콘텐츠로 이동

Optimum RBLN

Optimum RBLN은 허깅페이스 transformers와 diffusers 모델들이 RBLN NPU, 즉 ATOM (RBLN-CA02) 및 ATOM+ (RBLN-CA12)에서 실행 될 수 있도록 연결해주는 라이브러리 입니다. 이를 통해 허깅페이스 모델들의 다양한 다운스트림 태스크들을 단일 및 다중 NPU (Rebellions Scalable Design)에서 손쉽게 컴파일 및 추론 할 수 있습니다. 다음 표에 현재 Optimum RBLN이 지원하는 모델 목록이 나열되어 있습니다.

Transformers

Single NPU

Model Model Architecture Task
Phi-2 PhiForCausalLM
Gemma-2b GemmaForCausalLM
GPT2 GPT2LMHeadModel
GPT2-medium GPT2LMHeadModel
GPT2-large GPT2LMHeadModel
GPT2-xl GPT2LMHeadModel
T5-small T5ForConditionalGeneration
T5-base T5ForConditionalGeneration
T5-large T5ForConditionalGeneration
T5-3b T5ForConditionalGeneration
BART-base BartForConditionalGeneration
BART-large BartForConditionalGeneration
KoBART-base BartForConditionalGeneration
E5-base-4K BertModel
LaBSE BertModel
KR-SBERT-V40K-klueNLI-augSTS BertModel
BERT-base - BertForMaskedLM
- BertForQuestionAnswering
BERT-large - BertForMaskedLM
- BertForQuestionAnswering
DistilBERT-base DistilBertForQuestionAnswering
SecureBERT RobertaForMaskedLM
RoBERTa RobertaForSequenceClassification
BGE-Small-EN-v1.5 RBLNBertModel
BGE-Base-EN-v1.5 RBLNBertModel
BGE-Large-EN-v1.5 RBLNBertModel
BGE-M3 XLMRobertaModel
BGE-Reranker-V2-M3 XLMRobertaForSequenceClassification
BGE-Reranker-Base XLMRobertaForSequenceClassification
BGE-Reranker-Large XLMRobertaForSequenceClassification
Ko-Reranker XLMRobertaForSequenceClassification
Whisper-tiny WhisperForConditionalGeneration
Whisper-base WhisperForConditionalGeneration
Whisper-small WhisperForConditionalGeneration
Whisper-medium WhisperForConditionalGeneration
Whisper-large-v3 WhisperForConditionalGeneration
Whisper-large-v3-turbo WhisperForConditionalGeneration
Wav2Vec2 Wav2Vec2ForCTC
Audio-Spectogram-Transformer ASTForAudioClassification
DPT-large DPTForDepthEstimation
ViT-large ViTForImageClassification
ResNet50 ResNetForImageClassification

Multi-NPU (RSD)

Note

다중 NPU 기능은 ATOM+ (RBLN-CA12)에서만 지원됩니다. 지금 사용 중인 NPU의 종류는 rbln-stat 명령어로 확인할 수 있습니다.

Model Model Architecture Recommended # of NPUs Task
DeepSeek-R1-Distill-Llama-8b LlamaForCausalLM 8
DeepSeek-R1-Distill-Llama-70b LlamaForCausalLM 16
DeepSeek-R1-Distill-Qwen-1.5b Qwen2ForCausalLM 8
DeepSeek-R1-Distill-Qwen-7b Qwen2ForCausalLM 8
DeepSeek-R1-Distill-Qwen-14b Qwen2ForCausalLM 8
DeepSeek-R1-Distill-Qwen-32b Qwen2ForCausalLM 16
Llama3.3-70b LlamaForCausalLM 16
Llama3.2-3b LlamaForCausalLM 8
Llama3.1-70b LlamaForCausalLM 16
Llama3.1-8b LlamaForCausalLM 8
Llama3-8b LlamaForCausalLM 4 or 8
Llama3-8b + LoRA LlamaForCausalLM 4 or 8
Llama2-7b LlamaForCausalLM 4 or 8
Llama2-13b LlamaForCausalLM 4 or 8
Gemma-7b GemmaForCausalLM 4 or 8
Mistral-7b MistralForCausalLM 4 or 8
Qwen2-7b Qwen2ForCausalLM 4 or 8
Qwen2.5-7b Qwen2ForCausalLM 4 or 8
Qwen2.5-14b Qwen2ForCausalLM 4 or 8
Salamandra-7b LlamaForCausalLM 4 or 8
KONI-Llama3.1-8b LlamaForCausalLM 8
EXAONE-3.0-7.8b ExaoneForCausalLM 4 or 8
EXAONE-3.5-2.4b ExaoneForCausalLM 4
EXAONE-3.5-7.8b ExaoneForCausalLM 4 or 8
EXAONE-3.5-32b ExaoneForCausalLM 8 or 16
Mi:dm-7b MidmLMHeadModel 4 or 8
SOLAR-10.7b LlamaForCausalLM 4 or 8
EEVE-Korean-10.8b LlamaForCausalLM 4 or 8
Llava-v1.6-mistral-7b LlavaNextForConditionalGeneration 4 or 8

Diffusers

Note

윗 첨자 †가 표시된 모델은 크기가 커서 ATOM 하나에 담을 수 없습니다. 따라서 모델의 모듈들을 여러 개의 ATOM에 나누어 올려야합니다. 모델 모듈의 구체적인 분산 정보는 모델 예제 코드를 참조하십시오.

Model Model Architecture Task
Stable Diffusion
  • StableDiffusionPipeline
  • StableDiffusionImg2ImgPipeline
  • StableDiffusionInpaintPipeline
Stable Diffusion + LoRA
  • StableDiffusionPipeline
Stable Diffusion V3†
  • StableDiffusion3Pipeline
  • StableDiffusion3Img2ImgPipeline
  • StableDiffusion3InpaintPipeline
Stable Diffusion XL
  • StableDiffusionXLPipeline
  • StableDiffusionXLImg2ImgPipeline
  • StableDiffusionXLInpaintPipeline
Stable Diffusion XL + multi-LoRA
  • StableDiffusionXLPipeline
SDXL-turbo
  • StableDiffusionXLPipeline
  • StableDiffusionXLImg2ImgPipeline
Stable Diffusion + ControlNet
  • StableDiffusionControlNetPipeline
  • StableDiffusionControlNetImg2ImgPipeline
Stable Diffusion XL + ControlNet
  • StableDiffusionXLControlNetPipeline
  • StableDiffusionXLControlNetImg2ImgPipeline
Kandinsky V2.2
  • KandinskyV22PriorPipeline
  • KandinskyV22Pipeline
  • KandinskyV22Img2ImgPipeline
  • KandinskyV22InpaintPipeline
  • KandinskyV22CombinedPipeline
  • KandinskyV22Img2ImgCombinedPipeline
  • KandinskyV22InpaintCombinedPipeline