提交文件到 wan2.1-t2v-14b 项目

comfy/cldm/dit_embedder.py

+import math
+from typing import List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from torch import Tensor
+
+from comfy.ldm.modules.diffusionmodules.mmdit import DismantledBlock, PatchEmbed, VectorEmbedder, TimestepEmbedder, get_2d_sincos_pos_embed_torch
+
+
+class ControlNetEmbedder(nn.Module):
+
+    def __init__(
+        self,
+        img_size: int,
+        patch_size: int,
+        in_chans: int,
+        attention_head_dim: int,
+        num_attention_heads: int,
+        adm_in_channels: int,
+        num_layers: int,
+        main_model_double: int,
+        double_y_emb: bool,
+        device: torch.device,
+        dtype: torch.dtype,
+        pos_embed_max_size: Optional[int] = None,
+        operations = None,
+    ):
+        super().__init__()
+        self.main_model_double = main_model_double
+        self.dtype = dtype
+        self.hidden_size = num_attention_heads * attention_head_dim
+        self.patch_size = patch_size
+        self.x_embedder = PatchEmbed(
+            img_size=img_size,
+            patch_size=patch_size,
+            in_chans=in_chans,
+            embed_dim=self.hidden_size,
+            strict_img_size=pos_embed_max_size is None,
+            device=device,
+            dtype=dtype,
+            operations=operations,
+        )
+
+        self.t_embedder = TimestepEmbedder(self.hidden_size, dtype=dtype, device=device, operations=operations)
+
+        self.double_y_emb = double_y_emb
+        if self.double_y_emb:
+            self.orig_y_embedder = VectorEmbedder(
+                adm_in_channels, self.hidden_size, dtype, device, operations=operations
+            )
+            self.y_embedder = VectorEmbedder(
+                self.hidden_size, self.hidden_size, dtype, device, operations=operations
+            )
+        else:
+            self.y_embedder = VectorEmbedder(
+                adm_in_channels, self.hidden_size, dtype, device, operations=operations
+            )
+
+        self.transformer_blocks = nn.ModuleList(
+            DismantledBlock(
+                hidden_size=self.hidden_size, num_heads=num_attention_heads, qkv_bias=True,
+                dtype=dtype, device=device, operations=operations
+            )
+            for _ in range(num_layers)
+        )
+
+        # self.use_y_embedder = pooled_projection_dim != self.time_text_embed.text_embedder.linear_1.in_features
+        # TODO double check this logic when 8b
+        self.use_y_embedder = True
+
+        self.controlnet_blocks = nn.ModuleList([])
+        for _ in range(len(self.transformer_blocks)):
+            controlnet_block = operations.Linear(self.hidden_size, self.hidden_size, dtype=dtype, device=device)
+            self.controlnet_blocks.append(controlnet_block)
+
+        self.pos_embed_input = PatchEmbed(
+            img_size=img_size,
+            patch_size=patch_size,
+            in_chans=in_chans,
+            embed_dim=self.hidden_size,
+            strict_img_size=False,
+            device=device,
+            dtype=dtype,
+            operations=operations,
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        timesteps: torch.Tensor,
+        y: Optional[torch.Tensor] = None,
+        context: Optional[torch.Tensor] = None,
+        hint = None,
+    ) -> Tuple[Tensor, List[Tensor]]:
+        x_shape = list(x.shape)
+        x = self.x_embedder(x)
+        if not self.double_y_emb:
+            h = (x_shape[-2] + 1) // self.patch_size
+            w = (x_shape[-1] + 1) // self.patch_size
+            x += get_2d_sincos_pos_embed_torch(self.hidden_size, w, h, device=x.device)
+        c = self.t_embedder(timesteps, dtype=x.dtype)
+        if y is not None and self.y_embedder is not None:
+            if self.double_y_emb:
+                y = self.orig_y_embedder(y)
+            y = self.y_embedder(y)
+            c = c + y
+
+        x = x + self.pos_embed_input(hint)
+
+        block_out = ()
+
+        repeat = math.ceil(self.main_model_double / len(self.transformer_blocks))
+        for i in range(len(self.transformer_blocks)):
+            out = self.transformer_blocks[i](x, c)
+            if not self.double_y_emb:
+                x = out
+            block_out += (self.controlnet_blocks[i](out),) * repeat
+
+        return {"output": block_out}

comfy/cldm/mmdit.py

+import torch
+from typing import Optional
+import comfy.ldm.modules.diffusionmodules.mmdit
+
+class ControlNet(comfy.ldm.modules.diffusionmodules.mmdit.MMDiT):
+    def __init__(
+        self,
+        num_blocks = None,
+        control_latent_channels = None,
+        dtype = None,
+        device = None,
+        operations = None,
+        **kwargs,
+    ):
+        super().__init__(dtype=dtype, device=device, operations=operations, final_layer=False, num_blocks=num_blocks, **kwargs)
+        # controlnet_blocks
+        self.controlnet_blocks = torch.nn.ModuleList([])
+        for _ in range(len(self.joint_blocks)):
+            self.controlnet_blocks.append(operations.Linear(self.hidden_size, self.hidden_size, device=device, dtype=dtype))
+
+        if control_latent_channels is None:
+            control_latent_channels = self.in_channels
+
+        self.pos_embed_input = comfy.ldm.modules.diffusionmodules.mmdit.PatchEmbed(
+            None,
+            self.patch_size,
+            control_latent_channels,
+            self.hidden_size,
+            bias=True,
+            strict_img_size=False,
+            dtype=dtype,
+            device=device,
+            operations=operations
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        timesteps: torch.Tensor,
+        y: Optional[torch.Tensor] = None,
+        context: Optional[torch.Tensor] = None,
+        hint = None,
+    ) -> torch.Tensor:
+
+        #weird sd3 controlnet specific stuff
+        y = torch.zeros_like(y)
+
+        if self.context_processor is not None:
+            context = self.context_processor(context)
+
+        hw = x.shape[-2:]
+        x = self.x_embedder(x) + self.cropped_pos_embed(hw, device=x.device).to(dtype=x.dtype, device=x.device)
+        x += self.pos_embed_input(hint)
+
+        c = self.t_embedder(timesteps, dtype=x.dtype)
+        if y is not None and self.y_embedder is not None:
+            y = self.y_embedder(y)
+            c = c + y
+
+        if context is not None:
+            context = self.context_embedder(context)
+
+        output = []
+
+        blocks = len(self.joint_blocks)
+        for i in range(blocks):
+            context, x = self.joint_blocks[i](
+                context,
+                x,
+                c=c,
+                use_checkpoint=self.use_checkpoint,
+            )
+
+            out = self.controlnet_blocks[i](x)
+            count = self.depth // blocks
+            if i == blocks - 1:
+                count -= 1
+            for j in range(count):
+                output.append(out)
+
+        return {"output": output}

comfy/clip_config_bigg.json

+{
+  "architectures": [
+    "CLIPTextModel"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 0,
+  "dropout": 0.0,
+  "eos_token_id": 49407,
+  "hidden_act": "gelu",
+  "hidden_size": 1280,
+  "initializer_factor": 1.0,
+  "initializer_range": 0.02,
+  "intermediate_size": 5120,
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 77,
+  "model_type": "clip_text_model",
+  "num_attention_heads": 20,
+  "num_hidden_layers": 32,
+  "pad_token_id": 1,
+  "projection_dim": 1280,
+  "torch_dtype": "float32",
+  "vocab_size": 49408
+}

comfy/clip_vision.py

+from .utils import load_torch_file, transformers_convert, state_dict_prefix_replace
+import os
+import torch
+import json
+import logging
+
+import comfy.ops
+import comfy.model_patcher
+import comfy.model_management
+import comfy.utils
+import comfy.clip_model
+import comfy.image_encoders.dino2
+
+class Output:
+    def __getitem__(self, key):
+        return getattr(self, key)
+    def __setitem__(self, key, item):
+        setattr(self, key, item)
+
+def clip_preprocess(image, size=224, mean=[0.48145466, 0.4578275, 0.40821073], std=[0.26862954, 0.26130258, 0.27577711], crop=True):
+    image = image[:, :, :, :3] if image.shape[3] > 3 else image
+    mean = torch.tensor(mean, device=image.device, dtype=image.dtype)
+    std = torch.tensor(std, device=image.device, dtype=image.dtype)
+    image = image.movedim(-1, 1)
+    if not (image.shape[2] == size and image.shape[3] == size):
+        if crop:
+            scale = (size / min(image.shape[2], image.shape[3]))
+            scale_size = (round(scale * image.shape[2]), round(scale * image.shape[3]))
+        else:
+            scale_size = (size, size)
+
+        image = torch.nn.functional.interpolate(image, size=scale_size, mode="bicubic", antialias=True)
+        h = (image.shape[2] - size)//2
+        w = (image.shape[3] - size)//2
+        image = image[:,:,h:h+size,w:w+size]
+    image = torch.clip((255. * image), 0, 255).round() / 255.0
+    return (image - mean.view([3,1,1])) / std.view([3,1,1])
+
+IMAGE_ENCODERS = {
+    "clip_vision_model": comfy.clip_model.CLIPVisionModelProjection,
+    "siglip_vision_model": comfy.clip_model.CLIPVisionModelProjection,
+    "dinov2": comfy.image_encoders.dino2.Dinov2Model,
+}
+
+class ClipVisionModel():
+    def __init__(self, json_config):
+        with open(json_config) as f:
+            config = json.load(f)
+
+        self.image_size = config.get("image_size", 224)
+        self.image_mean = config.get("image_mean", [0.48145466, 0.4578275, 0.40821073])
+        self.image_std = config.get("image_std", [0.26862954, 0.26130258, 0.27577711])
+        model_class = IMAGE_ENCODERS.get(config.get("model_type", "clip_vision_model"))
+        self.load_device = comfy.model_management.text_encoder_device()
+        offload_device = comfy.model_management.text_encoder_offload_device()
+        self.dtype = comfy.model_management.text_encoder_dtype(self.load_device)
+        self.model = model_class(config, self.dtype, offload_device, comfy.ops.manual_cast)
+        self.model.eval()
+
+        self.patcher = comfy.model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
+
+    def load_sd(self, sd):
+        return self.model.load_state_dict(sd, strict=False)
+
+    def get_sd(self):
+        return self.model.state_dict()
+
+    def encode_image(self, image, crop=True):
+        comfy.model_management.load_model_gpu(self.patcher)
+        pixel_values = clip_preprocess(image.to(self.load_device), size=self.image_size, mean=self.image_mean, std=self.image_std, crop=crop).float()
+        out = self.model(pixel_values=pixel_values, intermediate_output=-2)
+
+        outputs = Output()
+        outputs["last_hidden_state"] = out[0].to(comfy.model_management.intermediate_device())
+        outputs["image_embeds"] = out[2].to(comfy.model_management.intermediate_device())
+        outputs["penultimate_hidden_states"] = out[1].to(comfy.model_management.intermediate_device())
+        outputs["mm_projected"] = out[3]
+        return outputs
+
+def convert_to_transformers(sd, prefix):
+    sd_k = sd.keys()
+    if "{}transformer.resblocks.0.attn.in_proj_weight".format(prefix) in sd_k:
+        keys_to_replace = {
+            "{}class_embedding".format(prefix): "vision_model.embeddings.class_embedding",
+            "{}conv1.weight".format(prefix): "vision_model.embeddings.patch_embedding.weight",
+            "{}positional_embedding".format(prefix): "vision_model.embeddings.position_embedding.weight",
+            "{}ln_post.bias".format(prefix): "vision_model.post_layernorm.bias",
+            "{}ln_post.weight".format(prefix): "vision_model.post_layernorm.weight",
+            "{}ln_pre.bias".format(prefix): "vision_model.pre_layrnorm.bias",
+            "{}ln_pre.weight".format(prefix): "vision_model.pre_layrnorm.weight",
+        }
+
+        for x in keys_to_replace:
+            if x in sd_k:
+                sd[keys_to_replace[x]] = sd.pop(x)
+
+        if "{}proj".format(prefix) in sd_k:
+            sd['visual_projection.weight'] = sd.pop("{}proj".format(prefix)).transpose(0, 1)
+
+        sd = transformers_convert(sd, prefix, "vision_model.", 48)
+    else:
+        replace_prefix = {prefix: ""}
+        sd = state_dict_prefix_replace(sd, replace_prefix)
+    return sd
+
+def load_clipvision_from_sd(sd, prefix="", convert_keys=False):
+    if convert_keys:
+        sd = convert_to_transformers(sd, prefix)
+    if "vision_model.encoder.layers.47.layer_norm1.weight" in sd:
+        json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_g.json")
+    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:
+            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:
+                json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_vitl_336.json")
+        else:
+            json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_vitl.json")
+    elif "embeddings.patch_embeddings.projection.weight" in sd:
+        json_config = os.path.join(os.path.join(os.path.dirname(os.path.realpath(__file__)), "image_encoders"), "dino2_giant.json")
+    else:
+        return None
+
+    clip = ClipVisionModel(json_config)
+    m, u = clip.load_sd(sd)
+    if len(m) > 0:
+        logging.warning("missing clip vision: {}".format(m))
+    u = set(u)
+    keys = list(sd.keys())
+    for k in keys:
+        if k not in u:
+            sd.pop(k)
+    return clip
+
+def load(ckpt_path):
+    sd = load_torch_file(ckpt_path)
+    if "visual.transformer.resblocks.0.attn.in_proj_weight" in sd:
+        return load_clipvision_from_sd(sd, prefix="visual.", convert_keys=True)
+    else:
+        return load_clipvision_from_sd(sd)

comfy/clip_vision_config_g.json

+{
+  "attention_dropout": 0.0,
+  "dropout": 0.0,
+  "hidden_act": "gelu",
+  "hidden_size": 1664,
+  "image_size": 224,
+  "initializer_factor": 1.0,
+  "initializer_range": 0.02,
+  "intermediate_size": 8192,
+  "layer_norm_eps": 1e-05,
+  "model_type": "clip_vision_model",
+  "num_attention_heads": 16,
+  "num_channels": 3,
+  "num_hidden_layers": 48,
+  "patch_size": 14,
+  "projection_dim": 1280,
+  "torch_dtype": "float32"
+}

comfy/clip_vision_config_h.json

+{
+  "attention_dropout": 0.0,
+  "dropout": 0.0,
+  "hidden_act": "gelu",
+  "hidden_size": 1280,
+  "image_size": 224,
+  "initializer_factor": 1.0,
+  "initializer_range": 0.02,
+  "intermediate_size": 5120,
+  "layer_norm_eps": 1e-05,
+  "model_type": "clip_vision_model",
+  "num_attention_heads": 16,
+  "num_channels": 3,
+  "num_hidden_layers": 32,
+  "patch_size": 14,
+  "projection_dim": 1024,
+  "torch_dtype": "float32"
+}

comfy/clip_vision_config_vitl.json

+{
+  "attention_dropout": 0.0,
+  "dropout": 0.0,
+  "hidden_act": "quick_gelu",
+  "hidden_size": 1024,
+  "image_size": 224,
+  "initializer_factor": 1.0,
+  "initializer_range": 0.02,
+  "intermediate_size": 4096,
+  "layer_norm_eps": 1e-05,
+  "model_type": "clip_vision_model",
+  "num_attention_heads": 16,
+  "num_channels": 3,
+  "num_hidden_layers": 24,
+  "patch_size": 14,
+  "projection_dim": 768,
+  "torch_dtype": "float32"
+}