from collections import OrderedDict, namedtuple from functools import wraps import os import torch MAX_CHECKPOINT_CACHE = int(os.environ.get("MAX_CHECKPOINT_CACHE", 1)) MAX_CONTROLNET_CACHE = int(os.environ.get("MAX_CONTROLNET_CACHE", 1)) MAX_VAE_CACHE = int(os.environ.get("MAX_VAE_CACHE", 1)) class Cache: def __init__(self, max_cache_size=0): self._max_size = max_cache_size self._cache = OrderedDict() @property def cache(self): return self._cache @property def size(self): return self._max_size def clear_cache(self): while len(self._cache) > self.size: k, _ = self._cache.popitem(last=False) print(f"[==Cache==] Removed {k} from cache.") def add_cache(self, k, item): """model patch change the key name of tensor, cache the names here""" keys = list(item.keys()) values = list(item.values()) self._cache[k] = (keys, values) self._cache.move_to_end(k) def get_cache(self, k): data = self._cache.get(k) if data: keys, values = data return dict(zip(keys, values)) return None caches_mapping = { "checkpoints": Cache(max_cache_size=MAX_CHECKPOINT_CACHE), "controlnet": Cache(max_cache_size=MAX_CONTROLNET_CACHE), "vae": Cache(max_cache_size=MAX_VAE_CACHE), } def checkpoint_cache_decorator(func): @wraps(func) def wrapper(ckpt, *args, device=None, **kwargs): model_dir, model_name = os.path.split(ckpt) dir_name = os.path.basename(model_dir) # get matched ckpt, vae, controlnet queue cache_mapper = caches_mapping.get(dir_name, None) if cache_mapper is not None: if cache_mapper.get_cache(ckpt) is not None: return cache_mapper.get_cache(ckpt) sd = func(ckpt, *args, device=device, **kwargs) if cache_mapper is not None: cache_mapper.add_cache(ckpt, sd) print(f"[==Cache==] Add {dir_name} {model_name} to cache") cache_mapper.clear_cache() return sd return wrapper