Host memory still occupied after huggingface model deleted

🤗Transformers
1

import torch
from transformers import AutoConfig, AutoTokenizer, BloomForCausalLM, AutoModelForCausalLM
import time
import psutil
import gc

def get_host_memory():
memory_allocated = round(psutil.Process().memory_info().rss / 1024**3, 3)
print(“cpu”," memory used total: ", memory_allocated, “GB”)

def load_and_run():
s_ = time.time()
config = AutoConfig.from_pretrained(‘bigscience/bloom-560m’)
model = AutoModelForCausalLM.from_config(config)
e_ = time.time()
print("model loading time: %f " % (e_ - s_))
get_host_memory()
del model
del config

load_and_run()
gc.collect()
for obj in gc.get_objects():
try:
if torch.is_tensor(obj) or (hasattr(obj, ‘data’) and torch.is_tensor(obj.data)):
print(type(obj), obj.size())
except:
pass
get_host_memory()

Expected behavior

right now:
after model loaded:
cpu memory used total: 3.327 GB
after model deleted:
cpu memory used total: 3.364 GB

expected, host memory should be freed after model deleted.

2

@sgugger could you help to check this quesiton? thank you!