File size: 2,263 Bytes
bfcf71e 44b21e0 bfcf71e 44b21e0 bfcf71e 44b21e0 bfcf71e |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 |
# Code to load a model.
import os
import warnings
import torch
import transformers
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
def load_model(repo_id, bnb=None, torch_dtype='auto'):
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# Try our best to get deterministic results.
if device.type == 'cuda':
# For determinism with CUDA >= 10.2, PyTorch says to use one of these.
os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8'
#os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':16:8'
torch.use_deterministic_algorithms(True)
# Ignore a spurious warning from huggingface_hub:
# https://github.com/huggingface/transformers/issues/30618
warnings.filterwarnings('ignore', message="`resume_download` is deprecated")
# Ignore a spurious warning from bitsandbytes.
warnings.filterwarnings('ignore', message="MatMul8bitLt: inputs will be cast from")
print(f'Loading model "{repo_id}" (bnb = "{bnb}")...')
# Ignore a spurious warning "Special tokens have been added..."
transformers.logging.set_verbosity_error()
tokenizer = AutoTokenizer.from_pretrained(repo_id, use_fast=True)
transformers.logging.set_verbosity_warning()
bnb_config = None
if bnb == 'nf8':
bnb_config = BitsAndBytesConfig(load_in_8bit=True)
if bnb == 'nf4':
bnb_config = BitsAndBytesConfig(load_in_4bit=True)
device_map = 'auto'
if device.type == 'cpu':
# BFloat16 is not supported on MPS
device_map = None
model = AutoModelForCausalLM.from_pretrained(
repo_id,
torch_dtype=torch_dtype,
device_map=device_map,
quantization_config=bnb_config,
)
# Disable gradients to save memory.
for param in model.parameters():
param.requires_grad = False
# Try our best to get deterministic results.
model.eval()
print('Done loading model.')
return model, tokenizer
def load_tokenizer(repo_id):
# Ignore a spurious warning "Special tokens have been added..."
transformers.logging.set_verbosity_error()
tokenizer = AutoTokenizer.from_pretrained(repo_id, use_fast=True)
transformers.logging.set_verbosity_warning()
return tokenizer |