Loading checkpoints with GPU slower than with CPU and moving to GPU?

[AlUlkesh]

This is probably dependent on which GPU and CPU are used, but on my system I get these numbers:

Two runs:

c:\tools\python>python safentensor_load2.py cpu ckpt
cpu: 1.8734321594238281
c:\tools\python>python safentensor_load2.py gpu ckpt
gpu: 2.2135305404663086

c:\tools\python>python safentensor_load2.py cpu safetensors
cpu: 1.740401029586792
c:\tools\python>python safentensor_load2.py gpu safetensors
gpu: 3.016714572906494


c:\tools\python>python safentensor_load2.py cpu ckpt
cpu: 1.8504252433776855
c:\tools\python>python safentensor_load2.py gpu ckpt
gpu: 2.2703230381011963

c:\tools\python>python safentensor_load2.py cpu safetensors
cpu: 1.7646374702453613
c:\tools\python>python safentensor_load2.py gpu safetensors
gpu: 3.029703140258789

Tested with:

import torch
import safetensors.torch
import time
import sys

checkpoint_file = r"s:\ai\#models\stable-diffusion\realisticVisionV13_v13.ckpt"
safetensor_file = r"s:\ai\#models\stable-diffusion\Realistic_Vision_V1.3.safetensors"

start_time = time.time()
if sys.argv[1] == "gpu":
    if sys.argv[2] == "ckpt":
        pl_sd = torch.load(checkpoint_file, map_location="cuda:0")
    elif sys.argv[2] == "safetensors":
        pl_sd = safetensors.torch.load_file(safetensor_file, device="cuda:0")
    else:
        raise Exception("Wrong first argument")
elif sys.argv[1] == "cpu":
    if sys.argv[2] == "ckpt":
        pl_sd = torch.load(checkpoint_file, map_location="cpu")
    elif sys.argv[2] == "safetensors":
        pl_sd = safetensors.torch.load_file(safetensor_file, device="cpu")
    else:
        raise Exception("Wrong first argument")
    loaded = {k: v.to("cuda:0") for k, v in pl_sd.items()}
print(f"{sys.argv[1]}: {time.time() - start_time}")    

Or do I have an error in my test code?

Running on a RTX 2060 and Ryzen 9 5950X.