diff --git a/comfy/clip_vision.py b/comfy/clip_vision.py index 87d32a66e..11bc57789 100644 --- a/comfy/clip_vision.py +++ b/comfy/clip_vision.py @@ -110,9 +110,13 @@ def load_clipvision_from_sd(sd, prefix="", convert_keys=False): elif "vision_model.encoder.layers.30.layer_norm1.weight" in sd: json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_h.json") elif "vision_model.encoder.layers.22.layer_norm1.weight" in sd: + embed_shape = sd["vision_model.embeddings.position_embedding.weight"].shape[0] if sd["vision_model.encoder.layers.0.layer_norm1.weight"].shape[0] == 1152: - json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip_384.json") - elif sd["vision_model.embeddings.position_embedding.weight"].shape[0] == 577: + if embed_shape == 729: + json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip_384.json") + elif embed_shape == 1024: + json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip_512.json") + elif embed_shape == 577: if "multi_modal_projector.linear_1.bias" in sd: json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_vitl_336_llava.json") else: diff --git a/comfy/clip_vision_siglip_512.json b/comfy/clip_vision_siglip_512.json new file mode 100644 index 000000000..7fb93ce15 --- /dev/null +++ b/comfy/clip_vision_siglip_512.json @@ -0,0 +1,13 @@ +{ + "num_channels": 3, + "hidden_act": "gelu_pytorch_tanh", + "hidden_size": 1152, + "image_size": 512, + "intermediate_size": 4304, + "model_type": "siglip_vision_model", + "num_attention_heads": 16, + "num_hidden_layers": 27, + "patch_size": 16, + "image_mean": [0.5, 0.5, 0.5], + "image_std": [0.5, 0.5, 0.5] +} diff --git a/comfy/ldm/modules/attention.py b/comfy/ldm/modules/attention.py index ede506463..45f9e311e 100644 --- a/comfy/ldm/modules/attention.py +++ b/comfy/ldm/modules/attention.py @@ -847,6 +847,7 @@ class SpatialTransformer(nn.Module): if not isinstance(context, list): context = [context] * len(self.transformer_blocks) b, c, h, w = x.shape + transformer_options["activations_shape"] = list(x.shape) x_in = x x = self.norm(x) if not self.use_linear: @@ -962,6 +963,7 @@ class SpatialVideoTransformer(SpatialTransformer): transformer_options={} ) -> torch.Tensor: _, _, h, w = x.shape + transformer_options["activations_shape"] = list(x.shape) x_in = x spatial_context = None if exists(context): diff --git a/comfy/model_management.py b/comfy/model_management.py index 84a260fc4..19e6c8dff 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -1237,6 +1237,8 @@ def soft_empty_cache(force=False): torch.xpu.empty_cache() elif is_ascend_npu(): torch.npu.empty_cache() + elif is_mlu(): + torch.mlu.empty_cache() elif torch.cuda.is_available(): torch.cuda.empty_cache() torch.cuda.ipc_collect() diff --git a/comfy/sd.py b/comfy/sd.py index d096f496c..4d3aef3e1 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -265,6 +265,7 @@ class VAE: self.process_input = lambda image: image * 2.0 - 1.0 self.process_output = lambda image: torch.clamp((image + 1.0) / 2.0, min=0.0, max=1.0) self.working_dtypes = [torch.bfloat16, torch.float32] + self.disable_offload = False self.downscale_index_formula = None self.upscale_index_formula = None @@ -337,6 +338,7 @@ class VAE: self.process_output = lambda audio: audio self.process_input = lambda audio: audio self.working_dtypes = [torch.float16, torch.bfloat16, torch.float32] + self.disable_offload = True elif "blocks.2.blocks.3.stack.5.weight" in sd or "decoder.blocks.2.blocks.3.stack.5.weight" in sd or "layers.4.layers.1.attn_block.attn.qkv.weight" in sd or "encoder.layers.4.layers.1.attn_block.attn.qkv.weight" in sd: #genmo mochi vae if "blocks.2.blocks.3.stack.5.weight" in sd: sd = comfy.utils.state_dict_prefix_replace(sd, {"": "decoder."}) @@ -515,7 +517,7 @@ class VAE: pixel_samples = None try: memory_used = self.memory_used_decode(samples_in.shape, self.vae_dtype) - model_management.load_models_gpu([self.patcher], memory_required=memory_used) + model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload) free_memory = model_management.get_free_memory(self.device) batch_number = int(free_memory / memory_used) batch_number = max(1, batch_number) @@ -544,7 +546,7 @@ class VAE: def decode_tiled(self, samples, tile_x=None, tile_y=None, overlap=None, tile_t=None, overlap_t=None): self.throw_exception_if_invalid() memory_used = self.memory_used_decode(samples.shape, self.vae_dtype) #TODO: calculate mem required for tile - model_management.load_models_gpu([self.patcher], memory_required=memory_used) + model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload) dims = samples.ndim - 2 args = {} if tile_x is not None: @@ -578,7 +580,7 @@ class VAE: pixel_samples = pixel_samples.movedim(1, 0).unsqueeze(0) try: memory_used = self.memory_used_encode(pixel_samples.shape, self.vae_dtype) - model_management.load_models_gpu([self.patcher], memory_required=memory_used) + model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload) free_memory = model_management.get_free_memory(self.device) batch_number = int(free_memory / max(1, memory_used)) batch_number = max(1, batch_number) @@ -612,7 +614,7 @@ class VAE: pixel_samples = pixel_samples.movedim(1, 0).unsqueeze(0) memory_used = self.memory_used_encode(pixel_samples.shape, self.vae_dtype) # TODO: calculate mem required for tile - model_management.load_models_gpu([self.patcher], memory_required=memory_used) + model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload) args = {} if tile_x is not None: diff --git a/comfy_extras/nodes_mask.py b/comfy_extras/nodes_mask.py index e1f0c8225..13d2b4bab 100644 --- a/comfy_extras/nodes_mask.py +++ b/comfy_extras/nodes_mask.py @@ -2,6 +2,7 @@ import numpy as np import scipy.ndimage import torch import comfy.utils +import node_helpers from nodes import MAX_RESOLUTION @@ -87,11 +88,7 @@ class ImageCompositeMasked: CATEGORY = "image" def composite(self, destination, source, x, y, resize_source, mask = None): - if destination.shape[-1] < source.shape[-1]: - source = source[...,:destination.shape[-1]] - elif destination.shape[-1] > source.shape[-1]: - destination = torch.nn.functional.pad(destination, (0, 1)) - destination[..., -1] = source[..., -1] + destination, source = node_helpers.image_alpha_fix(destination, source) destination = destination.clone().movedim(-1, 1) output = composite(destination, source.movedim(-1, 1), x, y, mask, 1, resize_source).movedim(1, -1) return (output,) diff --git a/comfy_extras/nodes_post_processing.py b/comfy_extras/nodes_post_processing.py index 68f6ef51e..5b9542015 100644 --- a/comfy_extras/nodes_post_processing.py +++ b/comfy_extras/nodes_post_processing.py @@ -6,7 +6,7 @@ import math import comfy.utils import comfy.model_management - +import node_helpers class Blend: def __init__(self): @@ -34,6 +34,7 @@ class Blend: CATEGORY = "image/postprocessing" def blend_images(self, image1: torch.Tensor, image2: torch.Tensor, blend_factor: float, blend_mode: str): + image1, image2 = node_helpers.image_alpha_fix(image1, image2) image2 = image2.to(image1.device) if image1.shape != image2.shape: image2 = image2.permute(0, 3, 1, 2) diff --git a/node_helpers.py b/node_helpers.py index 48da3b099..c3e1a14ca 100644 --- a/node_helpers.py +++ b/node_helpers.py @@ -44,3 +44,11 @@ def string_to_torch_dtype(string): return torch.float16 if string == "bf16": return torch.bfloat16 + +def image_alpha_fix(destination, source): + if destination.shape[-1] < source.shape[-1]: + source = source[...,:destination.shape[-1]] + elif destination.shape[-1] > source.shape[-1]: + destination = torch.nn.functional.pad(destination, (0, 1)) + destination[..., -1] = 1.0 + return destination, source