pip install transformers accelerate bitsandbytes
python - <<'PY'
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
m = 'meta-llama/Llama-3.1-8B-Instruct'
tok = AutoTokenizer.from_pretrained(m)
model = AutoModelForCausalLM.from_pretrained(m, torch_dtype=torch.float16, device_map='auto')
print(model.generate(**tok('Hello, world!', return_tensors='pt').to(model.device), max_new_tokens=32))
PY

pip install llama-index sentence-transformers faiss-cpu
python - <<'PY'
from llama_index.core import SimpleDirectoryReader, VectorStoreIndex
from llama_index.core.node_parser import SimpleNodeParser
from llama_index.core import Settings
from sentence_transformers import SentenceTransformer
Settings.embed_model = SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2')
docs = SimpleDirectoryReader('docs').load_data()
index = VectorStoreIndex.from_documents(docs)
print(index.as_query_engine().query('요약해줘'))
PY

pip install trl peft accelerate transformers datasets bitsandbytes
python - <<'PY'
from trl import SFTTrainer
from transformers import AutoTokenizer, AutoModelForCausalLM
model='meta-llama/Llama-3.1-8B'
tok=AutoTokenizer.from_pretrained(model)
mdl=AutoModelForCausalLM.from_pretrained(model, load_in_8bit=True, device_map='auto')
trainer=SFTTrainer(model=mdl, tokenizer=tok, train_dataset=[{'text':'hello'}], max_seq_length=512)
trainer.train()
PY

pip install diffusers accelerate transformers safetensors
python - <<'PY'
from diffusers import StableDiffusionPipeline
import torch
p = StableDiffusionPipeline.from_pretrained('runwayml/stable-diffusion-v1-5', torch_dtype=torch.float16).to('cuda')
img = p('a cozy workspace with a GPU').images[0]
img.save('out.png')
PY