Is there any way to make pinned CPU tensors released back to the OS immediately?

1

The pinned CPU tensors can’t be released back to the OS immediately.

import torch
import gc
import ctypes
import psutil
import os

def get_memory_usage():
    """Return current process RSS memory usage in MB."""
    process = psutil.Process(os.getpid())
    return process.memory_info().rss / (1024 * 1024)

def trim_memory():
    """Attempt to release unused memory back to the OS using malloc_trim."""
    libc = ctypes.CDLL("libc.so.6")
    libc.malloc_trim(0)

# Initial memory usage
print(f"[Before allocation] Memory usage: {get_memory_usage():.2f} MB")

# Allocate 1 GiB of pinned memory on CPU
x = torch.empty(1024 * 1024 * 1024, dtype=torch.uint8, device="cpu", pin_memory=True)
print(f"[After allocation] Memory usage: {get_memory_usage():.2f} MB")

# Delete the tensor
del x

# Clear cuda cache
torch.cuda.empty_cache()

# Run garbage collection
gc.collect()

# Try to trim memory
trim_memory()

print(f"[After del + gc + malloc_trim] Memory usage: {get_memory_usage():.2f} MB")
2

I believe torch.cuda.empty_cache is for the CUDA caching allocator, but pinned host memory would be allocated through the host caching allocator.
Can you try with torch._C._host_emptyCache() instead of torch.cuda.empty_cache() ?
Maybe you can raise an issue to create a public python API for it, since there seems to be reasonable use cases for it.