The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
EasyCache: Fix for mismatch in input/output channels with some models (#10788)
Jukka Seppänene1ab6bb394b8
diff --git a/comfy_extras/nodes_easycache.py b/comfy_extras/nodes_easycache.pyindex 1359e2f99..11b23ffdb 100644--- a/comfy_extras/nodes_easycache.py+++ b/comfy_extras/nodes_easycache.py@@ -11,13 +11,13 @@ if TYPE_CHECKING:def easycache_forward_wrapper(executor, *args, **kwargs):# get values from args- x: torch.Tensor = args[0]transformer_options: dict[str] = args[-1]if not isinstance(transformer_options, dict):transformer_options = kwargs.get("transformer_options")if not transformer_options:transformer_options = args[-2]easycache: EasyCacheHolder = transformer_options["easycache"]+ x: torch.Tensor = args[0][:, :easycache.output_channels]sigmas = transformer_options["sigmas"]uuids = transformer_options["uuids"]if sigmas is not None and easycache.is_past_end_timestep(sigmas):@@ -82,13 +82,13 @@ def easycache_forward_wrapper(executor, *args, **kwargs):def lazycache_predict_noise_wrapper(executor, *args, **kwargs):# get values from args- x: torch.Tensor = args[0]timestep: float = args[1]model_options: dict[str] = args[2]easycache: LazyCacheHolder = model_options["transformer_options"]["easycache"]if easycache.is_past_end_timestep(timestep):return executor(*args, **kwargs)# prepare next x_prev+ x: torch.Tensor = args[0][:, :easycache.output_channels]next_x_prev = xinput_change = Nonedo_easycache = easycache.should_do_easycache(timestep)@@ -173,7 +173,7 @@ def easycache_sample_wrapper(executor, *args, **kwargs):class EasyCacheHolder:- def __init__(self, reuse_threshold: float, start_percent: float, end_percent: float, subsample_factor: int, offload_cache_diff: bool, verbose: bool=False):+ def __init__(self, reuse_threshold: float, start_percent: float, end_percent: float, subsample_factor: int, offload_cache_diff: bool, verbose: bool=False, output_channels: int=None):self.name = "EasyCache"self.reuse_threshold = reuse_thresholdself.start_percent = start_percent@@ -202,6 +202,7 @@ class EasyCacheHolder:self.allow_mismatch = Trueself.cut_from_start = Trueself.state_metadata = None+ self.output_channels = output_channelsdef is_past_end_timestep(self, timestep: float) -> bool:return not (timestep[0] > self.end_t).item()@@ -264,7 +265,7 @@ class EasyCacheHolder:else:slicing.append(slice(None))batch_slice = batch_slice + slicing- x[batch_slice] += self.uuid_cache_diffs[uuid].to(x.device)+ x[tuple(batch_slice)] += self.uuid_cache_diffs[uuid].to(x.device)return xdef update_cache_diff(self, output: torch.Tensor, x: torch.Tensor, uuids: list[UUID]):@@ -283,7 +284,7 @@ class EasyCacheHolder:else:slicing.append(slice(None))skip_dim = False- x = x[slicing]+ x = x[tuple(slicing)]diff = output - xbatch_offset = diff.shape[0] // len(uuids)for i, uuid in enumerate(uuids):@@ -323,7 +324,7 @@ class EasyCacheHolder:return selfdef clone(self):- return EasyCacheHolder(self.reuse_threshold, self.start_percent, self.end_percent, self.subsample_factor, self.offload_cache_diff, self.verbose)+ return EasyCacheHolder(self.reuse_threshold, self.start_percent, self.end_percent, self.subsample_factor, self.offload_cache_diff, self.verbose, output_channels=self.output_channels)class EasyCacheNode(io.ComfyNode):@@ -350,7 +351,7 @@ class EasyCacheNode(io.ComfyNode):@classmethoddef execute(cls, model: io.Model.Type, reuse_threshold: float, start_percent: float, end_percent: float, verbose: bool) -> io.NodeOutput:model = model.clone()- model.model_options["transformer_options"]["easycache"] = EasyCacheHolder(reuse_threshold, start_percent, end_percent, subsample_factor=8, offload_cache_diff=False, verbose=verbose)+ model.model_options["transformer_options"]["easycache"] = EasyCacheHolder(reuse_threshold, start_percent, end_percent, subsample_factor=8, offload_cache_diff=False, verbose=verbose, output_channels=model.model.latent_format.latent_channels)model.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.OUTER_SAMPLE, "easycache", easycache_sample_wrapper)model.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.CALC_COND_BATCH, "easycache", easycache_calc_cond_batch_wrapper)model.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, "easycache", easycache_forward_wrapper)@@ -358,7 +359,7 @@ class EasyCacheNode(io.ComfyNode):class LazyCacheHolder:- def __init__(self, reuse_threshold: float, start_percent: float, end_percent: float, subsample_factor: int, offload_cache_diff: bool, verbose: bool=False):+ def __init__(self, reuse_threshold: float, start_percent: float, end_percent: float, subsample_factor: int, offload_cache_diff: bool, verbose: bool=False, output_channels: int=None):self.name = "LazyCache"self.reuse_threshold = reuse_thresholdself.start_percent = start_percent@@ -382,6 +383,7 @@ class LazyCacheHolder:self.approx_output_change_rates = []self.total_steps_skipped = 0self.state_metadata = None+ self.output_channels = output_channelsdef has_cache_diff(self) -> bool:return self.cache_diff is not None@@ -456,7 +458,7 @@ class LazyCacheHolder:return selfdef clone(self):- return LazyCacheHolder(self.reuse_threshold, self.start_percent, self.end_percent, self.subsample_factor, self.offload_cache_diff, self.verbose)+ return LazyCacheHolder(self.reuse_threshold, self.start_percent, self.end_percent, self.subsample_factor, self.offload_cache_diff, self.verbose, output_channels=self.output_channels)class LazyCacheNode(io.ComfyNode):@classmethod@@ -482,7 +484,7 @@ class LazyCacheNode(io.ComfyNode):@classmethoddef execute(cls, model: io.Model.Type, reuse_threshold: float, start_percent: float, end_percent: float, verbose: bool) -> io.NodeOutput:model = model.clone()- model.model_options["transformer_options"]["easycache"] = LazyCacheHolder(reuse_threshold, start_percent, end_percent, subsample_factor=8, offload_cache_diff=False, verbose=verbose)+ model.model_options["transformer_options"]["easycache"] = LazyCacheHolder(reuse_threshold, start_percent, end_percent, subsample_factor=8, offload_cache_diff=False, verbose=verbose, output_channels=model.model.latent_format.latent_channels)model.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.OUTER_SAMPLE, "lazycache", easycache_sample_wrapper)model.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.PREDICT_NOISE, "lazycache", lazycache_predict_noise_wrapper)return io.NodeOutput(model)