245 lines
9.2 KiB
Python
245 lines
9.2 KiB
Python
import torch
|
|
import comfy.sample
|
|
import comfy.model_management
|
|
import comfy.utils
|
|
import gc
|
|
import logging
|
|
import nodes
|
|
from typing import Dict, Union
|
|
import time
|
|
from contextlib import contextmanager
|
|
import psutil
|
|
|
|
|
|
class MemoryManager:
|
|
"""Manages memory resources for efficient video processing."""
|
|
|
|
def __init__(self, device=None, log_level: str = "INFO"):
|
|
self.logger = logging.getLogger("MemoryManager")
|
|
if not self.logger.handlers:
|
|
handler = logging.StreamHandler()
|
|
formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s')
|
|
handler.setFormatter(formatter)
|
|
self.logger.addHandler(handler)
|
|
self.logger.setLevel(getattr(logging, log_level))
|
|
|
|
self.device = device or comfy.model_management.get_torch_device()
|
|
self.logger.info(f"Using device: {self.device}")
|
|
|
|
# Memory thresholds (percentages)
|
|
self.warning_threshold = 85
|
|
self.critical_threshold = 95
|
|
|
|
def is_cuda_device(self) -> bool:
|
|
"""Check if the current device is a CUDA device."""
|
|
if isinstance(self.device, str):
|
|
return self.device.startswith("cuda")
|
|
elif isinstance(self.device, torch.device):
|
|
return self.device.type == "cuda"
|
|
return False
|
|
|
|
def get_memory_stats(self) -> Dict[str, Union[int, float]]:
|
|
"""Get current memory statistics for the device."""
|
|
stats = {}
|
|
|
|
if self.is_cuda_device() and torch.cuda.is_available():
|
|
try:
|
|
t = torch.cuda.get_device_properties(0)
|
|
stats["total"] = t.total_memory
|
|
stats["reserved"] = torch.cuda.memory_reserved(0)
|
|
stats["allocated"] = torch.cuda.memory_allocated(0)
|
|
stats["free"] = stats["total"] - stats["reserved"]
|
|
stats["usage_percent"] = (stats["allocated"] / stats["total"]) * 100
|
|
except Exception as e:
|
|
self.logger.error(f"Error getting CUDA memory stats: {e}")
|
|
stats = {"error": str(e)}
|
|
else:
|
|
# CPU memory stats
|
|
vm = psutil.virtual_memory()
|
|
stats["total"] = vm.total
|
|
stats["available"] = vm.available
|
|
stats["used"] = vm.used
|
|
stats["free"] = vm.free
|
|
stats["usage_percent"] = vm.percent
|
|
|
|
return stats
|
|
|
|
def is_memory_critical(self) -> bool:
|
|
"""Check if memory usage is at critical levels."""
|
|
stats = self.get_memory_stats()
|
|
if "error" in stats:
|
|
return True # Assume critical if we can't get stats
|
|
|
|
return stats.get("usage_percent", 0) > self.critical_threshold
|
|
|
|
@contextmanager
|
|
def track_memory(self, label: str = "Operation"):
|
|
"""Context manager to track memory usage before and after an operation."""
|
|
if self.is_cuda_device() and torch.cuda.is_available():
|
|
start_mem = torch.cuda.memory_allocated()
|
|
start_time = time.time()
|
|
try:
|
|
yield
|
|
finally:
|
|
end_mem = torch.cuda.memory_allocated()
|
|
end_time = time.time()
|
|
self.logger.info(f"{label} - Memory change: {(end_mem-start_mem)/1024**2:.2f}MB, Time: {end_time-start_time:.2f}s")
|
|
else:
|
|
start_time = time.time()
|
|
try:
|
|
yield
|
|
finally:
|
|
end_time = time.time()
|
|
self.logger.info(f"{label} - Time: {end_time-start_time:.2f}s")
|
|
|
|
def cleanup(self, force: bool = False):
|
|
"""Clean up memory resources."""
|
|
if self.is_cuda_device() and torch.cuda.is_available():
|
|
torch.cuda.empty_cache()
|
|
gc.collect()
|
|
|
|
if force:
|
|
# More aggressive cleanup
|
|
for obj in gc.get_objects():
|
|
try:
|
|
if torch.is_tensor(obj) and not obj.is_cuda:
|
|
del obj
|
|
except:
|
|
pass
|
|
gc.collect()
|
|
if self.is_cuda_device() and torch.cuda.is_available():
|
|
torch.cuda.empty_cache()
|
|
|
|
|
|
class WanVideoKsampler:
|
|
"""
|
|
Video K-sampler node with memory management for processing video latents.
|
|
"""
|
|
|
|
@classmethod
|
|
def INPUT_TYPES(cls):
|
|
return {
|
|
"required": {
|
|
"model": ("MODEL",),
|
|
"positive": ("CONDITIONING",),
|
|
"negative": ("CONDITIONING",),
|
|
"video_latents": ("LATENT",),
|
|
"seed": ("INT", {"default": 0, "min": 0, "max": 0xffffffffffffffff}),
|
|
"steps": ("INT", {"default": 20, "min": 1, "max": 10000}),
|
|
"cfg": ("FLOAT", {"default": 6.0, "min": 0.0, "max": 100.0}),
|
|
"sampler_name": (comfy.samplers.KSampler.SAMPLERS, ),
|
|
"scheduler": (comfy.samplers.KSampler.SCHEDULERS, ),
|
|
"denoise": ("FLOAT", {"default": 1, "min": 0.0, "max": 1.0, "step": 0.01}),
|
|
}
|
|
}
|
|
|
|
RETURN_TYPES = ("LATENT",)
|
|
FUNCTION = "sample"
|
|
CATEGORY = "sampling"
|
|
|
|
def __init__(self):
|
|
self.logger = logging.getLogger("WanVideoKsampler")
|
|
if not self.logger.handlers:
|
|
handler = logging.StreamHandler()
|
|
formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s')
|
|
handler.setFormatter(formatter)
|
|
self.logger.addHandler(handler)
|
|
self.logger.setLevel(logging.INFO)
|
|
|
|
# Initialize memory manager
|
|
self.memory_manager = None
|
|
|
|
def sample(
|
|
self,
|
|
model,
|
|
video_latents: Dict[str, torch.Tensor],
|
|
positive,
|
|
negative,
|
|
seed: int,
|
|
steps: int,
|
|
cfg: float,
|
|
sampler_name: str,
|
|
scheduler: str,
|
|
denoise: float
|
|
) -> Dict[str, torch.Tensor]:
|
|
"""
|
|
Sample video frames with memory management.
|
|
|
|
Args:
|
|
model: Diffusion model
|
|
video_latents: Dictionary containing latent tensors
|
|
positive: Positive conditioning
|
|
negative: Negative conditioning
|
|
seed: Random seed
|
|
steps: Number of sampling steps
|
|
cfg: Classifier-free guidance scale
|
|
sampler_name: Name of sampler to use
|
|
scheduler: Name of scheduler to use
|
|
denoise: Denoising strength
|
|
|
|
Returns:
|
|
Dictionary containing processed latent tensors
|
|
"""
|
|
start_time = time.time()
|
|
device = comfy.model_management.get_torch_device()
|
|
|
|
# Initialize memory manager if needed
|
|
if self.memory_manager is None:
|
|
self.memory_manager = MemoryManager(device)
|
|
|
|
# Log latent size for debugging
|
|
if isinstance(video_latents, dict) and 'samples' in video_latents:
|
|
latent_samples = video_latents['samples']
|
|
total_frames = latent_samples.shape[0]
|
|
self.logger.info(f"Processing latent shape: {latent_samples.shape}, total frames: {total_frames}")
|
|
else:
|
|
self.logger.error("Invalid latent format")
|
|
raise ValueError("Expected latent dictionary with 'samples' key")
|
|
|
|
self.logger.info(f"Processing with {steps} steps, {cfg} CFG, {sampler_name} sampler")
|
|
|
|
try:
|
|
# Process with memory tracking
|
|
with self.memory_manager.track_memory("Video processing"):
|
|
# Check memory usage before processing
|
|
memory_stats = self.memory_manager.get_memory_stats()
|
|
if "usage_percent" in memory_stats:
|
|
self.logger.info(f"Memory usage before processing: {memory_stats['usage_percent']:.1f}%")
|
|
|
|
# Apply sampling
|
|
result = nodes.common_ksampler(
|
|
model, seed, steps, cfg, sampler_name, scheduler,
|
|
positive, negative, video_latents, denoise=denoise
|
|
)
|
|
|
|
# Clear memory after processing
|
|
self.memory_manager.cleanup()
|
|
|
|
# Check memory usage after processing
|
|
memory_stats = self.memory_manager.get_memory_stats()
|
|
if "usage_percent" in memory_stats:
|
|
self.logger.info(f"Memory usage after processing: {memory_stats['usage_percent']:.1f}%")
|
|
|
|
end_time = time.time()
|
|
self.logger.info(f"Complete: {total_frames} frames in {end_time - start_time:.2f}s ({(end_time - start_time) / total_frames:.2f}s per frame)")
|
|
|
|
return result
|
|
|
|
except Exception as e:
|
|
self.logger.error(f"Error during processing: {str(e)}")
|
|
# Try to release memory
|
|
self.memory_manager.cleanup(force=True)
|
|
# Check if it's an out-of-memory error
|
|
if "CUDA out of memory" in str(e):
|
|
self.logger.error("Out of memory error. Consider reducing frame count or model complexity.")
|
|
raise e
|
|
|
|
|
|
# Node registration
|
|
NODE_CLASS_MAPPINGS = {
|
|
"WanVideoKsampler": WanVideoKsampler,
|
|
}
|
|
|
|
NODE_DISPLAY_NAME_MAPPINGS = {
|
|
"WanVideoKsampler": "Wan Video Ksampler",
|
|
} |