使用Optimum-habana对LLM模型训练推理

optimum-habana 是 Transformers 和 Diffusers 库与 Intel Gaudi AI 加速器（HPU） 之间的接口。它提供了一套工具，可轻松在单 HPU 和多 HPU 环境下针对不同下游任务进行模型加载、训练和推理。用户只需稍作修改，就可以在 Intel Gaudi 加速器上尝试数千个 Hugging Face 模型和相关任务。

1、官方已验证的模型和任务列表：

Transformers:

Architecture	Training	Inference	Tasks
BERT	✔️	✔️	text classificationquestion answeringlanguage modelingtext feature extraction
RoBERTa	✔️	✔️	question answeringlanguage modeling
ALBERT	✔️	✔️	question answeringlanguage modeling
DistilBERT	✔️	✔️	question answeringlanguage modeling
GPT2	✔️	✔️	language modelingtext generation
BLOOM(Z)		DeepSpeed	text generation
StarCoder / StarCoder2	✔️	Single-card	language modelingtext generation
GPT-J	DeepSpeed	Single cardDeepSpeed	language modelingtext generation
GPT-Neo		Single card	text generation
GPT-NeoX	DeepSpeed	DeepSpeed	language modelingtext generation
OPT		DeepSpeed	text generation
Llama 2 / CodeLlama / Llama 3 / Llama Guard / Granite	✔️	✔️	language modelingtext generationquestion answeringtext classification (Llama Guard)
StableLM		Single card	text generation
Falcon	LoRA	✔️	language modelingtext generation
CodeGen		Single card	text generation
MPT		Single card