 +2,9 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 # Adapted from https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct/blob/main/configuration_kimi_vl.py
 
+from transformers import DeepseekV2Config
 from transformers.configuration_utils import PretrainedConfig
 
-from vllm.transformers_utils.configs.deepseek_vl2 import DeepseekV2Config
 from vllm.transformers_utils.configs.moonvit import MoonViTConfig