[Misc] Use config definitions from Transformers library (#21913)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

vllm/model_executor/models/aimv2.py

@@ -8,7 +8,6 @@ from typing import Optional
 
 import torch
 import torch.nn as nn
-from transformers import PretrainedConfig
 
 from vllm.attention.layer import MultiHeadAttention
 from vllm.distributed import get_tensor_model_parallel_world_size
@@ -21,12 +20,13 @@ from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
 from vllm.model_executor.layers.quantization.base_config import (
     QuantizationConfig)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+from vllm.transformers_utils.configs.ovis import AIMv2Config
 
 
 class AIMv2SwiGLUFFN(nn.Module):
 
-    def __init__(self, config: PretrainedConfig,
+    def __init__(self, config: AIMv2Config, quant_config: QuantizationConfig,
-                 quant_config: QuantizationConfig, prefix: str):
+                 prefix: str):
         super().__init__()
         hidden_features = config.intermediate_size
         in_features = config.hidden_size
@@ -57,7 +57,7 @@ class AIMv2SwiGLUFFN(nn.Module):
 
 class AIMv2PatchEmbed(nn.Module):
 
-    def __init__(self, config: PretrainedConfig):
+    def __init__(self, config: AIMv2Config):
         super().__init__()
         self.proj = nn.Conv2d(
             config.num_channels,
@@ -75,7 +75,7 @@ class AIMv2PatchEmbed(nn.Module):
 
 class AIMv2ViTPreprocessor(nn.Module):
 
-    def __init__(self, config: PretrainedConfig):
+    def __init__(self, config: AIMv2Config):
         super().__init__()
         num_patches = (config.image_size // config.patch_size)**2
 
@@ -93,8 +93,8 @@ class AIMv2ViTPreprocessor(nn.Module):
 
 class AIMv2Attention(nn.Module):
 
-    def __init__(self, config: PretrainedConfig,
+    def __init__(self, config: AIMv2Config, quant_config: QuantizationConfig,
-                 quant_config: QuantizationConfig, prefix: str):
+                 prefix: str):
         super().__init__()
         self.config = config
         self.embed_dim = config.hidden_size
@@ -141,8 +141,8 @@ class AIMv2Attention(nn.Module):
 
 class AIMv2Block(nn.Module):
 
-    def __init__(self, config: PretrainedConfig,
+    def __init__(self, config: AIMv2Config, quant_config: QuantizationConfig,
-                 quant_config: QuantizationConfig, prefix: str):
+                 prefix: str):
         super().__init__()
         self.attn = AIMv2Attention(config,
                                    quant_config=quant_config,
@@ -163,7 +163,7 @@ class AIMv2Transformer(nn.Module):
 
     def __init__(
         self,
-        config: PretrainedConfig,
+        config: AIMv2Config,
         quant_config: QuantizationConfig,
         *,
         require_post_norm: Optional[bool] = None,
@@ -193,7 +193,7 @@ class AIMv2Transformer(nn.Module):
 class AIMv2Model(torch.nn.Module):
 
     def __init__(self,
-                 config: PretrainedConfig,
+                 config: AIMv2Config,
                  quant_config: QuantizationConfig,
                  *,
                  require_post_norm: Optional[bool] = None,

vllm/model_executor/models/commandr.py

@@ -27,7 +27,7 @@ from typing import Optional, Union
 
 import torch
 from torch import nn
-from transformers import CohereConfig
+from transformers import Cohere2Config, CohereConfig
 
 from vllm.attention import Attention
 from vllm.compilation.decorators import support_torch_compile
@@ -89,7 +89,7 @@ class CohereMLP(nn.Module):
 
     def __init__(
         self,
-        config: CohereConfig,
+        config: Union[CohereConfig, Cohere2Config],
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
     ):
@@ -124,7 +124,7 @@ class CohereAttention(nn.Module):
 
     def __init__(
         self,
-        config: CohereConfig,
+        config: Union[CohereConfig, Cohere2Config],
         cache_config: Optional[CacheConfig] = None,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
@@ -242,7 +242,7 @@ class CohereAttention(nn.Module):
 class CohereDecoderLayer(nn.Module):
 
     def __init__(self,
-                 config: CohereConfig,
+                 config: Union[CohereConfig, Cohere2Config],
                  cache_config: Optional[CacheConfig] = None,
                  quant_config: Optional[QuantizationConfig] = None,
                  prefix: str = ""):

vllm/model_executor/models/dbrx.py

@@ -6,7 +6,7 @@ from typing import Optional, Union
 
 import torch
 import torch.nn as nn
-from transformers import PretrainedConfig
+from transformers import DbrxConfig
 
 from vllm.attention import Attention
 from vllm.config import CacheConfig, VllmConfig
@@ -39,7 +39,7 @@ class DbrxRouter(nn.Module):
 
     def __init__(
         self,
-        config: PretrainedConfig,
+        config: DbrxConfig,
         params_dtype: Optional[torch.dtype] = None,
     ):
         super().__init__()
@@ -63,7 +63,7 @@ class DbrxExperts(FusedMoE):
 
     def __init__(
         self,
-        config: PretrainedConfig,
+        config: DbrxConfig,
         quant_config: Optional[QuantizationConfig] = None,
         params_dtype: Optional[torch.dtype] = None,
         prefix: str = "",
@@ -138,7 +138,7 @@ class DbrxMoE(nn.Module):
 
     def __init__(
         self,
-        config: PretrainedConfig,
+        config: DbrxConfig,
         quant_config: Optional[QuantizationConfig] = None,
         params_dtype: Optional[torch.dtype] = None,
         prefix: str = "",
@@ -169,7 +169,7 @@ class DbrxAttention(nn.Module):
 
     def __init__(
         self,
-        config: PretrainedConfig,
+        config: DbrxConfig,
         cache_config: Optional[CacheConfig] = None,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
@@ -249,7 +249,7 @@ class DbrxFusedNormAttention(nn.Module):
 
     def __init__(
         self,
-        config: PretrainedConfig,
+        config: DbrxConfig,
         cache_config: Optional[CacheConfig] = None,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
@@ -284,7 +284,7 @@ class DbrxBlock(nn.Module):
 
     def __init__(
         self,
-        config: PretrainedConfig,
+        config: DbrxConfig,
         cache_config: Optional[CacheConfig] = None,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",

vllm/model_executor/models/deepseek_v2.py

@@ -29,7 +29,7 @@ from typing import Any, Optional, Union
 
 import torch
 from torch import nn
-from transformers import PretrainedConfig
+from transformers import DeepseekV2Config, DeepseekV3Config
 
 from vllm.attention import Attention
 from vllm.compilation.decorators import support_torch_compile
@@ -100,7 +100,7 @@ class DeepseekV2MoE(nn.Module):
 
     def __init__(
         self,
-        config: PretrainedConfig,
+        config: Union[DeepseekV2Config, DeepseekV3Config],
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
         enable_eplb: bool = False,
@@ -221,7 +221,7 @@ class DeepseekV2Attention(nn.Module):
 
     def __init__(
         self,
-        config: PretrainedConfig,
+        config: Union[DeepseekV2Config, DeepseekV3Config],
         hidden_size: int,
         num_heads: int,
         qk_nope_head_dim: int,
@@ -373,7 +373,7 @@ class DeepseekV2MLAAttention(nn.Module):
 
     def __init__(
         self,
-        config: PretrainedConfig,
+        config: Union[DeepseekV2Config, DeepseekV3Config],
         hidden_size: int,
         num_heads: int,
         qk_nope_head_dim: int,
@@ -538,7 +538,7 @@ class DeepseekV2DecoderLayer(nn.Module):
 
     def __init__(
         self,
-        config: PretrainedConfig,
+        config: Union[DeepseekV2Config, DeepseekV3Config],
         prefix: str,
         model_config: ModelConfig,
         cache_config: Optional[CacheConfig] = None,
@@ -973,7 +973,10 @@ class DeepseekV3ForCausalLM(DeepseekV2ForCausalLM):
     pass
 
 
-def get_spec_layer_idx_from_weight_name(config: PretrainedConfig,
+# Compatibility with
+# https://huggingface.co/deepseek-ai/DeepSeek-V3-Base/blob/main/configuration_deepseek.py
+def get_spec_layer_idx_from_weight_name(config: Union[DeepseekV2Config,
+                                                      DeepseekV3Config],
                                         weight_name: str) -> Optional[int]:
     if (hasattr(config, "num_nextn_predict_layers")
             and config.num_nextn_predict_layers > 0):

vllm/model_executor/models/dots1.py

@@ -29,7 +29,7 @@ from typing import Any, Optional, Union
 
 import torch
 from torch import nn
-from transformers import PretrainedConfig
+from transformers import Dots1Config
 
 from vllm.attention import Attention
 from vllm.compilation.decorators import support_torch_compile
@@ -99,7 +99,7 @@ class Dots1MoE(nn.Module):
 
     def __init__(
         self,
-        config: PretrainedConfig,
+        config: Dots1Config,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
     ):
@@ -174,7 +174,7 @@ class Dots1Attention(nn.Module):
         hidden_size: int,
         num_heads: int,
         num_kv_heads: int,
-        config: PretrainedConfig,
+        config: Dots1Config,
         rope_theta: float = 10000,
         rope_scaling: Optional[dict[str, Any]] = None,
         max_position_embeddings: int = 8192,
@@ -260,7 +260,7 @@ class Dots1DecoderLayer(nn.Module):
 
     def __init__(
         self,
-        config: PretrainedConfig,
+        config: Dots1Config,
         prefix: str,
         model_config: ModelConfig,
         cache_config: Optional[CacheConfig] = None,

vllm/model_executor/models/exaone4.py

@@ -26,7 +26,7 @@ from typing import Any, Optional, Union
 
 import torch
 from torch import nn
-from transformers import PretrainedConfig
+from transformers import Exaone4Config
 
 from vllm.attention import Attention
 from vllm.compilation.decorators import support_torch_compile
@@ -96,7 +96,7 @@ class Exaone4Attention(nn.Module):
 
     def __init__(
         self,
-        config: PretrainedConfig,
+        config: Exaone4Config,
         hidden_size: int,
         num_heads: int,
         num_kv_heads: int,
@@ -224,7 +224,7 @@ class Exaone4DecoderLayer(nn.Module):
 
     def __init__(
         self,
-        config: PretrainedConfig,
+        config: Exaone4Config,
         cache_config: Optional[CacheConfig] = None,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",

vllm/model_executor/models/glm4_moe.py

@@ -28,7 +28,7 @@ from typing import Any, Optional, Union
 
 import torch
 from torch import nn
-from transformers import PretrainedConfig
+from transformers.models.glm4_moe import Glm4MoeConfig
 
 from vllm.attention import Attention
 from vllm.compilation.decorators import support_torch_compile
@@ -100,7 +100,7 @@ class Glm4MoE(nn.Module):
 
     def __init__(
         self,
-        config: PretrainedConfig,
+        config: Glm4MoeConfig,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
         enable_eplb: bool = False,
@@ -198,7 +198,7 @@ class Glm4MoeAttention(nn.Module):
 
     def __init__(
         self,
-        config: PretrainedConfig,
+        config: Glm4MoeConfig,
         hidden_size: int,
         num_heads: int,
         num_kv_heads: int,
@@ -297,7 +297,7 @@ class Glm4MoeDecoderLayer(nn.Module):
 
     def __init__(
         self,
-        config: PretrainedConfig,
+        config: Glm4MoeConfig,
         cache_config: Optional[CacheConfig] = None,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
@@ -681,7 +681,7 @@ class Glm4MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
         return self.model.get_expert_mapping()
 
 
-def get_spec_layer_idx_from_weight_name(config: PretrainedConfig,
+def get_spec_layer_idx_from_weight_name(config: Glm4MoeConfig,
                                         weight_name: str) -> Optional[int]:
     if hasattr(config,
                "num_nextn_predict_layers") and (config.num_nextn_predict_layers

vllm/model_executor/models/minimax_text_01.py

@@ -12,7 +12,7 @@ import torch.distributed
 import torch.nn.functional as F
 from einops import rearrange
 from torch import nn
-from transformers.configuration_utils import PretrainedConfig
+from transformers import MiniMaxConfig
 
 from vllm import envs
 from vllm.attention import Attention, AttentionMetadata
@@ -656,7 +656,7 @@ class MiniMaxText01DecoderLayer(nn.Module):
 
     def __init__(
         self,
-        config: PretrainedConfig,
+        config: MiniMaxConfig,
         cache_config: Optional[CacheConfig] = None,
         quant_config: Optional[QuantizationConfig] = None,
         expert_num: int = 1,
@@ -860,7 +860,7 @@ class MiniMaxText01Model(nn.Module):
 
     def __init__(
         self,
-        config: PretrainedConfig,
+        config: MiniMaxConfig,
         quant_config: Optional[QuantizationConfig] = None,
         cache_config: Optional[CacheConfig] = None,
         scheduler_config=None,

vllm/model_executor/models/olmoe.py

@@ -19,7 +19,7 @@ from typing import Any, Optional, Union
 
 import torch
 from torch import nn
-from transformers import PretrainedConfig
+from transformers import OlmoeConfig
 
 from vllm.attention import Attention
 from vllm.compilation.decorators import support_torch_compile
@@ -205,7 +205,7 @@ class OlmoeDecoderLayer(nn.Module):
 
     def __init__(
         self,
-        config: PretrainedConfig,
+        config: OlmoeConfig,
         cache_config: Optional[CacheConfig] = None,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",

vllm/model_executor/models/qwen2_moe.py

@@ -30,7 +30,7 @@ from typing import Any, Optional, Union
 import torch
 import torch.nn.functional as F
 from torch import nn
-from transformers import PretrainedConfig
+from transformers import Qwen2MoeConfig
 
 from vllm.attention import Attention
 from vllm.compilation.decorators import support_torch_compile
@@ -98,7 +98,7 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
 
     def __init__(
         self,
-        config: PretrainedConfig,
+        config: Qwen2MoeConfig,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
     ):
@@ -256,7 +256,7 @@ class Qwen2MoeDecoderLayer(nn.Module):
 
     def __init__(
         self,
-        config: PretrainedConfig,
+        config: Qwen2MoeConfig,
         cache_config: Optional[CacheConfig] = None,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",

vllm/model_executor/models/qwen3_moe.py

@@ -28,7 +28,7 @@ from typing import Any, Optional, Union
 
 import torch
 from torch import nn
-from transformers import PretrainedConfig
+from transformers import Qwen3MoeConfig
 
 from vllm.attention import Attention
 from vllm.compilation.decorators import support_torch_compile
@@ -101,7 +101,7 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
 
     def __init__(
         self,
-        config: PretrainedConfig,
+        config: Qwen3MoeConfig,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
         enable_eplb: bool = False,
@@ -278,7 +278,7 @@ class Qwen3MoeDecoderLayer(nn.Module):
 
     def __init__(
         self,
-        config: PretrainedConfig,
+        config: Qwen3MoeConfig,
         cache_config: Optional[CacheConfig] = None,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",