Meisoftcoltd
1c90c848e2
- Enforced a strict memory cleanup sequence in `MeisoftVRAMDefragmenter` (including IPC cleanup and CUDA synchronization). - Injected `PYTORCH_CUDA_ALLOC_CONF="max_split_size_mb:128"` in `__init__.py` to prevent VRAM micro-fragmentation. - Added `unload_detector` option to `VideoAnalyzerWithAudio` to eagerly destroy the face detector model and free VRAM.
54 lines
1.6 KiB
Python
54 lines
1.6 KiB
Python
import torch
|
|
import gc
|
|
import comfy.model_management as mm
|
|
from . import register_node
|
|
|
|
class AnyType(str):
|
|
def __ne__(self, __value: object) -> bool:
|
|
return False
|
|
|
|
any_type = AnyType("*")
|
|
|
|
@register_node
|
|
class MeisoftVRAMDefragmenter:
|
|
@classmethod
|
|
def INPUT_TYPES(s):
|
|
return {
|
|
"required": {
|
|
"anything": (any_type,),
|
|
"force_model_unload": ("BOOLEAN", {"default": False}),
|
|
}
|
|
}
|
|
|
|
RETURN_TYPES = (any_type,)
|
|
RETURN_NAMES = ("anything",)
|
|
FUNCTION = "defragment"
|
|
CATEGORY = "🔁 Sequential Batcher/Tools"
|
|
|
|
def defragment(self, anything, force_model_unload):
|
|
print(f"\n{'='*50}")
|
|
print(f"🧹 [DEBUG] NODO: VRAM Defragmenter (Pasarela)")
|
|
|
|
# Fase 1: Romper ciclos de referencia en Python
|
|
gc.collect()
|
|
|
|
# Fase 2: Evacuación suave nativa
|
|
mm.soft_empty_cache()
|
|
|
|
# Fase 3: Evacuación agresiva (Opcional)
|
|
if force_model_unload:
|
|
print(" -> 📦 Trasladando modelos inactivos a la RAM (unload_all_models)...")
|
|
mm.unload_all_models()
|
|
gc.collect() # Segunda barrida tras rotura de enlaces
|
|
|
|
# Fases 4, 5 y 6: Limpieza IPC, Vaciado CUDA y Sincronización
|
|
if torch.cuda.is_available():
|
|
print(" -> 🧱 Purgando memoria IPC y desfragmentando Caching Allocator...")
|
|
torch.cuda.ipc_collect()
|
|
torch.cuda.empty_cache()
|
|
torch.cuda.synchronize() # Barrera contra race conditions
|
|
|
|
print(" -> ✨ VRAM hiper-optimizada. Permitiendo paso al Sampler.")
|
|
print(f"{'='*50}\n")
|
|
return (anything,)
|