Merge branch 'master' into dr-support-pip-cm

comfy/ldm/chroma/layers.py

@@ -0,0 +1,183 @@
+import torch
+from torch import Tensor, nn
+
+from comfy.ldm.flux.math import attention
+from comfy.ldm.flux.layers import (
+    MLPEmbedder,
+    RMSNorm,
+    QKNorm,
+    SelfAttention,
+    ModulationOut,
+)
+
+
+
+class ChromaModulationOut(ModulationOut):
+    @classmethod
+    def from_offset(cls, tensor: torch.Tensor, offset: int = 0) -> ModulationOut:
+        return cls(
+            shift=tensor[:, offset : offset + 1, :],
+            scale=tensor[:, offset + 1 : offset + 2, :],
+            gate=tensor[:, offset + 2 : offset + 3, :],
+        )
+
+
+
+
+class Approximator(nn.Module):
+    def __init__(self, in_dim: int, out_dim: int, hidden_dim: int, n_layers = 5, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.in_proj = operations.Linear(in_dim, hidden_dim, bias=True, dtype=dtype, device=device)
+        self.layers = nn.ModuleList([MLPEmbedder(hidden_dim, hidden_dim, dtype=dtype, device=device, operations=operations) for x in range( n_layers)])
+        self.norms = nn.ModuleList([RMSNorm(hidden_dim, dtype=dtype, device=device, operations=operations) for x in range( n_layers)])
+        self.out_proj = operations.Linear(hidden_dim, out_dim, dtype=dtype, device=device)
+
+    @property
+    def device(self):
+        # Get the device of the module (assumes all parameters are on the same device)
+        return next(self.parameters()).device
+
+    def forward(self, x: Tensor) -> Tensor:
+        x = self.in_proj(x)
+
+        for layer, norms in zip(self.layers, self.norms):
+            x = x + layer(norms(x))
+
+        x = self.out_proj(x)
+
+        return x
+
+
+class DoubleStreamBlock(nn.Module):
+    def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, dtype=None, device=None, operations=None):
+        super().__init__()
+
+        mlp_hidden_dim = int(hidden_size * mlp_ratio)
+        self.num_heads = num_heads
+        self.hidden_size = hidden_size
+        self.img_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.img_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations)
+
+        self.img_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.img_mlp = nn.Sequential(
+            operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device),
+            nn.GELU(approximate="tanh"),
+            operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
+        )
+
+        self.txt_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.txt_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations)
+
+        self.txt_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.txt_mlp = nn.Sequential(
+            operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device),
+            nn.GELU(approximate="tanh"),
+            operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
+        )
+        self.flipped_img_txt = flipped_img_txt
+
+    def forward(self, img: Tensor, txt: Tensor, pe: Tensor, vec: Tensor, attn_mask=None):
+        (img_mod1, img_mod2), (txt_mod1, txt_mod2) = vec
+
+        # prepare image for attention
+        img_modulated = self.img_norm1(img)
+        img_modulated = (1 + img_mod1.scale) * img_modulated + img_mod1.shift
+        img_qkv = self.img_attn.qkv(img_modulated)
+        img_q, img_k, img_v = img_qkv.view(img_qkv.shape[0], img_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        img_q, img_k = self.img_attn.norm(img_q, img_k, img_v)
+
+        # prepare txt for attention
+        txt_modulated = self.txt_norm1(txt)
+        txt_modulated = (1 + txt_mod1.scale) * txt_modulated + txt_mod1.shift
+        txt_qkv = self.txt_attn.qkv(txt_modulated)
+        txt_q, txt_k, txt_v = txt_qkv.view(txt_qkv.shape[0], txt_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v)
+
+        # run actual attention
+        attn = attention(torch.cat((txt_q, img_q), dim=2),
+                         torch.cat((txt_k, img_k), dim=2),
+                         torch.cat((txt_v, img_v), dim=2),
+                         pe=pe, mask=attn_mask)
+
+        txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1] :]
+
+        # calculate the img bloks
+        img = img + img_mod1.gate * self.img_attn.proj(img_attn)
+        img = img + img_mod2.gate * self.img_mlp((1 + img_mod2.scale) * self.img_norm2(img) + img_mod2.shift)
+
+        # calculate the txt bloks
+        txt += txt_mod1.gate * self.txt_attn.proj(txt_attn)
+        txt += txt_mod2.gate * self.txt_mlp((1 + txt_mod2.scale) * self.txt_norm2(txt) + txt_mod2.shift)
+
+        if txt.dtype == torch.float16:
+            txt = torch.nan_to_num(txt, nan=0.0, posinf=65504, neginf=-65504)
+
+        return img, txt
+
+
+class SingleStreamBlock(nn.Module):
+    """
+    A DiT block with parallel linear layers as described in
+    https://arxiv.org/abs/2302.05442 and adapted modulation interface.
+    """
+
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        qk_scale: float = None,
+        dtype=None,
+        device=None,
+        operations=None
+    ):
+        super().__init__()
+        self.hidden_dim = hidden_size
+        self.num_heads = num_heads
+        head_dim = hidden_size // num_heads
+        self.scale = qk_scale or head_dim**-0.5
+
+        self.mlp_hidden_dim = int(hidden_size * mlp_ratio)
+        # qkv and mlp_in
+        self.linear1 = operations.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim, dtype=dtype, device=device)
+        # proj and mlp_out
+        self.linear2 = operations.Linear(hidden_size + self.mlp_hidden_dim, hidden_size, dtype=dtype, device=device)
+
+        self.norm = QKNorm(head_dim, dtype=dtype, device=device, operations=operations)
+
+        self.hidden_size = hidden_size
+        self.pre_norm = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+
+        self.mlp_act = nn.GELU(approximate="tanh")
+
+    def forward(self, x: Tensor, pe: Tensor, vec: Tensor, attn_mask=None) -> Tensor:
+        mod = vec
+        x_mod = (1 + mod.scale) * self.pre_norm(x) + mod.shift
+        qkv, mlp = torch.split(self.linear1(x_mod), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1)
+
+        q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        q, k = self.norm(q, k, v)
+
+        # compute attention
+        attn = attention(q, k, v, pe=pe, mask=attn_mask)
+        # compute activation in mlp stream, cat again and run second linear layer
+        output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2))
+        x += mod.gate * output
+        if x.dtype == torch.float16:
+            x = torch.nan_to_num(x, nan=0.0, posinf=65504, neginf=-65504)
+        return x
+
+
+class LastLayer(nn.Module):
+    def __init__(self, hidden_size: int, patch_size: int, out_channels: int, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.norm_final = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.linear = operations.Linear(hidden_size, out_channels, bias=True, dtype=dtype, device=device)
+
+    def forward(self, x: Tensor, vec: Tensor) -> Tensor:
+        shift, scale = vec
+        shift = shift.squeeze(1)
+        scale = scale.squeeze(1)
+        x = (1 + scale[:, None, :]) * self.norm_final(x) + shift[:, None, :]
+        x = self.linear(x)
+        return x

comfy/ldm/chroma/model.py

@@ -0,0 +1,271 @@
+#Original code can be found on: https://github.com/black-forest-labs/flux
+
+from dataclasses import dataclass
+
+import torch
+from torch import Tensor, nn
+from einops import rearrange, repeat
+import comfy.ldm.common_dit
+
+from comfy.ldm.flux.layers import (
+    EmbedND,
+    timestep_embedding,
+)
+
+from .layers import (
+    DoubleStreamBlock,
+    LastLayer,
+    SingleStreamBlock,
+    Approximator,
+    ChromaModulationOut,
+)
+
+
+@dataclass
+class ChromaParams:
+    in_channels: int
+    out_channels: int
+    context_in_dim: int
+    hidden_size: int
+    mlp_ratio: float
+    num_heads: int
+    depth: int
+    depth_single_blocks: int
+    axes_dim: list
+    theta: int
+    patch_size: int
+    qkv_bias: bool
+    in_dim: int
+    out_dim: int
+    hidden_dim: int
+    n_layers: int
+
+
+
+
+class Chroma(nn.Module):
+    """
+    Transformer model for flow matching on sequences.
+    """
+
+    def __init__(self, image_model=None, final_layer=True, dtype=None, device=None, operations=None, **kwargs):
+        super().__init__()
+        self.dtype = dtype
+        params = ChromaParams(**kwargs)
+        self.params = params
+        self.patch_size = params.patch_size
+        self.in_channels = params.in_channels
+        self.out_channels = params.out_channels
+        if params.hidden_size % params.num_heads != 0:
+            raise ValueError(
+                f"Hidden size {params.hidden_size} must be divisible by num_heads {params.num_heads}"
+            )
+        pe_dim = params.hidden_size // params.num_heads
+        if sum(params.axes_dim) != pe_dim:
+            raise ValueError(f"Got {params.axes_dim} but expected positional dim {pe_dim}")
+        self.hidden_size = params.hidden_size
+        self.num_heads = params.num_heads
+        self.in_dim = params.in_dim
+        self.out_dim = params.out_dim
+        self.hidden_dim = params.hidden_dim
+        self.n_layers = params.n_layers
+        self.pe_embedder = EmbedND(dim=pe_dim, theta=params.theta, axes_dim=params.axes_dim)
+        self.img_in = operations.Linear(self.in_channels, self.hidden_size, bias=True, dtype=dtype, device=device)
+        self.txt_in = operations.Linear(params.context_in_dim, self.hidden_size, dtype=dtype, device=device)
+        # set as nn identity for now, will overwrite it later.
+        self.distilled_guidance_layer = Approximator(
+                    in_dim=self.in_dim,
+                    hidden_dim=self.hidden_dim,
+                    out_dim=self.out_dim,
+                    n_layers=self.n_layers,
+                    dtype=dtype, device=device, operations=operations
+                )
+
+
+        self.double_blocks = nn.ModuleList(
+            [
+                DoubleStreamBlock(
+                    self.hidden_size,
+                    self.num_heads,
+                    mlp_ratio=params.mlp_ratio,
+                    qkv_bias=params.qkv_bias,
+                    dtype=dtype, device=device, operations=operations
+                )
+                for _ in range(params.depth)
+            ]
+        )
+
+        self.single_blocks = nn.ModuleList(
+            [
+                SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, dtype=dtype, device=device, operations=operations)
+                for _ in range(params.depth_single_blocks)
+            ]
+        )
+
+        if final_layer:
+            self.final_layer = LastLayer(self.hidden_size, 1, self.out_channels, dtype=dtype, device=device, operations=operations)
+
+        self.skip_mmdit = []
+        self.skip_dit = []
+        self.lite = False
+
+    def get_modulations(self, tensor: torch.Tensor, block_type: str, *, idx: int = 0):
+        # This function slices up the modulations tensor which has the following layout:
+        #   single     : num_single_blocks * 3 elements
+        #   double_img : num_double_blocks * 6 elements
+        #   double_txt : num_double_blocks * 6 elements
+        #   final      : 2 elements
+        if block_type == "final":
+            return (tensor[:, -2:-1, :], tensor[:, -1:, :])
+        single_block_count = self.params.depth_single_blocks
+        double_block_count = self.params.depth
+        offset = 3 * idx
+        if block_type == "single":
+            return ChromaModulationOut.from_offset(tensor, offset)
+        # Double block modulations are 6 elements so we double 3 * idx.
+        offset *= 2
+        if block_type in {"double_img", "double_txt"}:
+            # Advance past the single block modulations.
+            offset += 3 * single_block_count
+            if block_type == "double_txt":
+                # Advance past the double block img modulations.
+                offset += 6 * double_block_count
+            return (
+                ChromaModulationOut.from_offset(tensor, offset),
+                ChromaModulationOut.from_offset(tensor, offset + 3),
+            )
+        raise ValueError("Bad block_type")
+
+
+    def forward_orig(
+        self,
+        img: Tensor,
+        img_ids: Tensor,
+        txt: Tensor,
+        txt_ids: Tensor,
+        timesteps: Tensor,
+        guidance: Tensor = None,
+        control = None,
+        transformer_options={},
+        attn_mask: Tensor = None,
+    ) -> Tensor:
+        patches_replace = transformer_options.get("patches_replace", {})
+        if img.ndim != 3 or txt.ndim != 3:
+            raise ValueError("Input img and txt tensors must have 3 dimensions.")
+
+        # running on sequences img
+        img = self.img_in(img)
+
+        # distilled vector guidance
+        mod_index_length = 344
+        distill_timestep = timestep_embedding(timesteps.detach().clone(), 16).to(img.device, img.dtype)
+        # guidance = guidance *
+        distil_guidance = timestep_embedding(guidance.detach().clone(), 16).to(img.device, img.dtype)
+
+        # get all modulation index
+        modulation_index = timestep_embedding(torch.arange(mod_index_length), 32).to(img.device, img.dtype)
+        # we need to broadcast the modulation index here so each batch has all of the index
+        modulation_index = modulation_index.unsqueeze(0).repeat(img.shape[0], 1, 1).to(img.device, img.dtype)
+        # and we need to broadcast timestep and guidance along too
+        timestep_guidance = torch.cat([distill_timestep, distil_guidance], dim=1).unsqueeze(1).repeat(1, mod_index_length, 1).to(img.dtype).to(img.device, img.dtype)
+        # then and only then we could concatenate it together
+        input_vec = torch.cat([timestep_guidance, modulation_index], dim=-1).to(img.device, img.dtype)
+
+        mod_vectors = self.distilled_guidance_layer(input_vec)
+
+        txt = self.txt_in(txt)
+
+        ids = torch.cat((txt_ids, img_ids), dim=1)
+        pe = self.pe_embedder(ids)
+
+        blocks_replace = patches_replace.get("dit", {})
+        for i, block in enumerate(self.double_blocks):
+            if i not in self.skip_mmdit:
+                double_mod = (
+                    self.get_modulations(mod_vectors, "double_img", idx=i),
+                    self.get_modulations(mod_vectors, "double_txt", idx=i),
+                )
+                if ("double_block", i) in blocks_replace:
+                    def block_wrap(args):
+                        out = {}
+                        out["img"], out["txt"] = block(img=args["img"],
+                                                       txt=args["txt"],
+                                                       vec=args["vec"],
+                                                       pe=args["pe"],
+                                                       attn_mask=args.get("attn_mask"))
+                        return out
+
+                    out = blocks_replace[("double_block", i)]({"img": img,
+                                                               "txt": txt,
+                                                               "vec": double_mod,
+                                                               "pe": pe,
+                                                               "attn_mask": attn_mask},
+                                                              {"original_block": block_wrap})
+                    txt = out["txt"]
+                    img = out["img"]
+                else:
+                    img, txt = block(img=img,
+                                     txt=txt,
+                                     vec=double_mod,
+                                     pe=pe,
+                                     attn_mask=attn_mask)
+
+                if control is not None: # Controlnet
+                    control_i = control.get("input")
+                    if i < len(control_i):
+                        add = control_i[i]
+                        if add is not None:
+                            img += add
+
+        img = torch.cat((txt, img), 1)
+
+        for i, block in enumerate(self.single_blocks):
+            if i not in self.skip_dit:
+                single_mod = self.get_modulations(mod_vectors, "single", idx=i)
+                if ("single_block", i) in blocks_replace:
+                    def block_wrap(args):
+                        out = {}
+                        out["img"] = block(args["img"],
+                                           vec=args["vec"],
+                                           pe=args["pe"],
+                                           attn_mask=args.get("attn_mask"))
+                        return out
+
+                    out = blocks_replace[("single_block", i)]({"img": img,
+                                                               "vec": single_mod,
+                                                               "pe": pe,
+                                                               "attn_mask": attn_mask},
+                                                              {"original_block": block_wrap})
+                    img = out["img"]
+                else:
+                    img = block(img, vec=single_mod, pe=pe, attn_mask=attn_mask)
+
+                if control is not None: # Controlnet
+                    control_o = control.get("output")
+                    if i < len(control_o):
+                        add = control_o[i]
+                        if add is not None:
+                            img[:, txt.shape[1] :, ...] += add
+
+        img = img[:, txt.shape[1] :, ...]
+        final_mod = self.get_modulations(mod_vectors, "final")
+        img = self.final_layer(img, vec=final_mod)  # (N, T, patch_size ** 2 * out_channels)
+        return img
+
+    def forward(self, x, timestep, context, guidance, control=None, transformer_options={}, **kwargs):
+        bs, c, h, w = x.shape
+        patch_size = 2
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, (patch_size, patch_size))
+
+        img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size)
+
+        h_len = ((h + (patch_size // 2)) // patch_size)
+        w_len = ((w + (patch_size // 2)) // patch_size)
+        img_ids = torch.zeros((h_len, w_len, 3), device=x.device, dtype=x.dtype)
+        img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1)
+        img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0)
+        img_ids = repeat(img_ids, "h w c -> b (h w) c", b=bs)
+
+        txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype)
+        out = self.forward_orig(img, img_ids, context, txt_ids, timestep, guidance, control, transformer_options, attn_mask=kwargs.get("attention_mask", None))
+        return rearrange(out, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=h_len, w=w_len, ph=2, pw=2)[:,:,:h,:w]

comfy/ldm/lightricks/model.py

@@ -1,7 +1,6 @@
 import torch
 from torch import nn
 import comfy.ldm.modules.attention
-from comfy.ldm.genmo.joint_model.layers import RMSNorm
 import comfy.ldm.common_dit
 from einops import rearrange
 import math
@@ -262,8 +261,8 @@ class CrossAttention(nn.Module):
         self.heads = heads
         self.dim_head = dim_head
 
-        self.q_norm = RMSNorm(inner_dim, dtype=dtype, device=device)
-        self.k_norm = RMSNorm(inner_dim, dtype=dtype, device=device)
+        self.q_norm = operations.RMSNorm(inner_dim, dtype=dtype, device=device)
+        self.k_norm = operations.RMSNorm(inner_dim, dtype=dtype, device=device)
 
         self.to_q = operations.Linear(query_dim, inner_dim, bias=True, dtype=dtype, device=device)
         self.to_k = operations.Linear(context_dim, inner_dim, bias=True, dtype=dtype, device=device)

comfy/model_base.py

@@ -38,6 +38,7 @@ import comfy.ldm.lumina.model
 import comfy.ldm.wan.model
 import comfy.ldm.hunyuan3d.model
 import comfy.ldm.hidream.model
+import comfy.ldm.chroma.model
 
 import comfy.model_management
 import comfy.patcher_extension
@@ -786,8 +787,8 @@ class PixArt(BaseModel):
         return out
 
 class Flux(BaseModel):
-    def __init__(self, model_config, model_type=ModelType.FLUX, device=None):
-        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.flux.model.Flux)
+    def __init__(self, model_config, model_type=ModelType.FLUX, device=None, unet_model=comfy.ldm.flux.model.Flux):
+        super().__init__(model_config, model_type, device=device, unet_model=unet_model)
 
     def concat_cond(self, **kwargs):
         try:
@@ -1108,3 +1109,15 @@ class HiDream(BaseModel):
         if image_cond is not None:
             out['image_cond'] = comfy.conds.CONDNoiseShape(self.process_latent_in(image_cond))
         return out
+
+class Chroma(Flux):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.chroma.model.Chroma)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+
+        guidance = kwargs.get("guidance", 0)
+        if guidance is not None:
+            out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor([guidance]))
+        return out

comfy/model_detection.py

@@ -164,7 +164,9 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         if in_key in state_dict_keys:
             dit_config["in_channels"] = state_dict[in_key].shape[1] // (patch_size * patch_size)
         dit_config["out_channels"] = 16
-        dit_config["vec_in_dim"] = 768
+        vec_in_key = '{}vector_in.in_layer.weight'.format(key_prefix)
+        if vec_in_key in state_dict_keys:
+            dit_config["vec_in_dim"] = state_dict[vec_in_key].shape[1]
         dit_config["context_in_dim"] = 4096
         dit_config["hidden_size"] = 3072
         dit_config["mlp_ratio"] = 4.0
@@ -174,7 +176,16 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         dit_config["axes_dim"] = [16, 56, 56]
         dit_config["theta"] = 10000
         dit_config["qkv_bias"] = True
-        dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys
+        if '{}distilled_guidance_layer.0.norms.0.scale'.format(key_prefix) in state_dict_keys or '{}distilled_guidance_layer.norms.0.scale'.format(key_prefix) in state_dict_keys: #Chroma
+            dit_config["image_model"] = "chroma"
+            dit_config["in_channels"] = 64
+            dit_config["out_channels"] = 64
+            dit_config["in_dim"] = 64
+            dit_config["out_dim"] = 3072
+            dit_config["hidden_dim"] = 5120
+            dit_config["n_layers"] = 5
+        else:
+            dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys
         return dit_config
 
     if '{}t5_yproj.weight'.format(key_prefix) in state_dict_keys: #Genmo mochi preview

comfy/sd.py

@@ -714,6 +714,7 @@ class CLIPType(Enum):
     LUMINA2 = 12
     WAN = 13
     HIDREAM = 14
+    CHROMA = 15
 
 
 def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}):
@@ -818,7 +819,7 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
             elif clip_type == CLIPType.LTXV:
                 clip_target.clip = comfy.text_encoders.lt.ltxv_te(**t5xxl_detect(clip_data))
                 clip_target.tokenizer = comfy.text_encoders.lt.LTXVT5Tokenizer
-            elif clip_type == CLIPType.PIXART:
+            elif clip_type == CLIPType.PIXART or clip_type == CLIPType.CHROMA:
                 clip_target.clip = comfy.text_encoders.pixart_t5.pixart_te(**t5xxl_detect(clip_data))
                 clip_target.tokenizer = comfy.text_encoders.pixart_t5.PixArtTokenizer
             elif clip_type == CLIPType.WAN:

comfy/supported_models.py

@@ -1068,7 +1068,34 @@ class HiDream(supported_models_base.BASE):
     def clip_target(self, state_dict={}):
         return None #  TODO
 
+class Chroma(supported_models_base.BASE):
+    unet_config = {
+        "image_model": "chroma",
+    }
 
-models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, Lumina2, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, Hunyuan3Dv2mini, Hunyuan3Dv2, HiDream]
+    unet_extra_config = {
+    }
+
+    sampling_settings = {
+        "multiplier": 1.0,
+    }
+
+    latent_format = comfy.latent_formats.Flux
+
+    memory_usage_factor = 3.2
+
+    supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32]
+
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.Chroma(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        pref = self.text_encoder_key_prefix[0]
+        t5_detect = comfy.text_encoders.sd3_clip.t5_xxl_detect(state_dict, "{}t5xxl.transformer.".format(pref))
+        return supported_models_base.ClipTarget(comfy.text_encoders.pixart_t5.PixArtTokenizer, comfy.text_encoders.pixart_t5.pixart_te(**t5_detect))
+
+models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, Lumina2, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, Hunyuan3Dv2mini, Hunyuan3Dv2, HiDream, Chroma]
 
 models += [SVD_img2vid]

comfy/weight_adapter/boft.py

@@ -24,7 +24,7 @@ class BOFTAdapter(WeightAdapterBase):
     ) -> Optional["BOFTAdapter"]:
         if loaded_keys is None:
             loaded_keys = set()
-        blocks_name = "{}.boft_blocks".format(x)
+        blocks_name = "{}.oft_blocks".format(x)
         rescale_name = "{}.rescale".format(x)
 
         blocks = None
@@ -32,17 +32,18 @@ class BOFTAdapter(WeightAdapterBase):
             blocks = lora[blocks_name]
             if blocks.ndim == 4:
                 loaded_keys.add(blocks_name)
+            else:
+                blocks = None
+        if blocks is None:
+            return None
 
         rescale = None
         if rescale_name in lora.keys():
             rescale = lora[rescale_name]
             loaded_keys.add(rescale_name)
 
-        if blocks is not None:
-            weights = (blocks, rescale, alpha, dora_scale)
-            return cls(loaded_keys, weights)
-        else:
-            return None
+        weights = (blocks, rescale, alpha, dora_scale)
+        return cls(loaded_keys, weights)
 
     def calculate_weight(
         self,
@@ -71,7 +72,7 @@ class BOFTAdapter(WeightAdapterBase):
             # Get r
             I = torch.eye(boft_b, device=blocks.device, dtype=blocks.dtype)
             # for Q = -Q^T
-            q = blocks - blocks.transpose(1, 2)
+            q = blocks - blocks.transpose(-1, -2)
             normed_q = q
             if alpha > 0: # alpha in boft/bboft is for constraint
                 q_norm = torch.norm(q) + 1e-8
@@ -79,9 +80,8 @@ class BOFTAdapter(WeightAdapterBase):
                     normed_q = q * alpha / q_norm
             # use float() to prevent unsupported type in .inverse()
             r = (I + normed_q) @ (I - normed_q).float().inverse()
-            r = r.to(original_weight)
-
-            inp = org = original_weight
+            r = r.to(weight)
+            inp = org = weight
 
             r_b = boft_b//2
             for i in range(boft_m):
@@ -91,14 +91,14 @@ class BOFTAdapter(WeightAdapterBase):
                 if strength != 1:
                     bi = bi * strength + (1-strength) * I
                 inp = (
-                    inp.unflatten(-1, (-1, g, k))
-                    .transpose(-2, -1)
-                    .flatten(-3)
-                    .unflatten(-1, (-1, boft_b))
+                    inp.unflatten(0, (-1, g, k))
+                    .transpose(1, 2)
+                    .flatten(0, 2)
+                    .unflatten(0, (-1, boft_b))
                 )
-                inp = torch.einsum("b n m, b n ... -> b m ...", inp, bi)
+                inp = torch.einsum("b i j, b j ...-> b i ...", bi, inp)
                 inp = (
-                    inp.flatten(-2).unflatten(-1, (-1, k, g)).transpose(-2, -1).flatten(-3)
+                    inp.flatten(0, 1).unflatten(0, (-1, k, g)).transpose(1, 2).flatten(0, 2)
                 )
 
             if rescale is not None:
@@ -109,7 +109,7 @@ class BOFTAdapter(WeightAdapterBase):
             if dora_scale is not None:
                 weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function)
             else:
-                weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
+                weight += function((strength * lora_diff).type(weight.dtype))
         except Exception as e:
             logging.error("ERROR {} {} {}".format(self.name, key, e))
         return weight

comfy/weight_adapter/oft.py

@@ -32,17 +32,18 @@ class OFTAdapter(WeightAdapterBase):
             blocks = lora[blocks_name]
             if blocks.ndim == 3:
                 loaded_keys.add(blocks_name)
+            else:
+                blocks = None
+        if blocks is None:
+            return None
 
         rescale = None
         if rescale_name in lora.keys():
             rescale = lora[rescale_name]
             loaded_keys.add(rescale_name)
 
-        if blocks is not None:
-            weights = (blocks, rescale, alpha, dora_scale)
-            return cls(loaded_keys, weights)
-        else:
-            return None
+        weights = (blocks, rescale, alpha, dora_scale)
+        return cls(loaded_keys, weights)
 
     def calculate_weight(
         self,
@@ -79,16 +80,17 @@ class OFTAdapter(WeightAdapterBase):
                     normed_q = q * alpha / q_norm
             # use float() to prevent unsupported type in .inverse()
             r = (I + normed_q) @ (I - normed_q).float().inverse()
-            r = r.to(original_weight)
+            r = r.to(weight)
+            _, *shape = weight.shape
             lora_diff = torch.einsum(
                 "k n m, k n ... -> k m ...",
                 (r * strength) - strength * I,
-                original_weight,
-            )
+                weight.view(block_num, block_size, *shape),
+            ).view(-1, *shape)
             if dora_scale is not None:
                 weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function)
             else:
-                weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
+                weight += function((strength * lora_diff).type(weight.dtype))
         except Exception as e:
             logging.error("ERROR {} {} {}".format(self.name, key, e))
         return weight