Auto Classes¶

Auto Classes는 이름이나 경로를 기반으로 비슷한 트랜스포머 모델(가중치, 설정, 어휘 등)을 자동으로 불러옵니다. 이를 통해 사용자는 정확한 모델 아키텍처를 알지 못해도 모델을 쉽게 불러올 수 있습니다.

주요 클래스¶

RBLNAutoConfig: 오토 모델을 위한 설정 클래스
RBLNAutoModel: RBLN NPU를 사용하여 지원하는 트랜스포머 모델을 실행하기 위한 클래스
RBLNAutoModelForCTC: RBLN NPU를 사용하여 연결주의 시간 분류 (Connectionist Temporal Classification, CTC) 모델을 실행하기 위한 오토 모델 클래스
RBLNAutoModelForCausalLM: RBLN NPU를 사용하여 인과적 자연어 모델을 실행하기 위한 오토 모델 클래스
RBLNAutoModelForSeq2SeqLM: RBLN NPU를 사용하여 시퀀스에서 시퀀스를 생성하는 자연어 모델을 실행하기 위한 오토 모델 클래스
RBLNAutoModelForDepthEstimation: RBLN NPU를 사용하여 시퀀스에서 깊이를 추정하는 모델을 실행하기 위한 오토 모델 클래스
RBLNAutoModelForSequenceClassification: RBLN NPU를 사용하여 시퀀스 분류 모델을 실행하기 위한 오토 모델 클래스
RBLNAutoModelForSpeechSeq2Seq: RBLN NPU를 사용하여 음성 시퀀스에서 시퀀스를 생성하는 모델을 실행하기 위한 오토 모델 클래스
RBLNAutoModelForVision2Seq: RBLN NPU를 사용하여 이미지에서 시퀀스를 생성하는 모델을 실행하기 위한 오토 모델 클래스
RBLNAutoModelForImageTextToText: RBLN NPU를 사용하여 이미지 및 문자에서 문자를 생성하는 모델을 실행하기 위한 오토 모델 클래스
RBLNAutoModelForMaskedLM: RBLN NPU를 사용하여 마스킹 자연어 모델을 실행하기 위한 오토 모델 클래스
RBLNAutoModelForAudioClassification: RBLN NPU를 사용하여 오디오 분류 모델을 실행하기 위한 오토 모델 클래스
RBLNAutoModelForImageClassification: RBLN NPU를 사용하여 이미지 분류 모델을 실행하기 위한 오토 모델 클래스
RBLNAutoModelForQuestionAnswering: RBLN NPU를 사용하여 질문-답변 모델을 실행하기 위한 오토 모델 클래스
RBLNAutoModelForTextEncoding: RBLN NPU에서 문자 인코딩을 실행하기 위한 오토 모델 클래스
RBLNAutoModelForZeroShotObjectDetection: RBLN NPU에서 제로샷 객체 탐지를 실행하기 위한 오토 모델 클래스

커스텀 클래스 등록¶

AutoClass에는 커스텀 클래스를 추가하여 기능을 확장할 수 있는 register() 메서드가 있습니다. 예를 들어, RBLNMistralNeMoForTextUpsampler라는 커스텀 모델과 그 설정 클래스인 RBLNMistralNeMoForTextUpsamplerConfig를 만들었다면, 이들을 AutoClass에 추가하여 자동으로 불러올 수 있습니다.

컴파일

from cosmos_upsampler import RBLNMistralNeMoForTextUpsampler, RBLNMistralNeMoForTextUpsamplerConfig

model_id = "nvidia/Cosmos-UpsamplePrompt1-12B-Text2World"

# Register Custom Class
RBLNAutoModel.register(RBLNMistralNeMoForTextUpsampler, exist_ok=True)
RBLNAutoConfig.register(RBLNMistralNeMoForTextUpsamplerConfig, exist_ok=True)

# Compile model
model = RBLNMistralNeMoForTextUpsampler.from_pretrained(
    model_id=upsampler_model_id,
    export=True,
    rbln_config={
        "batch_size": 1,
        "max_seq_len": 1024,
        "tensor_parallel_size": 4,
        "create_runtimes": False,
    },
)

model.save_pretrained(os.path.basename(model_id))
...

추론

from cosmos_upsampler import RBLNMistralNeMoForTextUpsampler, RBLNMistralNeMoForTextUpsamplerConfig
...
model_id = "nvidia/Cosmos-UpsamplePrompt1-12B-Text2World"
model_dir = os.path.basename(model_id)

# Register Custom Class
RBLNAutoModel.register(RBLNMistralNeMoForTextUpsampler, exist_ok=True)
RBLNAutoConfig.register(RBLNMistralNeMoForTextUpsamplerConfig, exist_ok=True)

# Load from compiled model
model = RBLNMistralNeMoForTextUpsampler.from_pretrained(
    model_dir,
    export=False,
)
...

지원되는 모델¶

연결 시계열 분류 (CTC)

Model	Model Architecture	AutoClass
Wav2Vec2	Wav2Vec2ForCTC	RBLNAutoModelForCTC

인과 언어 모델 (CausalLM)

Model	Model Architecture	AutoClass
DeepSeek-R1-Distill-Llama-8b	LlamaForCausalLM	RBLNAutoModelForCausalLM
DeepSeek-R1-Distill-Llama-70b	LlamaForCausalLM	RBLNAutoModelForCausalLM
DeepSeek-R1-Distill-Qwen-1.5b	Qwen2ForCausalLM	RBLNAutoModelForCausalLM
DeepSeek-R1-Distill-Qwen-7b	Qwen2ForCausalLM	RBLNAutoModelForCausalLM
DeepSeek-R1-Distill-Qwen-14b	Qwen2ForCausalLM	RBLNAutoModelForCausalLM
DeepSeek-R1-Distill-Qwen-32b	Qwen2ForCausalLM	RBLNAutoModelForCausalLM
Llama3.3-70b	LlamaForCausalLM	RBLNAutoModelForCausalLM
Llama3.2-3b	LlamaForCausalLM	RBLNAutoModelForCausalLM
Llama3.1-70b	LlamaForCausalLM	RBLNAutoModelForCausalLM
Llama3.1-8b	LlamaForCausalLM	RBLNAutoModelForCausalLM
Llama3-8b	LlamaForCausalLM	RBLNAutoModelForCausalLM
Llama3-8b + LoRA	LlamaForCausalLM	RBLNAutoModelForCausalLM
Llama2-7b	LlamaForCausalLM	RBLNAutoModelForCausalLM
Llama2-13b	LlamaForCausalLM	RBLNAutoModelForCausalLM
Phi-2	PhiForCausalLM	RBLNAutoModelForCausalLM
Gemma-7b	GemmaForCausalLM	RBLNAutoModelForCausalLM
Gemma-2b	GemmaForCausalLM	RBLNAutoModelForCausalLM
OPT-2.7b	OPTForCausalLM	RBLNAutoModelForCausalLM
Mistral-7b	MistralForCausalLM	RBLNAutoModelForCausalLM
A.X-4.0-Light	Qwen2ForCausalLM	RBLNAutoModelForCausalLM
Qwen2-7b	Qwen2ForCausalLM	RBLNAutoModelForCausalLM
Qwen2.5-0.5b	Qwen2ForCausalLM	RBLNAutoModelForCausalLM
Qwen2.5-1.5b	Qwen2ForCausalLM	RBLNAutoModelForCausalLM
Qwen2.5-3b	Qwen2ForCausalLM	RBLNAutoModelForCausalLM
Qwen2.5-7b	Qwen2ForCausalLM	RBLNAutoModelForCausalLM
Qwen2.5-14b	Qwen2ForCausalLM	RBLNAutoModelForCausalLM
Qwen2.5-32b	Qwen2ForCausalLM	RBLNAutoModelForCausalLM
Qwen2.5-72b	Qwen2ForCausalLM	RBLNAutoModelForCausalLM
Qwen3-0.6b	Qwen3ForCausalLM	RBLNAutoModelForCausalLM
Qwen3-1.7b	Qwen3ForCausalLM	RBLNAutoModelForCausalLM
Qwen3-4b	Qwen3ForCausalLM	RBLNAutoModelForCausalLM
Qwen3-8b	Qwen3ForCausalLM	RBLNAutoModelForCausalLM
Qwen3-32b	Qwen3ForCausalLM	RBLNAutoModelForCausalLM
Midm-2.0-Mini	LlamaForCausalLM	RBLNAutoModelForCausalLM
Midm-2.0-Base	LlamaForCausalLM	RBLNAutoModelForCausalLM
Salamandra-7b	LlamaForCausalLM	RBLNAutoModelForCausalLM
KONI-Llama3.1-8b	LlamaForCausalLM	RBLNAutoModelForCausalLM
EXAONE-3.0-7.8b	ExaoneForCausalLM	RBLNAutoModelForCausalLM
EXAONE-3.5-2.4b	ExaoneForCausalLM	RBLNAutoModelForCausalLM
EXAONE-3.5-7.8b	ExaoneForCausalLM	RBLNAutoModelForCausalLM
EXAONE-3.5-32b	ExaoneForCausalLM	RBLNAutoModelForCausalLM
GPT2	GPT2LMHeadModel	RBLNAutoModelForCausalLM
GPT2-medium	GPT2LMHeadModel	RBLNAutoModelForCausalLM
GPT2-large	GPT2LMHeadModel	RBLNAutoModelForCausalLM
GPT2-xl	GPT2LMHeadModel	RBLNAutoModelForCausalLM
OPT-6.7b	OPTForCausalLM	RBLNAutoModelForCausalLM
SOLAR-10.7b	LlamaForCausalLM	RBLNAutoModelForCausalLM
EEVE-Korean-10.8b	LlamaForCausalLM	RBLNAutoModelForCausalLM

시퀀스-투-시퀀스 언어 모델 (Seq2SeqLM)

Model	Model Architecture	AutoClass
T5-11b	T5ForConditionalGeneration	RBLNAutoModelForSeq2SeqLM
T5-small	T5ForConditionalGeneration	RBLNAutoModelForSeq2SeqLM
T5-base	T5ForConditionalGeneration	RBLNAutoModelForSeq2SeqLM
T5-large	T5ForConditionalGeneration	RBLNAutoModelForSeq2SeqLM
T5-3b	T5ForConditionalGeneration	RBLNAutoModelForSeq2SeqLM
BART-base	BartForConditionalGeneration	RBLNAutoModelForSeq2SeqLM
BART-large	BartForConditionalGeneration	RBLNAutoModelForSeq2SeqLM
KoBART-base	BartForConditionalGeneration	RBLNAutoModelForSeq2SeqLM
Pegasus	PegasusForConditionalGeneration	RBLNAutoModelForSeq2SeqLM

깊이 추정 (DepthEstimation)

Model	Model Architecture	AutoClass
Depth-Anything-V2-Small	DepthAnythingForDepthEstimation	RBLNAutoModelForDepthEstimation
Depth-Anything-V2-Base	DepthAnythingForDepthEstimation	RBLNAutoModelForDepthEstimation
Depth-Anything-V2-Large	DepthAnythingForDepthEstimation	RBLNAutoModelForDepthEstimation
DPT-large	DPTForDepthEstimation	RBLNAutoModelForDepthEstimation

시퀀스 분류 (SequenceClassification)

Model	Model Architecture	AutoClass
RoBERTa	RobertaForSequenceClassification	RBLNAutoModelForSequenceClassification
BGE-Reranker-V2-M3	XLMRobertaForSequenceClassification	RBLNAutoModelForSequenceClassification
BGE-Reranker-Base	XLMRobertaForSequenceClassification	RBLNAutoModelForSequenceClassification
BGE-Reranker-Large	XLMRobertaForSequenceClassification	RBLNAutoModelForSequenceClassification
Ko-Reranker	XLMRobertaForSequenceClassification	RBLNAutoModelForSequenceClassification

음성 시퀀스에서 시퀀스 생성 (SpeechSeq2Seq)

Model	Model Architecture	AutoClass
Whisper-tiny	WhisperForConditionalGeneration	RBLNAutoModelForSpeechSeq2Seq
Whisper-base	WhisperForConditionalGeneration	RBLNAutoModelForSpeechSeq2Seq
Whisper-small	WhisperForConditionalGeneration	RBLNAutoModelForSpeechSeq2Seq
Whisper-medium	WhisperForConditionalGeneration	RBLNAutoModelForSpeechSeq2Seq
Whisper-large-v3	WhisperForConditionalGeneration	RBLNAutoModelForSpeechSeq2Seq
Whisper-large-v3-turbo	WhisperForConditionalGeneration	RBLNAutoModelForSpeechSeq2Seq

이미지에서 시퀀스 생성 (Vision2Seq)

Model	Model Architecture	AutoClass
Qwen2-VL-7b	Qwen2VLForConditionalGeneration	RBLNAutoModelForVision2Seq
Qwen2.5-VL-7b	Qwen2_5_VLForConditionalGeneration	RBLNAutoModelForVision2Seq
Idefics3-8B-Llama3	Idefics3ForConditionalGeneration	RBLNAutoModelForVision2Seq
Llava-v1.5-7b	LlavaForConditionalGeneration	RBLNAutoModelForVision2Seq
Llava-v1.6-mistral-7b	LlavaNextForConditionalGeneration	RBLNAutoModelForVision2Seq
Pixtral-12b	LlavaForConditionalGeneration	RBLNAutoModelForVision2Seq
BLIP2-6.7b	RBLNBlip2ForConditionalGeneration	RBLNAutoModelForVision2Seq
BLIP2-2.7b	RBLNBlip2ForConditionalGeneration	RBLNAutoModelForVision2Seq

이미지 및 문자에서 문자 생성 (ImageTextToText)

Model	Model Architecture	AutoClass
Gemma3-4b	Gemma3ForConditionalGeneration	RBLNAutoModelForImageTextToText
Gemma3-12b	Gemma3ForConditionalGeneration	RBLNAutoModelForImageTextToText
Gemma3-27b	Gemma3ForConditionalGeneration	RBLNAutoModelForImageTextToText
Qwen2-VL-7b	Qwen2VLForConditionalGeneration	RBLNAutoModelForImageTextToText
Qwen2.5-VL-7b	Qwen2_5_VLForConditionalGeneration	RBLNAutoModelForImageTextToText
Idefics3-8B-Llama3	Idefics3ForConditionalGeneration	RBLNAutoModelForImageTextToText
Llava-v1.5-7b	LlavaForConditionalGeneration	RBLNAutoModelForImageTextToText
Llava-v1.6-mistral-7b	LlavaNextForConditionalGeneration	RBLNAutoModelForImageTextToText
Pixtral-12b	LlavaForConditionalGeneration	RBLNAutoModelForImageTextToText
BLIP2-6.7b	RBLNBlip2ForConditionalGeneration	RBLNAutoModelForImageTextToText
BLIP2-2.7b	RBLNBlip2ForConditionalGeneration	RBLNAutoModelForImageTextToText

마스킹 자연어 모델 (MaskedLM)

Model	Model Architecture	AutoClass
BERT-base	BertForMaskedLM	RBLNAutoModelForMaskedLM
BERT-large	BertForMaskedLM	RBLNAutoModelForMaskedLM
SecureBERT	RobertaForMaskedLM	RBLNAutoModelForMaskedLM

음성 분류 (AudioClassification)

Model	Model Architecture	AutoClass
Audio-Spectogram-Transformer	ASTForAudioClassification	RBLNAutoModelForAudioClassification

이미지 분류 (ImageClassification)

Model	Model Architecture	AutoClass
ViT-large	ViTForImageClassification	RBLNAutoModelForImageClassification
ResNet50	ResNetForImageClassification	RBLNAutoModelForImageClassification

질문-답변 (QuestionAnswering)

Model	Model Architecture	AutoClass
BERT-base	BertForQuestionAnswering	RBLNAutoModelForQuestionAnswering
BERT-large	BertForQuestionAnswering	RBLNAutoModelForQuestionAnswering

문자 인코딩 (TextEncoding)

Model	Model Architecture	AutoClass
T5-Enc-11b	T5EncoderModel	RBLNAutoModelForTextEncoding
Qwen3-Embedding-4b	Qwen3Model	RBLNAutoModelForTextEncoding
Qwen3-Embedding-0.6b	Qwen3Model	RBLNAutoModelForTextEncoding
E5-base-4K	BertModel	RBLNAutoModelForTextEncoding
KR-SBERT-V40K-klueNLI-augSTS	BertModel	RBLNAutoModelForTextEncoding
BGE-Small-EN-v1.5	RBLNBertModel	RBLNAutoModelForTextEncoding
BGE-Large-EN-v1.5	RBLNBertModel	RBLNAutoModelForTextEncoding
BGE-M3/Dense-Embedding	XLMRobertaModel	RBLNAutoModelForTextEncoding
BGE-M3/Multi-Vector	XLMRobertaModel	RBLNAutoModelForTextEncoding
BGE-M3/Sparse-Embedding	XLMRobertaModel	RBLNAutoModelForTextEncoding

제로샷 객체 탐지 (Zero-Shot Object Detection)

Model	Model Architecture	AutoClass
GroundingDino-Tiny	GroundingDinoForObjectDetection	RBLNAutoModelForZeroShotObjectDetection
GroundingDino-Base	GroundingDinoForObjectDetection	RBLNAutoModelForZeroShotObjectDetection

API 참조¶

Classes¶

`RBLNAutoConfig` ¶

Resolver and factory for RBLN model configurations.

This class selects the concrete RBLNModelConfig subclass, validates the provided data, and returns a frozen configuration object that serves as the single source of truth during export and load. It does not define the schema or control model behavior.

Functions¶

`register(config, exist_ok=False)` `staticmethod` ¶

Register a new configuration for this class.

Parameters:

Name	Type	Description	Default
`config`	`RBLNModelConfig`	The config to register.	required
`exist_ok`	`bool`	Whether to allow registering an already registered model.	`False`

`load(path, passed_rbln_config=None, kwargs={}, return_unused_kwargs=False)` `staticmethod` ¶

Load RBLNModelConfig from a path. Class name is automatically inferred from the rbln_config.json file.

Parameters:

Name	Type	Description	Default
`path`	`str`	Path to the RBLNModelConfig.	required
`passed_rbln_config`	`Optional[RBLNModelConfig]`	RBLNModelConfig to pass its runtime options.	`None`

Returns:

Name	Type	Description
`RBLNModelConfig`	`Union[RBLNModelConfig, Tuple[RBLNModelConfig, Dict[str, Any]]]`	The loaded RBLNModelConfig.

Classes¶

`RBLNAutoModel` ¶

Bases: _BaseAutoModelClass

Automatically detect all supported transformers models.

Functions¶

`from_pretrained(model_id, export=None, rbln_config=None, **kwargs)` `classmethod` ¶

Load an RBLN-accelerated model from a pretrained checkpoint or a compiled RBLN artifact.

This convenience method determines the concrete RBLN* model class that matches the underlying HuggingFace architecture and dispatches to that class's from_pretrained() implementation. Depending on whether a compiled RBLN folder is detected (or if export=True is passed), it will either:

Compile from a HuggingFace checkpoint to an RBLN model
Or load an already-compiled RBLN model directory/repository

Parameters:

Name	Type	Description	Default
`model_id`	`Union[str, Path]`	HF repo id or local path. For compiled models, this should point to a directory (optionally under `subfolder`) that contains `*.rbln` files and `rbln_config.json`.	required
`export`	`bool`	Force compilation from a HuggingFace checkpoint. When `None`, this is inferred by checking whether compiled artifacts exist at `model_id`.	`None`
`rbln_config`	`Optional[Union[Dict, RBLNModelConfig]]`	RBLN compilation/runtime configuration. May be provided as a dictionary or as an instance of the specific model's config class (e.g., `RBLNLlamaForCausalLMConfig`).	`None`
`kwargs`		Additional keyword arguments. - Arguments prefixed with `rbln_` are forwarded to the RBLN config. - Remaining arguments are forwarded to the HuggingFace loader (e.g., `revision`, `token`, `trust_remote_code`, `cache_dir`, `subfolder`, `local_files_only`).	`{}`

Returns:

Type	Description
	An instantiated RBLN model ready for inference on RBLN NPUs.

`from_model(model, config=None, rbln_config=None, **kwargs)` `classmethod` ¶

Convert and compile an in-memory HuggingFace model into an RBLN model.

This method resolves the appropriate concrete RBLN* class from the input model's class name (e.g., LlamaForCausalLM -> RBLNLlamaForCausalLM) and then delegates to that class's from_model() implementation.

Parameters:

Name	Type	Description	Default
`model`	`PreTrainedModel`	A HuggingFace model instance to convert.	required
`config`	`Optional[PretrainedConfig]`	The configuration object associated with the model.	`None`
`rbln_config`	`Optional[Union[RBLNModelConfig, Dict]]`	RBLN compilation/runtime configuration. May be provided as a dictionary or as an instance of the specific model's config class.	`None`
`kwargs`	`Any`	Additional keyword arguments. - Arguments prefixed with `rbln_` are forwarded to the RBLN config.	`{}`

Returns:

Type	Description
`RBLNBaseModel`	An instantiated RBLN model ready for inference on RBLN NPUs.

`register(rbln_cls, exist_ok=False)` `staticmethod` ¶

Register a new RBLN model class.

Parameters:

Name	Type	Description	Default
`rbln_cls`	`Type[RBLNBaseModel]`	The RBLN model class to register.	required
`exist_ok`	`bool`	Whether to allow registering an already registered model.	`False`

`RBLNAutoModelForCTC` ¶

Bases: _BaseAutoModelClass

Automatically detect Connectionist Temporal Classification (CTC) head Models.

Functions¶

`from_pretrained(model_id, export=None, rbln_config=None, **kwargs)` `classmethod` ¶

Load an RBLN-accelerated model from a pretrained checkpoint or a compiled RBLN artifact.

This convenience method determines the concrete RBLN* model class that matches the underlying HuggingFace architecture and dispatches to that class's from_pretrained() implementation. Depending on whether a compiled RBLN folder is detected (or if export=True is passed), it will either:

Compile from a HuggingFace checkpoint to an RBLN model
Or load an already-compiled RBLN model directory/repository

Parameters:

Name	Type	Description	Default
`model_id`	`Union[str, Path]`	HF repo id or local path. For compiled models, this should point to a directory (optionally under `subfolder`) that contains `*.rbln` files and `rbln_config.json`.	required
`export`	`bool`	Force compilation from a HuggingFace checkpoint. When `None`, this is inferred by checking whether compiled artifacts exist at `model_id`.	`None`
`rbln_config`	`Optional[Union[Dict, RBLNModelConfig]]`	RBLN compilation/runtime configuration. May be provided as a dictionary or as an instance of the specific model's config class (e.g., `RBLNLlamaForCausalLMConfig`).	`None`
`kwargs`		Additional keyword arguments. - Arguments prefixed with `rbln_` are forwarded to the RBLN config. - Remaining arguments are forwarded to the HuggingFace loader (e.g., `revision`, `token`, `trust_remote_code`, `cache_dir`, `subfolder`, `local_files_only`).	`{}`

Returns:

Type	Description
	An instantiated RBLN model ready for inference on RBLN NPUs.

`from_model(model, config=None, rbln_config=None, **kwargs)` `classmethod` ¶

Convert and compile an in-memory HuggingFace model into an RBLN model.

This method resolves the appropriate concrete RBLN* class from the input model's class name (e.g., LlamaForCausalLM -> RBLNLlamaForCausalLM) and then delegates to that class's from_model() implementation.

Parameters:

Name	Type	Description	Default
`model`	`PreTrainedModel`	A HuggingFace model instance to convert.	required
`config`	`Optional[PretrainedConfig]`	The configuration object associated with the model.	`None`
`rbln_config`	`Optional[Union[RBLNModelConfig, Dict]]`	RBLN compilation/runtime configuration. May be provided as a dictionary or as an instance of the specific model's config class.	`None`
`kwargs`	`Any`	Additional keyword arguments. - Arguments prefixed with `rbln_` are forwarded to the RBLN config.	`{}`

Returns:

Type	Description
`RBLNBaseModel`	An instantiated RBLN model ready for inference on RBLN NPUs.

`register(rbln_cls, exist_ok=False)` `staticmethod` ¶

Register a new RBLN model class.

Parameters:

Name	Type	Description	Default
`rbln_cls`	`Type[RBLNBaseModel]`	The RBLN model class to register.	required
`exist_ok`	`bool`	Whether to allow registering an already registered model.	`False`

`RBLNAutoModelForCausalLM` ¶

Bases: _BaseAutoModelClass

Automatically detect Casual Language Models.

Functions¶

`from_pretrained(model_id, export=None, rbln_config=None, **kwargs)` `classmethod` ¶

Load an RBLN-accelerated model from a pretrained checkpoint or a compiled RBLN artifact.

This convenience method determines the concrete RBLN* model class that matches the underlying HuggingFace architecture and dispatches to that class's from_pretrained() implementation. Depending on whether a compiled RBLN folder is detected (or if export=True is passed), it will either:

Compile from a HuggingFace checkpoint to an RBLN model
Or load an already-compiled RBLN model directory/repository

Parameters:

Name	Type	Description	Default
`model_id`	`Union[str, Path]`	HF repo id or local path. For compiled models, this should point to a directory (optionally under `subfolder`) that contains `*.rbln` files and `rbln_config.json`.	required
`export`	`bool`	Force compilation from a HuggingFace checkpoint. When `None`, this is inferred by checking whether compiled artifacts exist at `model_id`.	`None`
`rbln_config`	`Optional[Union[Dict, RBLNModelConfig]]`	RBLN compilation/runtime configuration. May be provided as a dictionary or as an instance of the specific model's config class (e.g., `RBLNLlamaForCausalLMConfig`).	`None`
`kwargs`		Additional keyword arguments. - Arguments prefixed with `rbln_` are forwarded to the RBLN config. - Remaining arguments are forwarded to the HuggingFace loader (e.g., `revision`, `token`, `trust_remote_code`, `cache_dir`, `subfolder`, `local_files_only`).	`{}`

Returns:

Type	Description
	An instantiated RBLN model ready for inference on RBLN NPUs.

`from_model(model, config=None, rbln_config=None, **kwargs)` `classmethod` ¶

Convert and compile an in-memory HuggingFace model into an RBLN model.

This method resolves the appropriate concrete RBLN* class from the input model's class name (e.g., LlamaForCausalLM -> RBLNLlamaForCausalLM) and then delegates to that class's from_model() implementation.

Parameters:

Name	Type	Description	Default
`model`	`PreTrainedModel`	A HuggingFace model instance to convert.	required
`config`	`Optional[PretrainedConfig]`	The configuration object associated with the model.	`None`
`rbln_config`	`Optional[Union[RBLNModelConfig, Dict]]`	RBLN compilation/runtime configuration. May be provided as a dictionary or as an instance of the specific model's config class.	`None`
`kwargs`	`Any`	Additional keyword arguments. - Arguments prefixed with `rbln_` are forwarded to the RBLN config.	`{}`

Returns:

Type	Description
`RBLNBaseModel`	An instantiated RBLN model ready for inference on RBLN NPUs.

`register(rbln_cls, exist_ok=False)` `staticmethod` ¶

Register a new RBLN model class.

Parameters:

Name	Type	Description	Default
`rbln_cls`	`Type[RBLNBaseModel]`	The RBLN model class to register.	required
`exist_ok`	`bool`	Whether to allow registering an already registered model.	`False`

`RBLNAutoModelForSeq2SeqLM` ¶

Bases: _BaseAutoModelClass

Automatically detect Sequence to Sequence Language Models.

Functions¶

`from_pretrained(model_id, export=None, rbln_config=None, **kwargs)` `classmethod` ¶

Load an RBLN-accelerated model from a pretrained checkpoint or a compiled RBLN artifact.

This convenience method determines the concrete RBLN* model class that matches the underlying HuggingFace architecture and dispatches to that class's from_pretrained() implementation. Depending on whether a compiled RBLN folder is detected (or if export=True is passed), it will either:

Compile from a HuggingFace checkpoint to an RBLN model
Or load an already-compiled RBLN model directory/repository

Parameters:

Name	Type	Description	Default
`model_id`	`Union[str, Path]`	HF repo id or local path. For compiled models, this should point to a directory (optionally under `subfolder`) that contains `*.rbln` files and `rbln_config.json`.	required
`export`	`bool`	Force compilation from a HuggingFace checkpoint. When `None`, this is inferred by checking whether compiled artifacts exist at `model_id`.	`None`
`rbln_config`	`Optional[Union[Dict, RBLNModelConfig]]`	RBLN compilation/runtime configuration. May be provided as a dictionary or as an instance of the specific model's config class (e.g., `RBLNLlamaForCausalLMConfig`).	`None`
`kwargs`		Additional keyword arguments. - Arguments prefixed with `rbln_` are forwarded to the RBLN config. - Remaining arguments are forwarded to the HuggingFace loader (e.g., `revision`, `token`, `trust_remote_code`, `cache_dir`, `subfolder`, `local_files_only`).	`{}`

Returns:

Type	Description
	An instantiated RBLN model ready for inference on RBLN NPUs.

`from_model(model, config=None, rbln_config=None, **kwargs)` `classmethod` ¶

Convert and compile an in-memory HuggingFace model into an RBLN model.

This method resolves the appropriate concrete RBLN* class from the input model's class name (e.g., LlamaForCausalLM -> RBLNLlamaForCausalLM) and then delegates to that class's from_model() implementation.

Parameters:

Name	Type	Description	Default
`model`	`PreTrainedModel`	A HuggingFace model instance to convert.	required
`config`	`Optional[PretrainedConfig]`	The configuration object associated with the model.	`None`
`rbln_config`	`Optional[Union[RBLNModelConfig, Dict]]`	RBLN compilation/runtime configuration. May be provided as a dictionary or as an instance of the specific model's config class.	`None`
`kwargs`	`Any`	Additional keyword arguments. - Arguments prefixed with `rbln_` are forwarded to the RBLN config.	`{}`

Returns:

Type	Description
`RBLNBaseModel`	An instantiated RBLN model ready for inference on RBLN NPUs.

`register(rbln_cls, exist_ok=False)` `staticmethod` ¶

Register a new RBLN model class.

Parameters:

Name	Type	Description	Default
`rbln_cls`	`Type[RBLNBaseModel]`	The RBLN model class to register.	required
`exist_ok`	`bool`	Whether to allow registering an already registered model.	`False`

`RBLNAutoModelForSpeechSeq2Seq` ¶

Bases: _BaseAutoModelClass

Automatically detect Sequence to Sequence Generation Models.

Functions¶

`from_pretrained(model_id, export=None, rbln_config=None, **kwargs)` `classmethod` ¶

Load an RBLN-accelerated model from a pretrained checkpoint or a compiled RBLN artifact.

This convenience method determines the concrete RBLN* model class that matches the underlying HuggingFace architecture and dispatches to that class's from_pretrained() implementation. Depending on whether a compiled RBLN folder is detected (or if export=True is passed), it will either:

Compile from a HuggingFace checkpoint to an RBLN model
Or load an already-compiled RBLN model directory/repository

Parameters:

Name	Type	Description	Default
`model_id`	`Union[str, Path]`	HF repo id or local path. For compiled models, this should point to a directory (optionally under `subfolder`) that contains `*.rbln` files and `rbln_config.json`.	required
`export`	`bool`	Force compilation from a HuggingFace checkpoint. When `None`, this is inferred by checking whether compiled artifacts exist at `model_id`.	`None`
`rbln_config`	`Optional[Union[Dict, RBLNModelConfig]]`	RBLN compilation/runtime configuration. May be provided as a dictionary or as an instance of the specific model's config class (e.g., `RBLNLlamaForCausalLMConfig`).	`None`
`kwargs`		Additional keyword arguments. - Arguments prefixed with `rbln_` are forwarded to the RBLN config. - Remaining arguments are forwarded to the HuggingFace loader (e.g., `revision`, `token`, `trust_remote_code`, `cache_dir`, `subfolder`, `local_files_only`).	`{}`

Returns:

Type	Description
	An instantiated RBLN model ready for inference on RBLN NPUs.

`from_model(model, config=None, rbln_config=None, **kwargs)` `classmethod` ¶

Convert and compile an in-memory HuggingFace model into an RBLN model.

This method resolves the appropriate concrete RBLN* class from the input model's class name (e.g., LlamaForCausalLM -> RBLNLlamaForCausalLM) and then delegates to that class's from_model() implementation.

Parameters:

Name	Type	Description	Default
`model`	`PreTrainedModel`	A HuggingFace model instance to convert.	required
`config`	`Optional[PretrainedConfig]`	The configuration object associated with the model.	`None`
`rbln_config`	`Optional[Union[RBLNModelConfig, Dict]]`	RBLN compilation/runtime configuration. May be provided as a dictionary or as an instance of the specific model's config class.	`None`
`kwargs`	`Any`	Additional keyword arguments. - Arguments prefixed with `rbln_` are forwarded to the RBLN config.	`{}`

Returns:

Type	Description
`RBLNBaseModel`	An instantiated RBLN model ready for inference on RBLN NPUs.

`register(rbln_cls, exist_ok=False)` `staticmethod` ¶

Register a new RBLN model class.

Parameters:

Name	Type	Description	Default
`rbln_cls`	`Type[RBLNBaseModel]`	The RBLN model class to register.	required
`exist_ok`	`bool`	Whether to allow registering an already registered model.	`False`

`RBLNAutoModelForDepthEstimation` ¶

Bases: _BaseAutoModelClass

Automatically detect Speech Sequence to Sequence Language Models.

Functions¶

`from_pretrained(model_id, export=None, rbln_config=None, **kwargs)` `classmethod` ¶

Load an RBLN-accelerated model from a pretrained checkpoint or a compiled RBLN artifact.

This convenience method determines the concrete RBLN* model class that matches the underlying HuggingFace architecture and dispatches to that class's from_pretrained() implementation. Depending on whether a compiled RBLN folder is detected (or if export=True is passed), it will either:

Compile from a HuggingFace checkpoint to an RBLN model
Or load an already-compiled RBLN model directory/repository

Parameters:

Name	Type	Description	Default
`model_id`	`Union[str, Path]`	HF repo id or local path. For compiled models, this should point to a directory (optionally under `subfolder`) that contains `*.rbln` files and `rbln_config.json`.	required
`export`	`bool`	Force compilation from a HuggingFace checkpoint. When `None`, this is inferred by checking whether compiled artifacts exist at `model_id`.	`None`
`rbln_config`	`Optional[Union[Dict, RBLNModelConfig]]`	RBLN compilation/runtime configuration. May be provided as a dictionary or as an instance of the specific model's config class (e.g., `RBLNLlamaForCausalLMConfig`).	`None`
`kwargs`		Additional keyword arguments. - Arguments prefixed with `rbln_` are forwarded to the RBLN config. - Remaining arguments are forwarded to the HuggingFace loader (e.g., `revision`, `token`, `trust_remote_code`, `cache_dir`, `subfolder`, `local_files_only`).	`{}`

Returns:

Type	Description
	An instantiated RBLN model ready for inference on RBLN NPUs.

`from_model(model, config=None, rbln_config=None, **kwargs)` `classmethod` ¶

Convert and compile an in-memory HuggingFace model into an RBLN model.

This method resolves the appropriate concrete RBLN* class from the input model's class name (e.g., LlamaForCausalLM -> RBLNLlamaForCausalLM) and then delegates to that class's from_model() implementation.

Parameters:

Name	Type	Description	Default
`model`	`PreTrainedModel`	A HuggingFace model instance to convert.	required
`config`	`Optional[PretrainedConfig]`	The configuration object associated with the model.	`None`
`rbln_config`	`Optional[Union[RBLNModelConfig, Dict]]`	RBLN compilation/runtime configuration. May be provided as a dictionary or as an instance of the specific model's config class.	`None`
`kwargs`	`Any`	Additional keyword arguments. - Arguments prefixed with `rbln_` are forwarded to the RBLN config.	`{}`

Returns:

Type	Description
`RBLNBaseModel`	An instantiated RBLN model ready for inference on RBLN NPUs.

`register(rbln_cls, exist_ok=False)` `staticmethod` ¶

Register a new RBLN model class.

Parameters:

Name	Type	Description	Default
`rbln_cls`	`Type[RBLNBaseModel]`	The RBLN model class to register.	required
`exist_ok`	`bool`	Whether to allow registering an already registered model.	`False`

`RBLNAutoModelForSequenceClassification` ¶

Bases: _BaseAutoModelClass

Automatically detect Sequence Classification Models.

Functions¶

`from_pretrained(model_id, export=None, rbln_config=None, **kwargs)` `classmethod` ¶

Load an RBLN-accelerated model from a pretrained checkpoint or a compiled RBLN artifact.

This convenience method determines the concrete RBLN* model class that matches the underlying HuggingFace architecture and dispatches to that class's from_pretrained() implementation. Depending on whether a compiled RBLN folder is detected (or if export=True is passed), it will either:

Compile from a HuggingFace checkpoint to an RBLN model
Or load an already-compiled RBLN model directory/repository

Parameters:

Name	Type	Description	Default
`model_id`	`Union[str, Path]`	HF repo id or local path. For compiled models, this should point to a directory (optionally under `subfolder`) that contains `*.rbln` files and `rbln_config.json`.	required
`export`	`bool`	Force compilation from a HuggingFace checkpoint. When `None`, this is inferred by checking whether compiled artifacts exist at `model_id`.	`None`
`rbln_config`	`Optional[Union[Dict, RBLNModelConfig]]`	RBLN compilation/runtime configuration. May be provided as a dictionary or as an instance of the specific model's config class (e.g., `RBLNLlamaForCausalLMConfig`).	`None`
`kwargs`		Additional keyword arguments. - Arguments prefixed with `rbln_` are forwarded to the RBLN config. - Remaining arguments are forwarded to the HuggingFace loader (e.g., `revision`, `token`, `trust_remote_code`, `cache_dir`, `subfolder`, `local_files_only`).	`{}`

Returns:

Type	Description
	An instantiated RBLN model ready for inference on RBLN NPUs.

`from_model(model, config=None, rbln_config=None, **kwargs)` `classmethod` ¶

Convert and compile an in-memory HuggingFace model into an RBLN model.

This method resolves the appropriate concrete RBLN* class from the input model's class name (e.g., LlamaForCausalLM -> RBLNLlamaForCausalLM) and then delegates to that class's from_model() implementation.

Parameters:

Name	Type	Description	Default
`model`	`PreTrainedModel`	A HuggingFace model instance to convert.	required
`config`	`Optional[PretrainedConfig]`	The configuration object associated with the model.	`None`
`rbln_config`	`Optional[Union[RBLNModelConfig, Dict]]`	RBLN compilation/runtime configuration. May be provided as a dictionary or as an instance of the specific model's config class.	`None`
`kwargs`	`Any`	Additional keyword arguments. - Arguments prefixed with `rbln_` are forwarded to the RBLN config.	`{}`

Returns:

Type	Description
`RBLNBaseModel`	An instantiated RBLN model ready for inference on RBLN NPUs.

`register(rbln_cls, exist_ok=False)` `staticmethod` ¶

Register a new RBLN model class.

Parameters:

Name	Type	Description	Default
`rbln_cls`	`Type[RBLNBaseModel]`	The RBLN model class to register.	required
`exist_ok`	`bool`	Whether to allow registering an already registered model.	`False`

`RBLNAutoModelForVision2Seq` ¶

Bases: _BaseAutoModelClass

Automatically detect Vision to Sequence Generation Models.

Functions¶

`from_pretrained(model_id, export=None, rbln_config=None, **kwargs)` `classmethod` ¶

Load an RBLN-accelerated model from a pretrained checkpoint or a compiled RBLN artifact.

This convenience method determines the concrete RBLN* model class that matches the underlying HuggingFace architecture and dispatches to that class's from_pretrained() implementation. Depending on whether a compiled RBLN folder is detected (or if export=True is passed), it will either:

Compile from a HuggingFace checkpoint to an RBLN model
Or load an already-compiled RBLN model directory/repository

Parameters:

Name	Type	Description	Default
`model_id`	`Union[str, Path]`	HF repo id or local path. For compiled models, this should point to a directory (optionally under `subfolder`) that contains `*.rbln` files and `rbln_config.json`.	required
`export`	`bool`	Force compilation from a HuggingFace checkpoint. When `None`, this is inferred by checking whether compiled artifacts exist at `model_id`.	`None`
`rbln_config`	`Optional[Union[Dict, RBLNModelConfig]]`	RBLN compilation/runtime configuration. May be provided as a dictionary or as an instance of the specific model's config class (e.g., `RBLNLlamaForCausalLMConfig`).	`None`
`kwargs`		Additional keyword arguments. - Arguments prefixed with `rbln_` are forwarded to the RBLN config. - Remaining arguments are forwarded to the HuggingFace loader (e.g., `revision`, `token`, `trust_remote_code`, `cache_dir`, `subfolder`, `local_files_only`).	`{}`

Returns:

Type	Description
	An instantiated RBLN model ready for inference on RBLN NPUs.

`from_model(model, config=None, rbln_config=None, **kwargs)` `classmethod` ¶

Convert and compile an in-memory HuggingFace model into an RBLN model.

This method resolves the appropriate concrete RBLN* class from the input model's class name (e.g., LlamaForCausalLM -> RBLNLlamaForCausalLM) and then delegates to that class's from_model() implementation.

Parameters:

Name	Type	Description	Default
`model`	`PreTrainedModel`	A HuggingFace model instance to convert.	required
`config`	`Optional[PretrainedConfig]`	The configuration object associated with the model.	`None`
`rbln_config`	`Optional[Union[RBLNModelConfig, Dict]]`	RBLN compilation/runtime configuration. May be provided as a dictionary or as an instance of the specific model's config class.	`None`
`kwargs`	`Any`	Additional keyword arguments. - Arguments prefixed with `rbln_` are forwarded to the RBLN config.	`{}`

Returns:

Type	Description
`RBLNBaseModel`	An instantiated RBLN model ready for inference on RBLN NPUs.

`register(rbln_cls, exist_ok=False)` `staticmethod` ¶

Register a new RBLN model class.

Parameters:

Name	Type	Description	Default
`rbln_cls`	`Type[RBLNBaseModel]`	The RBLN model class to register.	required
`exist_ok`	`bool`	Whether to allow registering an already registered model.	`False`

`RBLNAutoModelForImageTextToText` ¶

Bases: _BaseAutoModelClass

Automatically detect Image and Text to Text Generation Models.

Functions¶

`from_pretrained(model_id, export=None, rbln_config=None, **kwargs)` `classmethod` ¶

Load an RBLN-accelerated model from a pretrained checkpoint or a compiled RBLN artifact.

This convenience method determines the concrete RBLN* model class that matches the underlying HuggingFace architecture and dispatches to that class's from_pretrained() implementation. Depending on whether a compiled RBLN folder is detected (or if export=True is passed), it will either:

Compile from a HuggingFace checkpoint to an RBLN model
Or load an already-compiled RBLN model directory/repository

Parameters:

Name	Type	Description	Default
`model_id`	`Union[str, Path]`	HF repo id or local path. For compiled models, this should point to a directory (optionally under `subfolder`) that contains `*.rbln` files and `rbln_config.json`.	required
`export`	`bool`	Force compilation from a HuggingFace checkpoint. When `None`, this is inferred by checking whether compiled artifacts exist at `model_id`.	`None`
`rbln_config`	`Optional[Union[Dict, RBLNModelConfig]]`	RBLN compilation/runtime configuration. May be provided as a dictionary or as an instance of the specific model's config class (e.g., `RBLNLlamaForCausalLMConfig`).	`None`
`kwargs`		Additional keyword arguments. - Arguments prefixed with `rbln_` are forwarded to the RBLN config. - Remaining arguments are forwarded to the HuggingFace loader (e.g., `revision`, `token`, `trust_remote_code`, `cache_dir`, `subfolder`, `local_files_only`).	`{}`

Returns:

Type	Description
	An instantiated RBLN model ready for inference on RBLN NPUs.

`from_model(model, config=None, rbln_config=None, **kwargs)` `classmethod` ¶

Convert and compile an in-memory HuggingFace model into an RBLN model.

This method resolves the appropriate concrete RBLN* class from the input model's class name (e.g., LlamaForCausalLM -> RBLNLlamaForCausalLM) and then delegates to that class's from_model() implementation.

Parameters:

Name	Type	Description	Default
`model`	`PreTrainedModel`	A HuggingFace model instance to convert.	required
`config`	`Optional[PretrainedConfig]`	The configuration object associated with the model.	`None`
`rbln_config`	`Optional[Union[RBLNModelConfig, Dict]]`	RBLN compilation/runtime configuration. May be provided as a dictionary or as an instance of the specific model's config class.	`None`
`kwargs`	`Any`	Additional keyword arguments. - Arguments prefixed with `rbln_` are forwarded to the RBLN config.	`{}`

Returns:

Type	Description
`RBLNBaseModel`	An instantiated RBLN model ready for inference on RBLN NPUs.

`register(rbln_cls, exist_ok=False)` `staticmethod` ¶

Register a new RBLN model class.

Parameters:

Name	Type	Description	Default
`rbln_cls`	`Type[RBLNBaseModel]`	The RBLN model class to register.	required
`exist_ok`	`bool`	Whether to allow registering an already registered model.	`False`

`RBLNAutoModelForMaskedLM` ¶

Bases: _BaseAutoModelClass

Automatically detect Masked Lanuage Models.

Functions¶

`from_pretrained(model_id, export=None, rbln_config=None, **kwargs)` `classmethod` ¶

Load an RBLN-accelerated model from a pretrained checkpoint or a compiled RBLN artifact.

This convenience method determines the concrete RBLN* model class that matches the underlying HuggingFace architecture and dispatches to that class's from_pretrained() implementation. Depending on whether a compiled RBLN folder is detected (or if export=True is passed), it will either:

Compile from a HuggingFace checkpoint to an RBLN model
Or load an already-compiled RBLN model directory/repository

Parameters:

Name	Type	Description	Default
`model_id`	`Union[str, Path]`	HF repo id or local path. For compiled models, this should point to a directory (optionally under `subfolder`) that contains `*.rbln` files and `rbln_config.json`.	required
`export`	`bool`	Force compilation from a HuggingFace checkpoint. When `None`, this is inferred by checking whether compiled artifacts exist at `model_id`.	`None`
`rbln_config`	`Optional[Union[Dict, RBLNModelConfig]]`	RBLN compilation/runtime configuration. May be provided as a dictionary or as an instance of the specific model's config class (e.g., `RBLNLlamaForCausalLMConfig`).	`None`
`kwargs`		Additional keyword arguments. - Arguments prefixed with `rbln_` are forwarded to the RBLN config. - Remaining arguments are forwarded to the HuggingFace loader (e.g., `revision`, `token`, `trust_remote_code`, `cache_dir`, `subfolder`, `local_files_only`).	`{}`

Returns:

Type	Description
	An instantiated RBLN model ready for inference on RBLN NPUs.

`from_model(model, config=None, rbln_config=None, **kwargs)` `classmethod` ¶

Convert and compile an in-memory HuggingFace model into an RBLN model.

This method resolves the appropriate concrete RBLN* class from the input model's class name (e.g., LlamaForCausalLM -> RBLNLlamaForCausalLM) and then delegates to that class's from_model() implementation.

Parameters:

Name	Type	Description	Default
`model`	`PreTrainedModel`	A HuggingFace model instance to convert.	required
`config`	`Optional[PretrainedConfig]`	The configuration object associated with the model.	`None`
`rbln_config`	`Optional[Union[RBLNModelConfig, Dict]]`	RBLN compilation/runtime configuration. May be provided as a dictionary or as an instance of the specific model's config class.	`None`
`kwargs`	`Any`	Additional keyword arguments. - Arguments prefixed with `rbln_` are forwarded to the RBLN config.	`{}`

Returns:

Type	Description
`RBLNBaseModel`	An instantiated RBLN model ready for inference on RBLN NPUs.

`register(rbln_cls, exist_ok=False)` `staticmethod` ¶

Register a new RBLN model class.

Parameters:

Name	Type	Description	Default
`rbln_cls`	`Type[RBLNBaseModel]`	The RBLN model class to register.	required
`exist_ok`	`bool`	Whether to allow registering an already registered model.	`False`

`RBLNAutoModelForAudioClassification` ¶

Bases: _BaseAutoModelClass

Automatically detect Audio Classification Models.

Functions¶

`from_pretrained(model_id, export=None, rbln_config=None, **kwargs)` `classmethod` ¶

Load an RBLN-accelerated model from a pretrained checkpoint or a compiled RBLN artifact.

This convenience method determines the concrete RBLN* model class that matches the underlying HuggingFace architecture and dispatches to that class's from_pretrained() implementation. Depending on whether a compiled RBLN folder is detected (or if export=True is passed), it will either:

Compile from a HuggingFace checkpoint to an RBLN model
Or load an already-compiled RBLN model directory/repository

Parameters:

Name	Type	Description	Default
`model_id`	`Union[str, Path]`	HF repo id or local path. For compiled models, this should point to a directory (optionally under `subfolder`) that contains `*.rbln` files and `rbln_config.json`.	required
`export`	`bool`	Force compilation from a HuggingFace checkpoint. When `None`, this is inferred by checking whether compiled artifacts exist at `model_id`.	`None`
`rbln_config`	`Optional[Union[Dict, RBLNModelConfig]]`	RBLN compilation/runtime configuration. May be provided as a dictionary or as an instance of the specific model's config class (e.g., `RBLNLlamaForCausalLMConfig`).	`None`
`kwargs`		Additional keyword arguments. - Arguments prefixed with `rbln_` are forwarded to the RBLN config. - Remaining arguments are forwarded to the HuggingFace loader (e.g., `revision`, `token`, `trust_remote_code`, `cache_dir`, `subfolder`, `local_files_only`).	`{}`

Returns:

Type	Description
	An instantiated RBLN model ready for inference on RBLN NPUs.

`from_model(model, config=None, rbln_config=None, **kwargs)` `classmethod` ¶

Convert and compile an in-memory HuggingFace model into an RBLN model.

This method resolves the appropriate concrete RBLN* class from the input model's class name (e.g., LlamaForCausalLM -> RBLNLlamaForCausalLM) and then delegates to that class's from_model() implementation.

Parameters:

Name	Type	Description	Default
`model`	`PreTrainedModel`	A HuggingFace model instance to convert.	required
`config`	`Optional[PretrainedConfig]`	The configuration object associated with the model.	`None`
`rbln_config`	`Optional[Union[RBLNModelConfig, Dict]]`	RBLN compilation/runtime configuration. May be provided as a dictionary or as an instance of the specific model's config class.	`None`
`kwargs`	`Any`	Additional keyword arguments. - Arguments prefixed with `rbln_` are forwarded to the RBLN config.	`{}`

Returns:

Type	Description
`RBLNBaseModel`	An instantiated RBLN model ready for inference on RBLN NPUs.

`register(rbln_cls, exist_ok=False)` `staticmethod` ¶

Register a new RBLN model class.

Parameters:

Name	Type	Description	Default
`rbln_cls`	`Type[RBLNBaseModel]`	The RBLN model class to register.	required
`exist_ok`	`bool`	Whether to allow registering an already registered model.	`False`

`RBLNAutoModelForImageClassification` ¶

Bases: _BaseAutoModelClass

Automatically detect Image Classification Models.

Functions¶

`from_pretrained(model_id, export=None, rbln_config=None, **kwargs)` `classmethod` ¶

Load an RBLN-accelerated model from a pretrained checkpoint or a compiled RBLN artifact.

This convenience method determines the concrete RBLN* model class that matches the underlying HuggingFace architecture and dispatches to that class's from_pretrained() implementation. Depending on whether a compiled RBLN folder is detected (or if export=True is passed), it will either:

Compile from a HuggingFace checkpoint to an RBLN model
Or load an already-compiled RBLN model directory/repository

Parameters:

Name	Type	Description	Default
`model_id`	`Union[str, Path]`	HF repo id or local path. For compiled models, this should point to a directory (optionally under `subfolder`) that contains `*.rbln` files and `rbln_config.json`.	required
`export`	`bool`	Force compilation from a HuggingFace checkpoint. When `None`, this is inferred by checking whether compiled artifacts exist at `model_id`.	`None`
`rbln_config`	`Optional[Union[Dict, RBLNModelConfig]]`	RBLN compilation/runtime configuration. May be provided as a dictionary or as an instance of the specific model's config class (e.g., `RBLNLlamaForCausalLMConfig`).	`None`
`kwargs`		Additional keyword arguments. - Arguments prefixed with `rbln_` are forwarded to the RBLN config. - Remaining arguments are forwarded to the HuggingFace loader (e.g., `revision`, `token`, `trust_remote_code`, `cache_dir`, `subfolder`, `local_files_only`).	`{}`

Returns:

Type	Description
	An instantiated RBLN model ready for inference on RBLN NPUs.

`from_model(model, config=None, rbln_config=None, **kwargs)` `classmethod` ¶

Convert and compile an in-memory HuggingFace model into an RBLN model.

This method resolves the appropriate concrete RBLN* class from the input model's class name (e.g., LlamaForCausalLM -> RBLNLlamaForCausalLM) and then delegates to that class's from_model() implementation.

Parameters:

Name	Type	Description	Default
`model`	`PreTrainedModel`	A HuggingFace model instance to convert.	required
`config`	`Optional[PretrainedConfig]`	The configuration object associated with the model.	`None`
`rbln_config`	`Optional[Union[RBLNModelConfig, Dict]]`	RBLN compilation/runtime configuration. May be provided as a dictionary or as an instance of the specific model's config class.	`None`
`kwargs`	`Any`	Additional keyword arguments. - Arguments prefixed with `rbln_` are forwarded to the RBLN config.	`{}`

Returns:

Type	Description
`RBLNBaseModel`	An instantiated RBLN model ready for inference on RBLN NPUs.

`register(rbln_cls, exist_ok=False)` `staticmethod` ¶

Register a new RBLN model class.

Parameters:

Name	Type	Description	Default
`rbln_cls`	`Type[RBLNBaseModel]`	The RBLN model class to register.	required
`exist_ok`	`bool`	Whether to allow registering an already registered model.	`False`

`RBLNAutoModelForQuestionAnswering` ¶

Bases: _BaseAutoModelClass

Automatically detect Question Answering Models.

Functions¶

`from_pretrained(model_id, export=None, rbln_config=None, **kwargs)` `classmethod` ¶

Load an RBLN-accelerated model from a pretrained checkpoint or a compiled RBLN artifact.

This convenience method determines the concrete RBLN* model class that matches the underlying HuggingFace architecture and dispatches to that class's from_pretrained() implementation. Depending on whether a compiled RBLN folder is detected (or if export=True is passed), it will either:

Compile from a HuggingFace checkpoint to an RBLN model
Or load an already-compiled RBLN model directory/repository

Parameters:

Name	Type	Description	Default
`model_id`	`Union[str, Path]`	HF repo id or local path. For compiled models, this should point to a directory (optionally under `subfolder`) that contains `*.rbln` files and `rbln_config.json`.	required
`export`	`bool`	Force compilation from a HuggingFace checkpoint. When `None`, this is inferred by checking whether compiled artifacts exist at `model_id`.	`None`
`rbln_config`	`Optional[Union[Dict, RBLNModelConfig]]`	RBLN compilation/runtime configuration. May be provided as a dictionary or as an instance of the specific model's config class (e.g., `RBLNLlamaForCausalLMConfig`).	`None`
`kwargs`		Additional keyword arguments. - Arguments prefixed with `rbln_` are forwarded to the RBLN config. - Remaining arguments are forwarded to the HuggingFace loader (e.g., `revision`, `token`, `trust_remote_code`, `cache_dir`, `subfolder`, `local_files_only`).	`{}`

Returns:

Type	Description
	An instantiated RBLN model ready for inference on RBLN NPUs.

`from_model(model, config=None, rbln_config=None, **kwargs)` `classmethod` ¶

Convert and compile an in-memory HuggingFace model into an RBLN model.

This method resolves the appropriate concrete RBLN* class from the input model's class name (e.g., LlamaForCausalLM -> RBLNLlamaForCausalLM) and then delegates to that class's from_model() implementation.

Parameters:

Name	Type	Description	Default
`model`	`PreTrainedModel`	A HuggingFace model instance to convert.	required
`config`	`Optional[PretrainedConfig]`	The configuration object associated with the model.	`None`
`rbln_config`	`Optional[Union[RBLNModelConfig, Dict]]`	RBLN compilation/runtime configuration. May be provided as a dictionary or as an instance of the specific model's config class.	`None`
`kwargs`	`Any`	Additional keyword arguments. - Arguments prefixed with `rbln_` are forwarded to the RBLN config.	`{}`

Returns:

Type	Description
`RBLNBaseModel`	An instantiated RBLN model ready for inference on RBLN NPUs.

`register(rbln_cls, exist_ok=False)` `staticmethod` ¶

Register a new RBLN model class.

Parameters:

Name	Type	Description	Default
`rbln_cls`	`Type[RBLNBaseModel]`	The RBLN model class to register.	required
`exist_ok`	`bool`	Whether to allow registering an already registered model.	`False`

`RBLNAutoModelForTextEncoding` ¶

Bases: _BaseAutoModelClass

Automatically detect Text Encoding Models.

Functions¶

`from_pretrained(model_id, export=None, rbln_config=None, **kwargs)` `classmethod` ¶

Load an RBLN-accelerated model from a pretrained checkpoint or a compiled RBLN artifact.

This convenience method determines the concrete RBLN* model class that matches the underlying HuggingFace architecture and dispatches to that class's from_pretrained() implementation. Depending on whether a compiled RBLN folder is detected (or if export=True is passed), it will either:

Compile from a HuggingFace checkpoint to an RBLN model
Or load an already-compiled RBLN model directory/repository

Parameters:

Name	Type	Description	Default
`model_id`	`Union[str, Path]`	HF repo id or local path. For compiled models, this should point to a directory (optionally under `subfolder`) that contains `*.rbln` files and `rbln_config.json`.	required
`export`	`bool`	Force compilation from a HuggingFace checkpoint. When `None`, this is inferred by checking whether compiled artifacts exist at `model_id`.	`None`
`rbln_config`	`Optional[Union[Dict, RBLNModelConfig]]`	RBLN compilation/runtime configuration. May be provided as a dictionary or as an instance of the specific model's config class (e.g., `RBLNLlamaForCausalLMConfig`).	`None`
`kwargs`		Additional keyword arguments. - Arguments prefixed with `rbln_` are forwarded to the RBLN config. - Remaining arguments are forwarded to the HuggingFace loader (e.g., `revision`, `token`, `trust_remote_code`, `cache_dir`, `subfolder`, `local_files_only`).	`{}`

Returns:

Type	Description
	An instantiated RBLN model ready for inference on RBLN NPUs.

`from_model(model, config=None, rbln_config=None, **kwargs)` `classmethod` ¶

Convert and compile an in-memory HuggingFace model into an RBLN model.

This method resolves the appropriate concrete RBLN* class from the input model's class name (e.g., LlamaForCausalLM -> RBLNLlamaForCausalLM) and then delegates to that class's from_model() implementation.

Parameters:

Name	Type	Description	Default
`model`	`PreTrainedModel`	A HuggingFace model instance to convert.	required
`config`	`Optional[PretrainedConfig]`	The configuration object associated with the model.	`None`
`rbln_config`	`Optional[Union[RBLNModelConfig, Dict]]`	RBLN compilation/runtime configuration. May be provided as a dictionary or as an instance of the specific model's config class.	`None`
`kwargs`	`Any`	Additional keyword arguments. - Arguments prefixed with `rbln_` are forwarded to the RBLN config.	`{}`

Returns:

Type	Description
`RBLNBaseModel`	An instantiated RBLN model ready for inference on RBLN NPUs.

`register(rbln_cls, exist_ok=False)` `staticmethod` ¶

Register a new RBLN model class.

Parameters:

Name	Type	Description	Default
`rbln_cls`	`Type[RBLNBaseModel]`	The RBLN model class to register.	required
`exist_ok`	`bool`	Whether to allow registering an already registered model.	`False`

`RBLNAutoModelForZeroShotObjectDetection` ¶

Bases: _BaseAutoModelClass

Automatically detect Zero Shot Object Detection Models.

Functions¶

`from_pretrained(model_id, export=None, rbln_config=None, **kwargs)` `classmethod` ¶

Load an RBLN-accelerated model from a pretrained checkpoint or a compiled RBLN artifact.

This convenience method determines the concrete RBLN* model class that matches the underlying HuggingFace architecture and dispatches to that class's from_pretrained() implementation. Depending on whether a compiled RBLN folder is detected (or if export=True is passed), it will either:

Compile from a HuggingFace checkpoint to an RBLN model
Or load an already-compiled RBLN model directory/repository

Parameters:

Name	Type	Description	Default
`model_id`	`Union[str, Path]`	HF repo id or local path. For compiled models, this should point to a directory (optionally under `subfolder`) that contains `*.rbln` files and `rbln_config.json`.	required
`export`	`bool`	Force compilation from a HuggingFace checkpoint. When `None`, this is inferred by checking whether compiled artifacts exist at `model_id`.	`None`
`rbln_config`	`Optional[Union[Dict, RBLNModelConfig]]`	RBLN compilation/runtime configuration. May be provided as a dictionary or as an instance of the specific model's config class (e.g., `RBLNLlamaForCausalLMConfig`).	`None`
`kwargs`		Additional keyword arguments. - Arguments prefixed with `rbln_` are forwarded to the RBLN config. - Remaining arguments are forwarded to the HuggingFace loader (e.g., `revision`, `token`, `trust_remote_code`, `cache_dir`, `subfolder`, `local_files_only`).	`{}`

Returns:

Type	Description
	An instantiated RBLN model ready for inference on RBLN NPUs.

`from_model(model, config=None, rbln_config=None, **kwargs)` `classmethod` ¶

Convert and compile an in-memory HuggingFace model into an RBLN model.

This method resolves the appropriate concrete RBLN* class from the input model's class name (e.g., LlamaForCausalLM -> RBLNLlamaForCausalLM) and then delegates to that class's from_model() implementation.

Parameters:

Name	Type	Description	Default
`model`	`PreTrainedModel`	A HuggingFace model instance to convert.	required
`config`	`Optional[PretrainedConfig]`	The configuration object associated with the model.	`None`
`rbln_config`	`Optional[Union[RBLNModelConfig, Dict]]`	RBLN compilation/runtime configuration. May be provided as a dictionary or as an instance of the specific model's config class.	`None`
`kwargs`	`Any`	Additional keyword arguments. - Arguments prefixed with `rbln_` are forwarded to the RBLN config.	`{}`

Returns:

Type	Description
`RBLNBaseModel`	An instantiated RBLN model ready for inference on RBLN NPUs.

`register(rbln_cls, exist_ok=False)` `staticmethod` ¶

Register a new RBLN model class.

Parameters:

Name	Type	Description	Default
`rbln_cls`	`Type[RBLNBaseModel]`	The RBLN model class to register.	required
`exist_ok`	`bool`	Whether to allow registering an already registered model.	`False`

Auto Classes¶

주요 클래스¶

커스텀 클래스 등록¶

지원되는 모델¶

API 참조¶

Classes¶

RBLNAutoConfig ¶

Functions¶

register(config, exist_ok=False) staticmethod ¶

load(path, passed_rbln_config=None, kwargs={}, return_unused_kwargs=False) staticmethod ¶

Classes¶

RBLNAutoModel ¶

Functions¶

from_pretrained(model_id, export=None, rbln_config=None, **kwargs) classmethod ¶

from_model(model, config=None, rbln_config=None, **kwargs) classmethod ¶

register(rbln_cls, exist_ok=False) staticmethod ¶

RBLNAutoModelForCTC ¶

Functions¶

from_pretrained(model_id, export=None, rbln_config=None, **kwargs) classmethod ¶

from_model(model, config=None, rbln_config=None, **kwargs) classmethod ¶

register(rbln_cls, exist_ok=False) staticmethod ¶

RBLNAutoModelForCausalLM ¶

Functions¶

from_pretrained(model_id, export=None, rbln_config=None, **kwargs) classmethod ¶

from_model(model, config=None, rbln_config=None, **kwargs) classmethod ¶

register(rbln_cls, exist_ok=False) staticmethod ¶

RBLNAutoModelForSeq2SeqLM ¶

Functions¶

from_pretrained(model_id, export=None, rbln_config=None, **kwargs) classmethod ¶

from_model(model, config=None, rbln_config=None, **kwargs) classmethod ¶

register(rbln_cls, exist_ok=False) staticmethod ¶

RBLNAutoModelForSpeechSeq2Seq ¶

Functions¶

from_pretrained(model_id, export=None, rbln_config=None, **kwargs) classmethod ¶

from_model(model, config=None, rbln_config=None, **kwargs) classmethod ¶

register(rbln_cls, exist_ok=False) staticmethod ¶

RBLNAutoModelForDepthEstimation ¶

Functions¶

from_pretrained(model_id, export=None, rbln_config=None, **kwargs) classmethod ¶

from_model(model, config=None, rbln_config=None, **kwargs) classmethod ¶

register(rbln_cls, exist_ok=False) staticmethod ¶

RBLNAutoModelForSequenceClassification ¶

Functions¶

from_pretrained(model_id, export=None, rbln_config=None, **kwargs) classmethod ¶

from_model(model, config=None, rbln_config=None, **kwargs) classmethod ¶

register(rbln_cls, exist_ok=False) staticmethod ¶

RBLNAutoModelForVision2Seq ¶

Functions¶

from_pretrained(model_id, export=None, rbln_config=None, **kwargs) classmethod ¶

from_model(model, config=None, rbln_config=None, **kwargs) classmethod ¶

register(rbln_cls, exist_ok=False) staticmethod ¶

RBLNAutoModelForImageTextToText ¶

Functions¶

from_pretrained(model_id, export=None, rbln_config=None, **kwargs) classmethod ¶

from_model(model, config=None, rbln_config=None, **kwargs) classmethod ¶

register(rbln_cls, exist_ok=False) staticmethod ¶

RBLNAutoModelForMaskedLM ¶

Functions¶

from_pretrained(model_id, export=None, rbln_config=None, **kwargs) classmethod ¶

from_model(model, config=None, rbln_config=None, **kwargs) classmethod ¶

register(rbln_cls, exist_ok=False) staticmethod ¶

RBLNAutoModelForAudioClassification ¶

Functions¶

from_pretrained(model_id, export=None, rbln_config=None, **kwargs) classmethod ¶

from_model(model, config=None, rbln_config=None, **kwargs) classmethod ¶

register(rbln_cls, exist_ok=False) staticmethod ¶

RBLNAutoModelForImageClassification ¶

Functions¶

from_pretrained(model_id, export=None, rbln_config=None, **kwargs) classmethod ¶

from_model(model, config=None, rbln_config=None, **kwargs) classmethod ¶

register(rbln_cls, exist_ok=False) staticmethod ¶

RBLNAutoModelForQuestionAnswering ¶

Functions¶

from_pretrained(model_id, export=None, rbln_config=None, **kwargs) classmethod ¶

from_model(model, config=None, rbln_config=None, **kwargs) classmethod ¶

register(rbln_cls, exist_ok=False) staticmethod ¶

RBLNAutoModelForTextEncoding ¶

Functions¶

from_pretrained(model_id, export=None, rbln_config=None, **kwargs) classmethod ¶

from_model(model, config=None, rbln_config=None, **kwargs) classmethod ¶

`RBLNAutoConfig` ¶

`register(config, exist_ok=False)` `staticmethod` ¶

`load(path, passed_rbln_config=None, kwargs={}, return_unused_kwargs=False)` `staticmethod` ¶

`RBLNAutoModel` ¶

`from_pretrained(model_id, export=None, rbln_config=None, **kwargs)` `classmethod` ¶

`from_model(model, config=None, rbln_config=None, **kwargs)` `classmethod` ¶

`register(rbln_cls, exist_ok=False)` `staticmethod` ¶

`RBLNAutoModelForCTC` ¶

`from_pretrained(model_id, export=None, rbln_config=None, **kwargs)` `classmethod` ¶

`from_model(model, config=None, rbln_config=None, **kwargs)` `classmethod` ¶

`register(rbln_cls, exist_ok=False)` `staticmethod` ¶

`RBLNAutoModelForCausalLM` ¶

`from_pretrained(model_id, export=None, rbln_config=None, **kwargs)` `classmethod` ¶

`from_model(model, config=None, rbln_config=None, **kwargs)` `classmethod` ¶

`register(rbln_cls, exist_ok=False)` `staticmethod` ¶

`RBLNAutoModelForSeq2SeqLM` ¶

`from_pretrained(model_id, export=None, rbln_config=None, **kwargs)` `classmethod` ¶

`from_model(model, config=None, rbln_config=None, **kwargs)` `classmethod` ¶

`register(rbln_cls, exist_ok=False)` `staticmethod` ¶

`RBLNAutoModelForSpeechSeq2Seq` ¶

`from_pretrained(model_id, export=None, rbln_config=None, **kwargs)` `classmethod` ¶

`from_model(model, config=None, rbln_config=None, **kwargs)` `classmethod` ¶

`register(rbln_cls, exist_ok=False)` `staticmethod` ¶

`RBLNAutoModelForDepthEstimation` ¶

`from_pretrained(model_id, export=None, rbln_config=None, **kwargs)` `classmethod` ¶

`from_model(model, config=None, rbln_config=None, **kwargs)` `classmethod` ¶

`register(rbln_cls, exist_ok=False)` `staticmethod` ¶

`RBLNAutoModelForSequenceClassification` ¶

`from_pretrained(model_id, export=None, rbln_config=None, **kwargs)` `classmethod` ¶

`from_model(model, config=None, rbln_config=None, **kwargs)` `classmethod` ¶

`register(rbln_cls, exist_ok=False)` `staticmethod` ¶

`RBLNAutoModelForVision2Seq` ¶

`from_pretrained(model_id, export=None, rbln_config=None, **kwargs)` `classmethod` ¶

`from_model(model, config=None, rbln_config=None, **kwargs)` `classmethod` ¶

`register(rbln_cls, exist_ok=False)` `staticmethod` ¶

`RBLNAutoModelForImageTextToText` ¶

`from_pretrained(model_id, export=None, rbln_config=None, **kwargs)` `classmethod` ¶

`from_model(model, config=None, rbln_config=None, **kwargs)` `classmethod` ¶

`register(rbln_cls, exist_ok=False)` `staticmethod` ¶

`RBLNAutoModelForMaskedLM` ¶

`from_pretrained(model_id, export=None, rbln_config=None, **kwargs)` `classmethod` ¶

`from_model(model, config=None, rbln_config=None, **kwargs)` `classmethod` ¶

`register(rbln_cls, exist_ok=False)` `staticmethod` ¶

`RBLNAutoModelForAudioClassification` ¶

`from_pretrained(model_id, export=None, rbln_config=None, **kwargs)` `classmethod` ¶

`from_model(model, config=None, rbln_config=None, **kwargs)` `classmethod` ¶

`register(rbln_cls, exist_ok=False)` `staticmethod` ¶

`RBLNAutoModelForImageClassification` ¶

`from_pretrained(model_id, export=None, rbln_config=None, **kwargs)` `classmethod` ¶

`from_model(model, config=None, rbln_config=None, **kwargs)` `classmethod` ¶

`register(rbln_cls, exist_ok=False)` `staticmethod` ¶

`RBLNAutoModelForQuestionAnswering` ¶

`from_pretrained(model_id, export=None, rbln_config=None, **kwargs)` `classmethod` ¶

`from_model(model, config=None, rbln_config=None, **kwargs)` `classmethod` ¶

`register(rbln_cls, exist_ok=False)` `staticmethod` ¶

`RBLNAutoModelForTextEncoding` ¶

`from_pretrained(model_id, export=None, rbln_config=None, **kwargs)` `classmethod` ¶

`from_model(model, config=None, rbln_config=None, **kwargs)` `classmethod` ¶

`register(rbln_cls, exist_ok=False)` `staticmethod` ¶

`RBLNAutoModelForZeroShotObjectDetection` ¶

`from_pretrained(model_id, export=None, rbln_config=None, **kwargs)` `classmethod` ¶

`from_model(model, config=None, rbln_config=None, **kwargs)` `classmethod` ¶

`register(rbln_cls, exist_ok=False)` `staticmethod` ¶