safetensors 是一种用于安全存储张量(与 pickle 相反)的新型简单格式,并且仍然很快(零拷贝)。
safetensors 真的很快。
pip install safetensors
conda install -c huggingface safetensors
from safetensors import safe_open
tensors = {}
with safe_open("model.safetensors", framework="pt", device=0) as f:
for k in f.keys():
tensors[k] = f.get_tensor(k)
仅加载部分张量(在多个GPU上运行时很有趣):
from safetensors import safe_open
tensors = {}
with safe_open("model.safetensors", framework="pt", device=0) as f:
tensor_slice = f.get_slice("embedding")
vocab_size, hidden_dim = tensor_slice.get_shape()
tensor = tensor_slice[:, :hidden_dim]
import torch
from safetensors.torch import save_file
tensors = {
"embedding": torch.zeros((2, 2)),
"attention": torch.zeros((2, 3))
}
save_file(tensors, "model.safetensors")
safetensors 真的很快。让我们通过加载 gpt2 权重将其进行比较。要运行 GPU 基准测试,请确保您的机器具有 GPU,或者您已选择是否使用的是 Google Colab。
在开始之前,请确保已安装所有必要的库:
pip install safetensors huggingface_hub torch
让我们从导入所有将使用的包开始:
import os
import datetime
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
import torch
Download safetensors & torch weights for gpt2:
sf_filename = hf_hub_download("gpt2", filename="model.safetensors")
pt_filename = hf_hub_download("gpt2", filename="pytorch_model.bin")
start_st = datetime.datetime.now()
weights = load_file(sf_filename, device="cpu")
load_time_st = datetime.datetime.now() - start_st
print(f"Loaded safetensors {load_time_st}")
输出结果为:
Loaded safetensors 0:00:00.026842
start_pt = datetime.datetime.now()
weights = torch.load(pt_filename, map_location="cpu")
load_time_pt = datetime.datetime.now() - start_pt
print(f"Loaded pytorch {load_time_pt}")
输出结果为:
Loaded pytorch 0:00:00.182266
print(f"on CPU, safetensors is faster than pytorch by: {load_time_pt/load_time_st:.1f} X")
输出结果为:
on CPU, safetensors is faster than pytorch by: 6.8 X
这种加速是由于该库通过直接映射文件来避免不必要的副本。实际上可以在 torch 上完成。 当前显示的加速比已打开:
os.environ["SAFETENSORS_FAST_GPU"] = "1"
torch.zeros((2, 2)).cuda()
start_st = datetime.datetime.now()
weights = load_file(sf_filename, device="cuda:0")
load_time_st = datetime.datetime.now() - start_st
print(f"Loaded safetensors {load_time_st}")
start_pt = datetime.datetime.now()
weights = torch.load(pt_filename, map_location="cuda:0")
load_time_pt = datetime.datetime.now() - start_pt
print(f"Loaded pytorch {load_time_pt}")
print(f"on GPU, safetensors is faster than pytorch by: {load_time_pt/load_time_st:.1f} X")
输出结果为:
Loaded safetensors 0:00:00.497415
Loaded pytorch 0:00:00.250602
on GPU, safetensors is faster than pytorch by: 0.5 X
加速有效是因为此库能够跳过不必要的 CPU 分配。不幸的是,据我们所知,它无法在纯 pytorch 中复制。该库的工作原理是内存映射文件,使用 pytorch 创建空张量,并直接调用以直接在 GPU 上移动张量。
显卡:GTX 3060