MagicNodes / mod /mg_cleanup.py
DZRobo
Add MG_CleanUp node and improve memory management
9068525
raw
history blame
12.7 kB
import os
import gc
import time
import platform
import ctypes
from ctypes import wintypes
import torch
import torch.nn.functional as F
import comfy.model_management as model_management
import comfy.sample as _sample
import comfy.samplers as _samplers
import comfy.utils as _utils
try:
import psutil # type: ignore
except Exception: # pragma: no cover
psutil = None # type: ignore
def _get_ram_mb() -> float:
try:
if psutil is not None:
p = psutil.Process(os.getpid())
rss = float(p.memory_info().rss)
try:
private = getattr(p.memory_full_info(), "private", None)
if isinstance(private, (int, float)) and private > 0:
rss = float(private)
except Exception:
pass
return rss / (1024.0 * 1024.0)
except Exception:
pass
return 0.0
def _get_vram_mb_per_device() -> list[tuple[int, float, float]]:
out = []
try:
if torch.cuda.is_available():
for d in range(torch.cuda.device_count()):
try:
reserved = float(torch.cuda.memory_reserved(d)) / (1024.0 * 1024.0)
allocated = float(torch.cuda.memory_allocated(d)) / (1024.0 * 1024.0)
except Exception:
reserved = 0.0
allocated = 0.0
out.append((d, reserved, allocated))
except Exception:
pass
return out
def _trim_working_set_windows():
try:
if platform.system().lower().startswith("win"):
kernel32 = ctypes.windll.kernel32 # type: ignore[attr-defined]
proc = kernel32.GetCurrentProcess()
kernel32.SetProcessWorkingSetSize(proc, ctypes.c_size_t(-1), ctypes.c_size_t(-1))
except Exception:
pass
def _enable_win_privileges(names):
"""Best-effort enable a set of Windows privileges for the current process."""
try:
if not platform.system().lower().startswith('win'):
return False
advapi32 = ctypes.windll.advapi32 # type: ignore[attr-defined]
kernel32 = ctypes.windll.kernel32 # type: ignore[attr-defined]
token = wintypes.HANDLE()
TOKEN_ADJUST_PRIVILEGES = 0x20
TOKEN_QUERY = 0x8
if not advapi32.OpenProcessToken(kernel32.GetCurrentProcess(), TOKEN_ADJUST_PRIVILEGES | TOKEN_QUERY, ctypes.byref(token)):
return False
class LUID(ctypes.Structure):
_fields_ = [("LowPart", wintypes.DWORD), ("HighPart", wintypes.LONG)]
class LUID_AND_ATTRIBUTES(ctypes.Structure):
_fields_ = [("Luid", LUID), ("Attributes", wintypes.DWORD)]
class TOKEN_PRIVILEGES(ctypes.Structure):
_fields_ = [("PrivilegeCount", wintypes.DWORD), ("Privileges", LUID_AND_ATTRIBUTES * 1)]
SE_PRIVILEGE_ENABLED = 0x2
success = False
for name in names:
luid = LUID()
if not advapi32.LookupPrivilegeValueW(None, ctypes.c_wchar_p(name), ctypes.byref(luid)):
continue
tp = TOKEN_PRIVILEGES()
tp.PrivilegeCount = 1
tp.Privileges[0].Luid = luid
tp.Privileges[0].Attributes = SE_PRIVILEGE_ENABLED
if advapi32.AdjustTokenPrivileges(token, False, ctypes.byref(tp), 0, None, None):
success = True
return success
except Exception:
return False
def _system_cache_trim_windows():
"""Attempt to purge standby/file caches on Windows (requires privileges)."""
try:
if not platform.system().lower().startswith('win'):
return False
_enable_win_privileges([
'SeIncreaseQuotaPrivilege',
'SeProfileSingleProcessPrivilege',
'SeDebugPrivilege',
])
try:
kernel32 = ctypes.windll.kernel32 # type: ignore[attr-defined]
SIZE_T = ctypes.c_size_t
kernel32.SetSystemFileCacheSize(SIZE_T(-1), SIZE_T(-1), wintypes.DWORD(0))
except Exception:
pass
try:
ntdll = ctypes.windll.ntdll # type: ignore[attr-defined]
SystemMemoryListInformation = 0x50
MemoryPurgeStandbyList = ctypes.c_ulong(4)
ntdll.NtSetSystemInformation(wintypes.ULONG(SystemMemoryListInformation), ctypes.byref(MemoryPurgeStandbyList), ctypes.sizeof(MemoryPurgeStandbyList))
except Exception:
pass
return True
except Exception:
return False
def cleanup_memory(sync_cuda: bool = True, hard_trim: bool = True) -> dict:
"""Run a best-effort cleanup of RAM/VRAM. Returns stats dict with before/after deltas."""
stats: dict = {"ram_before_mb": 0.0, "ram_after_mb": 0.0, "ram_freed_mb": 0.0, "gpu": []}
stats["ram_before_mb"] = _get_ram_mb()
gpu_before = _get_vram_mb_per_device()
try:
if sync_cuda and torch.cuda.is_available():
torch.cuda.synchronize()
except Exception:
pass
try:
import comfy.model_management as mm
if hasattr(mm, 'soft_empty_cache'):
mm.soft_empty_cache()
except Exception:
pass
try:
gc.collect()
except Exception:
pass
try:
if torch.cuda.is_available():
torch.cuda.empty_cache()
torch.cuda.ipc_collect()
except Exception:
pass
try:
time.sleep(0)
except Exception:
pass
if hard_trim:
try:
import comfy.model_management as mm
if hasattr(mm, 'unload_all_models'):
mm.unload_all_models()
except Exception:
pass
try:
for _ in range(2):
time.sleep(0)
gc.collect()
except Exception:
pass
try:
if hasattr(_utils, 'cleanup_lru_caches'):
_utils.cleanup_lru_caches()
except Exception:
pass
try:
_trim_working_set_windows()
psapi = ctypes.windll.psapi # type: ignore[attr-defined]
kernel32 = ctypes.windll.kernel32 # type: ignore[attr-defined]
psapi.EmptyWorkingSet(kernel32.GetCurrentProcess())
except Exception:
pass
try:
if platform.system().lower().startswith('linux'):
libc = ctypes.CDLL('libc.so.6')
libc.malloc_trim(0)
except Exception:
pass
try:
_system_cache_trim_windows()
except Exception:
pass
stats["ram_after_mb"] = _get_ram_mb()
stats["ram_freed_mb"] = max(0.0, stats["ram_before_mb"] - stats["ram_after_mb"])
gpu_after = _get_vram_mb_per_device()
device_map = {d: (r, a) for d, r, a in gpu_before}
gpu_stats = []
for d, r_after, a_after in gpu_after:
r_before, a_before = device_map.get(d, (0.0, 0.0))
gpu_stats.append({
"device": d,
"reserved_before_mb": r_before,
"reserved_after_mb": r_after,
"reserved_freed_mb": max(0.0, r_before - r_after),
"allocated_before_mb": a_before,
"allocated_after_mb": a_after,
"allocated_freed_mb": max(0.0, a_before - a_after),
})
stats["gpu"] = gpu_stats
return stats
class MG_CleanUp:
@classmethod
def INPUT_TYPES(cls):
return {
"required": {
"samples": ("LATENT", {}),
},
"optional": {
"hard_trim": ("BOOLEAN", {"default": True, "tooltip": "Aggressively free RAM/VRAM and ask OS to return pages to the system."}),
"sync_cuda": ("BOOLEAN", {"default": True, "tooltip": "Synchronize CUDA before cleanup to flush pending kernels."}),
"hires_only_threshold": ("INT", {"default": 0, "min": 0, "max": 16384, "step": 64, "tooltip": "Apply only when latent longest side >= threshold (0 == always)."}),
}
}
RETURN_TYPES = ("LATENT", "IMAGE")
RETURN_NAMES = ("samples", "Preview")
FUNCTION = "apply"
CATEGORY = "MagicNodes"
def apply(self, samples, hard_trim=True, sync_cuda=True, hires_only_threshold=0,
model=None, positive=None, negative=None, vae=None):
img_prev = None
try:
if (model is not None) and (positive is not None) and (negative is not None) and (vae is not None):
lat = samples.get("samples", None)
if lat is not None and isinstance(lat, torch.Tensor) and lat.ndim == 4:
z = lat
B, C, H, W = z.shape
target = 32
z_ds = z if (H == target and W == target) else F.interpolate(z, size=(target, target), mode='bilinear', align_corners=False)
lat_img = _sample.fix_empty_latent_channels(model, z_ds) if hasattr(_sample, 'fix_empty_latent_channels') else z_ds
batch_inds = samples.get("batch_index", None)
noise = _sample.prepare_noise(lat_img, int(0), batch_inds)
steps = 1
out = _sample.sample(
model, noise, int(steps), float(1.0), "ddim", "normal",
positive, negative, lat_img,
denoise=float(0.10), disable_noise=False, start_step=None, last_step=None,
force_full_denoise=False, noise_mask=None, callback=None,
disable_pbar=not _utils.PROGRESS_BAR_ENABLED, seed=int(0)
)
try:
img_prev = vae.decode(out)
if len(img_prev.shape) == 5:
img_prev = img_prev.reshape(-1, img_prev.shape[-3], img_prev.shape[-2], img_prev.shape[-1])
except Exception:
img_prev = None
except Exception:
img_prev = None
try:
do_cleanup = True
try:
if int(hires_only_threshold) > 0:
z = samples.get("samples", None)
if z is not None and hasattr(z, "shape") and len(z.shape) >= 4:
_, _, H, W = z.shape
if max(int(H), int(W)) < int(hires_only_threshold):
do_cleanup = False
except Exception:
pass
if do_cleanup:
print("=== CleanUP RAM and GPU ===")
stats = cleanup_memory(sync_cuda=bool(sync_cuda), hard_trim=bool(hard_trim))
try:
print(f"RAM freed: {stats['ram_freed_mb']:.1f} MB (before {stats['ram_before_mb']:.1f} -> after {stats['ram_after_mb']:.1f})")
except Exception:
pass
try:
for g in stats.get("gpu", []):
print(
f"GPU{g['device']}: reserved freed {g['reserved_freed_mb']:.1f} MB ( {g['reserved_before_mb']:.1f} -> {g['reserved_after_mb']:.1f} ), "
f"allocated freed {g['allocated_freed_mb']:.1f} MB ( {g['allocated_before_mb']:.1f} -> {g['allocated_after_mb']:.1f} )"
)
except Exception:
pass
# Second pass after short delay to catch late releasers
try:
time.sleep(0.150)
stats2 = cleanup_memory(sync_cuda=False, hard_trim=bool(hard_trim))
if stats2 and float(stats2.get('ram_freed_mb', 0.0)) > 0.0:
print(f"2nd pass: RAM freed +{stats2['ram_freed_mb']:.1f} MB")
try:
for g in stats2.get('gpu', []):
if float(g.get('reserved_freed_mb', 0.0)) > 0.0 or float(g.get('allocated_freed_mb', 0.0)) > 0.0:
print(f"2nd pass GPU{g['device']}: reserved +{g['reserved_freed_mb']:.1f} MB, allocated +{g['allocated_freed_mb']:.1f} MB")
except Exception:
pass
except Exception:
pass
print("done.")
except Exception:
pass
if img_prev is None:
try:
device = model_management.intermediate_device() if hasattr(model_management, 'intermediate_device') else 'cpu'
img_prev = torch.zeros((1, 32, 32, 3), dtype=torch.float32, device=device)
except Exception:
img_prev = torch.zeros((1, 32, 32, 3))
return (samples, img_prev)