Skip to content

vllm.model_executor.models.registry

Whenever you add an architecture to this page, please also update tests/models/registry.py with example HuggingFace models for it.

ModelRegistry module-attribute 

ModelRegistry = _ModelRegistry(
    {
        model_arch: (
            _LazyRegisteredModel(
                module_name=f"vllm.model_executor.models.{mod_relname}",
                class_name=cls_name,
            )
        )
        for (model_arch, (mod_relname, cls_name)) in (
            items()
        )
    }
)

_CROSS_ENCODER_MODELS module-attribute 

_CROSS_ENCODER_MODELS = {
    "BertForSequenceClassification": (
        "bert",
        "BertForSequenceClassification",
    ),
    "RobertaForSequenceClassification": (
        "roberta",
        "RobertaForSequenceClassification",
    ),
    "XLMRobertaForSequenceClassification": (
        "roberta",
        "RobertaForSequenceClassification",
    ),
    "ModernBertForSequenceClassification": (
        "modernbert",
        "ModernBertForSequenceClassification",
    ),
    "JinaVLForRanking": (
        "jina_vl",
        "JinaVLForSequenceClassification",
    ),
}

_EMBEDDING_MODELS module-attribute 

_EMBEDDING_MODELS = {
    "BertModel": ("bert", "BertEmbeddingModel"),
    "DeciLMForCausalLM": (
        "nemotron_nas",
        "DeciLMForCausalLM",
    ),
    "Gemma2Model": ("gemma2", "Gemma2ForCausalLM"),
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
    "GPT2ForSequenceClassification": (
        "gpt2",
        "GPT2ForSequenceClassification",
    ),
    "GritLM": ("gritlm", "GritLM"),
    "GteModel": ("bert_with_rope", "SnowflakeGteNewModel"),
    "GteNewModel": ("bert_with_rope", "GteNewModel"),
    "InternLM2ForRewardModel": (
        "internlm2",
        "InternLM2ForRewardModel",
    ),
    "JambaForSequenceClassification": (
        "jamba",
        "JambaForSequenceClassification",
    ),
    "LlamaModel": ("llama", "LlamaForCausalLM"),
    None: {
        k: (mod, arch)
        for (k, (mod, arch)) in (items())
        if arch == "LlamaForCausalLM"
    },
    "MistralModel": ("llama", "LlamaForCausalLM"),
    "ModernBertModel": ("modernbert", "ModernBertModel"),
    "NomicBertModel": ("bert_with_rope", "NomicBertModel"),
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
    "Qwen2Model": ("qwen2", "Qwen2ForCausalLM"),
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
    "Qwen2ForRewardModel": (
        "qwen2_rm",
        "Qwen2ForRewardModel",
    ),
    "Qwen2ForProcessRewardModel": (
        "qwen2_rm",
        "Qwen2ForProcessRewardModel",
    ),
    "RobertaForMaskedLM": (
        "roberta",
        "RobertaEmbeddingModel",
    ),
    "RobertaModel": ("roberta", "RobertaEmbeddingModel"),
    "TeleChat2ForCausalLM": (
        "telechat2",
        "TeleChat2ForCausalLM",
    ),
    "XLMRobertaModel": ("roberta", "RobertaEmbeddingModel"),
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
    ),
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
    "Qwen2VLForConditionalGeneration": (
        "qwen2_vl",
        "Qwen2VLForConditionalGeneration",
    ),
    "PrithviGeoSpatialMAE": (
        "prithvi_geospatial_mae",
        "PrithviGeoSpatialMAE",
    ),
}

_MULTIMODAL_MODELS module-attribute 

_MULTIMODAL_MODELS = {
    "AriaForConditionalGeneration": (
        "aria",
        "AriaForConditionalGeneration",
    ),
    "AyaVisionForConditionalGeneration": (
        "aya_vision",
        "AyaVisionForConditionalGeneration",
    ),
    "Blip2ForConditionalGeneration": (
        "blip2",
        "Blip2ForConditionalGeneration",
    ),
    "ChameleonForConditionalGeneration": (
        "chameleon",
        "ChameleonForConditionalGeneration",
    ),
    "Cohere2VisionForConditionalGeneration": (
        "cohere2_vision",
        "Cohere2VisionForConditionalGeneration",
    ),
    "DeepseekVLV2ForCausalLM": (
        "deepseek_vl2",
        "DeepseekVLV2ForCausalLM",
    ),
    "FuyuForCausalLM": ("fuyu", "FuyuForCausalLM"),
    "Gemma3ForConditionalGeneration": (
        "gemma3_mm",
        "Gemma3ForConditionalGeneration",
    ),
    "Gemma3nForConditionalGeneration": (
        "gemma3n_mm",
        "Gemma3nForConditionalGeneration",
    ),
    "GLM4VForCausalLM": ("glm4v", "GLM4VForCausalLM"),
    "Glm4vForConditionalGeneration": (
        "glm4_1v",
        "Glm4vForConditionalGeneration",
    ),
    "Glm4vMoeForConditionalGeneration": (
        "glm4_1v",
        "Glm4vMoeForConditionalGeneration",
    ),
    "GraniteSpeechForConditionalGeneration": (
        "granite_speech",
        "GraniteSpeechForConditionalGeneration",
    ),
    "H2OVLChatModel": ("h2ovl", "H2OVLChatModel"),
    "InternVLChatModel": ("internvl", "InternVLChatModel"),
    "InternS1ForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
    ),
    "Idefics3ForConditionalGeneration": (
        "idefics3",
        "Idefics3ForConditionalGeneration",
    ),
    "SmolVLMForConditionalGeneration": (
        "smolvlm",
        "SmolVLMForConditionalGeneration",
    ),
    "KeyeForConditionalGeneration": (
        "keye",
        "KeyeForConditionalGeneration",
    ),
    "RForConditionalGeneration": (
        "rvl",
        "RForConditionalGeneration",
    ),
    "KimiVLForConditionalGeneration": (
        "kimi_vl",
        "KimiVLForConditionalGeneration",
    ),
    "Llama_Nemotron_Nano_VL": (
        "nemotron_vl",
        "LlamaNemotronVLChatModel",
    ),
    "LlavaForConditionalGeneration": (
        "llava",
        "LlavaForConditionalGeneration",
    ),
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
    ),
    "LlavaNextVideoForConditionalGeneration": (
        "llava_next_video",
        "LlavaNextVideoForConditionalGeneration",
    ),
    "LlavaOnevisionForConditionalGeneration": (
        "llava_onevision",
        "LlavaOnevisionForConditionalGeneration",
    ),
    "MantisForConditionalGeneration": (
        "llava",
        "MantisForConditionalGeneration",
    ),
    "MiniMaxVL01ForConditionalGeneration": (
        "minimax_vl_01",
        "MiniMaxVL01ForConditionalGeneration",
    ),
    "MiniCPMO": ("minicpmo", "MiniCPMO"),
    "MiniCPMV": ("minicpmv", "MiniCPMV"),
    "Mistral3ForConditionalGeneration": (
        "mistral3",
        "Mistral3ForConditionalGeneration",
    ),
    "MolmoForCausalLM": ("molmo", "MolmoForCausalLM"),
    "NVLM_D": ("nvlm_d", "NVLM_D_Model"),
    "Ovis": ("ovis", "Ovis"),
    "Ovis2_5": ("ovis2_5", "Ovis2_5"),
    "PaliGemmaForConditionalGeneration": (
        "paligemma",
        "PaliGemmaForConditionalGeneration",
    ),
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
    "Phi4MMForCausalLM": ("phi4mm", "Phi4MMForCausalLM"),
    "Phi4MultimodalForCausalLM": (
        "phi4_multimodal",
        "Phi4MultimodalForCausalLM",
    ),
    "PixtralForConditionalGeneration": (
        "pixtral",
        "PixtralForConditionalGeneration",
    ),
    "QwenVLForConditionalGeneration": (
        "qwen_vl",
        "QwenVLForConditionalGeneration",
    ),
    "Qwen2VLForConditionalGeneration": (
        "qwen2_vl",
        "Qwen2VLForConditionalGeneration",
    ),
    "Qwen2_5_VLForConditionalGeneration": (
        "qwen2_5_vl",
        "Qwen2_5_VLForConditionalGeneration",
    ),
    "Qwen2AudioForConditionalGeneration": (
        "qwen2_audio",
        "Qwen2AudioForConditionalGeneration",
    ),
    "Qwen2_5OmniModel": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
    ),
    "Qwen2_5OmniForConditionalGeneration": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
    ),
    "UltravoxModel": ("ultravox", "UltravoxModel"),
    "Step3VLForConditionalGeneration": (
        "step3_vl",
        "Step3VLForConditionalGeneration",
    ),
    "TarsierForConditionalGeneration": (
        "tarsier",
        "TarsierForConditionalGeneration",
    ),
    "Tarsier2ForConditionalGeneration": (
        "qwen2_vl",
        "Tarsier2ForConditionalGeneration",
    ),
    "VoxtralForConditionalGeneration": (
        "voxtral",
        "VoxtralForConditionalGeneration",
    ),
    "DonutForConditionalGeneration": (
        "donut",
        "DonutForConditionalGeneration",
    ),
    "Florence2ForConditionalGeneration": (
        "florence2",
        "Florence2ForConditionalGeneration",
    ),
    "MllamaForConditionalGeneration": (
        "mllama",
        "MllamaForConditionalGeneration",
    ),
    "Llama4ForConditionalGeneration": (
        "mllama4",
        "Llama4ForConditionalGeneration",
    ),
    "SkyworkR1VChatModel": (
        "skyworkr1v",
        "SkyworkR1VChatModel",
    ),
    "WhisperForConditionalGeneration": (
        "whisper",
        "WhisperForConditionalGeneration",
    ),
}

_PREVIOUSLY_SUPPORTED_MODELS module-attribute 

_PREVIOUSLY_SUPPORTED_MODELS = {
    "Phi3SmallForCausalLM": "0.9.2"
}

_SPECULATIVE_DECODING_MODELS module-attribute 

_SPECULATIVE_DECODING_MODELS = {
    "MiMoMTPModel": ("mimo_mtp", "MiMoMTP"),
    "EagleLlamaForCausalLM": (
        "llama_eagle",
        "EagleLlamaForCausalLM",
    ),
    "EagleLlama4ForCausalLM": (
        "llama4_eagle",
        "EagleLlama4ForCausalLM",
    ),
    "EagleMiniCPMForCausalLM": (
        "minicpm_eagle",
        "EagleMiniCPMForCausalLM",
    ),
    "Eagle3LlamaForCausalLM": (
        "llama_eagle3",
        "Eagle3LlamaForCausalLM",
    ),
    "EagleDeepSeekMTPModel": (
        "deepseek_eagle",
        "EagleDeepseekV3ForCausalLM",
    ),
    "DeepSeekMTPModel": ("deepseek_mtp", "DeepSeekMTP"),
    "ErnieMTPModel": ("ernie_mtp", "ErnieMTP"),
    "Glm4MoeMTPModel": ("glm4_moe_mtp", "Glm4MoeMTP"),
    "MedusaModel": ("medusa", "Medusa"),
}

_SUBPROCESS_COMMAND module-attribute 

_SUBPROCESS_COMMAND = [
    executable,
    "-m",
    "vllm.model_executor.models.registry",
]

_T module-attribute 

_T = TypeVar('_T')

_TEXT_GENERATION_MODELS module-attribute 

_TEXT_GENERATION_MODELS = {
    "AquilaModel": ("llama", "LlamaForCausalLM"),
    "AquilaForCausalLM": ("llama", "LlamaForCausalLM"),
    "ArceeForCausalLM": ("arcee", "ArceeForCausalLM"),
    "ArcticForCausalLM": ("arctic", "ArcticForCausalLM"),
    "MiniMaxForCausalLM": (
        "minimax_text_01",
        "MiniMaxText01ForCausalLM",
    ),
    "MiniMaxText01ForCausalLM": (
        "minimax_text_01",
        "MiniMaxText01ForCausalLM",
    ),
    "MiniMaxM1ForCausalLM": (
        "minimax_text_01",
        "MiniMaxText01ForCausalLM",
    ),
    "BaiChuanForCausalLM": (
        "baichuan",
        "BaiChuanForCausalLM",
    ),
    "BaichuanForCausalLM": (
        "baichuan",
        "BaichuanForCausalLM",
    ),
    "BailingMoeForCausalLM": (
        "bailing_moe",
        "BailingMoeForCausalLM",
    ),
    "BambaForCausalLM": ("bamba", "BambaForCausalLM"),
    "BloomForCausalLM": ("bloom", "BloomForCausalLM"),
    "ChatGLMModel": ("chatglm", "ChatGLMForCausalLM"),
    "ChatGLMForConditionalGeneration": (
        "chatglm",
        "ChatGLMForCausalLM",
    ),
    "CohereForCausalLM": ("commandr", "CohereForCausalLM"),
    "Cohere2ForCausalLM": ("commandr", "CohereForCausalLM"),
    "DbrxForCausalLM": ("dbrx", "DbrxForCausalLM"),
    "DeciLMForCausalLM": (
        "nemotron_nas",
        "DeciLMForCausalLM",
    ),
    "DeepseekForCausalLM": (
        "deepseek",
        "DeepseekForCausalLM",
    ),
    "DeepseekV2ForCausalLM": (
        "deepseek_v2",
        "DeepseekV2ForCausalLM",
    ),
    "DeepseekV3ForCausalLM": (
        "deepseek_v2",
        "DeepseekV3ForCausalLM",
    ),
    "Dots1ForCausalLM": ("dots1", "Dots1ForCausalLM"),
    "Ernie4_5ForCausalLM": (
        "ernie45",
        "Ernie4_5ForCausalLM",
    ),
    "Ernie4_5_MoeForCausalLM": (
        "ernie45_moe",
        "Ernie4_5_MoeForCausalLM",
    ),
    "ExaoneForCausalLM": ("exaone", "ExaoneForCausalLM"),
    "Exaone4ForCausalLM": ("exaone4", "Exaone4ForCausalLM"),
    "FalconForCausalLM": ("falcon", "FalconForCausalLM"),
    "Fairseq2LlamaForCausalLM": (
        "fairseq2_llama",
        "Fairseq2LlamaForCausalLM",
    ),
    "GemmaForCausalLM": ("gemma", "GemmaForCausalLM"),
    "Gemma2ForCausalLM": ("gemma2", "Gemma2ForCausalLM"),
    "Gemma3ForCausalLM": ("gemma3", "Gemma3ForCausalLM"),
    "Gemma3nForCausalLM": ("gemma3n", "Gemma3nForCausalLM"),
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
    "Glm4ForCausalLM": ("glm4", "Glm4ForCausalLM"),
    "Glm4MoeForCausalLM": (
        "glm4_moe",
        "Glm4MoeForCausalLM",
    ),
    "GptOssForCausalLM": ("gpt_oss", "GptOssForCausalLM"),
    "GPT2LMHeadModel": ("gpt2", "GPT2LMHeadModel"),
    "GPTBigCodeForCausalLM": (
        "gpt_bigcode",
        "GPTBigCodeForCausalLM",
    ),
    "GPTJForCausalLM": ("gpt_j", "GPTJForCausalLM"),
    "GPTNeoXForCausalLM": (
        "gpt_neox",
        "GPTNeoXForCausalLM",
    ),
    "GraniteForCausalLM": ("granite", "GraniteForCausalLM"),
    "GraniteMoeForCausalLM": (
        "granitemoe",
        "GraniteMoeForCausalLM",
    ),
    "GraniteMoeHybridForCausalLM": (
        "granitemoehybrid",
        "GraniteMoeHybridForCausalLM",
    ),
    "GraniteMoeSharedForCausalLM": (
        "granitemoeshared",
        "GraniteMoeSharedForCausalLM",
    ),
    "GritLM": ("gritlm", "GritLM"),
    "Grok1ModelForCausalLM": ("grok1", "Grok1ForCausalLM"),
    "HunYuanMoEV1ForCausalLM": (
        "hunyuan_v1",
        "HunYuanMoEV1ForCausalLM",
    ),
    "HunYuanDenseV1ForCausalLM": (
        "hunyuan_v1",
        "HunYuanDenseV1ForCausalLM",
    ),
    "HCXVisionForCausalLM": (
        "hyperclovax_vision",
        "HCXVisionForCausalLM",
    ),
    "InternLMForCausalLM": ("llama", "LlamaForCausalLM"),
    "InternLM2ForCausalLM": (
        "internlm2",
        "InternLM2ForCausalLM",
    ),
    "InternLM2VEForCausalLM": (
        "internlm2_ve",
        "InternLM2VEForCausalLM",
    ),
    "InternLM3ForCausalLM": ("llama", "LlamaForCausalLM"),
    "JAISLMHeadModel": ("jais", "JAISLMHeadModel"),
    "JambaForCausalLM": ("jamba", "JambaForCausalLM"),
    "Lfm2ForCausalLM": ("lfm2", "Lfm2ForCausalLM"),
    "LlamaForCausalLM": ("llama", "LlamaForCausalLM"),
    "Llama4ForCausalLM": ("llama4", "Llama4ForCausalLM"),
    "LLaMAForCausalLM": ("llama", "LlamaForCausalLM"),
    "MambaForCausalLM": ("mamba", "MambaForCausalLM"),
    "FalconMambaForCausalLM": ("mamba", "MambaForCausalLM"),
    "FalconH1ForCausalLM": (
        "falcon_h1",
        "FalconH1ForCausalLM",
    ),
    "Mamba2ForCausalLM": ("mamba2", "Mamba2ForCausalLM"),
    "MiniCPMForCausalLM": ("minicpm", "MiniCPMForCausalLM"),
    "MiniCPM3ForCausalLM": (
        "minicpm3",
        "MiniCPM3ForCausalLM",
    ),
    "MistralForCausalLM": ("llama", "LlamaForCausalLM"),
    "MixtralForCausalLM": ("mixtral", "MixtralForCausalLM"),
    "QuantMixtralForCausalLM": (
        "mixtral_quant",
        "MixtralForCausalLM",
    ),
    "MptForCausalLM": ("mpt", "MPTForCausalLM"),
    "MPTForCausalLM": ("mpt", "MPTForCausalLM"),
    "MiMoForCausalLM": ("mimo", "MiMoForCausalLM"),
    "NemotronForCausalLM": (
        "nemotron",
        "NemotronForCausalLM",
    ),
    "NemotronHForCausalLM": (
        "nemotron_h",
        "NemotronHForCausalLM",
    ),
    "OlmoForCausalLM": ("olmo", "OlmoForCausalLM"),
    "Olmo2ForCausalLM": ("olmo2", "Olmo2ForCausalLM"),
    "OlmoeForCausalLM": ("olmoe", "OlmoeForCausalLM"),
    "OPTForCausalLM": ("opt", "OPTForCausalLM"),
    "OrionForCausalLM": ("orion", "OrionForCausalLM"),
    "PersimmonForCausalLM": (
        "persimmon",
        "PersimmonForCausalLM",
    ),
    "PhiForCausalLM": ("phi", "PhiForCausalLM"),
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
    "PhiMoEForCausalLM": ("phimoe", "PhiMoEForCausalLM"),
    "Phi4FlashForCausalLM": (
        "phi4flash",
        "Phi4FlashForCausalLM",
    ),
    "Plamo2ForCausalLM": ("plamo2", "Plamo2ForCausalLM"),
    "QWenLMHeadModel": ("qwen", "QWenLMHeadModel"),
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
    "Qwen2MoeForCausalLM": (
        "qwen2_moe",
        "Qwen2MoeForCausalLM",
    ),
    "Qwen3ForCausalLM": ("qwen3", "Qwen3ForCausalLM"),
    "Qwen3MoeForCausalLM": (
        "qwen3_moe",
        "Qwen3MoeForCausalLM",
    ),
    "RWForCausalLM": ("falcon", "FalconForCausalLM"),
    "SeedOssForCausalLM": (
        "seed_oss",
        "SeedOssForCausalLM",
    ),
    "Step3TextForCausalLM": (
        "step3_text",
        "Step3TextForCausalLM",
    ),
    "StableLMEpochForCausalLM": (
        "stablelm",
        "StablelmForCausalLM",
    ),
    "StableLmForCausalLM": (
        "stablelm",
        "StablelmForCausalLM",
    ),
    "Starcoder2ForCausalLM": (
        "starcoder2",
        "Starcoder2ForCausalLM",
    ),
    "SolarForCausalLM": ("solar", "SolarForCausalLM"),
    "TeleChat2ForCausalLM": (
        "telechat2",
        "TeleChat2ForCausalLM",
    ),
    "TeleFLMForCausalLM": ("teleflm", "TeleFLMForCausalLM"),
    "XverseForCausalLM": ("llama", "LlamaForCausalLM"),
    "Zamba2ForCausalLM": ("zamba2", "Zamba2ForCausalLM"),
    "BartModel": ("bart", "BartForConditionalGeneration"),
    "BartForConditionalGeneration": (
        "bart",
        "BartForConditionalGeneration",
    ),
    "MBartForConditionalGeneration": (
        "bart",
        "MBartForConditionalGeneration",
    ),
}

_TRANSFORMERS_BACKEND_MODELS module-attribute 

_TRANSFORMERS_BACKEND_MODELS = {
    "TransformersModel": (
        "transformers",
        "TransformersModel",
    ),
    "TransformersForCausalLM": (
        "transformers",
        "TransformersForCausalLM",
    ),
    "TransformersForMultimodalLM": (
        "transformers",
        "TransformersForMultimodalLM",
    ),
}

_TRANSFORMERS_SUPPORTED_MODELS module-attribute 

_TRANSFORMERS_SUPPORTED_MODELS = {
    "SmolLM3ForCausalLM": (
        "transformers",
        "TransformersForCausalLM",
    ),
    "Emu3ForConditionalGeneration": (
        "transformers",
        "TransformersForMultimodalLM",
    ),
}

_VLLM_MODELS module-attribute 

_VLLM_MODELS = {
    None: _TEXT_GENERATION_MODELS,
    None: _EMBEDDING_MODELS,
    None: _CROSS_ENCODER_MODELS,
    None: _MULTIMODAL_MODELS,
    None: _SPECULATIVE_DECODING_MODELS,
    None: _TRANSFORMERS_SUPPORTED_MODELS,
    None: _TRANSFORMERS_BACKEND_MODELS,
}

logger module-attribute 

logger = init_logger(__name__)

_BaseRegisteredModel

Bases: ABC

Source code in vllm/model_executor/models/registry.py 
356
357
358
359
360
361
362
363
364
class _BaseRegisteredModel(ABC):

    @abstractmethod
    def inspect_model_cls(self) -> _ModelInfo:
        raise NotImplementedError

    @abstractmethod
    def load_model_cls(self) -> type[nn.Module]:
        raise NotImplementedError

inspect_model_cls abstractmethod 

inspect_model_cls() -> _ModelInfo
Source code in vllm/model_executor/models/registry.py 
358
359
360
@abstractmethod
def inspect_model_cls(self) -> _ModelInfo:
    raise NotImplementedError

load_model_cls abstractmethod 

load_model_cls() -> type[Module]
Source code in vllm/model_executor/models/registry.py 
362
363
364
@abstractmethod
def load_model_cls(self) -> type[nn.Module]:
    raise NotImplementedError

_LazyRegisteredModel dataclass

Bases: _BaseRegisteredModel

Represents a model that has not been imported in the main process.

Source code in vllm/model_executor/models/registry.py 
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
@dataclass(frozen=True)
class _LazyRegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has not been imported in the main process.
    """
    module_name: str
    class_name: str

    # Performed in another process to avoid initializing CUDA
    def inspect_model_cls(self) -> _ModelInfo:
        return _run_in_subprocess(
            lambda: _ModelInfo.from_model_cls(self.load_model_cls()))

    def load_model_cls(self) -> type[nn.Module]:
        mod = importlib.import_module(self.module_name)
        return getattr(mod, self.class_name)

class_name instance-attribute 

class_name: str

module_name instance-attribute 

module_name: str

__init__ 

__init__(module_name: str, class_name: str) -> None

inspect_model_cls 

inspect_model_cls() -> _ModelInfo
Source code in vllm/model_executor/models/registry.py 
399
400
401
def inspect_model_cls(self) -> _ModelInfo:
    return _run_in_subprocess(
        lambda: _ModelInfo.from_model_cls(self.load_model_cls()))

load_model_cls 

load_model_cls() -> type[Module]
Source code in vllm/model_executor/models/registry.py 
403
404
405
def load_model_cls(self) -> type[nn.Module]:
    mod = importlib.import_module(self.module_name)
    return getattr(mod, self.class_name)

_ModelInfo dataclass

Source code in vllm/model_executor/models/registry.py 
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
@dataclass(frozen=True)
class _ModelInfo:
    architecture: str
    is_text_generation_model: bool
    is_pooling_model: bool
    default_pooling_type: str
    supports_cross_encoding: bool
    supports_multimodal: bool
    supports_multimodal_raw_input: bool
    supports_pp: bool
    has_inner_state: bool
    is_attention_free: bool
    is_hybrid: bool
    has_noops: bool
    supports_transcription: bool
    supports_transcription_only: bool
    supports_v0_only: bool

    @staticmethod
    def from_model_cls(model: type[nn.Module]) -> "_ModelInfo":
        return _ModelInfo(
            architecture=model.__name__,
            is_text_generation_model=is_text_generation_model(model),
            is_pooling_model=is_pooling_model(model),
            default_pooling_type=get_default_pooling_type(model),
            supports_cross_encoding=supports_cross_encoding(model),
            supports_multimodal=supports_multimodal(model),
            supports_multimodal_raw_input=supports_multimodal_raw_input(model),
            supports_pp=supports_pp(model),
            has_inner_state=has_inner_state(model),
            is_attention_free=is_attention_free(model),
            is_hybrid=is_hybrid(model),
            supports_transcription=supports_transcription(model),
            supports_transcription_only=(supports_transcription(model) and
                                         model.supports_transcription_only),
            supports_v0_only=supports_v0_only(model),
            has_noops=has_noops(model),
        )

architecture instance-attribute 

architecture: str

default_pooling_type instance-attribute 

default_pooling_type: str

has_inner_state instance-attribute 

has_inner_state: bool

has_noops instance-attribute 

has_noops: bool

is_attention_free instance-attribute 

is_attention_free: bool

is_hybrid instance-attribute 

is_hybrid: bool

is_pooling_model instance-attribute 

is_pooling_model: bool

is_text_generation_model instance-attribute 

is_text_generation_model: bool

supports_cross_encoding instance-attribute 

supports_cross_encoding: bool

supports_multimodal instance-attribute 

supports_multimodal: bool

supports_multimodal_raw_input instance-attribute 

supports_multimodal_raw_input: bool

supports_pp instance-attribute 

supports_pp: bool

supports_transcription instance-attribute 

supports_transcription: bool

supports_transcription_only instance-attribute 

supports_transcription_only: bool

supports_v0_only instance-attribute 

supports_v0_only: bool

__init__ 

__init__(
    architecture: str,
    is_text_generation_model: bool,
    is_pooling_model: bool,
    default_pooling_type: str,
    supports_cross_encoding: bool,
    supports_multimodal: bool,
    supports_multimodal_raw_input: bool,
    supports_pp: bool,
    has_inner_state: bool,
    is_attention_free: bool,
    is_hybrid: bool,
    has_noops: bool,
    supports_transcription: bool,
    supports_transcription_only: bool,
    supports_v0_only: bool,
) -> None

from_model_cls staticmethod 

from_model_cls(model: type[Module]) -> _ModelInfo
Source code in vllm/model_executor/models/registry.py 
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
@staticmethod
def from_model_cls(model: type[nn.Module]) -> "_ModelInfo":
    return _ModelInfo(
        architecture=model.__name__,
        is_text_generation_model=is_text_generation_model(model),
        is_pooling_model=is_pooling_model(model),
        default_pooling_type=get_default_pooling_type(model),
        supports_cross_encoding=supports_cross_encoding(model),
        supports_multimodal=supports_multimodal(model),
        supports_multimodal_raw_input=supports_multimodal_raw_input(model),
        supports_pp=supports_pp(model),
        has_inner_state=has_inner_state(model),
        is_attention_free=is_attention_free(model),
        is_hybrid=is_hybrid(model),
        supports_transcription=supports_transcription(model),
        supports_transcription_only=(supports_transcription(model) and
                                     model.supports_transcription_only),
        supports_v0_only=supports_v0_only(model),
        has_noops=has_noops(model),
    )

_ModelRegistry dataclass

Source code in vllm/model_executor/models/registry.py 
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
@dataclass
class _ModelRegistry:
    # Keyed by model_arch
    models: dict[str, _BaseRegisteredModel] = field(default_factory=dict)

    def get_supported_archs(self) -> Set[str]:
        return self.models.keys()

    def register_model(
        self,
        model_arch: str,
        model_cls: Union[type[nn.Module], str],
    ) -> None:
        """
        Register an external model to be used in vLLM.

        `model_cls` can be either:

        - A [`torch.nn.Module`][] class directly referencing the model.
        - A string in the format `<module>:<class>` which can be used to
          lazily import the model. This is useful to avoid initializing CUDA
          when importing the model and thus the related error
          `RuntimeError: Cannot re-initialize CUDA in forked subprocess`.
        """
        if not isinstance(model_arch, str):
            msg = f"`model_arch` should be a string, not a {type(model_arch)}"
            raise TypeError(msg)

        if model_arch in self.models:
            logger.warning(
                "Model architecture %s is already registered, and will be "
                "overwritten by the new model class %s.", model_arch,
                model_cls)

        if isinstance(model_cls, str):
            split_str = model_cls.split(":")
            if len(split_str) != 2:
                msg = "Expected a string in the format `<module>:<class>`"
                raise ValueError(msg)

            model = _LazyRegisteredModel(*split_str)
        elif isinstance(model_cls, type) and issubclass(model_cls, nn.Module):
            model = _RegisteredModel.from_model_cls(model_cls)
        else:
            msg = ("`model_cls` should be a string or PyTorch model class, "
                   f"not a {type(model_arch)}")
            raise TypeError(msg)

        self.models[model_arch] = model

    def _raise_for_unsupported(self, architectures: list[str]):
        all_supported_archs = self.get_supported_archs()

        if any(arch in all_supported_archs for arch in architectures):
            raise ValueError(
                f"Model architectures {architectures} failed "
                "to be inspected. Please check the logs for more details.")

        for arch in architectures:
            if arch in _PREVIOUSLY_SUPPORTED_MODELS:
                previous_version = _PREVIOUSLY_SUPPORTED_MODELS[arch]

                raise ValueError(
                    f"Model architecture {arch} was supported in vLLM until "
                    f"v{previous_version}, and is not supported anymore. "
                    "Please use an older version of vLLM if you want to "
                    "use this model architecture.")

        raise ValueError(
            f"Model architectures {architectures} are not supported for now. "
            f"Supported architectures: {all_supported_archs}")

    def _try_load_model_cls(self,
                            model_arch: str) -> Optional[type[nn.Module]]:
        if model_arch not in self.models:
            return None

        return _try_load_model_cls(model_arch, self.models[model_arch])

    def _try_inspect_model_cls(self, model_arch: str) -> Optional[_ModelInfo]:
        if model_arch not in self.models:
            return None

        return _try_inspect_model_cls(model_arch, self.models[model_arch])

    def _try_resolve_transformers(
        self,
        architecture: str,
        model_config: ModelConfig,
    ) -> Optional[str]:
        if architecture in _TRANSFORMERS_BACKEND_MODELS:
            return architecture

        auto_map: dict[str, str] = getattr(model_config.hf_config, "auto_map",
                                           None) or dict()

        # Make sure that config class is always initialized before model class,
        # otherwise the model class won't be able to access the config class,
        # the expected auto_map should have correct order like:
        # "auto_map": {
        #     "AutoConfig": "<your-repo-name>--<config-name>",
        #     "AutoModel": "<your-repo-name>--<config-name>",
        #     "AutoModelFor<Task>": "<your-repo-name>--<config-name>",
        # },
        for prefix in ("AutoConfig", "AutoModel"):
            for name, module in auto_map.items():
                if name.startswith(prefix):
                    try_get_class_from_dynamic_module(
                        module,
                        model_config.model,
                        revision=model_config.revision,
                        warn_on_fail=False,
                    )

        model_module = getattr(transformers, architecture, None)

        if model_module is None:
            for name, module in auto_map.items():
                if name.startswith("AutoModel"):
                    model_module = try_get_class_from_dynamic_module(
                        module,
                        model_config.model,
                        revision=model_config.revision,
                        warn_on_fail=True,
                    )
                    if model_module is not None:
                        break
            else:
                if model_config.model_impl != ModelImpl.TRANSFORMERS:
                    return None

                raise ValueError(
                    f"Cannot find model module. {architecture!r} is not a "
                    "registered model in the Transformers library (only "
                    "relevant if the model is meant to be in Transformers) "
                    "and 'AutoModel' is not present in the model config's "
                    "'auto_map' (relevant if the model is custom).")

        if not model_module.is_backend_compatible():
            if model_config.model_impl != ModelImpl.TRANSFORMERS:
                return None

            raise ValueError(
                f"The Transformers implementation of {architecture!r} "
                "is not compatible with vLLM.")

        return model_config._get_transformers_backend_cls()

    def _normalize_arch(
        self,
        architecture: str,
        model_config: ModelConfig,
    ) -> str:
        if architecture in self.models:
            return architecture

        # This may be called in order to resolve runner_type and convert_type
        # in the first place, in which case we consider the default match
        match = try_match_architecture_defaults(
            architecture,
            runner_type=getattr(model_config, "runner_type", None),
            convert_type=getattr(model_config, "convert_type", None),
        )
        if match:
            suffix, _ = match

            # Get the name of the base model to convert
            for repl_suffix, _ in iter_architecture_defaults():
                base_arch = architecture.replace(suffix, repl_suffix)
                if base_arch in self.models:
                    return base_arch

        return architecture

    def inspect_model_cls(
        self,
        architectures: Union[str, list[str]],
        model_config: ModelConfig,
    ) -> tuple[_ModelInfo, str]:
        if isinstance(architectures, str):
            architectures = [architectures]
        if not architectures:
            raise ValueError("No model architectures are specified")

        # Require transformers impl
        if model_config.model_impl == ModelImpl.TRANSFORMERS:
            arch = self._try_resolve_transformers(architectures[0],
                                                  model_config)
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)

        # Fallback to transformers impl (after resolving convert_type)
        if (all(arch not in self.models for arch in architectures)
                and model_config.model_impl == ModelImpl.AUTO
                and getattr(model_config, "convert_type", "none") == "none"):
            arch = self._try_resolve_transformers(architectures[0],
                                                  model_config)
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)

        for arch in architectures:
            normalized_arch = self._normalize_arch(arch, model_config)
            model_info = self._try_inspect_model_cls(normalized_arch)
            if model_info is not None:
                return (model_info, arch)

        # Fallback to transformers impl (before resolving runner_type)
        if (all(arch not in self.models for arch in architectures)
                and model_config.model_impl == ModelImpl.AUTO):
            arch = self._try_resolve_transformers(architectures[0],
                                                  model_config)
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)

        return self._raise_for_unsupported(architectures)

    def resolve_model_cls(
        self,
        architectures: Union[str, list[str]],
        model_config: ModelConfig,
    ) -> tuple[type[nn.Module], str]:
        if isinstance(architectures, str):
            architectures = [architectures]
        if not architectures:
            raise ValueError("No model architectures are specified")

        # Require transformers impl
        if model_config.model_impl == ModelImpl.TRANSFORMERS:
            arch = self._try_resolve_transformers(architectures[0],
                                                  model_config)
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)

        # Fallback to transformers impl (after resolving convert_type)
        if (all(arch not in self.models for arch in architectures)
                and model_config.model_impl == ModelImpl.AUTO
                and getattr(model_config, "convert_type", "none") == "none"):
            arch = self._try_resolve_transformers(architectures[0],
                                                  model_config)
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)

        for arch in architectures:
            normalized_arch = self._normalize_arch(arch, model_config)
            model_cls = self._try_load_model_cls(normalized_arch)
            if model_cls is not None:
                return (model_cls, arch)

        # Fallback to transformers impl (before resolving runner_type)
        if (all(arch not in self.models for arch in architectures)
                and model_config.model_impl == ModelImpl.AUTO):
            arch = self._try_resolve_transformers(architectures[0],
                                                  model_config)
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)

        return self._raise_for_unsupported(architectures)

    def is_text_generation_model(
        self,
        architectures: Union[str, list[str]],
        model_config: ModelConfig,
    ) -> bool:
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
        return model_cls.is_text_generation_model

    def is_pooling_model(
        self,
        architectures: Union[str, list[str]],
        model_config: ModelConfig,
    ) -> bool:
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
        return model_cls.is_pooling_model

    def is_cross_encoder_model(
        self,
        architectures: Union[str, list[str]],
        model_config: ModelConfig,
    ) -> bool:
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
        return model_cls.supports_cross_encoding

    def is_multimodal_model(
        self,
        architectures: Union[str, list[str]],
        model_config: ModelConfig,
    ) -> bool:
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
        return model_cls.supports_multimodal

    def supports_multimodal_raw_input(
        self,
        architectures: Union[str, list[str]],
        model_config: ModelConfig,
    ) -> bool:
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
        return model_cls.supports_multimodal_raw_input

    def is_pp_supported_model(
        self,
        architectures: Union[str, list[str]],
        model_config: ModelConfig,
    ) -> bool:
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
        return model_cls.supports_pp

    def model_has_inner_state(
        self,
        architectures: Union[str, list[str]],
        model_config: ModelConfig,
    ) -> bool:
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
        return model_cls.has_inner_state

    def is_attention_free_model(
        self,
        architectures: Union[str, list[str]],
        model_config: ModelConfig,
    ) -> bool:
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
        return model_cls.is_attention_free

    def is_hybrid_model(
        self,
        architectures: Union[str, list[str]],
        model_config: ModelConfig,
    ) -> bool:
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
        return model_cls.is_hybrid

    def is_noops_model(
        self,
        architectures: Union[str, list[str]],
        model_config: ModelConfig,
    ) -> bool:
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
        return model_cls.has_noops

    def is_transcription_model(
        self,
        architectures: Union[str, list[str]],
        model_config: ModelConfig,
    ) -> bool:
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
        return model_cls.supports_transcription

    def is_transcription_only_model(
        self,
        architectures: Union[str, list[str]],
        model_config: ModelConfig,
    ) -> bool:
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
        return model_cls.supports_transcription_only

    def is_v1_compatible(
        self,
        architectures: Union[str, list[str]],
        model_config: ModelConfig,
    ) -> bool:
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
        return not model_cls.supports_v0_only

models class-attribute instance-attribute 

models: dict[str, _BaseRegisteredModel] = field(
    default_factory=dict
)

__init__ 

__init__(
    models: dict[str, _BaseRegisteredModel] = dict(),
) -> None

_normalize_arch 

_normalize_arch(
    architecture: str, model_config: ModelConfig
) -> str
Source code in vllm/model_executor/models/registry.py 
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
def _normalize_arch(
    self,
    architecture: str,
    model_config: ModelConfig,
) -> str:
    if architecture in self.models:
        return architecture

    # This may be called in order to resolve runner_type and convert_type
    # in the first place, in which case we consider the default match
    match = try_match_architecture_defaults(
        architecture,
        runner_type=getattr(model_config, "runner_type", None),
        convert_type=getattr(model_config, "convert_type", None),
    )
    if match:
        suffix, _ = match

        # Get the name of the base model to convert
        for repl_suffix, _ in iter_architecture_defaults():
            base_arch = architecture.replace(suffix, repl_suffix)
            if base_arch in self.models:
                return base_arch

    return architecture

_raise_for_unsupported 

_raise_for_unsupported(architectures: list[str])
Source code in vllm/model_executor/models/registry.py 
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
def _raise_for_unsupported(self, architectures: list[str]):
    all_supported_archs = self.get_supported_archs()

    if any(arch in all_supported_archs for arch in architectures):
        raise ValueError(
            f"Model architectures {architectures} failed "
            "to be inspected. Please check the logs for more details.")

    for arch in architectures:
        if arch in _PREVIOUSLY_SUPPORTED_MODELS:
            previous_version = _PREVIOUSLY_SUPPORTED_MODELS[arch]

            raise ValueError(
                f"Model architecture {arch} was supported in vLLM until "
                f"v{previous_version}, and is not supported anymore. "
                "Please use an older version of vLLM if you want to "
                "use this model architecture.")

    raise ValueError(
        f"Model architectures {architectures} are not supported for now. "
        f"Supported architectures: {all_supported_archs}")

_try_inspect_model_cls 

_try_inspect_model_cls(
    model_arch: str,
) -> Optional[_ModelInfo]
Source code in vllm/model_executor/models/registry.py 
515
516
517
518
519
def _try_inspect_model_cls(self, model_arch: str) -> Optional[_ModelInfo]:
    if model_arch not in self.models:
        return None

    return _try_inspect_model_cls(model_arch, self.models[model_arch])

_try_load_model_cls 

_try_load_model_cls(
    model_arch: str,
) -> Optional[type[Module]]
Source code in vllm/model_executor/models/registry.py 
508
509
510
511
512
513
def _try_load_model_cls(self,
                        model_arch: str) -> Optional[type[nn.Module]]:
    if model_arch not in self.models:
        return None

    return _try_load_model_cls(model_arch, self.models[model_arch])

_try_resolve_transformers 

_try_resolve_transformers(
    architecture: str, model_config: ModelConfig
) -> Optional[str]
Source code in vllm/model_executor/models/registry.py 
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
def _try_resolve_transformers(
    self,
    architecture: str,
    model_config: ModelConfig,
) -> Optional[str]:
    if architecture in _TRANSFORMERS_BACKEND_MODELS:
        return architecture

    auto_map: dict[str, str] = getattr(model_config.hf_config, "auto_map",
                                       None) or dict()

    # Make sure that config class is always initialized before model class,
    # otherwise the model class won't be able to access the config class,
    # the expected auto_map should have correct order like:
    # "auto_map": {
    #     "AutoConfig": "<your-repo-name>--<config-name>",
    #     "AutoModel": "<your-repo-name>--<config-name>",
    #     "AutoModelFor<Task>": "<your-repo-name>--<config-name>",
    # },
    for prefix in ("AutoConfig", "AutoModel"):
        for name, module in auto_map.items():
            if name.startswith(prefix):
                try_get_class_from_dynamic_module(
                    module,
                    model_config.model,
                    revision=model_config.revision,
                    warn_on_fail=False,
                )

    model_module = getattr(transformers, architecture, None)

    if model_module is None:
        for name, module in auto_map.items():
            if name.startswith("AutoModel"):
                model_module = try_get_class_from_dynamic_module(
                    module,
                    model_config.model,
                    revision=model_config.revision,
                    warn_on_fail=True,
                )
                if model_module is not None:
                    break
        else:
            if model_config.model_impl != ModelImpl.TRANSFORMERS:
                return None

            raise ValueError(
                f"Cannot find model module. {architecture!r} is not a "
                "registered model in the Transformers library (only "
                "relevant if the model is meant to be in Transformers) "
                "and 'AutoModel' is not present in the model config's "
                "'auto_map' (relevant if the model is custom).")

    if not model_module.is_backend_compatible():
        if model_config.model_impl != ModelImpl.TRANSFORMERS:
            return None

        raise ValueError(
            f"The Transformers implementation of {architecture!r} "
            "is not compatible with vLLM.")

    return model_config._get_transformers_backend_cls()

get_supported_archs 

get_supported_archs() -> Set[str]
Source code in vllm/model_executor/models/registry.py 
441
442
def get_supported_archs(self) -> Set[str]:
    return self.models.keys()

inspect_model_cls 

inspect_model_cls(
    architectures: Union[str, list[str]],
    model_config: ModelConfig,
) -> tuple[_ModelInfo, str]
Source code in vllm/model_executor/models/registry.py 
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
def inspect_model_cls(
    self,
    architectures: Union[str, list[str]],
    model_config: ModelConfig,
) -> tuple[_ModelInfo, str]:
    if isinstance(architectures, str):
        architectures = [architectures]
    if not architectures:
        raise ValueError("No model architectures are specified")

    # Require transformers impl
    if model_config.model_impl == ModelImpl.TRANSFORMERS:
        arch = self._try_resolve_transformers(architectures[0],
                                              model_config)
        if arch is not None:
            model_info = self._try_inspect_model_cls(arch)
            if model_info is not None:
                return (model_info, arch)

    # Fallback to transformers impl (after resolving convert_type)
    if (all(arch not in self.models for arch in architectures)
            and model_config.model_impl == ModelImpl.AUTO
            and getattr(model_config, "convert_type", "none") == "none"):
        arch = self._try_resolve_transformers(architectures[0],
                                              model_config)
        if arch is not None:
            model_info = self._try_inspect_model_cls(arch)
            if model_info is not None:
                return (model_info, arch)

    for arch in architectures:
        normalized_arch = self._normalize_arch(arch, model_config)
        model_info = self._try_inspect_model_cls(normalized_arch)
        if model_info is not None:
            return (model_info, arch)

    # Fallback to transformers impl (before resolving runner_type)
    if (all(arch not in self.models for arch in architectures)
            and model_config.model_impl == ModelImpl.AUTO):
        arch = self._try_resolve_transformers(architectures[0],
                                              model_config)
        if arch is not None:
            model_info = self._try_inspect_model_cls(arch)
            if model_info is not None:
                return (model_info, arch)

    return self._raise_for_unsupported(architectures)

is_attention_free_model 

is_attention_free_model(
    architectures: Union[str, list[str]],
    model_config: ModelConfig,
) -> bool
Source code in vllm/model_executor/models/registry.py 
762
763
764
765
766
767
768
def is_attention_free_model(
    self,
    architectures: Union[str, list[str]],
    model_config: ModelConfig,
) -> bool:
    model_cls, _ = self.inspect_model_cls(architectures, model_config)
    return model_cls.is_attention_free

is_cross_encoder_model 

is_cross_encoder_model(
    architectures: Union[str, list[str]],
    model_config: ModelConfig,
) -> bool
Source code in vllm/model_executor/models/registry.py 
722
723
724
725
726
727
728
def is_cross_encoder_model(
    self,
    architectures: Union[str, list[str]],
    model_config: ModelConfig,
) -> bool:
    model_cls, _ = self.inspect_model_cls(architectures, model_config)
    return model_cls.supports_cross_encoding

is_hybrid_model 

is_hybrid_model(
    architectures: Union[str, list[str]],
    model_config: ModelConfig,
) -> bool
Source code in vllm/model_executor/models/registry.py 
770
771
772
773
774
775
776
def is_hybrid_model(
    self,
    architectures: Union[str, list[str]],
    model_config: ModelConfig,
) -> bool:
    model_cls, _ = self.inspect_model_cls(architectures, model_config)
    return model_cls.is_hybrid

is_multimodal_model 

is_multimodal_model(
    architectures: Union[str, list[str]],
    model_config: ModelConfig,
) -> bool
Source code in vllm/model_executor/models/registry.py 
730
731
732
733
734
735
736
def is_multimodal_model(
    self,
    architectures: Union[str, list[str]],
    model_config: ModelConfig,
) -> bool:
    model_cls, _ = self.inspect_model_cls(architectures, model_config)
    return model_cls.supports_multimodal

is_noops_model 

is_noops_model(
    architectures: Union[str, list[str]],
    model_config: ModelConfig,
) -> bool
Source code in vllm/model_executor/models/registry.py 
778
779
780
781
782
783
784
def is_noops_model(
    self,
    architectures: Union[str, list[str]],
    model_config: ModelConfig,
) -> bool:
    model_cls, _ = self.inspect_model_cls(architectures, model_config)
    return model_cls.has_noops

is_pooling_model 

is_pooling_model(
    architectures: Union[str, list[str]],
    model_config: ModelConfig,
) -> bool
Source code in vllm/model_executor/models/registry.py 
714
715
716
717
718
719
720
def is_pooling_model(
    self,
    architectures: Union[str, list[str]],
    model_config: ModelConfig,
) -> bool:
    model_cls, _ = self.inspect_model_cls(architectures, model_config)
    return model_cls.is_pooling_model

is_pp_supported_model 

is_pp_supported_model(
    architectures: Union[str, list[str]],
    model_config: ModelConfig,
) -> bool
Source code in vllm/model_executor/models/registry.py 
746
747
748
749
750
751
752
def is_pp_supported_model(
    self,
    architectures: Union[str, list[str]],
    model_config: ModelConfig,
) -> bool:
    model_cls, _ = self.inspect_model_cls(architectures, model_config)
    return model_cls.supports_pp

is_text_generation_model 

is_text_generation_model(
    architectures: Union[str, list[str]],
    model_config: ModelConfig,
) -> bool
Source code in vllm/model_executor/models/registry.py 
706
707
708
709
710
711
712
def is_text_generation_model(
    self,
    architectures: Union[str, list[str]],
    model_config: ModelConfig,
) -> bool:
    model_cls, _ = self.inspect_model_cls(architectures, model_config)
    return model_cls.is_text_generation_model

is_transcription_model 

is_transcription_model(
    architectures: Union[str, list[str]],
    model_config: ModelConfig,
) -> bool
Source code in vllm/model_executor/models/registry.py 
786
787
788
789
790
791
792
def is_transcription_model(
    self,
    architectures: Union[str, list[str]],
    model_config: ModelConfig,
) -> bool:
    model_cls, _ = self.inspect_model_cls(architectures, model_config)
    return model_cls.supports_transcription

is_transcription_only_model 

is_transcription_only_model(
    architectures: Union[str, list[str]],
    model_config: ModelConfig,
) -> bool
Source code in vllm/model_executor/models/registry.py 
794
795
796
797
798
799
800
def is_transcription_only_model(
    self,
    architectures: Union[str, list[str]],
    model_config: ModelConfig,
) -> bool:
    model_cls, _ = self.inspect_model_cls(architectures, model_config)
    return model_cls.supports_transcription_only

is_v1_compatible 

is_v1_compatible(
    architectures: Union[str, list[str]],
    model_config: ModelConfig,
) -> bool
Source code in vllm/model_executor/models/registry.py 
802
803
804
805
806
807
808
def is_v1_compatible(
    self,
    architectures: Union[str, list[str]],
    model_config: ModelConfig,
) -> bool:
    model_cls, _ = self.inspect_model_cls(architectures, model_config)
    return not model_cls.supports_v0_only

model_has_inner_state 

model_has_inner_state(
    architectures: Union[str, list[str]],
    model_config: ModelConfig,
) -> bool
Source code in vllm/model_executor/models/registry.py 
754
755
756
757
758
759
760
def model_has_inner_state(
    self,
    architectures: Union[str, list[str]],
    model_config: ModelConfig,
) -> bool:
    model_cls, _ = self.inspect_model_cls(architectures, model_config)
    return model_cls.has_inner_state

register_model 

register_model(
    model_arch: str, model_cls: Union[type[Module], str]
) -> None

Register an external model to be used in vLLM.

model_cls can be either:

  • A torch.nn.Module class directly referencing the model.
  • A string in the format <module>:<class> which can be used to lazily import the model. This is useful to avoid initializing CUDA when importing the model and thus the related error RuntimeError: Cannot re-initialize CUDA in forked subprocess.
Source code in vllm/model_executor/models/registry.py 
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
def register_model(
    self,
    model_arch: str,
    model_cls: Union[type[nn.Module], str],
) -> None:
    """
    Register an external model to be used in vLLM.

    `model_cls` can be either:

    - A [`torch.nn.Module`][] class directly referencing the model.
    - A string in the format `<module>:<class>` which can be used to
      lazily import the model. This is useful to avoid initializing CUDA
      when importing the model and thus the related error
      `RuntimeError: Cannot re-initialize CUDA in forked subprocess`.
    """
    if not isinstance(model_arch, str):
        msg = f"`model_arch` should be a string, not a {type(model_arch)}"
        raise TypeError(msg)

    if model_arch in self.models:
        logger.warning(
            "Model architecture %s is already registered, and will be "
            "overwritten by the new model class %s.", model_arch,
            model_cls)

    if isinstance(model_cls, str):
        split_str = model_cls.split(":")
        if len(split_str) != 2:
            msg = "Expected a string in the format `<module>:<class>`"
            raise ValueError(msg)

        model = _LazyRegisteredModel(*split_str)
    elif isinstance(model_cls, type) and issubclass(model_cls, nn.Module):
        model = _RegisteredModel.from_model_cls(model_cls)
    else:
        msg = ("`model_cls` should be a string or PyTorch model class, "
               f"not a {type(model_arch)}")
        raise TypeError(msg)

    self.models[model_arch] = model

resolve_model_cls 

resolve_model_cls(
    architectures: Union[str, list[str]],
    model_config: ModelConfig,
) -> tuple[type[Module], str]
Source code in vllm/model_executor/models/registry.py 
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
def resolve_model_cls(
    self,
    architectures: Union[str, list[str]],
    model_config: ModelConfig,
) -> tuple[type[nn.Module], str]:
    if isinstance(architectures, str):
        architectures = [architectures]
    if not architectures:
        raise ValueError("No model architectures are specified")

    # Require transformers impl
    if model_config.model_impl == ModelImpl.TRANSFORMERS:
        arch = self._try_resolve_transformers(architectures[0],
                                              model_config)
        if arch is not None:
            model_cls = self._try_load_model_cls(arch)
            if model_cls is not None:
                return (model_cls, arch)

    # Fallback to transformers impl (after resolving convert_type)
    if (all(arch not in self.models for arch in architectures)
            and model_config.model_impl == ModelImpl.AUTO
            and getattr(model_config, "convert_type", "none") == "none"):
        arch = self._try_resolve_transformers(architectures[0],
                                              model_config)
        if arch is not None:
            model_cls = self._try_load_model_cls(arch)
            if model_cls is not None:
                return (model_cls, arch)

    for arch in architectures:
        normalized_arch = self._normalize_arch(arch, model_config)
        model_cls = self._try_load_model_cls(normalized_arch)
        if model_cls is not None:
            return (model_cls, arch)

    # Fallback to transformers impl (before resolving runner_type)
    if (all(arch not in self.models for arch in architectures)
            and model_config.model_impl == ModelImpl.AUTO):
        arch = self._try_resolve_transformers(architectures[0],
                                              model_config)
        if arch is not None:
            model_cls = self._try_load_model_cls(arch)
            if model_cls is not None:
                return (model_cls, arch)

    return self._raise_for_unsupported(architectures)

supports_multimodal_raw_input 

supports_multimodal_raw_input(
    architectures: Union[str, list[str]],
    model_config: ModelConfig,
) -> bool
Source code in vllm/model_executor/models/registry.py 
738
739
740
741
742
743
744
def supports_multimodal_raw_input(
    self,
    architectures: Union[str, list[str]],
    model_config: ModelConfig,
) -> bool:
    model_cls, _ = self.inspect_model_cls(architectures, model_config)
    return model_cls.supports_multimodal_raw_input

_RegisteredModel dataclass

Bases: _BaseRegisteredModel

Represents a model that has already been imported in the main process.

Source code in vllm/model_executor/models/registry.py 
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
@dataclass(frozen=True)
class _RegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has already been imported in the main process.
    """

    interfaces: _ModelInfo
    model_cls: type[nn.Module]

    @staticmethod
    def from_model_cls(model_cls: type[nn.Module]):
        return _RegisteredModel(
            interfaces=_ModelInfo.from_model_cls(model_cls),
            model_cls=model_cls,
        )

    def inspect_model_cls(self) -> _ModelInfo:
        return self.interfaces

    def load_model_cls(self) -> type[nn.Module]:
        return self.model_cls

interfaces instance-attribute 

interfaces: _ModelInfo

model_cls instance-attribute 

model_cls: type[Module]

__init__ 

__init__(
    interfaces: _ModelInfo, model_cls: type[Module]
) -> None

from_model_cls staticmethod 

from_model_cls(model_cls: type[Module])
Source code in vllm/model_executor/models/registry.py 
376
377
378
379
380
381
@staticmethod
def from_model_cls(model_cls: type[nn.Module]):
    return _RegisteredModel(
        interfaces=_ModelInfo.from_model_cls(model_cls),
        model_cls=model_cls,
    )

inspect_model_cls 

inspect_model_cls() -> _ModelInfo
Source code in vllm/model_executor/models/registry.py 
383
384
def inspect_model_cls(self) -> _ModelInfo:
    return self.interfaces

load_model_cls 

load_model_cls() -> type[Module]
Source code in vllm/model_executor/models/registry.py 
386
387
def load_model_cls(self) -> type[nn.Module]:
    return self.model_cls

_run 

_run() -> None
Source code in vllm/model_executor/models/registry.py 
851
852
853
854
855
856
857
858
859
860
861
def _run() -> None:
    # Setup plugins
    from vllm.plugins import load_general_plugins
    load_general_plugins()

    fn, output_file = pickle.loads(sys.stdin.buffer.read())

    result = fn()

    with open(output_file, "wb") as f:
        f.write(pickle.dumps(result))

_run_in_subprocess 

_run_in_subprocess(fn: Callable[[], _T]) -> _T
Source code in vllm/model_executor/models/registry.py 
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
def _run_in_subprocess(fn: Callable[[], _T]) -> _T:
    # NOTE: We use a temporary directory instead of a temporary file to avoid
    # issues like https://stackoverflow.com/questions/23212435/permission-denied-to-write-to-my-temporary-file
    with tempfile.TemporaryDirectory() as tempdir:
        output_filepath = os.path.join(tempdir, "registry_output.tmp")

        # `cloudpickle` allows pickling lambda functions directly
        import cloudpickle
        input_bytes = cloudpickle.dumps((fn, output_filepath))

        # cannot use `sys.executable __file__` here because the script
        # contains relative imports
        returned = subprocess.run(_SUBPROCESS_COMMAND,
                                  input=input_bytes,
                                  capture_output=True)

        # check if the subprocess is successful
        try:
            returned.check_returncode()
        except Exception as e:
            # wrap raised exception to provide more information
            raise RuntimeError(f"Error raised in subprocess:\n"
                               f"{returned.stderr.decode()}") from e

        with open(output_filepath, "rb") as f:
            return pickle.load(f)

_try_inspect_model_cls cached 

_try_inspect_model_cls(
    model_arch: str, model: _BaseRegisteredModel
) -> Optional[_ModelInfo]
Source code in vllm/model_executor/models/registry.py 
423
424
425
426
427
428
429
430
431
432
433
@lru_cache(maxsize=128)
def _try_inspect_model_cls(
    model_arch: str,
    model: _BaseRegisteredModel,
) -> Optional[_ModelInfo]:
    try:
        return model.inspect_model_cls()
    except Exception:
        logger.exception("Error in inspecting model architecture '%s'",
                         model_arch)
        return None

_try_load_model_cls cached 

_try_load_model_cls(
    model_arch: str, model: _BaseRegisteredModel
) -> Optional[type[Module]]
Source code in vllm/model_executor/models/registry.py 
408
409
410
411
412
413
414
415
416
417
418
419
420
@lru_cache(maxsize=128)
def _try_load_model_cls(
    model_arch: str,
    model: _BaseRegisteredModel,
) -> Optional[type[nn.Module]]:
    from vllm.platforms import current_platform
    current_platform.verify_model_arch(model_arch)
    try:
        return model.load_model_cls()
    except Exception:
        logger.exception("Error in loading model architecture '%s'",
                         model_arch)
        return None