sync with repo 28.08

comfy/cli_args.py

@@ -92,6 +92,10 @@ class LatentPreviewMethod(enum.Enum):
 
 parser.add_argument("--preview-method", type=LatentPreviewMethod, default=LatentPreviewMethod.NoPreviews, help="Default preview method for sampler nodes.", action=EnumAction)
 
+cache_group = parser.add_mutually_exclusive_group()
+cache_group.add_argument("--cache-classic", action="store_true", help="Use the old style (aggressive) caching.")
+cache_group.add_argument("--cache-lru", type=int, default=0, help="Use LRU caching with a maximum of N node results cached. May use more RAM/VRAM.")
+
 attn_group = parser.add_mutually_exclusive_group()
 attn_group.add_argument("--use-split-cross-attention", action="store_true", help="Use the split cross attention optimization. Ignored when xformers is used.")
 attn_group.add_argument("--use-quad-cross-attention", action="store_true", help="Use the sub-quadratic cross attention optimization . Ignored when xformers is used.")
@@ -112,10 +116,14 @@ vram_group.add_argument("--lowvram", action="store_true", help="Split the unet i
 vram_group.add_argument("--novram", action="store_true", help="When lowvram isn't enough.")
 vram_group.add_argument("--cpu", action="store_true", help="To use the CPU for everything (slow).")
 
+parser.add_argument("--reserve-vram", type=float, default=None, help="Set the amount of vram in GB you want to reserve for use by your OS/other software. By default some amount is reverved depending on your OS.")
+
+
 parser.add_argument("--default-hashing-function", type=str, choices=['md5', 'sha1', 'sha256', 'sha512'], default='sha256', help="Allows you to choose the hash function to use for duplicate filename / contents comparison. Default is sha256.")
 
 parser.add_argument("--disable-smart-memory", action="store_true", help="Force ComfyUI to agressively offload to regular ram instead of keeping models in vram when it can.")
 parser.add_argument("--deterministic", action="store_true", help="Make pytorch use slower deterministic algorithms when it can. Note that this might not make images deterministic in all cases.")
+parser.add_argument("--fast", action="store_true", help="Enable some untested and potentially quality deteriorating optimizations.")
 
 parser.add_argument("--dont-print-server", action="store_true", help="Don't print server output.")
 parser.add_argument("--quick-test-for-ci", action="store_true", help="Quick test for CI.")

comfy/clip_model.py

@@ -88,10 +88,11 @@ class CLIPTextModel_(torch.nn.Module):
         heads = config_dict["num_attention_heads"]
         intermediate_size = config_dict["intermediate_size"]
         intermediate_activation = config_dict["hidden_act"]
+        num_positions = config_dict["max_position_embeddings"]
         self.eos_token_id = config_dict["eos_token_id"]
 
         super().__init__()
-        self.embeddings = CLIPEmbeddings(embed_dim, dtype=dtype, device=device, operations=operations)
+        self.embeddings = CLIPEmbeddings(embed_dim, num_positions=num_positions, dtype=dtype, device=device, operations=operations)
         self.encoder = CLIPEncoder(num_layers, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations)
         self.final_layer_norm = operations.LayerNorm(embed_dim, dtype=dtype, device=device)
 
@@ -123,7 +124,6 @@ class CLIPTextModel(torch.nn.Module):
         self.text_model = CLIPTextModel_(config_dict, dtype, device, operations)
         embed_dim = config_dict["hidden_size"]
         self.text_projection = operations.Linear(embed_dim, embed_dim, bias=False, dtype=dtype, device=device)
-        self.text_projection.weight.copy_(torch.eye(embed_dim))
         self.dtype = dtype
 
     def get_input_embeddings(self):

comfy/controlnet.py

@@ -1,4 +1,24 @@
+"""
+    This file is part of ComfyUI.
+    Copyright (C) 2024 Comfy
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+"""
+
+
 import torch
+from enum import Enum
 import math
 import os
 import logging
@@ -13,6 +33,8 @@ import comfy.cldm.cldm
 import comfy.t2i_adapter.adapter
 import comfy.ldm.cascade.controlnet
 import comfy.cldm.mmdit
+import comfy.ldm.hydit.controlnet
+import comfy.ldm.flux.controlnet_xlabs
 
 
 def broadcast_image_to(tensor, target_batch_size, batched_number):
@@ -33,6 +55,10 @@ def broadcast_image_to(tensor, target_batch_size, batched_number):
     else:
         return torch.cat([tensor] * batched_number, dim=0)
 
+class StrengthType(Enum):
+    CONSTANT = 1
+    LINEAR_UP = 2
+
 class ControlBase:
     def __init__(self, device=None):
         self.cond_hint_original = None
@@ -51,6 +77,8 @@ class ControlBase:
             device = comfy.model_management.get_torch_device()
         self.device = device
         self.previous_controlnet = None
+        self.extra_conds = []
+        self.strength_type = StrengthType.CONSTANT
 
     def set_cond_hint(self, cond_hint, strength=1.0, timestep_percent_range=(0.0, 1.0), vae=None):
         self.cond_hint_original = cond_hint
@@ -93,6 +121,8 @@ class ControlBase:
         c.latent_format = self.latent_format
         c.extra_args = self.extra_args.copy()
         c.vae = self.vae
+        c.extra_conds = self.extra_conds.copy()
+        c.strength_type = self.strength_type
 
     def inference_memory_requirements(self, dtype):
         if self.previous_controlnet is not None:
@@ -113,7 +143,10 @@ class ControlBase:
 
                     if x not in applied_to: #memory saving strategy, allow shared tensors and only apply strength to shared tensors once
                         applied_to.add(x)
-                        x *= self.strength
+                        if self.strength_type == StrengthType.CONSTANT:
+                            x *= self.strength
+                        elif self.strength_type == StrengthType.LINEAR_UP:
+                            x *= (self.strength ** float(len(control_output) - i))
 
                     if x.dtype != output_dtype:
                         x = x.to(output_dtype)
@@ -142,7 +175,7 @@ class ControlBase:
 
 
 class ControlNet(ControlBase):
-    def __init__(self, control_model=None, global_average_pooling=False, compression_ratio=8, latent_format=None, device=None, load_device=None, manual_cast_dtype=None):
+    def __init__(self, control_model=None, global_average_pooling=False, compression_ratio=8, latent_format=None, device=None, load_device=None, manual_cast_dtype=None, extra_conds=["y"], strength_type=StrengthType.CONSTANT):
         super().__init__(device)
         self.control_model = control_model
         self.load_device = load_device
@@ -154,6 +187,8 @@ class ControlNet(ControlBase):
         self.model_sampling_current = None
         self.manual_cast_dtype = manual_cast_dtype
         self.latent_format = latent_format
+        self.extra_conds += extra_conds
+        self.strength_type = strength_type
 
     def get_control(self, x_noisy, t, cond, batched_number):
         control_prev = None
@@ -191,13 +226,16 @@ class ControlNet(ControlBase):
             self.cond_hint = broadcast_image_to(self.cond_hint, x_noisy.shape[0], batched_number)
 
         context = cond.get('crossattn_controlnet', cond['c_crossattn'])
-        y = cond.get('y', None)
-        if y is not None:
-            y = y.to(dtype)
+        extra = self.extra_args.copy()
+        for c in self.extra_conds:
+            temp = cond.get(c, None)
+            if temp is not None:
+                extra[c] = temp.to(dtype)
+
         timestep = self.model_sampling_current.timestep(t)
         x_noisy = self.model_sampling_current.calculate_input(t, x_noisy)
 
-        control = self.control_model(x=x_noisy.to(dtype), hint=self.cond_hint, timesteps=timestep.float(), context=context.to(dtype), y=y, **self.extra_args)
+        control = self.control_model(x=x_noisy.to(dtype), hint=self.cond_hint, timesteps=timestep.to(dtype), context=context.to(dtype), **extra)
         return self.control_merge(control, control_prev, output_dtype)
 
     def copy(self):
@@ -286,6 +324,7 @@ class ControlLora(ControlNet):
         ControlBase.__init__(self, device)
         self.control_weights = control_weights
         self.global_average_pooling = global_average_pooling
+        self.extra_conds += ["y"]
 
     def pre_run(self, model, percent_to_timestep_function):
         super().pre_run(model, percent_to_timestep_function)
@@ -338,12 +377,8 @@ class ControlLora(ControlNet):
     def inference_memory_requirements(self, dtype):
         return comfy.utils.calculate_parameters(self.control_weights) * comfy.model_management.dtype_size(dtype) + ControlBase.inference_memory_requirements(self, dtype)
 
-def load_controlnet_mmdit(sd):
-    new_sd = comfy.model_detection.convert_diffusers_mmdit(sd, "")
-    model_config = comfy.model_detection.model_config_from_unet(new_sd, "", True)
-    num_blocks = comfy.model_detection.count_blocks(new_sd, 'joint_blocks.{}.')
-    for k in sd:
-        new_sd[k] = sd[k]
+def controlnet_config(sd):
+    model_config = comfy.model_detection.model_config_from_unet(sd, "", True)
 
     supported_inference_dtypes = model_config.supported_inference_dtypes
 
@@ -356,14 +391,28 @@ def load_controlnet_mmdit(sd):
     else:
         operations = comfy.ops.disable_weight_init
 
-    control_model = comfy.cldm.mmdit.ControlNet(num_blocks=num_blocks, operations=operations, device=load_device, dtype=unet_dtype, **controlnet_config)
-    missing, unexpected = control_model.load_state_dict(new_sd, strict=False)
+    offload_device = comfy.model_management.unet_offload_device()
+    return model_config, operations, load_device, unet_dtype, manual_cast_dtype, offload_device
+
+def controlnet_load_state_dict(control_model, sd):
+    missing, unexpected = control_model.load_state_dict(sd, strict=False)
 
     if len(missing) > 0:
         logging.warning("missing controlnet keys: {}".format(missing))
 
     if len(unexpected) > 0:
         logging.debug("unexpected controlnet keys: {}".format(unexpected))
+    return control_model
+
+def load_controlnet_mmdit(sd):
+    new_sd = comfy.model_detection.convert_diffusers_mmdit(sd, "")
+    model_config, operations, load_device, unet_dtype, manual_cast_dtype, offload_device = controlnet_config(new_sd)
+    num_blocks = comfy.model_detection.count_blocks(new_sd, 'joint_blocks.{}.')
+    for k in sd:
+        new_sd[k] = sd[k]
+
+    control_model = comfy.cldm.mmdit.ControlNet(num_blocks=num_blocks, operations=operations, device=offload_device, dtype=unet_dtype, **model_config.unet_config)
+    control_model = controlnet_load_state_dict(control_model, new_sd)
 
     latent_format = comfy.latent_formats.SD3()
     latent_format.shift_factor = 0 #SD3 controlnet weirdness
@@ -371,8 +420,31 @@ def load_controlnet_mmdit(sd):
     return control
 
 
+def load_controlnet_hunyuandit(controlnet_data):
+    model_config, operations, load_device, unet_dtype, manual_cast_dtype, offload_device = controlnet_config(controlnet_data)
+
+    control_model = comfy.ldm.hydit.controlnet.HunYuanControlNet(operations=operations, device=offload_device, dtype=unet_dtype)
+    control_model = controlnet_load_state_dict(control_model, controlnet_data)
+
+    latent_format = comfy.latent_formats.SDXL()
+    extra_conds = ['text_embedding_mask', 'encoder_hidden_states_t5', 'text_embedding_mask_t5', 'image_meta_size', 'style', 'cos_cis_img', 'sin_cis_img']
+    control = ControlNet(control_model, compression_ratio=1, latent_format=latent_format, load_device=load_device, manual_cast_dtype=manual_cast_dtype, extra_conds=extra_conds, strength_type=StrengthType.CONSTANT)
+    return control
+
+def load_controlnet_flux_xlabs(sd):
+    model_config, operations, load_device, unet_dtype, manual_cast_dtype, offload_device = controlnet_config(sd)
+    control_model = comfy.ldm.flux.controlnet_xlabs.ControlNetFlux(operations=operations, device=offload_device, dtype=unet_dtype, **model_config.unet_config)
+    control_model = controlnet_load_state_dict(control_model, sd)
+    extra_conds = ['y', 'guidance']
+    control = ControlNet(control_model, load_device=load_device, manual_cast_dtype=manual_cast_dtype, extra_conds=extra_conds)
+    return control
+
+
 def load_controlnet(ckpt_path, model=None):
     controlnet_data = comfy.utils.load_torch_file(ckpt_path, safe_load=True)
+    if 'after_proj_list.18.bias' in controlnet_data.keys(): #Hunyuan DiT
+        return load_controlnet_hunyuandit(controlnet_data)
+
     if "lora_controlnet" in controlnet_data:
         return ControlLora(controlnet_data)
 
@@ -430,7 +502,10 @@ def load_controlnet(ckpt_path, model=None):
             logging.warning("leftover keys: {}".format(leftover_keys))
         controlnet_data = new_sd
     elif "controlnet_blocks.0.weight" in controlnet_data: #SD3 diffusers format
-        return load_controlnet_mmdit(controlnet_data)
+        if "double_blocks.0.img_attn.norm.key_norm.scale" in controlnet_data:
+            return load_controlnet_flux_xlabs(controlnet_data)
+        else:
+            return load_controlnet_mmdit(controlnet_data)
 
     pth_key = 'control_model.zero_convs.0.0.weight'
     pth = False
@@ -462,6 +537,7 @@ def load_controlnet(ckpt_path, model=None):
     if manual_cast_dtype is not None:
         controlnet_config["operations"] = comfy.ops.manual_cast
     controlnet_config["dtype"] = unet_dtype
+    controlnet_config["device"] = comfy.model_management.unet_offload_device()
     controlnet_config.pop("out_channels")
     controlnet_config["hint_channels"] = controlnet_data["{}input_hint_block.0.weight".format(prefix)].shape[1]
     control_model = comfy.cldm.cldm.ControlNet(**controlnet_config)

comfy/diffusers_load.py

@@ -22,7 +22,7 @@ def load_diffusers(model_path, output_vae=True, output_clip=True, embedding_dire
     if text_encoder2_path is not None:
         text_encoder_paths.append(text_encoder2_path)
 
-    unet = comfy.sd.load_unet(unet_path)
+    unet = comfy.sd.load_diffusion_model(unet_path)
 
     clip = None
     if output_clip:

comfy/float.py

@@ -0,0 +1,62 @@
+import torch
+import math
+
+def calc_mantissa(abs_x, exponent, normal_mask, MANTISSA_BITS, EXPONENT_BIAS, generator=None):
+    mantissa_scaled = torch.where(
+        normal_mask,
+        (abs_x / (2.0 ** (exponent - EXPONENT_BIAS)) - 1.0) * (2**MANTISSA_BITS),
+        (abs_x / (2.0 ** (-EXPONENT_BIAS + 1 - MANTISSA_BITS)))
+    )
+
+    mantissa_scaled += torch.rand(mantissa_scaled.size(), dtype=mantissa_scaled.dtype, layout=mantissa_scaled.layout, device=mantissa_scaled.device, generator=generator)
+    return mantissa_scaled.floor() / (2**MANTISSA_BITS)
+
+#Not 100% sure about this
+def manual_stochastic_round_to_float8(x, dtype, generator=None):
+    if dtype == torch.float8_e4m3fn:
+        EXPONENT_BITS, MANTISSA_BITS, EXPONENT_BIAS = 4, 3, 7
+    elif dtype == torch.float8_e5m2:
+        EXPONENT_BITS, MANTISSA_BITS, EXPONENT_BIAS = 5, 2, 15
+    else:
+        raise ValueError("Unsupported dtype")
+
+    x = x.half()
+    sign = torch.sign(x)
+    abs_x = x.abs()
+    sign = torch.where(abs_x == 0, 0, sign)
+
+    # Combine exponent calculation and clamping
+    exponent = torch.clamp(
+        torch.floor(torch.log2(abs_x)) + EXPONENT_BIAS,
+        0, 2**EXPONENT_BITS - 1
+    )
+
+    # Combine mantissa calculation and rounding
+    normal_mask = ~(exponent == 0)
+
+    abs_x[:] = calc_mantissa(abs_x, exponent, normal_mask, MANTISSA_BITS, EXPONENT_BIAS, generator=generator)
+
+    sign *= torch.where(
+        normal_mask,
+        (2.0 ** (exponent - EXPONENT_BIAS)) * (1.0 + abs_x),
+        (2.0 ** (-EXPONENT_BIAS + 1)) * abs_x
+    )
+    del abs_x
+
+    return sign.to(dtype=dtype)
+
+
+
+def stochastic_rounding(value, dtype, seed=0):
+    if dtype == torch.float32:
+        return value.to(dtype=torch.float32)
+    if dtype == torch.float16:
+        return value.to(dtype=torch.float16)
+    if dtype == torch.bfloat16:
+        return value.to(dtype=torch.bfloat16)
+    if dtype == torch.float8_e4m3fn or dtype == torch.float8_e5m2:
+        generator = torch.Generator(device=value.device)
+        generator.manual_seed(seed)
+        return manual_stochastic_round_to_float8(value, dtype, generator=generator)
+
+    return value.to(dtype=dtype)

comfy/k_diffusion/sampling.py

@@ -9,6 +9,7 @@ from tqdm.auto import trange, tqdm
 from . import utils
 from . import deis
 import comfy.model_patcher
+import comfy.model_sampling
 
 def append_zero(x):
     return torch.cat([x, x.new_zeros([1])])
@@ -509,6 +510,9 @@ def sample_dpm_adaptive(model, x, sigma_min, sigma_max, extra_args=None, callbac
 
 @torch.no_grad()
 def sample_dpmpp_2s_ancestral(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
+    if isinstance(model.inner_model.inner_model.model_sampling, comfy.model_sampling.CONST):
+        return sample_dpmpp_2s_ancestral_RF(model, x, sigmas, extra_args, callback, disable, eta, s_noise, noise_sampler)
+
     """Ancestral sampling with DPM-Solver++(2S) second-order steps."""
     extra_args = {} if extra_args is None else extra_args
     noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
@@ -541,6 +545,55 @@ def sample_dpmpp_2s_ancestral(model, x, sigmas, extra_args=None, callback=None,
     return x
 
 
+@torch.no_grad()
+def sample_dpmpp_2s_ancestral_RF(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
+    """Ancestral sampling with DPM-Solver++(2S) second-order steps."""
+    extra_args = {} if extra_args is None else extra_args
+    noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
+    s_in = x.new_ones([x.shape[0]])
+    sigma_fn = lambda lbda: (lbda.exp() + 1) ** -1
+    lambda_fn = lambda sigma: ((1-sigma)/sigma).log()
+
+    # logged_x = x.unsqueeze(0)
+
+    for i in trange(len(sigmas) - 1, disable=disable):
+        denoised = model(x, sigmas[i] * s_in, **extra_args)
+        downstep_ratio = 1 + (sigmas[i+1]/sigmas[i] - 1) * eta
+        sigma_down = sigmas[i+1] * downstep_ratio
+        alpha_ip1 = 1 - sigmas[i+1]
+        alpha_down = 1 - sigma_down
+        renoise_coeff = (sigmas[i+1]**2 - sigma_down**2*alpha_ip1**2/alpha_down**2)**0.5
+        # sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
+        if sigmas[i + 1] == 0:
+            # Euler method
+            d = to_d(x, sigmas[i], denoised)
+            dt = sigma_down - sigmas[i]
+            x = x + d * dt
+        else:
+            # DPM-Solver++(2S)
+            if sigmas[i] == 1.0:
+                sigma_s = 0.9999
+            else:
+                t_i, t_down = lambda_fn(sigmas[i]), lambda_fn(sigma_down)
+                r = 1 / 2
+                h = t_down - t_i
+                s = t_i + r * h
+                sigma_s = sigma_fn(s)
+            # sigma_s = sigmas[i+1]
+            sigma_s_i_ratio = sigma_s / sigmas[i]
+            u = sigma_s_i_ratio * x + (1 - sigma_s_i_ratio) * denoised
+            D_i = model(u, sigma_s * s_in, **extra_args)
+            sigma_down_i_ratio = sigma_down / sigmas[i]
+            x = sigma_down_i_ratio * x + (1 - sigma_down_i_ratio) * D_i
+            # print("sigma_i", sigmas[i], "sigma_ip1", sigmas[i+1],"sigma_down", sigma_down, "sigma_down_i_ratio", sigma_down_i_ratio, "sigma_s_i_ratio", sigma_s_i_ratio, "renoise_coeff", renoise_coeff)
+        # Noise addition
+        if sigmas[i + 1] > 0 and eta > 0:
+            x = (alpha_ip1/alpha_down) * x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * renoise_coeff
+        # logged_x = torch.cat((logged_x, x.unsqueeze(0)), dim=0)
+    return x
+
 @torch.no_grad()
 def sample_dpmpp_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=1 / 2):
     """DPM-Solver++ (stochastic)."""

comfy/latent_formats.py

@@ -141,6 +141,7 @@ class StableAudio1(LatentFormat):
     latent_channels = 64
 
 class Flux(SD3):
+    latent_channels = 16
     def __init__(self):
         self.scale_factor = 0.3611
         self.shift_factor = 0.1159
@@ -162,6 +163,7 @@ class Flux(SD3):
             [-0.0005, -0.0530, -0.0020],
             [-0.1273, -0.0932, -0.0680]
         ]
+        self.taesd_decoder_name = "taef1_decoder"
 
     def process_in(self, latent):
         return (latent - self.shift_factor) * self.scale_factor

comfy/ldm/aura/mmdit.py

@@ -9,6 +9,7 @@ import torch.nn.functional as F
 
 from comfy.ldm.modules.attention import optimized_attention
 import comfy.ops
+import comfy.ldm.common_dit
 
 def modulate(x, shift, scale):
     return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
@@ -407,10 +408,7 @@ class MMDiT(nn.Module):
 
     def patchify(self, x):
         B, C, H, W = x.size()
-        pad_h = (self.patch_size - H % self.patch_size) % self.patch_size
-        pad_w = (self.patch_size - W % self.patch_size) % self.patch_size
-
-        x = torch.nn.functional.pad(x, (0, pad_w, 0, pad_h), mode='circular')
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size))
         x = x.view(
             B,
             C,

comfy/ldm/common_dit.py

@@ -0,0 +1,8 @@
+import torch
+
+def pad_to_patch_size(img, patch_size=(2, 2), padding_mode="circular"):
+    if padding_mode == "circular" and torch.jit.is_tracing() or torch.jit.is_scripting():
+        padding_mode = "reflect"
+    pad_h = (patch_size[0] - img.shape[-2] % patch_size[0]) % patch_size[0]
+    pad_w = (patch_size[1] - img.shape[-1] % patch_size[1]) % patch_size[1]
+    return torch.nn.functional.pad(img, (0, pad_w, 0, pad_h), mode=padding_mode)

comfy/ldm/flux/controlnet_xlabs.py

@@ -0,0 +1,104 @@
+#Original code can be found on: https://github.com/XLabs-AI/x-flux/blob/main/src/flux/controlnet.py
+
+import torch
+from torch import Tensor, nn
+from einops import rearrange, repeat
+
+from .layers import (DoubleStreamBlock, EmbedND, LastLayer,
+                                 MLPEmbedder, SingleStreamBlock,
+                                 timestep_embedding)
+
+from .model import Flux
+import comfy.ldm.common_dit
+
+
+class ControlNetFlux(Flux):
+    def __init__(self, image_model=None, dtype=None, device=None, operations=None, **kwargs):
+        super().__init__(final_layer=False, dtype=dtype, device=device, operations=operations, **kwargs)
+
+        # add ControlNet blocks
+        self.controlnet_blocks = nn.ModuleList([])
+        for _ in range(self.params.depth):
+            controlnet_block = operations.Linear(self.hidden_size, self.hidden_size, dtype=dtype, device=device)
+            # controlnet_block = zero_module(controlnet_block)
+            self.controlnet_blocks.append(controlnet_block)
+        self.pos_embed_input = operations.Linear(self.in_channels, self.hidden_size, bias=True, dtype=dtype, device=device)
+        self.gradient_checkpointing = False
+        self.input_hint_block = nn.Sequential(
+            operations.Conv2d(3, 16, 3, padding=1, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device)
+        )
+
+    def forward_orig(
+        self,
+        img: Tensor,
+        img_ids: Tensor,
+        controlnet_cond: Tensor,
+        txt: Tensor,
+        txt_ids: Tensor,
+        timesteps: Tensor,
+        y: Tensor,
+        guidance: Tensor = None,
+    ) -> Tensor:
+        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)
+        controlnet_cond = self.input_hint_block(controlnet_cond)
+        controlnet_cond = rearrange(controlnet_cond, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=2, pw=2)
+        controlnet_cond = self.pos_embed_input(controlnet_cond)
+        img = img + controlnet_cond
+        vec = self.time_in(timestep_embedding(timesteps, 256))
+        if self.params.guidance_embed:
+            vec = vec + self.guidance_in(timestep_embedding(guidance, 256))
+        vec = vec + self.vector_in(y)
+        txt = self.txt_in(txt)
+
+        ids = torch.cat((txt_ids, img_ids), dim=1)
+        pe = self.pe_embedder(ids)
+
+        block_res_samples = ()
+
+        for block in self.double_blocks:
+            img, txt = block(img=img, txt=txt, vec=vec, pe=pe)
+            block_res_samples = block_res_samples + (img,)
+
+        controlnet_block_res_samples = ()
+        for block_res_sample, controlnet_block in zip(block_res_samples, self.controlnet_blocks):
+            block_res_sample = controlnet_block(block_res_sample)
+            controlnet_block_res_samples = controlnet_block_res_samples + (block_res_sample,)
+
+        return {"input": (controlnet_block_res_samples * 10)[:19]}
+
+    def forward(self, x, timesteps, context, y, guidance=None, hint=None, **kwargs):
+        hint = hint * 2.0 - 1.0
+
+        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)[:, None]
+        img_ids[..., 2] = img_ids[..., 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype)[None, :]
+        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)
+        return self.forward_orig(img, img_ids, hint, context, txt_ids, timesteps, y, guidance)

comfy/ldm/flux/layers.py

@@ -2,12 +2,12 @@ import math
 from dataclasses import dataclass
 
 import torch
-from einops import rearrange
 from torch import Tensor, nn
 
 from .math import attention, rope
 import comfy.ops
 
+
 class EmbedND(nn.Module):
     def __init__(self, dim: int, theta: int, axes_dim: list):
         super().__init__()
@@ -36,9 +36,7 @@ def timestep_embedding(t: Tensor, dim, max_period=10000, time_factor: float = 10
     """
     t = time_factor * t
     half = dim // 2
-    freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half).to(
-        t.device
-    )
+    freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32, device=t.device) / half)
 
     args = t[:, None].float() * freqs[None]
     embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
@@ -48,7 +46,6 @@ def timestep_embedding(t: Tensor, dim, max_period=10000, time_factor: float = 10
         embedding = embedding.to(t)
     return embedding
 
-
 class MLPEmbedder(nn.Module):
     def __init__(self, in_dim: int, hidden_dim: int, dtype=None, device=None, operations=None):
         super().__init__()
@@ -66,10 +63,8 @@ class RMSNorm(torch.nn.Module):
         self.scale = nn.Parameter(torch.empty((dim), dtype=dtype, device=device))
 
     def forward(self, x: Tensor):
-        x_dtype = x.dtype
-        x = x.float()
         rrms = torch.rsqrt(torch.mean(x**2, dim=-1, keepdim=True) + 1e-6)
-        return (x * rrms).to(dtype=x_dtype) * comfy.ops.cast_to(self.scale, dtype=x_dtype, device=x.device)
+        return (x * rrms) * comfy.ops.cast_to(self.scale, dtype=x.dtype, device=x.device)
 
 
 class QKNorm(torch.nn.Module):
@@ -94,14 +89,6 @@ class SelfAttention(nn.Module):
         self.norm = QKNorm(head_dim, dtype=dtype, device=device, operations=operations)
         self.proj = operations.Linear(dim, dim, dtype=dtype, device=device)
 
-    def forward(self, x: Tensor, pe: Tensor) -> Tensor:
-        qkv = self.qkv(x)
-        q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
-        q, k = self.norm(q, k, v)
-        x = attention(q, k, v, pe=pe)
-        x = self.proj(x)
-        return x
-
 
 @dataclass
 class ModulationOut:
@@ -163,22 +150,21 @@ class DoubleStreamBlock(nn.Module):
         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 = rearrange(img_qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
+        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 = rearrange(txt_qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
+        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
-        q = torch.cat((txt_q, img_q), dim=2)
-        k = torch.cat((txt_k, img_k), dim=2)
-        v = torch.cat((txt_v, img_v), dim=2)
+        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)
 
-        attn = attention(q, k, v, pe=pe)
         txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1] :]
 
         # calculate the img bloks
@@ -186,8 +172,12 @@ class DoubleStreamBlock(nn.Module):
         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 + txt_mod1.gate * self.txt_attn.proj(txt_attn)
-        txt = txt + txt_mod2.gate * self.txt_mlp((1 + txt_mod2.scale) * self.txt_norm2(txt) + txt_mod2.shift)
+        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
 
 
@@ -232,14 +222,17 @@ class SingleStreamBlock(nn.Module):
         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 = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
+        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)
         # compute activation in mlp stream, cat again and run second linear layer
         output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2))
-        return x + mod.gate * output
+        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):

comfy/ldm/flux/math.py

@@ -14,7 +14,7 @@ def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor) -> Tensor:
 
 def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
     assert dim % 2 == 0
-    if comfy.model_management.is_device_mps(pos.device):
+    if comfy.model_management.is_device_mps(pos.device) or comfy.model_management.is_intel_xpu():
         device = torch.device("cpu")
     else:
         device = pos.device

comfy/ldm/flux/model.py

@@ -15,6 +15,7 @@ from .layers import (
 )
 
 from einops import rearrange, repeat
+import comfy.ldm.common_dit
 
 @dataclass
 class FluxParams:
@@ -37,12 +38,12 @@ class Flux(nn.Module):
     Transformer model for flow matching on sequences.
     """
 
-    def __init__(self, image_model=None, dtype=None, device=None, operations=None, **kwargs):
+    def __init__(self, image_model=None, final_layer=True, dtype=None, device=None, operations=None, **kwargs):
         super().__init__()
         self.dtype = dtype
         params = FluxParams(**kwargs)
         self.params = params
-        self.in_channels = params.in_channels
+        self.in_channels = params.in_channels * 2 * 2
         self.out_channels = self.in_channels
         if params.hidden_size % params.num_heads != 0:
             raise ValueError(
@@ -82,7 +83,8 @@ class Flux(nn.Module):
             ]
         )
 
-        self.final_layer = LastLayer(self.hidden_size, 1, self.out_channels, dtype=dtype, device=device, operations=operations)
+        if final_layer:
+            self.final_layer = LastLayer(self.hidden_size, 1, self.out_channels, dtype=dtype, device=device, operations=operations)
 
     def forward_orig(
         self,
@@ -93,6 +95,7 @@ class Flux(nn.Module):
         timesteps: Tensor,
         y: Tensor,
         guidance: Tensor = None,
+        control=None,
     ) -> Tensor:
         if img.ndim != 3 or txt.ndim != 3:
             raise ValueError("Input img and txt tensors must have 3 dimensions.")
@@ -111,24 +114,37 @@ class Flux(nn.Module):
         ids = torch.cat((txt_ids, img_ids), dim=1)
         pe = self.pe_embedder(ids)
 
-        for block in self.double_blocks:
+        for i, block in enumerate(self.double_blocks):
             img, txt = block(img=img, txt=txt, vec=vec, pe=pe)
 
+            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 block in self.single_blocks:
+
+        for i, block in enumerate(self.single_blocks):
             img = block(img, vec=vec, pe=pe)
+
+            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] :, ...]
 
         img = self.final_layer(img, vec)  # (N, T, patch_size ** 2 * out_channels)
         return img
 
-    def forward(self, x, timestep, context, y, guidance, **kwargs):
+    def forward(self, x, timestep, context, y, guidance, control=None, **kwargs):
         bs, c, h, w = x.shape
         patch_size = 2
-        pad_h = (patch_size - h % 2) % patch_size
-        pad_w = (patch_size - w % 2) % patch_size
-
-        x = torch.nn.functional.pad(x, (0, pad_w, 0, pad_h), mode='circular')
+        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)
 
@@ -140,5 +156,5 @@ class Flux(nn.Module):
         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, y, guidance)
+        out = self.forward_orig(img, img_ids, context, txt_ids, timestep, y, guidance, control)
         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/hydit/attn_layers.py

@@ -47,7 +47,7 @@ def reshape_for_broadcast(freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]], x
 
 
 def rotate_half(x):
-    x_real, x_imag = x.float().reshape(*x.shape[:-1], -1, 2).unbind(-1)  # [B, S, H, D//2]
+    x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1)  # [B, S, H, D//2]
     return torch.stack([-x_imag, x_real], dim=-1).flatten(3)
 
 
@@ -78,10 +78,9 @@ def apply_rotary_emb(
     xk_out = None
     if isinstance(freqs_cis, tuple):
         cos, sin = reshape_for_broadcast(freqs_cis, xq, head_first)    # [S, D]
-        cos, sin = cos.to(xq.device), sin.to(xq.device)
-        xq_out = (xq.float() * cos + rotate_half(xq.float()) * sin).type_as(xq)
+        xq_out = (xq * cos + rotate_half(xq) * sin)
         if xk is not None:
-            xk_out = (xk.float() * cos + rotate_half(xk.float()) * sin).type_as(xk)
+            xk_out = (xk * cos + rotate_half(xk) * sin)
     else:
         xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2))  # [B, S, H, D//2]
         freqs_cis = reshape_for_broadcast(freqs_cis, xq_, head_first).to(xq.device)   # [S, D//2] --> [1, S, 1, D//2]

comfy/ldm/hydit/controlnet.py

@@ -0,0 +1,321 @@
+from typing import Any, Optional
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from torch.utils import checkpoint
+
+from comfy.ldm.modules.diffusionmodules.mmdit import (
+    Mlp,
+    TimestepEmbedder,
+    PatchEmbed,
+    RMSNorm,
+)
+from comfy.ldm.modules.diffusionmodules.util import timestep_embedding
+from .poolers import AttentionPool
+
+import comfy.latent_formats
+from .models import HunYuanDiTBlock, calc_rope
+
+from .posemb_layers import get_2d_rotary_pos_embed, get_fill_resize_and_crop
+
+
+class HunYuanControlNet(nn.Module):
+    """
+    HunYuanDiT: Diffusion model with a Transformer backbone.
+
+    Inherit ModelMixin and ConfigMixin to be compatible with the sampler StableDiffusionPipeline of diffusers.
+
+    Inherit PeftAdapterMixin to be compatible with the PEFT training pipeline.
+
+    Parameters
+    ----------
+    args: argparse.Namespace
+        The arguments parsed by argparse.
+    input_size: tuple
+        The size of the input image.
+    patch_size: int
+        The size of the patch.
+    in_channels: int
+        The number of input channels.
+    hidden_size: int
+        The hidden size of the transformer backbone.
+    depth: int
+        The number of transformer blocks.
+    num_heads: int
+        The number of attention heads.
+    mlp_ratio: float
+        The ratio of the hidden size of the MLP in the transformer block.
+    log_fn: callable
+        The logging function.
+    """
+
+    def __init__(
+        self,
+        input_size: tuple = 128,
+        patch_size: int = 2,
+        in_channels: int = 4,
+        hidden_size: int = 1408,
+        depth: int = 40,
+        num_heads: int = 16,
+        mlp_ratio: float = 4.3637,
+        text_states_dim=1024,
+        text_states_dim_t5=2048,
+        text_len=77,
+        text_len_t5=256,
+        qk_norm=True,  # See http://arxiv.org/abs/2302.05442 for details.
+        size_cond=False,
+        use_style_cond=False,
+        learn_sigma=True,
+        norm="layer",
+        log_fn: callable = print,
+        attn_precision=None,
+        dtype=None,
+        device=None,
+        operations=None,
+        **kwargs,
+    ):
+        super().__init__()
+        self.log_fn = log_fn
+        self.depth = depth
+        self.learn_sigma = learn_sigma
+        self.in_channels = in_channels
+        self.out_channels = in_channels * 2 if learn_sigma else in_channels
+        self.patch_size = patch_size
+        self.num_heads = num_heads
+        self.hidden_size = hidden_size
+        self.text_states_dim = text_states_dim
+        self.text_states_dim_t5 = text_states_dim_t5
+        self.text_len = text_len
+        self.text_len_t5 = text_len_t5
+        self.size_cond = size_cond
+        self.use_style_cond = use_style_cond
+        self.norm = norm
+        self.dtype = dtype
+        self.latent_format = comfy.latent_formats.SDXL
+
+        self.mlp_t5 = nn.Sequential(
+            nn.Linear(
+                self.text_states_dim_t5,
+                self.text_states_dim_t5 * 4,
+                bias=True,
+                dtype=dtype,
+                device=device,
+            ),
+            nn.SiLU(),
+            nn.Linear(
+                self.text_states_dim_t5 * 4,
+                self.text_states_dim,
+                bias=True,
+                dtype=dtype,
+                device=device,
+            ),
+        )
+        # learnable replace
+        self.text_embedding_padding = nn.Parameter(
+            torch.randn(
+                self.text_len + self.text_len_t5,
+                self.text_states_dim,
+                dtype=dtype,
+                device=device,
+            )
+        )
+
+        # Attention pooling
+        pooler_out_dim = 1024
+        self.pooler = AttentionPool(
+            self.text_len_t5,
+            self.text_states_dim_t5,
+            num_heads=8,
+            output_dim=pooler_out_dim,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+
+        # Dimension of the extra input vectors
+        self.extra_in_dim = pooler_out_dim
+
+        if self.size_cond:
+            # Image size and crop size conditions
+            self.extra_in_dim += 6 * 256
+
+        if self.use_style_cond:
+            # Here we use a default learned embedder layer for future extension.
+            self.style_embedder = nn.Embedding(
+                1, hidden_size, dtype=dtype, device=device
+            )
+            self.extra_in_dim += hidden_size
+
+        # Text embedding for `add`
+        self.x_embedder = PatchEmbed(
+            input_size,
+            patch_size,
+            in_channels,
+            hidden_size,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+        self.t_embedder = TimestepEmbedder(
+            hidden_size, dtype=dtype, device=device, operations=operations
+        )
+        self.extra_embedder = nn.Sequential(
+            operations.Linear(
+                self.extra_in_dim, hidden_size * 4, dtype=dtype, device=device
+            ),
+            nn.SiLU(),
+            operations.Linear(
+                hidden_size * 4, hidden_size, bias=True, dtype=dtype, device=device
+            ),
+        )
+
+        # Image embedding
+        num_patches = self.x_embedder.num_patches
+
+        # HUnYuanDiT Blocks
+        self.blocks = nn.ModuleList(
+            [
+                HunYuanDiTBlock(
+                    hidden_size=hidden_size,
+                    c_emb_size=hidden_size,
+                    num_heads=num_heads,
+                    mlp_ratio=mlp_ratio,
+                    text_states_dim=self.text_states_dim,
+                    qk_norm=qk_norm,
+                    norm_type=self.norm,
+                    skip=False,
+                    attn_precision=attn_precision,
+                    dtype=dtype,
+                    device=device,
+                    operations=operations,
+                )
+                for _ in range(19)
+            ]
+        )
+
+        # Input zero linear for the first block
+        self.before_proj = operations.Linear(self.hidden_size, self.hidden_size, dtype=dtype, device=device)
+
+
+        # Output zero linear for the every block
+        self.after_proj_list = nn.ModuleList(
+            [
+
+                    operations.Linear(
+                        self.hidden_size, self.hidden_size, dtype=dtype, device=device
+                    )
+                for _ in range(len(self.blocks))
+            ]
+        )
+
+    def forward(
+        self,
+        x,
+        hint,
+        timesteps,
+        context,#encoder_hidden_states=None,
+        text_embedding_mask=None,
+        encoder_hidden_states_t5=None,
+        text_embedding_mask_t5=None,
+        image_meta_size=None,
+        style=None,
+        return_dict=False,
+        **kwarg,
+    ):
+        """
+        Forward pass of the encoder.
+
+        Parameters
+        ----------
+        x: torch.Tensor
+            (B, D, H, W)
+        t: torch.Tensor
+            (B)
+        encoder_hidden_states: torch.Tensor
+            CLIP text embedding, (B, L_clip, D)
+        text_embedding_mask: torch.Tensor
+            CLIP text embedding mask, (B, L_clip)
+        encoder_hidden_states_t5: torch.Tensor
+            T5 text embedding, (B, L_t5, D)
+        text_embedding_mask_t5: torch.Tensor
+            T5 text embedding mask, (B, L_t5)
+        image_meta_size: torch.Tensor
+            (B, 6)
+        style: torch.Tensor
+            (B)
+        cos_cis_img: torch.Tensor
+        sin_cis_img: torch.Tensor
+        return_dict: bool
+            Whether to return a dictionary.
+        """
+        condition = hint
+        if condition.shape[0] == 1:
+            condition = torch.repeat_interleave(condition, x.shape[0], dim=0)
+
+        text_states = context  # 2,77,1024
+        text_states_t5 = encoder_hidden_states_t5  # 2,256,2048
+        text_states_mask = text_embedding_mask.bool()  # 2,77
+        text_states_t5_mask = text_embedding_mask_t5.bool()  # 2,256
+        b_t5, l_t5, c_t5 = text_states_t5.shape
+        text_states_t5 = self.mlp_t5(text_states_t5.view(-1, c_t5)).view(b_t5, l_t5, -1)
+
+        padding = comfy.ops.cast_to_input(self.text_embedding_padding, text_states)
+
+        text_states[:, -self.text_len :] = torch.where(
+            text_states_mask[:, -self.text_len :].unsqueeze(2),
+            text_states[:, -self.text_len :],
+            padding[: self.text_len],
+        )
+        text_states_t5[:, -self.text_len_t5 :] = torch.where(
+            text_states_t5_mask[:, -self.text_len_t5 :].unsqueeze(2),
+            text_states_t5[:, -self.text_len_t5 :],
+            padding[self.text_len :],
+        )
+
+        text_states = torch.cat([text_states, text_states_t5], dim=1)  # 2,205，1024
+
+        # _, _, oh, ow = x.shape
+        # th, tw = oh // self.patch_size, ow // self.patch_size
+
+        # Get image RoPE embedding according to `reso`lution.
+        freqs_cis_img = calc_rope(
+            x, self.patch_size, self.hidden_size // self.num_heads
+        )  # (cos_cis_img, sin_cis_img)
+
+        # ========================= Build time and image embedding =========================
+        t = self.t_embedder(timesteps, dtype=self.dtype)
+        x = self.x_embedder(x)
+
+        # ========================= Concatenate all extra vectors =========================
+        # Build text tokens with pooling
+        extra_vec = self.pooler(encoder_hidden_states_t5)
+
+        # Build image meta size tokens if applicable
+        # if image_meta_size is not None:
+        #     image_meta_size = timestep_embedding(image_meta_size.view(-1), 256)   # [B * 6, 256]
+        #     if image_meta_size.dtype != self.dtype:
+        #         image_meta_size = image_meta_size.half()
+        #     image_meta_size = image_meta_size.view(-1, 6 * 256)
+        #     extra_vec = torch.cat([extra_vec, image_meta_size], dim=1)  # [B, D + 6 * 256]
+
+        # Build style tokens
+        if style is not None:
+            style_embedding = self.style_embedder(style)
+            extra_vec = torch.cat([extra_vec, style_embedding], dim=1)
+
+        # Concatenate all extra vectors
+        c = t + self.extra_embedder(extra_vec)  # [B, D]
+
+        # ========================= Deal with Condition =========================
+        condition = self.x_embedder(condition)
+
+        # ========================= Forward pass through HunYuanDiT blocks =========================
+        controls = []
+        x = x + self.before_proj(condition)  # add condition
+        for layer, block in enumerate(self.blocks):
+            x = block(x, c, text_states, freqs_cis_img)
+            controls.append(self.after_proj_list[layer](x))  # zero linear for output
+
+        return {"output": controls}

comfy/ldm/hydit/models.py

@@ -21,6 +21,7 @@ def calc_rope(x, patch_size, head_size):
     sub_args = [start, stop, (th, tw)]
     # head_size = HUNYUAN_DIT_CONFIG['DiT-g/2']['hidden_size'] // HUNYUAN_DIT_CONFIG['DiT-g/2']['num_heads']
     rope = get_2d_rotary_pos_embed(head_size, *sub_args)
+    rope = (rope[0].to(x), rope[1].to(x))
     return rope
 
 
@@ -91,6 +92,8 @@ class HunYuanDiTBlock(nn.Module):
         # Long Skip Connection
         if self.skip_linear is not None:
             cat = torch.cat([x, skip], dim=-1)
+            if cat.dtype != x.dtype:
+                cat = cat.to(x.dtype)
             cat = self.skip_norm(cat)
             x = self.skip_linear(cat)
 
@@ -362,6 +365,8 @@ class HunYuanDiT(nn.Module):
         c = t + self.extra_embedder(extra_vec)  # [B, D]
 
         controls = None
+        if control:
+            controls = control.get("output", None)
         # ========================= Forward pass through HunYuanDiT blocks =========================
         skips = []
         for layer, block in enumerate(self.blocks):

comfy/ldm/modules/attention.py

@@ -358,7 +358,7 @@ def attention_xformers(q, k, v, heads, mask=None, attn_precision=None, skip_resh
             disabled_xformers = True
 
     if disabled_xformers:
-        return attention_pytorch(q, k, v, heads, mask)
+        return attention_pytorch(q, k, v, heads, mask, skip_reshape=skip_reshape)
 
     if skip_reshape:
          q, k, v = map(

comfy/ldm/modules/diffusionmodules/mmdit.py

@@ -9,6 +9,7 @@ from .. import attention
 from einops import rearrange, repeat
 from .util import timestep_embedding
 import comfy.ops
+import comfy.ldm.common_dit
 
 def default(x, y):
     if x is not None:
@@ -111,9 +112,7 @@ class PatchEmbed(nn.Module):
         #             f"Input width ({W}) should be divisible by patch size ({self.patch_size[1]})."
         #         )
         if self.dynamic_img_pad:
-            pad_h = (self.patch_size[0] - H % self.patch_size[0]) % self.patch_size[0]
-            pad_w = (self.patch_size[1] - W % self.patch_size[1]) % self.patch_size[1]
-            x = torch.nn.functional.pad(x, (0, pad_w, 0, pad_h), mode=self.padding_mode)
+            x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size, padding_mode=self.padding_mode)
         x = self.proj(x)
         if self.flatten:
             x = x.flatten(2).transpose(1, 2)  # NCHW -> NLC

comfy/lora.py

@@ -1,5 +1,27 @@
+"""
+    This file is part of ComfyUI.
+    Copyright (C) 2024 Comfy
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+"""
+
+from __future__ import annotations
 import comfy.utils
+import comfy.model_management
+import comfy.model_base
 import logging
+import torch
 
 LORA_CLIP_MAP = {
     "mlp.fc1": "mlp_fc1",
@@ -218,11 +240,17 @@ def model_lora_keys_clip(model, key_map={}):
                     lora_key = "lora_prior_te_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c]) #cascade lora: TODO put lora key prefix in the model config
                     key_map[lora_key] = k
 
-    for k in sdk: #OneTrainer SD3 lora
-        if k.startswith("t5xxl.transformer.") and k.endswith(".weight"):
-            l_key = k[len("t5xxl.transformer."):-len(".weight")]
-            lora_key = "lora_te3_{}".format(l_key.replace(".", "_"))
-            key_map[lora_key] = k
+    for k in sdk:
+        if k.endswith(".weight"):
+            if k.startswith("t5xxl.transformer."):#OneTrainer SD3 lora
+                l_key = k[len("t5xxl.transformer."):-len(".weight")]
+                lora_key = "lora_te3_{}".format(l_key.replace(".", "_"))
+                key_map[lora_key] = k
+            elif k.startswith("hydit_clip.transformer.bert."): #HunyuanDiT Lora
+                l_key = k[len("hydit_clip.transformer.bert."):-len(".weight")]
+                lora_key = "lora_te1_{}".format(l_key.replace(".", "_"))
+                key_map[lora_key] = k
+
 
     k = "clip_g.transformer.text_projection.weight"
     if k in sdk:
@@ -245,6 +273,7 @@ def model_lora_keys_unet(model, key_map={}):
             key_lora = k[len("diffusion_model."):-len(".weight")].replace(".", "_")
             key_map["lora_unet_{}".format(key_lora)] = k
             key_map["lora_prior_unet_{}".format(key_lora)] = k #cascade lora: TODO put lora key prefix in the model config
+            key_map["{}".format(k[:-len(".weight")])] = k #generic lora format without any weird key names
 
     diffusers_keys = comfy.utils.unet_to_diffusers(model.model_config.unet_config)
     for k in diffusers_keys:
@@ -288,4 +317,240 @@ def model_lora_keys_unet(model, key_map={}):
                 key_lora = k[len("diffusion_model."):-len(".weight")]
                 key_map["base_model.model.{}".format(key_lora)] = k #official hunyuan lora format
 
+    if isinstance(model, comfy.model_base.Flux): #Diffusers lora Flux
+        diffusers_keys = comfy.utils.flux_to_diffusers(model.model_config.unet_config, output_prefix="diffusion_model.")
+        for k in diffusers_keys:
+            if k.endswith(".weight"):
+                to = diffusers_keys[k]
+                key_map["transformer.{}".format(k[:-len(".weight")])] = to #simpletrainer and probably regular diffusers flux lora format
+                key_map["lycoris_{}".format(k[:-len(".weight")].replace(".", "_"))] = to #simpletrainer lycoris
+
     return key_map
+
+
+def weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype):
+    dora_scale = comfy.model_management.cast_to_device(dora_scale, weight.device, intermediate_dtype)
+    lora_diff *= alpha
+    weight_calc = weight + lora_diff.type(weight.dtype)
+    weight_norm = (
+        weight_calc.transpose(0, 1)
+        .reshape(weight_calc.shape[1], -1)
+        .norm(dim=1, keepdim=True)
+        .reshape(weight_calc.shape[1], *[1] * (weight_calc.dim() - 1))
+        .transpose(0, 1)
+    )
+
+    weight_calc *= (dora_scale / weight_norm).type(weight.dtype)
+    if strength != 1.0:
+        weight_calc -= weight
+        weight += strength * (weight_calc)
+    else:
+        weight[:] = weight_calc
+    return weight
+
+def pad_tensor_to_shape(tensor: torch.Tensor, new_shape: list[int]) -> torch.Tensor:
+    """
+    Pad a tensor to a new shape with zeros.
+
+    Args:
+        tensor (torch.Tensor): The original tensor to be padded.
+        new_shape (List[int]): The desired shape of the padded tensor.
+
+    Returns:
+        torch.Tensor: A new tensor padded with zeros to the specified shape.
+
+    Note:
+        If the new shape is smaller than the original tensor in any dimension,
+        the original tensor will be truncated in that dimension.
+    """
+    if any([new_shape[i] < tensor.shape[i] for i in range(len(new_shape))]):
+        raise ValueError("The new shape must be larger than the original tensor in all dimensions")
+
+    if len(new_shape) != len(tensor.shape):
+        raise ValueError("The new shape must have the same number of dimensions as the original tensor")
+
+    # Create a new tensor filled with zeros
+    padded_tensor = torch.zeros(new_shape, dtype=tensor.dtype, device=tensor.device)
+
+    # Create slicing tuples for both tensors
+    orig_slices = tuple(slice(0, dim) for dim in tensor.shape)
+    new_slices = tuple(slice(0, dim) for dim in tensor.shape)
+
+    # Copy the original tensor into the new tensor
+    padded_tensor[new_slices] = tensor[orig_slices]
+
+    return padded_tensor
+
+def calculate_weight(patches, weight, key, intermediate_dtype=torch.float32):
+    for p in patches:
+        strength = p[0]
+        v = p[1]
+        strength_model = p[2]
+        offset = p[3]
+        function = p[4]
+        if function is None:
+            function = lambda a: a
+
+        old_weight = None
+        if offset is not None:
+            old_weight = weight
+            weight = weight.narrow(offset[0], offset[1], offset[2])
+
+        if strength_model != 1.0:
+            weight *= strength_model
+
+        if isinstance(v, list):
+            v = (calculate_weight(v[1:], v[0].clone(), key, intermediate_dtype=intermediate_dtype), )
+
+        if len(v) == 1:
+            patch_type = "diff"
+        elif len(v) == 2:
+            patch_type = v[0]
+            v = v[1]
+
+        if patch_type == "diff":
+            diff: torch.Tensor = v[0]
+            # An extra flag to pad the weight if the diff's shape is larger than the weight
+            do_pad_weight = len(v) > 1 and v[1]['pad_weight']
+            if do_pad_weight and diff.shape != weight.shape:
+                logging.info("Pad weight {} from {} to shape: {}".format(key, weight.shape, diff.shape))
+                weight = pad_tensor_to_shape(weight, diff.shape)
+
+            if strength != 0.0:
+                if diff.shape != weight.shape:
+                    logging.warning("WARNING SHAPE MISMATCH {} WEIGHT NOT MERGED {} != {}".format(key, diff.shape, weight.shape))
+                else:
+                    weight += function(strength * comfy.model_management.cast_to_device(diff, weight.device, weight.dtype))
+        elif patch_type == "lora": #lora/locon
+            mat1 = comfy.model_management.cast_to_device(v[0], weight.device, intermediate_dtype)
+            mat2 = comfy.model_management.cast_to_device(v[1], weight.device, intermediate_dtype)
+            dora_scale = v[4]
+            if v[2] is not None:
+                alpha = v[2] / mat2.shape[0]
+            else:
+                alpha = 1.0
+
+            if v[3] is not None:
+                #locon mid weights, hopefully the math is fine because I didn't properly test it
+                mat3 = comfy.model_management.cast_to_device(v[3], weight.device, intermediate_dtype)
+                final_shape = [mat2.shape[1], mat2.shape[0], mat3.shape[2], mat3.shape[3]]
+                mat2 = torch.mm(mat2.transpose(0, 1).flatten(start_dim=1), mat3.transpose(0, 1).flatten(start_dim=1)).reshape(final_shape).transpose(0, 1)
+            try:
+                lora_diff = torch.mm(mat1.flatten(start_dim=1), mat2.flatten(start_dim=1)).reshape(weight.shape)
+                if dora_scale is not None:
+                    weight = function(weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype))
+                else:
+                    weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
+            except Exception as e:
+                logging.error("ERROR {} {} {}".format(patch_type, key, e))
+        elif patch_type == "lokr":
+            w1 = v[0]
+            w2 = v[1]
+            w1_a = v[3]
+            w1_b = v[4]
+            w2_a = v[5]
+            w2_b = v[6]
+            t2 = v[7]
+            dora_scale = v[8]
+            dim = None
+
+            if w1 is None:
+                dim = w1_b.shape[0]
+                w1 = torch.mm(comfy.model_management.cast_to_device(w1_a, weight.device, intermediate_dtype),
+                                comfy.model_management.cast_to_device(w1_b, weight.device, intermediate_dtype))
+            else:
+                w1 = comfy.model_management.cast_to_device(w1, weight.device, intermediate_dtype)
+
+            if w2 is None:
+                dim = w2_b.shape[0]
+                if t2 is None:
+                    w2 = torch.mm(comfy.model_management.cast_to_device(w2_a, weight.device, intermediate_dtype),
+                                    comfy.model_management.cast_to_device(w2_b, weight.device, intermediate_dtype))
+                else:
+                    w2 = torch.einsum('i j k l, j r, i p -> p r k l',
+                                        comfy.model_management.cast_to_device(t2, weight.device, intermediate_dtype),
+                                        comfy.model_management.cast_to_device(w2_b, weight.device, intermediate_dtype),
+                                        comfy.model_management.cast_to_device(w2_a, weight.device, intermediate_dtype))
+            else:
+                w2 = comfy.model_management.cast_to_device(w2, weight.device, intermediate_dtype)
+
+            if len(w2.shape) == 4:
+                w1 = w1.unsqueeze(2).unsqueeze(2)
+            if v[2] is not None and dim is not None:
+                alpha = v[2] / dim
+            else:
+                alpha = 1.0
+
+            try:
+                lora_diff = torch.kron(w1, w2).reshape(weight.shape)
+                if dora_scale is not None:
+                    weight = function(weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype))
+                else:
+                    weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
+            except Exception as e:
+                logging.error("ERROR {} {} {}".format(patch_type, key, e))
+        elif patch_type == "loha":
+            w1a = v[0]
+            w1b = v[1]
+            if v[2] is not None:
+                alpha = v[2] / w1b.shape[0]
+            else:
+                alpha = 1.0
+
+            w2a = v[3]
+            w2b = v[4]
+            dora_scale = v[7]
+            if v[5] is not None: #cp decomposition
+                t1 = v[5]
+                t2 = v[6]
+                m1 = torch.einsum('i j k l, j r, i p -> p r k l',
+                                    comfy.model_management.cast_to_device(t1, weight.device, intermediate_dtype),
+                                    comfy.model_management.cast_to_device(w1b, weight.device, intermediate_dtype),
+                                    comfy.model_management.cast_to_device(w1a, weight.device, intermediate_dtype))
+
+                m2 = torch.einsum('i j k l, j r, i p -> p r k l',
+                                    comfy.model_management.cast_to_device(t2, weight.device, intermediate_dtype),
+                                    comfy.model_management.cast_to_device(w2b, weight.device, intermediate_dtype),
+                                    comfy.model_management.cast_to_device(w2a, weight.device, intermediate_dtype))
+            else:
+                m1 = torch.mm(comfy.model_management.cast_to_device(w1a, weight.device, intermediate_dtype),
+                                comfy.model_management.cast_to_device(w1b, weight.device, intermediate_dtype))
+                m2 = torch.mm(comfy.model_management.cast_to_device(w2a, weight.device, intermediate_dtype),
+                                comfy.model_management.cast_to_device(w2b, weight.device, intermediate_dtype))
+
+            try:
+                lora_diff = (m1 * m2).reshape(weight.shape)
+                if dora_scale is not None:
+                    weight = function(weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype))
+                else:
+                    weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
+            except Exception as e:
+                logging.error("ERROR {} {} {}".format(patch_type, key, e))
+        elif patch_type == "glora":
+            if v[4] is not None:
+                alpha = v[4] / v[0].shape[0]
+            else:
+                alpha = 1.0
+
+            dora_scale = v[5]
+
+            a1 = comfy.model_management.cast_to_device(v[0].flatten(start_dim=1), weight.device, intermediate_dtype)
+            a2 = comfy.model_management.cast_to_device(v[1].flatten(start_dim=1), weight.device, intermediate_dtype)
+            b1 = comfy.model_management.cast_to_device(v[2].flatten(start_dim=1), weight.device, intermediate_dtype)
+            b2 = comfy.model_management.cast_to_device(v[3].flatten(start_dim=1), weight.device, intermediate_dtype)
+
+            try:
+                lora_diff = (torch.mm(b2, b1) + torch.mm(torch.mm(weight.flatten(start_dim=1), a2), a1)).reshape(weight.shape)
+                if dora_scale is not None:
+                    weight = function(weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype))
+                else:
+                    weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
+            except Exception as e:
+                logging.error("ERROR {} {} {}".format(patch_type, key, e))
+        else:
+            logging.warning("patch type not recognized {} {}".format(patch_type, key))
+
+        if old_weight is not None:
+            weight = old_weight
+
+    return weight

comfy/model_base.py

@@ -1,3 +1,21 @@
+"""
+    This file is part of ComfyUI.
+    Copyright (C) 2024 Comfy
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+"""
+
 import torch
 import logging
 from comfy.ldm.modules.diffusionmodules.openaimodel import UNetModel, Timestep
@@ -74,16 +92,18 @@ class BaseModel(torch.nn.Module):
         self.latent_format = model_config.latent_format
         self.model_config = model_config
         self.manual_cast_dtype = model_config.manual_cast_dtype
+        self.device = device
 
         if not unet_config.get("disable_unet_model_creation", False):
-            if self.manual_cast_dtype is not None:
-                operations = comfy.ops.manual_cast
+            if model_config.custom_operations is None:
+                operations = comfy.ops.pick_operations(unet_config.get("dtype", None), self.manual_cast_dtype)
             else:
-                operations = comfy.ops.disable_weight_init
+                operations = model_config.custom_operations
             self.diffusion_model = unet_model(**unet_config, device=device, operations=operations)
             if comfy.model_management.force_channels_last():
                 self.diffusion_model.to(memory_format=torch.channels_last)
                 logging.debug("using channels last mode for diffusion model")
+            logging.info("model weight dtype {}, manual cast: {}".format(self.get_dtype(), self.manual_cast_dtype))
         self.model_type = model_type
         self.model_sampling = model_sampling(model_config, model_type)
 
@@ -94,6 +114,7 @@ class BaseModel(torch.nn.Module):
         self.concat_keys = ()
         logging.info("model_type {}".format(model_type.name))
         logging.debug("adm {}".format(self.adm_channels))
+        self.memory_usage_factor = model_config.memory_usage_factor
 
     def apply_model(self, x, t, c_concat=None, c_crossattn=None, control=None, transformer_options={}, **kwargs):
         sigma = t
@@ -252,11 +273,11 @@ class BaseModel(torch.nn.Module):
                 dtype = self.manual_cast_dtype
             #TODO: this needs to be tweaked
             area = input_shape[0] * math.prod(input_shape[2:])
-            return (area * comfy.model_management.dtype_size(dtype) / 50) * (1024 * 1024)
+            return (area * comfy.model_management.dtype_size(dtype) * 0.01 * self.memory_usage_factor) * (1024 * 1024)
         else:
             #TODO: this formula might be too aggressive since I tweaked the sub-quad and split algorithms to use less memory.
             area = input_shape[0] * math.prod(input_shape[2:])
-            return (area * 0.3) * (1024 * 1024)
+            return (area * 0.15 * self.memory_usage_factor) * (1024 * 1024)
 
 
 def unclip_adm(unclip_conditioning, device, noise_augmentor, noise_augment_merge=0.0, seed=None):
@@ -354,6 +375,7 @@ class SDXL(BaseModel):
         flat = torch.flatten(torch.cat(out)).unsqueeze(dim=0).repeat(clip_pooled.shape[0], 1)
         return torch.cat((clip_pooled.to(flat.device), flat), dim=1)
 
+
 class SVD_img2vid(BaseModel):
     def __init__(self, model_config, model_type=ModelType.V_PREDICTION_EDM, device=None):
         super().__init__(model_config, model_type, device=device)
@@ -594,17 +616,6 @@ class SD3(BaseModel):
             out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
         return out
 
-    def memory_required(self, input_shape):
-        if comfy.model_management.xformers_enabled() or comfy.model_management.pytorch_attention_flash_attention():
-            dtype = self.get_dtype()
-            if self.manual_cast_dtype is not None:
-                dtype = self.manual_cast_dtype
-            #TODO: this probably needs to be tweaked
-            area = input_shape[0] * input_shape[2] * input_shape[3]
-            return (area * comfy.model_management.dtype_size(dtype) * 0.012) * (1024 * 1024)
-        else:
-            area = input_shape[0] * input_shape[2] * input_shape[3]
-            return (area * 0.3) * (1024 * 1024)
 
 class AuraFlow(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
@@ -702,15 +713,3 @@ class Flux(BaseModel):
             out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
         out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor([kwargs.get("guidance", 3.5)]))
         return out
-
-    def memory_required(self, input_shape):
-        if comfy.model_management.xformers_enabled() or comfy.model_management.pytorch_attention_flash_attention():
-            dtype = self.get_dtype()
-            if self.manual_cast_dtype is not None:
-                dtype = self.manual_cast_dtype
-            #TODO: this probably needs to be tweaked
-            area = input_shape[0] * input_shape[2] * input_shape[3]
-            return (area * comfy.model_management.dtype_size(dtype) * 0.020) * (1024 * 1024)
-        else:
-            area = input_shape[0] * input_shape[2] * input_shape[3]
-            return (area * 0.3) * (1024 * 1024)

comfy/model_detection.py

@@ -131,14 +131,14 @@ def detect_unet_config(state_dict, key_prefix):
     if '{}double_blocks.0.img_attn.norm.key_norm.scale'.format(key_prefix) in state_dict_keys: #Flux
         dit_config = {}
         dit_config["image_model"] = "flux"
-        dit_config["in_channels"] = 64
+        dit_config["in_channels"] = 16
         dit_config["vec_in_dim"] = 768
         dit_config["context_in_dim"] = 4096
         dit_config["hidden_size"] = 3072
         dit_config["mlp_ratio"] = 4.0
         dit_config["num_heads"] = 24
-        dit_config["depth"] = 19
-        dit_config["depth_single_blocks"] = 38
+        dit_config["depth"] = count_blocks(state_dict_keys, '{}double_blocks.'.format(key_prefix) + '{}.')
+        dit_config["depth_single_blocks"] = count_blocks(state_dict_keys, '{}single_blocks.'.format(key_prefix) + '{}.')
         dit_config["axes_dim"] = [16, 56, 56]
         dit_config["theta"] = 10000
         dit_config["qkv_bias"] = True
@@ -472,9 +472,15 @@ def unet_config_from_diffusers_unet(state_dict, dtype=None):
             'transformer_depth': [0, 1, 1], 'channel_mult': [1, 2, 4], 'transformer_depth_middle': -2, 'use_linear_in_transformer': False,
             'context_dim': 768, 'num_head_channels': 64, 'transformer_depth_output': [0, 0, 1, 1, 1, 1],
             'use_temporal_attention': False, 'use_temporal_resblock': False}
+    
+    SD15_diffusers_inpaint = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False, 'adm_in_channels': None,
+            'dtype': dtype, 'in_channels': 9, 'model_channels': 320, 'num_res_blocks': [2, 2, 2, 2], 'transformer_depth': [1, 1, 1, 1, 1, 1, 0, 0],
+            'channel_mult': [1, 2, 4, 4], 'transformer_depth_middle': 1, 'use_linear_in_transformer': False, 'context_dim': 768, 'num_heads': 8,
+            'transformer_depth_output': [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0],
+            'use_temporal_attention': False, 'use_temporal_resblock': False}  
 
 
-    supported_models = [SDXL, SDXL_refiner, SD21, SD15, SD21_uncliph, SD21_unclipl, SDXL_mid_cnet, SDXL_small_cnet, SDXL_diffusers_inpaint, SSD_1B, Segmind_Vega, KOALA_700M, KOALA_1B, SD09_XS, SD_XS, SDXL_diffusers_ip2p]
+    supported_models = [SDXL, SDXL_refiner, SD21, SD15, SD21_uncliph, SD21_unclipl, SDXL_mid_cnet, SDXL_small_cnet, SDXL_diffusers_inpaint, SSD_1B, Segmind_Vega, KOALA_700M, KOALA_1B, SD09_XS, SD_XS, SDXL_diffusers_ip2p, SD15_diffusers_inpaint]
 
     for unet_config in supported_models:
         matches = True
@@ -495,7 +501,12 @@ def model_config_from_diffusers_unet(state_dict):
 def convert_diffusers_mmdit(state_dict, output_prefix=""):
     out_sd = {}
 
-    if 'transformer_blocks.0.attn.add_q_proj.weight' in state_dict: #SD3
+    if 'transformer_blocks.0.attn.norm_added_k.weight' in state_dict: #Flux
+        depth = count_blocks(state_dict, 'transformer_blocks.{}.')
+        depth_single_blocks = count_blocks(state_dict, 'single_transformer_blocks.{}.')
+        hidden_size = state_dict["x_embedder.bias"].shape[0]
+        sd_map = comfy.utils.flux_to_diffusers({"depth": depth, "depth_single_blocks": depth_single_blocks, "hidden_size": hidden_size}, output_prefix=output_prefix)
+    elif 'transformer_blocks.0.attn.add_q_proj.weight' in state_dict: #SD3
         num_blocks = count_blocks(state_dict, 'transformer_blocks.{}.')
         depth = state_dict["pos_embed.proj.weight"].shape[0] // 64
         sd_map = comfy.utils.mmdit_to_diffusers({"depth": depth, "num_blocks": num_blocks}, output_prefix=output_prefix)
@@ -521,7 +532,12 @@ def convert_diffusers_mmdit(state_dict, output_prefix=""):
                     old_weight = out_sd.get(t[0], None)
                     if old_weight is None:
                         old_weight = torch.empty_like(weight)
-                        old_weight = old_weight.repeat([3] + [1] * (len(old_weight.shape) - 1))
+                    if old_weight.shape[offset[0]] < offset[1] + offset[2]:
+                        exp = list(weight.shape)
+                        exp[offset[0]] = offset[1] + offset[2]
+                        new = torch.empty(exp, device=weight.device, dtype=weight.dtype)
+                        new[:old_weight.shape[0]] = old_weight
+                        old_weight = new
 
                     w = old_weight.narrow(offset[0], offset[1], offset[2])
                 else:

comfy/model_management.py

@@ -1,3 +1,21 @@
+"""
+    This file is part of ComfyUI.
+    Copyright (C) 2024 Comfy
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+"""
+
 import psutil
 import logging
 from enum import Enum
@@ -26,9 +44,14 @@ cpu_state = CPUState.GPU
 
 total_vram = 0
 
-lowvram_available = True
 xpu_available = False
+try:
+    torch_version = torch.version.__version__
+    xpu_available = (int(torch_version[0]) < 2 or (int(torch_version[0]) == 2 and int(torch_version[2]) <= 4)) and torch.xpu.is_available()
+except:
+    pass
 
+lowvram_available = True
 if args.deterministic:
     logging.info("Using deterministic algorithms for pytorch")
     torch.use_deterministic_algorithms(True, warn_only=True)
@@ -48,10 +71,10 @@ if args.directml is not None:
 
 try:
     import intel_extension_for_pytorch as ipex
-    if torch.xpu.is_available():
-        xpu_available = True
+    _ = torch.xpu.device_count()
+    xpu_available = torch.xpu.is_available()
 except:
-    pass
+    xpu_available = xpu_available or (hasattr(torch, "xpu") and torch.xpu.is_available())
 
 try:
     if torch.backends.mps.is_available():
@@ -171,7 +194,6 @@ VAE_DTYPES = [torch.float32]
 
 try:
     if is_nvidia():
-        torch_version = torch.version.__version__
         if int(torch_version[0]) >= 2:
             if ENABLE_PYTORCH_ATTENTION == False and args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
                 ENABLE_PYTORCH_ATTENTION = True
@@ -273,9 +295,12 @@ class LoadedModel:
     def model_memory(self):
         return self.model.model_size()
 
+    def model_offloaded_memory(self):
+        return self.model.model_size() - self.model.loaded_size()
+
     def model_memory_required(self, device):
-        if device == self.model.current_device:
-            return 0
+        if device == self.model.current_loaded_device():
+            return self.model_offloaded_memory()
         else:
             return self.model_memory()
 
@@ -287,39 +312,77 @@ class LoadedModel:
 
         load_weights = not self.weights_loaded
 
-        try:
-            if lowvram_model_memory > 0 and load_weights:
-                self.real_model = self.model.patch_model_lowvram(device_to=patch_model_to, lowvram_model_memory=lowvram_model_memory, force_patch_weights=force_patch_weights)
-            else:
-                self.real_model = self.model.patch_model(device_to=patch_model_to, patch_weights=load_weights)
-        except Exception as e:
-            self.model.unpatch_model(self.model.offload_device)
-            self.model_unload()
-            raise e
+        if self.model.loaded_size() > 0:
+            use_more_vram = lowvram_model_memory
+            if use_more_vram == 0:
+                use_more_vram = 1e32
+            self.model_use_more_vram(use_more_vram)
+        else:
+            try:
+                self.real_model = self.model.patch_model(device_to=patch_model_to, lowvram_model_memory=lowvram_model_memory, load_weights=load_weights, force_patch_weights=force_patch_weights)
+            except Exception as e:
+                self.model.unpatch_model(self.model.offload_device)
+                self.model_unload()
+                raise e
 
-        if is_intel_xpu() and not args.disable_ipex_optimize:
-            self.real_model = ipex.optimize(self.real_model.eval(), graph_mode=True, concat_linear=True)
+        if is_intel_xpu() and not args.disable_ipex_optimize and self.real_model is not None:
+            with torch.no_grad():
+                self.real_model = ipex.optimize(self.real_model.eval(), inplace=True, graph_mode=True, concat_linear=True)
 
         self.weights_loaded = True
         return self.real_model
 
     def should_reload_model(self, force_patch_weights=False):
-        if force_patch_weights and self.model.lowvram_patch_counter > 0:
+        if force_patch_weights and self.model.lowvram_patch_counter() > 0:
             return True
         return False
 
-    def model_unload(self, unpatch_weights=True):
+    def model_unload(self, memory_to_free=None, unpatch_weights=True):
+        if memory_to_free is not None:
+            if memory_to_free < self.model.loaded_size():
+                freed = self.model.partially_unload(self.model.offload_device, memory_to_free)
+                if freed >= memory_to_free:
+                    return False
         self.model.unpatch_model(self.model.offload_device, unpatch_weights=unpatch_weights)
         self.model.model_patches_to(self.model.offload_device)
         self.weights_loaded = self.weights_loaded and not unpatch_weights
         self.real_model = None
+        return True
+
+    def model_use_more_vram(self, extra_memory):
+        return self.model.partially_load(self.device, extra_memory)
 
     def __eq__(self, other):
         return self.model is other.model
 
+def use_more_memory(extra_memory, loaded_models, device):
+    for m in loaded_models:
+        if m.device == device:
+            extra_memory -= m.model_use_more_vram(extra_memory)
+            if extra_memory <= 0:
+                break
+
+def offloaded_memory(loaded_models, device):
+    offloaded_mem = 0
+    for m in loaded_models:
+        if m.device == device:
+            offloaded_mem += m.model_offloaded_memory()
+    return offloaded_mem
+
 def minimum_inference_memory():
     return (1024 * 1024 * 1024) * 1.2
 
+EXTRA_RESERVED_VRAM = 200 * 1024 * 1024
+if any(platform.win32_ver()):
+    EXTRA_RESERVED_VRAM = 500 * 1024 * 1024 #Windows is higher because of the shared vram issue
+
+if args.reserve_vram is not None:
+    EXTRA_RESERVED_VRAM = args.reserve_vram * 1024 * 1024 * 1024
+    logging.debug("Reserving {}MB vram for other applications.".format(EXTRA_RESERVED_VRAM / (1024 * 1024)))
+
+def extra_reserved_memory():
+    return EXTRA_RESERVED_VRAM
+
 def unload_model_clones(model, unload_weights_only=True, force_unload=True):
     to_unload = []
     for i in range(len(current_loaded_models)):
@@ -342,6 +405,8 @@ def unload_model_clones(model, unload_weights_only=True, force_unload=True):
     if not force_unload:
         if unload_weights_only and unload_weight == False:
             return None
+    else:
+        unload_weight = True
 
     for i in to_unload:
         logging.debug("unload clone {} {}".format(i, unload_weight))
@@ -352,6 +417,7 @@ def unload_model_clones(model, unload_weights_only=True, force_unload=True):
 def free_memory(memory_required, device, keep_loaded=[]):
     unloaded_model = []
     can_unload = []
+    unloaded_models = []
 
     for i in range(len(current_loaded_models) -1, -1, -1):
         shift_model = current_loaded_models[i]
@@ -362,14 +428,18 @@ def free_memory(memory_required, device, keep_loaded=[]):
 
     for x in sorted(can_unload):
         i = x[-1]
+        memory_to_free = None
         if not DISABLE_SMART_MEMORY:
-            if get_free_memory(device) > memory_required:
+            free_mem = get_free_memory(device)
+            if free_mem > memory_required:
                 break
-        current_loaded_models[i].model_unload()
-        unloaded_model.append(i)
+            memory_to_free = memory_required - free_mem
+        logging.debug(f"Unloading {current_loaded_models[i].model.model.__class__.__name__}")
+        if current_loaded_models[i].model_unload(memory_to_free):
+            unloaded_model.append(i)
 
     for i in sorted(unloaded_model, reverse=True):
-        current_loaded_models.pop(i)
+        unloaded_models.append(current_loaded_models.pop(i))
 
     if len(unloaded_model) > 0:
         soft_empty_cache()
@@ -378,16 +448,17 @@ def free_memory(memory_required, device, keep_loaded=[]):
             mem_free_total, mem_free_torch = get_free_memory(device, torch_free_too=True)
             if mem_free_torch > mem_free_total * 0.25:
                 soft_empty_cache()
+    return unloaded_models
 
-def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None):
+def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None, force_full_load=False):
     global vram_state
 
     inference_memory = minimum_inference_memory()
-    extra_mem = max(inference_memory, memory_required)
+    extra_mem = max(inference_memory, memory_required + extra_reserved_memory())
     if minimum_memory_required is None:
         minimum_memory_required = extra_mem
     else:
-        minimum_memory_required = max(inference_memory, minimum_memory_required)
+        minimum_memory_required = max(inference_memory, minimum_memory_required + extra_reserved_memory())
 
     models = set(models)
 
@@ -420,25 +491,36 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
         devs = set(map(lambda a: a.device, models_already_loaded))
         for d in devs:
             if d != torch.device("cpu"):
-                free_memory(extra_mem, d, models_already_loaded)
-        return
+                free_memory(extra_mem + offloaded_memory(models_already_loaded, d), d, models_already_loaded)
+                free_mem = get_free_memory(d)
+                if free_mem < minimum_memory_required:
+                    logging.info("Unloading models for lowram load.") #TODO: partial model unloading when this case happens, also handle the opposite case where models can be unlowvramed.
+                    models_to_load = free_memory(minimum_memory_required, d)
+                    logging.info("{} models unloaded.".format(len(models_to_load)))
+                else:
+                    use_more_memory(free_mem - minimum_memory_required, models_already_loaded, d)
+        if len(models_to_load) == 0:
+            return
 
     logging.info(f"Loading {len(models_to_load)} new model{'s' if len(models_to_load) > 1 else ''}")
 
     total_memory_required = {}
     for loaded_model in models_to_load:
-        if unload_model_clones(loaded_model.model, unload_weights_only=True, force_unload=False) == True:#unload clones where the weights are different
-            total_memory_required[loaded_model.device] = total_memory_required.get(loaded_model.device, 0) + loaded_model.model_memory_required(loaded_model.device)
+        unload_model_clones(loaded_model.model, unload_weights_only=True, force_unload=False) #unload clones where the weights are different
+        total_memory_required[loaded_model.device] = total_memory_required.get(loaded_model.device, 0) + loaded_model.model_memory_required(loaded_model.device)
 
-    for device in total_memory_required:
-        if device != torch.device("cpu"):
-            free_memory(total_memory_required[device] * 1.3 + extra_mem, device, models_already_loaded)
+    for loaded_model in models_already_loaded:
+        total_memory_required[loaded_model.device] = total_memory_required.get(loaded_model.device, 0) + loaded_model.model_memory_required(loaded_model.device)
 
     for loaded_model in models_to_load:
         weights_unloaded = unload_model_clones(loaded_model.model, unload_weights_only=False, force_unload=False) #unload the rest of the clones where the weights can stay loaded
         if weights_unloaded is not None:
             loaded_model.weights_loaded = not weights_unloaded
 
+    for device in total_memory_required:
+        if device != torch.device("cpu"):
+            free_memory(total_memory_required[device] * 1.1 + extra_mem, device, models_already_loaded)
+
     for loaded_model in models_to_load:
         model = loaded_model.model
         torch_dev = model.load_device
@@ -447,10 +529,10 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
         else:
             vram_set_state = vram_state
         lowvram_model_memory = 0
-        if lowvram_available and (vram_set_state == VRAMState.LOW_VRAM or vram_set_state == VRAMState.NORMAL_VRAM):
+        if lowvram_available and (vram_set_state == VRAMState.LOW_VRAM or vram_set_state == VRAMState.NORMAL_VRAM) and not force_full_load:
             model_size = loaded_model.model_memory_required(torch_dev)
             current_free_mem = get_free_memory(torch_dev)
-            lowvram_model_memory = int(max(64 * (1024 * 1024), (current_free_mem - minimum_memory_required)))
+            lowvram_model_memory = max(64 * (1024 * 1024), (current_free_mem - minimum_memory_required), min(current_free_mem * 0.4, current_free_mem - minimum_inference_memory()))
             if model_size <= lowvram_model_memory: #only switch to lowvram if really necessary
                 lowvram_model_memory = 0
 
@@ -459,6 +541,14 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
 
         cur_loaded_model = loaded_model.model_load(lowvram_model_memory, force_patch_weights=force_patch_weights)
         current_loaded_models.insert(0, loaded_model)
+
+
+    devs = set(map(lambda a: a.device, models_already_loaded))
+    for d in devs:
+        if d != torch.device("cpu"):
+            free_mem = get_free_memory(d)
+            if free_mem > minimum_memory_required:
+                use_more_memory(free_mem - minimum_memory_required, models_already_loaded, d)
     return
 
 
@@ -478,7 +568,9 @@ def loaded_models(only_currently_used=False):
 def cleanup_models(keep_clone_weights_loaded=False):
     to_delete = []
     for i in range(len(current_loaded_models)):
-        if sys.getrefcount(current_loaded_models[i].model) <= 2:
+        #TODO: very fragile function needs improvement
+        num_refs = sys.getrefcount(current_loaded_models[i].model)
+        if num_refs <= 2:
             if not keep_clone_weights_loaded:
                 to_delete = [i] + to_delete
             #TODO: find a less fragile way to do this.
@@ -527,6 +619,9 @@ def unet_inital_load_device(parameters, dtype):
     else:
         return cpu_dev
 
+def maximum_vram_for_weights(device=None):
+    return (get_total_memory(device) * 0.88 - minimum_inference_memory())
+
 def unet_dtype(device=None, model_params=0, supported_dtypes=[torch.float16, torch.bfloat16, torch.float32]):
     if args.bf16_unet:
         return torch.bfloat16
@@ -536,12 +631,37 @@ def unet_dtype(device=None, model_params=0, supported_dtypes=[torch.float16, tor
         return torch.float8_e4m3fn
     if args.fp8_e5m2_unet:
         return torch.float8_e5m2
-    if should_use_fp16(device=device, model_params=model_params, manual_cast=True):
-        if torch.float16 in supported_dtypes:
-            return torch.float16
-    if should_use_bf16(device, model_params=model_params, manual_cast=True):
-        if torch.bfloat16 in supported_dtypes:
-            return torch.bfloat16
+
+    fp8_dtype = None
+    try:
+        for dtype in [torch.float8_e4m3fn, torch.float8_e5m2]:
+            if dtype in supported_dtypes:
+                fp8_dtype = dtype
+                break
+    except:
+        pass
+
+    if fp8_dtype is not None:
+        free_model_memory = maximum_vram_for_weights(device)
+        if model_params * 2 > free_model_memory:
+            return fp8_dtype
+
+    for dt in supported_dtypes:
+        if dt == torch.float16 and should_use_fp16(device=device, model_params=model_params):
+            if torch.float16 in supported_dtypes:
+                return torch.float16
+        if dt == torch.bfloat16 and should_use_bf16(device, model_params=model_params):
+            if torch.bfloat16 in supported_dtypes:
+                return torch.bfloat16
+
+    for dt in supported_dtypes:
+        if dt == torch.float16 and should_use_fp16(device=device, model_params=model_params, manual_cast=True):
+            if torch.float16 in supported_dtypes:
+                return torch.float16
+        if dt == torch.bfloat16 and should_use_bf16(device, model_params=model_params, manual_cast=True):
+            if torch.bfloat16 in supported_dtypes:
+                return torch.bfloat16
+
     return torch.float32
 
 # None means no manual cast
@@ -557,13 +677,14 @@ def unet_manual_cast(weight_dtype, inference_device, supported_dtypes=[torch.flo
     if bf16_supported and weight_dtype == torch.bfloat16:
         return None
 
-    if fp16_supported and torch.float16 in supported_dtypes:
-        return torch.float16
+    fp16_supported = should_use_fp16(inference_device, prioritize_performance=True)
+    for dt in supported_dtypes:
+        if dt == torch.float16 and fp16_supported:
+            return torch.float16
+        if dt == torch.bfloat16 and bf16_supported:
+            return torch.bfloat16
 
-    elif bf16_supported and torch.bfloat16 in supported_dtypes:
-        return torch.bfloat16
-    else:
-        return torch.float32
+    return torch.float32
 
 def text_encoder_offload_device():
     if args.gpu_only:
@@ -582,6 +703,20 @@ def text_encoder_device():
     else:
         return torch.device("cpu")
 
+def text_encoder_initial_device(load_device, offload_device, model_size=0):
+    if load_device == offload_device or model_size <= 1024 * 1024 * 1024:
+        return offload_device
+
+    if is_device_mps(load_device):
+        return offload_device
+
+    mem_l = get_free_memory(load_device)
+    mem_o = get_free_memory(offload_device)
+    if mem_l > (mem_o * 0.5) and model_size * 1.2 < mem_l:
+        return load_device
+    else:
+        return offload_device
+
 def text_encoder_dtype(device=None):
     if args.fp8_e4m3fn_text_enc:
         return torch.float8_e4m3fn
@@ -758,7 +893,8 @@ def pytorch_attention_flash_attention():
 def force_upcast_attention_dtype():
     upcast = args.force_upcast_attention
     try:
-        if platform.mac_ver()[0] in ['14.5']: #black image bug on OSX Sonoma 14.5
+        macos_version = tuple(int(n) for n in platform.mac_ver()[0].split("."))
+        if (14, 5) <= macos_version < (14, 7):  # black image bug on recent versions of MacOS
             upcast = True
     except:
         pass
@@ -854,24 +990,21 @@ def should_use_fp16(device=None, model_params=0, prioritize_performance=True, ma
     if torch.version.hip:
         return True
 
-    props = torch.cuda.get_device_properties("cuda")
+    props = torch.cuda.get_device_properties(device)
     if props.major >= 8:
         return True
 
     if props.major < 6:
         return False
 
-    fp16_works = False
-    #FP16 is confirmed working on a 1080 (GP104) but it's a bit slower than FP32 so it should only be enabled
-    #when the model doesn't actually fit on the card
-    #TODO: actually test if GP106 and others have the same type of behavior
+    #FP16 is confirmed working on a 1080 (GP104) and on latest pytorch actually seems faster than fp32
     nvidia_10_series = ["1080", "1070", "titan x", "p3000", "p3200", "p4000", "p4200", "p5000", "p5200", "p6000", "1060", "1050", "p40", "p100", "p6", "p4"]
     for x in nvidia_10_series:
         if x in props.name.lower():
-            fp16_works = True
+            return True
 
-    if fp16_works or manual_cast:
-        free_model_memory = (get_free_memory() * 0.9 - minimum_inference_memory())
+    if manual_cast:
+        free_model_memory = maximum_vram_for_weights(device)
         if (not prioritize_performance) or model_params * 4 > free_model_memory:
             return True
 
@@ -910,9 +1043,6 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, ma
     if is_intel_xpu():
         return True
 
-    if device is None:
-        device = torch.device("cuda")
-
     props = torch.cuda.get_device_properties(device)
     if props.major >= 8:
         return True
@@ -920,12 +1050,22 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, ma
     bf16_works = torch.cuda.is_bf16_supported()
 
     if bf16_works or manual_cast:
-        free_model_memory = (get_free_memory() * 0.9 - minimum_inference_memory())
+        free_model_memory = maximum_vram_for_weights(device)
         if (not prioritize_performance) or model_params * 4 > free_model_memory:
             return True
 
     return False
 
+def supports_fp8_compute(device=None):
+    props = torch.cuda.get_device_properties(device)
+    if props.major >= 9:
+        return True
+    if props.major < 8:
+        return False
+    if props.minor < 9:
+        return False
+    return True
+
 def soft_empty_cache(force=False):
     global cpu_state
     if cpu_state == CPUState.MPS:

comfy/model_patcher.py

@@ -1,34 +1,47 @@
+"""
+    This file is part of ComfyUI.
+    Copyright (C) 2024 Comfy
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+"""
+
 import torch
 import copy
 import inspect
 import logging
 import uuid
+import collections
+import math
 
 import comfy.utils
+import comfy.float
 import comfy.model_management
+import comfy.lora
 from comfy.types import UnetWrapperFunction
 
-
-def weight_decompose(dora_scale, weight, lora_diff, alpha, strength):
-    dora_scale = comfy.model_management.cast_to_device(dora_scale, weight.device, torch.float32)
-    lora_diff *= alpha
-    weight_calc = weight + lora_diff.type(weight.dtype)
-    weight_norm = (
-        weight_calc.transpose(0, 1)
-        .reshape(weight_calc.shape[1], -1)
-        .norm(dim=1, keepdim=True)
-        .reshape(weight_calc.shape[1], *[1] * (weight_calc.dim() - 1))
-        .transpose(0, 1)
-    )
-
-    weight_calc *= (dora_scale / weight_norm).type(weight.dtype)
-    if strength != 1.0:
-        weight_calc -= weight
-        weight += strength * (weight_calc)
-    else:
-        weight[:] = weight_calc
-    return weight
-
+def string_to_seed(data):
+    crc = 0xFFFFFFFF
+    for byte in data:
+        if isinstance(byte, str):
+            byte = ord(byte)
+        crc ^= byte
+        for _ in range(8):
+            if crc & 1:
+                crc = (crc >> 1) ^ 0xEDB88320
+            else:
+                crc >>= 1
+    return crc ^ 0xFFFFFFFF
 
 def set_model_options_patch_replace(model_options, patch, name, block_name, number, transformer_index=None):
     to = model_options["transformer_options"].copy()
@@ -63,10 +76,30 @@ def set_model_options_pre_cfg_function(model_options, pre_cfg_function, disable_
         model_options["disable_cfg1_optimization"] = True
     return model_options
 
+def wipe_lowvram_weight(m):
+    if hasattr(m, "prev_comfy_cast_weights"):
+        m.comfy_cast_weights = m.prev_comfy_cast_weights
+        del m.prev_comfy_cast_weights
+    m.weight_function = None
+    m.bias_function = None
+
+class LowVramPatch:
+    def __init__(self, key, patches):
+        self.key = key
+        self.patches = patches
+    def __call__(self, weight):
+        return comfy.lora.calculate_weight(self.patches[self.key], weight, self.key, intermediate_dtype=weight.dtype)
+
 class ModelPatcher:
-    def __init__(self, model, load_device, offload_device, size=0, current_device=None, weight_inplace_update=False):
+    def __init__(self, model, load_device, offload_device, size=0, weight_inplace_update=False):
         self.size = size
         self.model = model
+        if not hasattr(self.model, 'device'):
+            logging.debug("Model doesn't have a device attribute.")
+            self.model.device = offload_device
+        elif self.model.device is None:
+            self.model.device = offload_device
+
         self.patches = {}
         self.backup = {}
         self.object_patches = {}
@@ -75,24 +108,32 @@ class ModelPatcher:
         self.model_size()
         self.load_device = load_device
         self.offload_device = offload_device
-        if current_device is None:
-            self.current_device = self.offload_device
-        else:
-            self.current_device = current_device
-
         self.weight_inplace_update = weight_inplace_update
-        self.model_lowvram = False
-        self.lowvram_patch_counter = 0
         self.patches_uuid = uuid.uuid4()
 
+        if not hasattr(self.model, 'model_loaded_weight_memory'):
+            self.model.model_loaded_weight_memory = 0
+
+        if not hasattr(self.model, 'lowvram_patch_counter'):
+            self.model.lowvram_patch_counter = 0
+
+        if not hasattr(self.model, 'model_lowvram'):
+            self.model.model_lowvram = False
+
     def model_size(self):
         if self.size > 0:
             return self.size
         self.size = comfy.model_management.module_size(self.model)
         return self.size
 
+    def loaded_size(self):
+        return self.model.model_loaded_weight_memory
+
+    def lowvram_patch_counter(self):
+        return self.model.lowvram_patch_counter
+
     def clone(self):
-        n = ModelPatcher(self.model, self.load_device, self.offload_device, self.size, self.current_device, weight_inplace_update=self.weight_inplace_update)
+        n = ModelPatcher(self.model, self.load_device, self.offload_device, self.size, weight_inplace_update=self.weight_inplace_update)
         n.patches = {}
         for k in self.patches:
             n.patches[k] = self.patches[k][:]
@@ -264,67 +305,52 @@ class ModelPatcher:
                     sd.pop(k)
         return sd
 
-    def patch_weight_to_device(self, key, device_to=None):
+    def patch_weight_to_device(self, key, device_to=None, inplace_update=False):
         if key not in self.patches:
             return
 
         weight = comfy.utils.get_attr(self.model, key)
 
-        inplace_update = self.weight_inplace_update
+        inplace_update = self.weight_inplace_update or inplace_update
 
         if key not in self.backup:
-            self.backup[key] = weight.to(device=self.offload_device, copy=inplace_update)
+            self.backup[key] = collections.namedtuple('Dimension', ['weight', 'inplace_update'])(weight.to(device=self.offload_device, copy=inplace_update), inplace_update)
 
         if device_to is not None:
             temp_weight = comfy.model_management.cast_to_device(weight, device_to, torch.float32, copy=True)
         else:
             temp_weight = weight.to(torch.float32, copy=True)
-        out_weight = self.calculate_weight(self.patches[key], temp_weight, key).to(weight.dtype)
+        out_weight = comfy.lora.calculate_weight(self.patches[key], temp_weight, key)
+        out_weight = comfy.float.stochastic_rounding(out_weight, weight.dtype, seed=string_to_seed(key))
         if inplace_update:
             comfy.utils.copy_to_param(self.model, key, out_weight)
         else:
             comfy.utils.set_attr_param(self.model, key, out_weight)
 
-    def patch_model(self, device_to=None, patch_weights=True):
-        for k in self.object_patches:
-            old = comfy.utils.set_attr(self.model, k, self.object_patches[k])
-            if k not in self.object_patches_backup:
-                self.object_patches_backup[k] = old
-
-        if patch_weights:
-            model_sd = self.model_state_dict()
-            for key in self.patches:
-                if key not in model_sd:
-                    logging.warning("could not patch. key doesn't exist in model: {}".format(key))
-                    continue
-
-                self.patch_weight_to_device(key, device_to)
-
-            if device_to is not None:
-                self.model.to(device_to)
-                self.current_device = device_to
-
-        return self.model
-
-    def patch_model_lowvram(self, device_to=None, lowvram_model_memory=0, force_patch_weights=False):
-        self.patch_model(device_to, patch_weights=False)
-
-        logging.info("loading in lowvram mode {}".format(lowvram_model_memory/(1024 * 1024)))
-        class LowVramPatch:
-            def __init__(self, key, model_patcher):
-                self.key = key
-                self.model_patcher = model_patcher
-            def __call__(self, weight):
-                return self.model_patcher.calculate_weight(self.model_patcher.patches[self.key], weight, self.key)
-
+    def load(self, device_to=None, lowvram_model_memory=0, force_patch_weights=False, full_load=False):
         mem_counter = 0
         patch_counter = 0
+        lowvram_counter = 0
+        loading = []
         for n, m in self.model.named_modules():
+            if hasattr(m, "comfy_cast_weights") or hasattr(m, "weight"):
+                loading.append((comfy.model_management.module_size(m), n, m))
+
+        load_completely = []
+        loading.sort(reverse=True)
+        for x in loading:
+            n = x[1]
+            m = x[2]
+            module_mem = x[0]
+
             lowvram_weight = False
-            if hasattr(m, "comfy_cast_weights"):
-                module_mem = comfy.model_management.module_size(m)
+
+            if not full_load and hasattr(m, "comfy_cast_weights"):
                 if mem_counter + module_mem >= lowvram_model_memory:
                     lowvram_weight = True
+                    lowvram_counter += 1
+                    if hasattr(m, "prev_comfy_cast_weights"): #Already lowvramed
+                        continue
 
             weight_key = "{}.weight".format(n)
             bias_key = "{}.bias".format(n)
@@ -334,227 +360,173 @@ class ModelPatcher:
                     if force_patch_weights:
                         self.patch_weight_to_device(weight_key)
                     else:
-                        m.weight_function = LowVramPatch(weight_key, self)
+                        m.weight_function = LowVramPatch(weight_key, self.patches)
                         patch_counter += 1
                 if bias_key in self.patches:
                     if force_patch_weights:
                         self.patch_weight_to_device(bias_key)
                     else:
-                        m.bias_function = LowVramPatch(bias_key, self)
+                        m.bias_function = LowVramPatch(bias_key, self.patches)
                         patch_counter += 1
 
                 m.prev_comfy_cast_weights = m.comfy_cast_weights
                 m.comfy_cast_weights = True
             else:
+                if hasattr(m, "comfy_cast_weights"):
+                    if m.comfy_cast_weights:
+                        wipe_lowvram_weight(m)
+
                 if hasattr(m, "weight"):
-                    self.patch_weight_to_device(weight_key, device_to)
-                    self.patch_weight_to_device(bias_key, device_to)
-                    m.to(device_to)
-                    mem_counter += comfy.model_management.module_size(m)
-                    logging.debug("lowvram: loaded module regularly {} {}".format(n, m))
+                    mem_counter += module_mem
+                    load_completely.append((module_mem, n, m))
 
-        self.model_lowvram = True
-        self.lowvram_patch_counter = patch_counter
+        load_completely.sort(reverse=True)
+        for x in load_completely:
+            n = x[1]
+            m = x[2]
+            weight_key = "{}.weight".format(n)
+            bias_key = "{}.bias".format(n)
+            if hasattr(m, "comfy_patched_weights"):
+                if m.comfy_patched_weights == True:
+                    continue
+
+            self.patch_weight_to_device(weight_key, device_to=device_to)
+            self.patch_weight_to_device(bias_key, device_to=device_to)
+            logging.debug("lowvram: loaded module regularly {} {}".format(n, m))
+            m.comfy_patched_weights = True
+
+        for x in load_completely:
+            x[2].to(device_to)
+
+        if lowvram_counter > 0:
+            logging.info("loaded partially {} {} {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), patch_counter))
+            self.model.model_lowvram = True
+        else:
+            logging.info("loaded completely {} {} {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), full_load))
+            self.model.model_lowvram = False
+            if full_load:
+                self.model.to(device_to)
+                mem_counter = self.model_size()
+
+        self.model.lowvram_patch_counter += patch_counter
+        self.model.device = device_to
+        self.model.model_loaded_weight_memory = mem_counter
+
+    def patch_model(self, device_to=None, lowvram_model_memory=0, load_weights=True, force_patch_weights=False):
+        for k in self.object_patches:
+            old = comfy.utils.set_attr(self.model, k, self.object_patches[k])
+            if k not in self.object_patches_backup:
+                self.object_patches_backup[k] = old
+
+        if lowvram_model_memory == 0:
+            full_load = True
+        else:
+            full_load = False
+
+        if load_weights:
+            self.load(device_to, lowvram_model_memory=lowvram_model_memory, force_patch_weights=force_patch_weights, full_load=full_load)
         return self.model
 
-    def calculate_weight(self, patches, weight, key):
-        for p in patches:
-            strength = p[0]
-            v = p[1]
-            strength_model = p[2]
-            offset = p[3]
-            function = p[4]
-            if function is None:
-                function = lambda a: a
-
-            old_weight = None
-            if offset is not None:
-                old_weight = weight
-                weight = weight.narrow(offset[0], offset[1], offset[2])
-
-            if strength_model != 1.0:
-                weight *= strength_model
-
-            if isinstance(v, list):
-                v = (self.calculate_weight(v[1:], v[0].clone(), key), )
-
-            if len(v) == 1:
-                patch_type = "diff"
-            elif len(v) == 2:
-                patch_type = v[0]
-                v = v[1]
-
-            if patch_type == "diff":
-                w1 = v[0]
-                if strength != 0.0:
-                    if w1.shape != weight.shape:
-                        logging.warning("WARNING SHAPE MISMATCH {} WEIGHT NOT MERGED {} != {}".format(key, w1.shape, weight.shape))
-                    else:
-                        weight += function(strength * comfy.model_management.cast_to_device(w1, weight.device, weight.dtype))
-            elif patch_type == "lora": #lora/locon
-                mat1 = comfy.model_management.cast_to_device(v[0], weight.device, torch.float32)
-                mat2 = comfy.model_management.cast_to_device(v[1], weight.device, torch.float32)
-                dora_scale = v[4]
-                if v[2] is not None:
-                    alpha = v[2] / mat2.shape[0]
-                else:
-                    alpha = 1.0
-
-                if v[3] is not None:
-                    #locon mid weights, hopefully the math is fine because I didn't properly test it
-                    mat3 = comfy.model_management.cast_to_device(v[3], weight.device, torch.float32)
-                    final_shape = [mat2.shape[1], mat2.shape[0], mat3.shape[2], mat3.shape[3]]
-                    mat2 = torch.mm(mat2.transpose(0, 1).flatten(start_dim=1), mat3.transpose(0, 1).flatten(start_dim=1)).reshape(final_shape).transpose(0, 1)
-                try:
-                    lora_diff = torch.mm(mat1.flatten(start_dim=1), mat2.flatten(start_dim=1)).reshape(weight.shape)
-                    if dora_scale is not None:
-                        weight = function(weight_decompose(dora_scale, weight, lora_diff, alpha, strength))
-                    else:
-                        weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
-                except Exception as e:
-                    logging.error("ERROR {} {} {}".format(patch_type, key, e))
-            elif patch_type == "lokr":
-                w1 = v[0]
-                w2 = v[1]
-                w1_a = v[3]
-                w1_b = v[4]
-                w2_a = v[5]
-                w2_b = v[6]
-                t2 = v[7]
-                dora_scale = v[8]
-                dim = None
-
-                if w1 is None:
-                    dim = w1_b.shape[0]
-                    w1 = torch.mm(comfy.model_management.cast_to_device(w1_a, weight.device, torch.float32),
-                                  comfy.model_management.cast_to_device(w1_b, weight.device, torch.float32))
-                else:
-                    w1 = comfy.model_management.cast_to_device(w1, weight.device, torch.float32)
-
-                if w2 is None:
-                    dim = w2_b.shape[0]
-                    if t2 is None:
-                        w2 = torch.mm(comfy.model_management.cast_to_device(w2_a, weight.device, torch.float32),
-                                      comfy.model_management.cast_to_device(w2_b, weight.device, torch.float32))
-                    else:
-                        w2 = torch.einsum('i j k l, j r, i p -> p r k l',
-                                          comfy.model_management.cast_to_device(t2, weight.device, torch.float32),
-                                          comfy.model_management.cast_to_device(w2_b, weight.device, torch.float32),
-                                          comfy.model_management.cast_to_device(w2_a, weight.device, torch.float32))
-                else:
-                    w2 = comfy.model_management.cast_to_device(w2, weight.device, torch.float32)
-
-                if len(w2.shape) == 4:
-                    w1 = w1.unsqueeze(2).unsqueeze(2)
-                if v[2] is not None and dim is not None:
-                    alpha = v[2] / dim
-                else:
-                    alpha = 1.0
-
-                try:
-                    lora_diff = torch.kron(w1, w2).reshape(weight.shape)
-                    if dora_scale is not None:
-                        weight = function(weight_decompose(dora_scale, weight, lora_diff, alpha, strength))
-                    else:
-                        weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
-                except Exception as e:
-                    logging.error("ERROR {} {} {}".format(patch_type, key, e))
-            elif patch_type == "loha":
-                w1a = v[0]
-                w1b = v[1]
-                if v[2] is not None:
-                    alpha = v[2] / w1b.shape[0]
-                else:
-                    alpha = 1.0
-
-                w2a = v[3]
-                w2b = v[4]
-                dora_scale = v[7]
-                if v[5] is not None: #cp decomposition
-                    t1 = v[5]
-                    t2 = v[6]
-                    m1 = torch.einsum('i j k l, j r, i p -> p r k l',
-                                      comfy.model_management.cast_to_device(t1, weight.device, torch.float32),
-                                      comfy.model_management.cast_to_device(w1b, weight.device, torch.float32),
-                                      comfy.model_management.cast_to_device(w1a, weight.device, torch.float32))
-
-                    m2 = torch.einsum('i j k l, j r, i p -> p r k l',
-                                      comfy.model_management.cast_to_device(t2, weight.device, torch.float32),
-                                      comfy.model_management.cast_to_device(w2b, weight.device, torch.float32),
-                                      comfy.model_management.cast_to_device(w2a, weight.device, torch.float32))
-                else:
-                    m1 = torch.mm(comfy.model_management.cast_to_device(w1a, weight.device, torch.float32),
-                                  comfy.model_management.cast_to_device(w1b, weight.device, torch.float32))
-                    m2 = torch.mm(comfy.model_management.cast_to_device(w2a, weight.device, torch.float32),
-                                  comfy.model_management.cast_to_device(w2b, weight.device, torch.float32))
-
-                try:
-                    lora_diff = (m1 * m2).reshape(weight.shape)
-                    if dora_scale is not None:
-                        weight = function(weight_decompose(dora_scale, weight, lora_diff, alpha, strength))
-                    else:
-                        weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
-                except Exception as e:
-                    logging.error("ERROR {} {} {}".format(patch_type, key, e))
-            elif patch_type == "glora":
-                if v[4] is not None:
-                    alpha = v[4] / v[0].shape[0]
-                else:
-                    alpha = 1.0
-
-                dora_scale = v[5]
-
-                a1 = comfy.model_management.cast_to_device(v[0].flatten(start_dim=1), weight.device, torch.float32)
-                a2 = comfy.model_management.cast_to_device(v[1].flatten(start_dim=1), weight.device, torch.float32)
-                b1 = comfy.model_management.cast_to_device(v[2].flatten(start_dim=1), weight.device, torch.float32)
-                b2 = comfy.model_management.cast_to_device(v[3].flatten(start_dim=1), weight.device, torch.float32)
-
-                try:
-                    lora_diff = (torch.mm(b2, b1) + torch.mm(torch.mm(weight.flatten(start_dim=1), a2), a1)).reshape(weight.shape)
-                    if dora_scale is not None:
-                        weight = function(weight_decompose(dora_scale, weight, lora_diff, alpha, strength))
-                    else:
-                        weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
-                except Exception as e:
-                    logging.error("ERROR {} {} {}".format(patch_type, key, e))
-            else:
-                logging.warning("patch type not recognized {} {}".format(patch_type, key))
-
-            if old_weight is not None:
-                weight = old_weight
-
-        return weight
-
     def unpatch_model(self, device_to=None, unpatch_weights=True):
         if unpatch_weights:
-            if self.model_lowvram:
+            if self.model.model_lowvram:
                 for m in self.model.modules():
-                    if hasattr(m, "prev_comfy_cast_weights"):
-                        m.comfy_cast_weights = m.prev_comfy_cast_weights
-                        del m.prev_comfy_cast_weights
-                    m.weight_function = None
-                    m.bias_function = None
+                    wipe_lowvram_weight(m)
 
-                self.model_lowvram = False
-                self.lowvram_patch_counter = 0
+                self.model.model_lowvram = False
+                self.model.lowvram_patch_counter = 0
 
             keys = list(self.backup.keys())
 
-            if self.weight_inplace_update:
-                for k in keys:
-                    comfy.utils.copy_to_param(self.model, k, self.backup[k])
-            else:
-                for k in keys:
-                    comfy.utils.set_attr_param(self.model, k, self.backup[k])
+            for k in keys:
+                bk = self.backup[k]
+                if bk.inplace_update:
+                    comfy.utils.copy_to_param(self.model, k, bk.weight)
+                else:
+                    comfy.utils.set_attr_param(self.model, k, bk.weight)
 
             self.backup.clear()
 
             if device_to is not None:
                 self.model.to(device_to)
-                self.current_device = device_to
+                self.model.device = device_to
+            self.model.model_loaded_weight_memory = 0
+
+            for m in self.model.modules():
+                if hasattr(m, "comfy_patched_weights"):
+                    del m.comfy_patched_weights
 
         keys = list(self.object_patches_backup.keys())
         for k in keys:
             comfy.utils.set_attr(self.model, k, self.object_patches_backup[k])
 
         self.object_patches_backup.clear()
+
+    def partially_unload(self, device_to, memory_to_free=0):
+        memory_freed = 0
+        patch_counter = 0
+        unload_list = []
+
+        for n, m in self.model.named_modules():
+            shift_lowvram = False
+            if hasattr(m, "comfy_cast_weights"):
+                module_mem = comfy.model_management.module_size(m)
+                unload_list.append((module_mem, n, m))
+
+        unload_list.sort()
+        for unload in unload_list:
+            if memory_to_free < memory_freed:
+                break
+            module_mem = unload[0]
+            n = unload[1]
+            m = unload[2]
+            weight_key = "{}.weight".format(n)
+            bias_key = "{}.bias".format(n)
+
+            if hasattr(m, "comfy_patched_weights") and m.comfy_patched_weights == True:
+                for key in [weight_key, bias_key]:
+                    bk = self.backup.get(key, None)
+                    if bk is not None:
+                        if bk.inplace_update:
+                            comfy.utils.copy_to_param(self.model, key, bk.weight)
+                        else:
+                            comfy.utils.set_attr_param(self.model, key, bk.weight)
+                        self.backup.pop(key)
+
+                m.to(device_to)
+                if weight_key in self.patches:
+                    m.weight_function = LowVramPatch(weight_key, self.patches)
+                    patch_counter += 1
+                if bias_key in self.patches:
+                    m.bias_function = LowVramPatch(bias_key, self.patches)
+                    patch_counter += 1
+
+                m.prev_comfy_cast_weights = m.comfy_cast_weights
+                m.comfy_cast_weights = True
+                m.comfy_patched_weights = False
+                memory_freed += module_mem
+                logging.debug("freed {}".format(n))
+
+        self.model.model_lowvram = True
+        self.model.lowvram_patch_counter += patch_counter
+        self.model.model_loaded_weight_memory -= memory_freed
+        return memory_freed
+
+    def partially_load(self, device_to, extra_memory=0):
+        self.unpatch_model(unpatch_weights=False)
+        self.patch_model(load_weights=False)
+        full_load = False
+        if self.model.model_lowvram == False:
+            return 0
+        if self.model.model_loaded_weight_memory + extra_memory > self.model_size():
+            full_load = True
+        current_used = self.model.model_loaded_weight_memory
+        self.load(device_to, lowvram_model_memory=current_used + extra_memory, full_load=full_load)
+        return self.model.model_loaded_weight_memory - current_used
+
+    def current_loaded_device(self):
+        return self.model.device
+
+    def calculate_weight(self, patches, weight, key, intermediate_dtype=torch.float32):
+        print("WARNING the ModelPatcher.calculate_weight function is deprecated, please use: comfy.lora.calculate_weight instead")
+        return comfy.lora.calculate_weight(patches, weight, key, intermediate_dtype=intermediate_dtype)

comfy/ops.py

@@ -18,29 +18,42 @@
 
 import torch
 import comfy.model_management
+from comfy.cli_args import args
 
+def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False):
+    if device is None or weight.device == device:
+        if not copy:
+            if dtype is None or weight.dtype == dtype:
+                return weight
+        return weight.to(dtype=dtype, copy=copy)
 
-def cast_to(weight, dtype=None, device=None, non_blocking=False):
-    return weight.to(device=device, dtype=dtype, non_blocking=non_blocking)
+    r = torch.empty_like(weight, dtype=dtype, device=device)
+    r.copy_(weight, non_blocking=non_blocking)
+    return r
 
-def cast_to_input(weight, input, non_blocking=False):
-    return cast_to(weight, input.dtype, input.device, non_blocking=non_blocking)
+def cast_to_input(weight, input, non_blocking=False, copy=True):
+    return cast_to(weight, input.dtype, input.device, non_blocking=non_blocking, copy=copy)
 
-def cast_bias_weight(s, input=None, dtype=None, device=None):
+def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None):
     if input is not None:
         if dtype is None:
             dtype = input.dtype
+        if bias_dtype is None:
+            bias_dtype = dtype
         if device is None:
             device = input.device
 
     bias = None
-    non_blocking = comfy.model_management.device_should_use_non_blocking(device)
+    non_blocking = comfy.model_management.device_supports_non_blocking(device)
     if s.bias is not None:
-        bias = cast_to(s.bias, dtype, device, non_blocking=non_blocking)
-        if s.bias_function is not None:
+        has_function = s.bias_function is not None
+        bias = cast_to(s.bias, bias_dtype, device, non_blocking=non_blocking, copy=has_function)
+        if has_function:
             bias = s.bias_function(bias)
-    weight = cast_to(s.weight, dtype, device, non_blocking=non_blocking)
-    if s.weight_function is not None:
+
+    has_function = s.weight_function is not None
+    weight = cast_to(s.weight, dtype, device, non_blocking=non_blocking, copy=has_function)
+    if has_function:
         weight = s.weight_function(weight)
     return weight, bias
 
@@ -238,3 +251,59 @@ class manual_cast(disable_weight_init):
 
     class Embedding(disable_weight_init.Embedding):
         comfy_cast_weights = True
+
+
+def fp8_linear(self, input):
+    dtype = self.weight.dtype
+    if dtype not in [torch.float8_e4m3fn]:
+        return None
+
+    if len(input.shape) == 3:
+        inn = input.reshape(-1, input.shape[2]).to(dtype)
+        non_blocking = comfy.model_management.device_supports_non_blocking(input.device)
+        w, bias = cast_bias_weight(self, input, dtype=dtype, bias_dtype=input.dtype)
+        w = w.t()
+
+        scale_weight = self.scale_weight
+        scale_input = self.scale_input
+        if scale_weight is None:
+            scale_weight = torch.ones((1), device=input.device, dtype=torch.float32)
+            if scale_input is None:
+                scale_input = scale_weight
+        if scale_input is None:
+            scale_input = torch.ones((1), device=input.device, dtype=torch.float32)
+
+        if bias is not None:
+            o = torch._scaled_mm(inn, w, out_dtype=input.dtype, bias=bias, scale_a=scale_input, scale_b=scale_weight)
+        else:
+            o = torch._scaled_mm(inn, w, out_dtype=input.dtype, scale_a=scale_input, scale_b=scale_weight)
+
+        if isinstance(o, tuple):
+            o = o[0]
+
+        return o.reshape((-1, input.shape[1], self.weight.shape[0]))
+    return None
+
+class fp8_ops(manual_cast):
+    class Linear(manual_cast.Linear):
+        def reset_parameters(self):
+            self.scale_weight = None
+            self.scale_input = None
+            return None
+
+        def forward_comfy_cast_weights(self, input):
+            out = fp8_linear(self, input)
+            if out is not None:
+                return out
+
+            weight, bias = cast_bias_weight(self, input)
+            return torch.nn.functional.linear(input, weight, bias)
+
+
+def pick_operations(weight_dtype, compute_dtype, load_device=None):
+    if compute_dtype is None or weight_dtype == compute_dtype:
+        return disable_weight_init
+    if args.fast:
+        if comfy.model_management.supports_fp8_compute(load_device):
+            return fp8_ops
+    return manual_cast

comfy/samplers.py

@@ -171,7 +171,7 @@ def calc_cond_batch(model, conds, x_in, timestep, model_options):
         for i in range(1, len(to_batch_temp) + 1):
             batch_amount = to_batch_temp[:len(to_batch_temp)//i]
             input_shape = [len(batch_amount) * first_shape[0]] + list(first_shape)[1:]
-            if model.memory_required(input_shape) < free_memory:
+            if model.memory_required(input_shape) * 1.5 < free_memory:
                 to_batch = batch_amount
                 break
 

comfy/sd.py

@@ -24,6 +24,7 @@ import comfy.text_encoders.sa_t5
 import comfy.text_encoders.aura_t5
 import comfy.text_encoders.hydit
 import comfy.text_encoders.flux
+import comfy.text_encoders.long_clipl
 
 import comfy.model_patcher
 import comfy.lora
@@ -62,7 +63,7 @@ def load_lora_for_models(model, clip, lora, strength_model, strength_clip):
 
 
 class CLIP:
-    def __init__(self, target=None, embedding_directory=None, no_init=False, tokenizer_data={}):
+    def __init__(self, target=None, embedding_directory=None, no_init=False, tokenizer_data={}, parameters=0, model_options={}):
         if no_init:
             return
         params = target.params.copy()
@@ -71,20 +72,29 @@ class CLIP:
 
         load_device = model_management.text_encoder_device()
         offload_device = model_management.text_encoder_offload_device()
-        params['device'] = offload_device
-        dtype = model_management.text_encoder_dtype(load_device)
+        dtype = model_options.get("dtype", None)
+        if dtype is None:
+            dtype = model_management.text_encoder_dtype(load_device)
+
         params['dtype'] = dtype
+        params['device'] = model_management.text_encoder_initial_device(load_device, offload_device, parameters * model_management.dtype_size(dtype))
+        params['model_options'] = model_options
 
         self.cond_stage_model = clip(**(params))
 
         for dt in self.cond_stage_model.dtypes:
             if not model_management.supports_cast(load_device, dt):
                 load_device = offload_device
+                if params['device'] != offload_device:
+                    self.cond_stage_model.to(offload_device)
+                    logging.warning("Had to shift TE back.")
 
         self.tokenizer = tokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
         self.patcher = comfy.model_patcher.ModelPatcher(self.cond_stage_model, load_device=load_device, offload_device=offload_device)
+        if params['device'] == load_device:
+            model_management.load_models_gpu([self.patcher], force_full_load=True)
         self.layer_idx = None
-        logging.debug("CLIP model load device: {}, offload device: {}".format(load_device, offload_device))
+        logging.debug("CLIP model load device: {}, offload device: {}, current: {}".format(load_device, offload_device, params['device']))
 
     def clone(self):
         n = CLIP(no_init=True)
@@ -390,11 +400,14 @@ class CLIPType(Enum):
     HUNYUAN_DIT = 5
     FLUX = 6
 
-def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION):
+def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}):
     clip_data = []
     for p in ckpt_paths:
         clip_data.append(comfy.utils.load_torch_file(p, safe_load=True))
+    return load_text_encoder_state_dicts(clip_data, embedding_directory=embedding_directory, clip_type=clip_type, model_options=model_options)
 
+def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}):
+    clip_data = state_dicts
     class EmptyClass:
         pass
 
@@ -431,8 +444,13 @@ def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DI
             clip_target.clip = comfy.text_encoders.sa_t5.SAT5Model
             clip_target.tokenizer = comfy.text_encoders.sa_t5.SAT5Tokenizer
         else:
-            clip_target.clip = sd1_clip.SD1ClipModel
-            clip_target.tokenizer = sd1_clip.SD1Tokenizer
+            w = clip_data[0].get("text_model.embeddings.position_embedding.weight", None)
+            if w is not None and w.shape[0] == 248:
+                clip_target.clip = comfy.text_encoders.long_clipl.LongClipModel
+                clip_target.tokenizer = comfy.text_encoders.long_clipl.LongClipTokenizer
+            else:
+                clip_target.clip = sd1_clip.SD1ClipModel
+                clip_target.tokenizer = sd1_clip.SD1Tokenizer
     elif len(clip_data) == 2:
         if clip_type == CLIPType.SD3:
             clip_target.clip = comfy.text_encoders.sd3_clip.sd3_clip(clip_l=True, clip_g=True, t5=False)
@@ -456,7 +474,11 @@ def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DI
         clip_target.clip = comfy.text_encoders.sd3_clip.SD3ClipModel
         clip_target.tokenizer = comfy.text_encoders.sd3_clip.SD3Tokenizer
 
-    clip = CLIP(clip_target, embedding_directory=embedding_directory)
+    parameters = 0
+    for c in clip_data:
+        parameters += comfy.utils.calculate_parameters(c)
+
+    clip = CLIP(clip_target, embedding_directory=embedding_directory, parameters=parameters, model_options=model_options)
     for c in clip_data:
         m, u = clip.load_sd(c)
         if len(m) > 0:
@@ -498,25 +520,39 @@ def load_checkpoint(config_path=None, ckpt_path=None, output_vae=True, output_cl
 
     return (model, clip, vae)
 
-def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, output_clipvision=False, embedding_directory=None, output_model=True):
+def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, output_clipvision=False, embedding_directory=None, output_model=True, model_options={}, te_model_options={}):
     sd = comfy.utils.load_torch_file(ckpt_path)
-    sd_keys = sd.keys()
+    out = load_state_dict_guess_config(sd, output_vae, output_clip, output_clipvision, embedding_directory, output_model, model_options, te_model_options=te_model_options)
+    if out is None:
+        raise RuntimeError("ERROR: Could not detect model type of: {}".format(ckpt_path))
+    return out
+
+def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_clipvision=False, embedding_directory=None, output_model=True, model_options={}, te_model_options={}):
     clip = None
     clipvision = None
     vae = None
     model = None
     model_patcher = None
-    clip_target = None
 
     diffusion_model_prefix = model_detection.unet_prefix_from_state_dict(sd)
     parameters = comfy.utils.calculate_parameters(sd, diffusion_model_prefix)
+    weight_dtype = comfy.utils.weight_dtype(sd, diffusion_model_prefix)
     load_device = model_management.get_torch_device()
 
     model_config = model_detection.model_config_from_unet(sd, diffusion_model_prefix)
     if model_config is None:
-        raise RuntimeError("ERROR: Could not detect model type of: {}".format(ckpt_path))
+        return None
+
+    unet_weight_dtype = list(model_config.supported_inference_dtypes)
+    if weight_dtype is not None:
+        unet_weight_dtype.append(weight_dtype)
+
+    model_config.custom_operations = model_options.get("custom_operations", None)
+    unet_dtype = model_options.get("weight_dtype", None)
+
+    if unet_dtype is None:
+        unet_dtype = model_management.unet_dtype(model_params=parameters, supported_dtypes=unet_weight_dtype)
 
-    unet_dtype = model_management.unet_dtype(model_params=parameters, supported_dtypes=model_config.supported_inference_dtypes)
     manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device, model_config.supported_inference_dtypes)
     model_config.set_inference_dtype(unet_dtype, manual_cast_dtype)
 
@@ -540,7 +576,8 @@ def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, o
         if clip_target is not None:
             clip_sd = model_config.process_clip_state_dict(sd)
             if len(clip_sd) > 0:
-                clip = CLIP(clip_target, embedding_directory=embedding_directory, tokenizer_data=clip_sd)
+                parameters = comfy.utils.calculate_parameters(clip_sd)
+                clip = CLIP(clip_target, embedding_directory=embedding_directory, tokenizer_data=clip_sd, parameters=parameters, model_options=te_model_options)
                 m, u = clip.load_sd(clip_sd, full_model=True)
                 if len(m) > 0:
                     m_filter = list(filter(lambda a: ".logit_scale" not in a and ".transformer.text_projection.weight" not in a, m))
@@ -559,15 +596,16 @@ def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, o
         logging.debug("left over keys: {}".format(left_over))
 
     if output_model:
-        model_patcher = comfy.model_patcher.ModelPatcher(model, load_device=load_device, offload_device=model_management.unet_offload_device(), current_device=inital_load_device)
+        model_patcher = comfy.model_patcher.ModelPatcher(model, load_device=load_device, offload_device=model_management.unet_offload_device())
         if inital_load_device != torch.device("cpu"):
             logging.info("loaded straight to GPU")
-            model_management.load_model_gpu(model_patcher)
+            model_management.load_models_gpu([model_patcher], force_full_load=True)
 
     return (model_patcher, clip, vae, clipvision)
 
 
-def load_unet_state_dict(sd, dtype=None): #load unet in diffusers or regular format
+def load_diffusion_model_state_dict(sd, model_options={}): #load unet in diffusers or regular format
+    dtype = model_options.get("dtype", None)
 
     #Allow loading unets from checkpoint files
     diffusion_model_prefix = model_detection.unet_prefix_from_state_dict(sd)
@@ -609,6 +647,7 @@ def load_unet_state_dict(sd, dtype=None): #load unet in diffusers or regular for
 
     manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device, model_config.supported_inference_dtypes)
     model_config.set_inference_dtype(unet_dtype, manual_cast_dtype)
+    model_config.custom_operations = model_options.get("custom_operations", None)
     model = model_config.get_model(new_sd, "")
     model = model.to(offload_device)
     model.load_model_weights(new_sd, "")
@@ -617,24 +656,36 @@ def load_unet_state_dict(sd, dtype=None): #load unet in diffusers or regular for
         logging.info("left over keys in unet: {}".format(left_over))
     return comfy.model_patcher.ModelPatcher(model, load_device=load_device, offload_device=offload_device)
 
-def load_unet(unet_path, dtype=None):
+
+def load_diffusion_model(unet_path, model_options={}):
     sd = comfy.utils.load_torch_file(unet_path)
-    model = load_unet_state_dict(sd, dtype=dtype)
+    model = load_diffusion_model_state_dict(sd, model_options=model_options)
     if model is None:
         logging.error("ERROR UNSUPPORTED UNET {}".format(unet_path))
         raise RuntimeError("ERROR: Could not detect model type of: {}".format(unet_path))
     return model
 
+def load_unet(unet_path, dtype=None):
+    print("WARNING: the load_unet function has been deprecated and will be removed please switch to: load_diffusion_model")
+    return load_diffusion_model(unet_path, model_options={"dtype": dtype})
+
+def load_unet_state_dict(sd, dtype=None):
+    print("WARNING: the load_unet_state_dict function has been deprecated and will be removed please switch to: load_diffusion_model_state_dict")
+    return load_diffusion_model_state_dict(sd, model_options={"dtype": dtype})
+
 def save_checkpoint(output_path, model, clip=None, vae=None, clip_vision=None, metadata=None, extra_keys={}):
     clip_sd = None
     load_models = [model]
     if clip is not None:
         load_models.append(clip.load_model())
         clip_sd = clip.get_sd()
+    vae_sd = None
+    if vae is not None:
+        vae_sd = vae.get_sd()
 
     model_management.load_models_gpu(load_models, force_patch_weights=True)
     clip_vision_sd = clip_vision.get_sd() if clip_vision is not None else None
-    sd = model.model.state_dict_for_saving(clip_sd, vae.get_sd(), clip_vision_sd)
+    sd = model.model.state_dict_for_saving(clip_sd, vae_sd, clip_vision_sd)
     for k in extra_keys:
         sd[k] = extra_keys[k]
 

comfy/sd1_clip.py

@@ -75,7 +75,6 @@ class ClipTokenWeightEncoder:
         return r
 
 class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
-    """Uses the CLIP transformer encoder for text (from huggingface)"""
     LAYERS = [
         "last",
         "pooled",
@@ -84,7 +83,7 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
     def __init__(self, version="openai/clip-vit-large-patch14", device="cpu", max_length=77,
                  freeze=True, layer="last", layer_idx=None, textmodel_json_config=None, dtype=None, model_class=comfy.clip_model.CLIPTextModel,
                  special_tokens={"start": 49406, "end": 49407, "pad": 49407}, layer_norm_hidden_state=True, enable_attention_masks=False, zero_out_masked=False,
-                 return_projected_pooled=True, return_attention_masks=False):  # clip-vit-base-patch32
+                 return_projected_pooled=True, return_attention_masks=False, model_options={}):  # clip-vit-base-patch32
         super().__init__()
         assert layer in self.LAYERS
 
@@ -94,7 +93,11 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
         with open(textmodel_json_config) as f:
             config = json.load(f)
 
-        self.operations = comfy.ops.manual_cast
+        operations = model_options.get("custom_operations", None)
+        if operations is None:
+            operations = comfy.ops.manual_cast
+
+        self.operations = operations
         self.transformer = model_class(config, dtype, device, self.operations)
         self.num_layers = self.transformer.num_layers
 
@@ -313,6 +316,17 @@ def expand_directory_list(directories):
             dirs.add(root)
     return list(dirs)
 
+def bundled_embed(embed, prefix, suffix): #bundled embedding in lora format
+    i = 0
+    out_list = []
+    for k in embed:
+        if k.startswith(prefix) and k.endswith(suffix):
+            out_list.append(embed[k])
+    if len(out_list) == 0:
+        return None
+
+    return torch.cat(out_list, dim=0)
+
 def load_embed(embedding_name, embedding_directory, embedding_size, embed_key=None):
     if isinstance(embedding_directory, str):
         embedding_directory = [embedding_directory]
@@ -379,8 +393,12 @@ def load_embed(embedding_name, embedding_directory, embedding_size, embed_key=No
         elif embed_key is not None and embed_key in embed:
             embed_out = embed[embed_key]
         else:
-            values = embed.values()
-            embed_out = next(iter(values))
+            embed_out = bundled_embed(embed, 'bundle_emb.', '.string_to_param.*')
+            if embed_out is None:
+                embed_out = bundled_embed(embed, 'bundle_emb.', '.{}'.format(embed_key))
+            if embed_out is None:
+                values = embed.values()
+                embed_out = next(iter(values))
     return embed_out
 
 class SDTokenizer:
@@ -537,8 +555,12 @@ class SD1Tokenizer:
     def state_dict(self):
         return {}
 
+class SD1CheckpointClipModel(SDClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        super().__init__(device=device, return_projected_pooled=False, dtype=dtype, model_options=model_options)
+
 class SD1ClipModel(torch.nn.Module):
-    def __init__(self, device="cpu", dtype=None, clip_name="l", clip_model=SDClipModel, name=None, **kwargs):
+    def __init__(self, device="cpu", dtype=None, model_options={}, clip_name="l", clip_model=SD1CheckpointClipModel, name=None, **kwargs):
         super().__init__()
 
         if name is not None:
@@ -548,7 +570,7 @@ class SD1ClipModel(torch.nn.Module):
             self.clip_name = clip_name
             self.clip = "clip_{}".format(self.clip_name)
 
-        setattr(self, self.clip, clip_model(device=device, dtype=dtype, **kwargs))
+        setattr(self, self.clip, clip_model(device=device, dtype=dtype, model_options=model_options, **kwargs))
 
         self.dtypes = set()
         if dtype is not None:

comfy/sdxl_clip.py

@@ -3,14 +3,14 @@ import torch
 import os
 
 class SDXLClipG(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", max_length=77, freeze=True, layer="penultimate", layer_idx=None, dtype=None):
+    def __init__(self, device="cpu", max_length=77, freeze=True, layer="penultimate", layer_idx=None, dtype=None, model_options={}):
         if layer == "penultimate":
             layer="hidden"
             layer_idx=-2
 
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_config_bigg.json")
         super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype,
-                         special_tokens={"start": 49406, "end": 49407, "pad": 0}, layer_norm_hidden_state=False)
+                         special_tokens={"start": 49406, "end": 49407, "pad": 0}, layer_norm_hidden_state=False, return_projected_pooled=True, model_options=model_options)
 
     def load_sd(self, sd):
         return super().load_sd(sd)
@@ -38,10 +38,10 @@ class SDXLTokenizer:
         return {}
 
 class SDXLClipModel(torch.nn.Module):
-    def __init__(self, device="cpu", dtype=None):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
         super().__init__()
-        self.clip_l = sd1_clip.SDClipModel(layer="hidden", layer_idx=-2, device=device, dtype=dtype, layer_norm_hidden_state=False)
-        self.clip_g = SDXLClipG(device=device, dtype=dtype)
+        self.clip_l = sd1_clip.SDClipModel(layer="hidden", layer_idx=-2, device=device, dtype=dtype, layer_norm_hidden_state=False, model_options=model_options)
+        self.clip_g = SDXLClipG(device=device, dtype=dtype, model_options=model_options)
         self.dtypes = set([dtype])
 
     def set_clip_options(self, options):
@@ -66,8 +66,8 @@ class SDXLClipModel(torch.nn.Module):
             return self.clip_l.load_sd(sd)
 
 class SDXLRefinerClipModel(sd1_clip.SD1ClipModel):
-    def __init__(self, device="cpu", dtype=None):
-        super().__init__(device=device, dtype=dtype, clip_name="g", clip_model=SDXLClipG)
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        super().__init__(device=device, dtype=dtype, clip_name="g", clip_model=SDXLClipG, model_options=model_options)
 
 
 class StableCascadeClipGTokenizer(sd1_clip.SDTokenizer):
@@ -79,14 +79,14 @@ class StableCascadeTokenizer(sd1_clip.SD1Tokenizer):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="g", tokenizer=StableCascadeClipGTokenizer)
 
 class StableCascadeClipG(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", max_length=77, freeze=True, layer="hidden", layer_idx=-1, dtype=None):
+    def __init__(self, device="cpu", max_length=77, freeze=True, layer="hidden", layer_idx=-1, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_config_bigg.json")
         super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype,
-                         special_tokens={"start": 49406, "end": 49407, "pad": 49407}, layer_norm_hidden_state=False, enable_attention_masks=True)
+                         special_tokens={"start": 49406, "end": 49407, "pad": 49407}, layer_norm_hidden_state=False, enable_attention_masks=True, return_projected_pooled=True, model_options=model_options)
 
     def load_sd(self, sd):
         return super().load_sd(sd)
 
 class StableCascadeClipModel(sd1_clip.SD1ClipModel):
-    def __init__(self, device="cpu", dtype=None):
-        super().__init__(device=device, dtype=dtype, clip_name="g", clip_model=StableCascadeClipG)
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        super().__init__(device=device, dtype=dtype, clip_name="g", clip_model=StableCascadeClipG, model_options=model_options)

comfy/supported_models.py

@@ -31,6 +31,7 @@ class SD15(supported_models_base.BASE):
     }
 
     latent_format = latent_formats.SD15
+    memory_usage_factor = 1.0
 
     def process_clip_state_dict(self, state_dict):
         k = list(state_dict.keys())
@@ -77,6 +78,7 @@ class SD20(supported_models_base.BASE):
     }
 
     latent_format = latent_formats.SD15
+    memory_usage_factor = 1.0
 
     def model_type(self, state_dict, prefix=""):
         if self.unet_config["in_channels"] == 4: #SD2.0 inpainting models are not v prediction
@@ -140,6 +142,7 @@ class SDXLRefiner(supported_models_base.BASE):
     }
 
     latent_format = latent_formats.SDXL
+    memory_usage_factor = 1.0
 
     def get_model(self, state_dict, prefix="", device=None):
         return model_base.SDXLRefiner(self, device=device)
@@ -178,6 +181,8 @@ class SDXL(supported_models_base.BASE):
 
     latent_format = latent_formats.SDXL
 
+    memory_usage_factor = 0.8
+
     def model_type(self, state_dict, prefix=""):
         if 'edm_mean' in state_dict and 'edm_std' in state_dict: #Playground V2.5
             self.latent_format = latent_formats.SDXL_Playground_2_5()
@@ -505,6 +510,9 @@ class SD3(supported_models_base.BASE):
 
     unet_extra_config = {}
     latent_format = latent_formats.SD3
+
+    memory_usage_factor = 1.2
+
     text_encoder_key_prefix = ["text_encoders."]
 
     def get_model(self, state_dict, prefix="", device=None):
@@ -631,7 +639,10 @@ class Flux(supported_models_base.BASE):
 
     unet_extra_config = {}
     latent_format = latent_formats.Flux
-    supported_inference_dtypes = [torch.bfloat16, torch.float32]
+
+    memory_usage_factor = 2.8
+
+    supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32]
 
     vae_key_prefix = ["vae."]
     text_encoder_key_prefix = ["text_encoders."]
@@ -641,7 +652,12 @@ class Flux(supported_models_base.BASE):
         return out
 
     def clip_target(self, state_dict={}):
-        return supported_models_base.ClipTarget(comfy.text_encoders.flux.FluxTokenizer, comfy.text_encoders.flux.FluxClipModel)
+        pref = self.text_encoder_key_prefix[0]
+        t5_key = "{}t5xxl.transformer.encoder.final_layer_norm.weight".format(pref)
+        dtype_t5 = None
+        if t5_key in state_dict:
+            dtype_t5 = state_dict[t5_key].dtype
+        return supported_models_base.ClipTarget(comfy.text_encoders.flux.FluxTokenizer, comfy.text_encoders.flux.flux_clip(dtype_t5=dtype_t5))
 
 class FluxSchnell(Flux):
     unet_config = {

comfy/supported_models_base.py

@@ -1,3 +1,21 @@
+"""
+    This file is part of ComfyUI.
+    Copyright (C) 2024 Comfy
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+"""
+
 import torch
 from . import model_base
 from . import utils
@@ -27,7 +45,10 @@ class BASE:
     text_encoder_key_prefix = ["cond_stage_model."]
     supported_inference_dtypes = [torch.float16, torch.bfloat16, torch.float32]
 
+    memory_usage_factor = 2.0
+
     manual_cast_dtype = None
+    custom_operations = None
 
     @classmethod
     def matches(s, unet_config, state_dict=None):

comfy/text_encoders/aura_t5.py

@@ -4,9 +4,9 @@ import comfy.text_encoders.t5
 import os
 
 class PT5XlModel(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_pile_config_xl.json")
-        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 2, "pad": 1}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=True, zero_out_masked=True)
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 2, "pad": 1}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=True, zero_out_masked=True, model_options=model_options)
 
 class PT5XlTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -18,5 +18,5 @@ class AuraT5Tokenizer(sd1_clip.SD1Tokenizer):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="pile_t5xl", tokenizer=PT5XlTokenizer)
 
 class AuraT5Model(sd1_clip.SD1ClipModel):
-    def __init__(self, device="cpu", dtype=None, **kwargs):
-        super().__init__(device=device, dtype=dtype, name="pile_t5xl", clip_model=PT5XlModel, **kwargs)
+    def __init__(self, device="cpu", dtype=None, model_options={}, **kwargs):
+        super().__init__(device=device, dtype=dtype, model_options=model_options, name="pile_t5xl", clip_model=PT5XlModel, **kwargs)

comfy/text_encoders/flux.py

@@ -6,9 +6,9 @@ import torch
 import os
 
 class T5XXLModel(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_config_xxl.json")
-        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5)
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, model_options=model_options)
 
 class T5XXLTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -35,11 +35,11 @@ class FluxTokenizer:
 
 
 class FluxClipModel(torch.nn.Module):
-    def __init__(self, dtype_t5=None, device="cpu", dtype=None):
+    def __init__(self, dtype_t5=None, device="cpu", dtype=None, model_options={}):
         super().__init__()
         dtype_t5 = comfy.model_management.pick_weight_dtype(dtype_t5, dtype, device)
-        self.clip_l = sd1_clip.SDClipModel(device=device, dtype=dtype, return_projected_pooled=False)
-        self.t5xxl = T5XXLModel(device=device, dtype=dtype_t5)
+        self.clip_l = sd1_clip.SDClipModel(device=device, dtype=dtype, return_projected_pooled=False, model_options=model_options)
+        self.t5xxl = T5XXLModel(device=device, dtype=dtype_t5, model_options=model_options)
         self.dtypes = set([dtype, dtype_t5])
 
     def set_clip_options(self, options):
@@ -52,9 +52,9 @@ class FluxClipModel(torch.nn.Module):
 
     def encode_token_weights(self, token_weight_pairs):
         token_weight_pairs_l = token_weight_pairs["l"]
-        token_weight_pars_t5 = token_weight_pairs["t5xxl"]
+        token_weight_pairs_t5 = token_weight_pairs["t5xxl"]
 
-        t5_out, t5_pooled = self.t5xxl.encode_token_weights(token_weight_pars_t5)
+        t5_out, t5_pooled = self.t5xxl.encode_token_weights(token_weight_pairs_t5)
         l_out, l_pooled = self.clip_l.encode_token_weights(token_weight_pairs_l)
         return t5_out, l_pooled
 
@@ -66,6 +66,6 @@ class FluxClipModel(torch.nn.Module):
 
 def flux_clip(dtype_t5=None):
     class FluxClipModel_(FluxClipModel):
-        def __init__(self, device="cpu", dtype=None):
-            super().__init__(dtype_t5=dtype_t5, device=device, dtype=dtype)
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            super().__init__(dtype_t5=dtype_t5, device=device, dtype=dtype, model_options=model_options)
     return FluxClipModel_

comfy/text_encoders/hydit.py

@@ -7,9 +7,9 @@ import os
 import torch
 
 class HyditBertModel(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "hydit_clip.json")
-        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"start": 101, "end": 102, "pad": 0}, model_class=BertModel, enable_attention_masks=True, return_attention_masks=True)
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"start": 101, "end": 102, "pad": 0}, model_class=BertModel, enable_attention_masks=True, return_attention_masks=True, model_options=model_options)
 
 class HyditBertTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -18,9 +18,9 @@ class HyditBertTokenizer(sd1_clip.SDTokenizer):
 
 
 class MT5XLModel(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "mt5_config_xl.json")
-        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=True, return_attention_masks=True)
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=True, return_attention_masks=True, model_options=model_options)
 
 class MT5XLTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -50,10 +50,10 @@ class HyditTokenizer:
         return {"mt5xl.spiece_model": self.mt5xl.state_dict()["spiece_model"]}
 
 class HyditModel(torch.nn.Module):
-    def __init__(self, device="cpu", dtype=None):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
         super().__init__()
-        self.hydit_clip = HyditBertModel(dtype=dtype)
-        self.mt5xl = MT5XLModel(dtype=dtype)
+        self.hydit_clip = HyditBertModel(dtype=dtype, model_options=model_options)
+        self.mt5xl = MT5XLModel(dtype=dtype, model_options=model_options)
 
         self.dtypes = set()
         if dtype is not None:

comfy/text_encoders/long_clipl.json

@@ -0,0 +1,25 @@
+{
+  "_name_or_path": "openai/clip-vit-large-patch14",
+  "architectures": [
+    "CLIPTextModel"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 0,
+  "dropout": 0.0,
+  "eos_token_id": 49407,
+  "hidden_act": "quick_gelu",
+  "hidden_size": 768,
+  "initializer_factor": 1.0,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 248,
+  "model_type": "clip_text_model",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "pad_token_id": 1,
+  "projection_dim": 768,
+  "torch_dtype": "float32",
+  "transformers_version": "4.24.0",
+  "vocab_size": 49408
+}

comfy/text_encoders/long_clipl.py

@@ -0,0 +1,19 @@
+from comfy import sd1_clip
+import os
+
+class LongClipTokenizer_(sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(max_length=248, embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
+
+class LongClipModel_(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "long_clipl.json")
+        super().__init__(device=device, textmodel_json_config=textmodel_json_config, return_projected_pooled=False, dtype=dtype, model_options=model_options)
+
+class LongClipTokenizer(sd1_clip.SD1Tokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, tokenizer=LongClipTokenizer_)
+
+class LongClipModel(sd1_clip.SD1ClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}, **kwargs):
+        super().__init__(device=device, dtype=dtype, model_options=model_options, clip_model=LongClipModel_, **kwargs)

comfy/text_encoders/sa_t5.py

@@ -4,9 +4,9 @@ import comfy.text_encoders.t5
 import os
 
 class T5BaseModel(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_config_base.json")
-        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=True, zero_out_masked=True)
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, model_options=model_options, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=True, zero_out_masked=True)
 
 class T5BaseTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -18,5 +18,5 @@ class SAT5Tokenizer(sd1_clip.SD1Tokenizer):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="t5base", tokenizer=T5BaseTokenizer)
 
 class SAT5Model(sd1_clip.SD1ClipModel):
-    def __init__(self, device="cpu", dtype=None, **kwargs):
-        super().__init__(device=device, dtype=dtype, name="t5base", clip_model=T5BaseModel, **kwargs)
+    def __init__(self, device="cpu", dtype=None, model_options={}, **kwargs):
+        super().__init__(device=device, dtype=dtype, model_options=model_options, name="t5base", clip_model=T5BaseModel, **kwargs)

comfy/text_encoders/sd2_clip.py

@@ -2,13 +2,13 @@ from comfy import sd1_clip
 import os
 
 class SD2ClipHModel(sd1_clip.SDClipModel):
-    def __init__(self, arch="ViT-H-14", device="cpu", max_length=77, freeze=True, layer="penultimate", layer_idx=None, dtype=None):
+    def __init__(self, arch="ViT-H-14", device="cpu", max_length=77, freeze=True, layer="penultimate", layer_idx=None, dtype=None, model_options={}):
         if layer == "penultimate":
             layer="hidden"
             layer_idx=-2
 
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd2_clip_config.json")
-        super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"start": 49406, "end": 49407, "pad": 0})
+        super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"start": 49406, "end": 49407, "pad": 0}, return_projected_pooled=True, model_options=model_options)
 
 class SD2ClipHTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, tokenizer_path=None, embedding_directory=None, tokenizer_data={}):
@@ -19,5 +19,5 @@ class SD2Tokenizer(sd1_clip.SD1Tokenizer):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="h", tokenizer=SD2ClipHTokenizer)
 
 class SD2ClipModel(sd1_clip.SD1ClipModel):
-    def __init__(self, device="cpu", dtype=None, **kwargs):
-        super().__init__(device=device, dtype=dtype, clip_name="h", clip_model=SD2ClipHModel, **kwargs)
+    def __init__(self, device="cpu", dtype=None, model_options={}, **kwargs):
+        super().__init__(device=device, dtype=dtype, model_options=model_options, clip_name="h", clip_model=SD2ClipHModel, **kwargs)

comfy/text_encoders/sd3_clip.py

@@ -8,14 +8,14 @@ import comfy.model_management
 import logging
 
 class T5XXLModel(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_config_xxl.json")
-        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5)
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, model_options=model_options)
 
 class T5XXLTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_tokenizer")
-        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=4096, embedding_key='t5xxl', tokenizer_class=T5TokenizerFast, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=77)
+        super().__init__(tokenizer_path, embedding_directory=embedding_directory, pad_with_end=False, embedding_size=4096, embedding_key='t5xxl', tokenizer_class=T5TokenizerFast, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=77)
 
 
 class SD3Tokenizer:
@@ -38,24 +38,24 @@ class SD3Tokenizer:
         return {}
 
 class SD3ClipModel(torch.nn.Module):
-    def __init__(self, clip_l=True, clip_g=True, t5=True, dtype_t5=None, device="cpu", dtype=None):
+    def __init__(self, clip_l=True, clip_g=True, t5=True, dtype_t5=None, device="cpu", dtype=None, model_options={}):
         super().__init__()
         self.dtypes = set()
         if clip_l:
-            self.clip_l = sd1_clip.SDClipModel(layer="hidden", layer_idx=-2, device=device, dtype=dtype, layer_norm_hidden_state=False, return_projected_pooled=False)
+            self.clip_l = sd1_clip.SDClipModel(layer="hidden", layer_idx=-2, device=device, dtype=dtype, layer_norm_hidden_state=False, return_projected_pooled=False, model_options=model_options)
             self.dtypes.add(dtype)
         else:
             self.clip_l = None
 
         if clip_g:
-            self.clip_g = sdxl_clip.SDXLClipG(device=device, dtype=dtype)
+            self.clip_g = sdxl_clip.SDXLClipG(device=device, dtype=dtype, model_options=model_options)
             self.dtypes.add(dtype)
         else:
             self.clip_g = None
 
         if t5:
             dtype_t5 = comfy.model_management.pick_weight_dtype(dtype_t5, dtype, device)
-            self.t5xxl = T5XXLModel(device=device, dtype=dtype_t5)
+            self.t5xxl = T5XXLModel(device=device, dtype=dtype_t5, model_options=model_options)
             self.dtypes.add(dtype_t5)
         else:
             self.t5xxl = None
@@ -81,7 +81,7 @@ class SD3ClipModel(torch.nn.Module):
     def encode_token_weights(self, token_weight_pairs):
         token_weight_pairs_l = token_weight_pairs["l"]
         token_weight_pairs_g = token_weight_pairs["g"]
-        token_weight_pars_t5 = token_weight_pairs["t5xxl"]
+        token_weight_pairs_t5 = token_weight_pairs["t5xxl"]
         lg_out = None
         pooled = None
         out = None
@@ -108,7 +108,7 @@ class SD3ClipModel(torch.nn.Module):
             pooled = torch.cat((l_pooled, g_pooled), dim=-1)
 
         if self.t5xxl is not None:
-            t5_out, t5_pooled = self.t5xxl.encode_token_weights(token_weight_pars_t5)
+            t5_out, t5_pooled = self.t5xxl.encode_token_weights(token_weight_pairs_t5)
             if lg_out is not None:
                 out = torch.cat([lg_out, t5_out], dim=-2)
             else:
@@ -132,6 +132,6 @@ class SD3ClipModel(torch.nn.Module):
 
 def sd3_clip(clip_l=True, clip_g=True, t5=True, dtype_t5=None):
     class SD3ClipModel_(SD3ClipModel):
-        def __init__(self, device="cpu", dtype=None):
-            super().__init__(clip_l=clip_l, clip_g=clip_g, t5=t5, dtype_t5=dtype_t5, device=device, dtype=dtype)
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            super().__init__(clip_l=clip_l, clip_g=clip_g, t5=t5, dtype_t5=dtype_t5, device=device, dtype=dtype, model_options=model_options)
     return SD3ClipModel_

comfy/utils.py

@@ -1,3 +1,22 @@
+"""
+    This file is part of ComfyUI.
+    Copyright (C) 2024 Comfy
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+"""
+
+
 import torch
 import math
 import struct
@@ -40,9 +59,22 @@ def calculate_parameters(sd, prefix=""):
     params = 0
     for k in sd.keys():
         if k.startswith(prefix):
-            params += sd[k].nelement()
+            w = sd[k]
+            params += w.nelement()
     return params
 
+def weight_dtype(sd, prefix=""):
+    dtypes = {}
+    for k in sd.keys():
+        if k.startswith(prefix):
+            w = sd[k]
+            dtypes[w.dtype] = dtypes.get(w.dtype, 0) + 1
+
+    if len(dtypes) == 0:
+        return None
+
+    return max(dtypes, key=dtypes.get)
+
 def state_dict_key_replace(state_dict, keys_to_replace):
     for x in keys_to_replace:
         if x in state_dict:
@@ -402,6 +434,110 @@ def auraflow_to_diffusers(mmdit_config, output_prefix=""):
 
     return key_map
 
+def flux_to_diffusers(mmdit_config, output_prefix=""):
+    n_double_layers = mmdit_config.get("depth", 0)
+    n_single_layers = mmdit_config.get("depth_single_blocks", 0)
+    hidden_size = mmdit_config.get("hidden_size", 0)
+
+    key_map = {}
+    for index in range(n_double_layers):
+        prefix_from = "transformer_blocks.{}".format(index)
+        prefix_to = "{}double_blocks.{}".format(output_prefix, index)
+
+        for end in ("weight", "bias"):
+            k = "{}.attn.".format(prefix_from)
+            qkv = "{}.img_attn.qkv.{}".format(prefix_to, end)
+            key_map["{}to_q.{}".format(k, end)] = (qkv, (0, 0, hidden_size))
+            key_map["{}to_k.{}".format(k, end)] = (qkv, (0, hidden_size, hidden_size))
+            key_map["{}to_v.{}".format(k, end)] = (qkv, (0, hidden_size * 2, hidden_size))
+
+            k = "{}.attn.".format(prefix_from)
+            qkv = "{}.txt_attn.qkv.{}".format(prefix_to, end)
+            key_map["{}add_q_proj.{}".format(k, end)] = (qkv, (0, 0, hidden_size))
+            key_map["{}add_k_proj.{}".format(k, end)] = (qkv, (0, hidden_size, hidden_size))
+            key_map["{}add_v_proj.{}".format(k, end)] = (qkv, (0, hidden_size * 2, hidden_size))
+
+        block_map = {
+                        "attn.to_out.0.weight": "img_attn.proj.weight",
+                        "attn.to_out.0.bias": "img_attn.proj.bias",
+                        "norm1.linear.weight": "img_mod.lin.weight",
+                        "norm1.linear.bias": "img_mod.lin.bias",
+                        "norm1_context.linear.weight": "txt_mod.lin.weight",
+                        "norm1_context.linear.bias": "txt_mod.lin.bias",
+                        "attn.to_add_out.weight": "txt_attn.proj.weight",
+                        "attn.to_add_out.bias": "txt_attn.proj.bias",
+                        "ff.net.0.proj.weight": "img_mlp.0.weight",
+                        "ff.net.0.proj.bias": "img_mlp.0.bias",
+                        "ff.net.2.weight": "img_mlp.2.weight",
+                        "ff.net.2.bias": "img_mlp.2.bias",
+                        "ff_context.net.0.proj.weight": "txt_mlp.0.weight",
+                        "ff_context.net.0.proj.bias": "txt_mlp.0.bias",
+                        "ff_context.net.2.weight": "txt_mlp.2.weight",
+                        "ff_context.net.2.bias": "txt_mlp.2.bias",
+                        "attn.norm_q.weight": "img_attn.norm.query_norm.scale",
+                        "attn.norm_k.weight": "img_attn.norm.key_norm.scale",
+                        "attn.norm_added_q.weight": "txt_attn.norm.query_norm.scale",
+                        "attn.norm_added_k.weight": "txt_attn.norm.key_norm.scale",
+                    }
+
+        for k in block_map:
+            key_map["{}.{}".format(prefix_from, k)] = "{}.{}".format(prefix_to, block_map[k])
+
+    for index in range(n_single_layers):
+        prefix_from = "single_transformer_blocks.{}".format(index)
+        prefix_to = "{}single_blocks.{}".format(output_prefix, index)
+
+        for end in ("weight", "bias"):
+            k = "{}.attn.".format(prefix_from)
+            qkv = "{}.linear1.{}".format(prefix_to, end)
+            key_map["{}to_q.{}".format(k, end)] = (qkv, (0, 0, hidden_size))
+            key_map["{}to_k.{}".format(k, end)] = (qkv, (0, hidden_size, hidden_size))
+            key_map["{}to_v.{}".format(k, end)] = (qkv, (0, hidden_size * 2, hidden_size))
+            key_map["{}.proj_mlp.{}".format(prefix_from, end)] = (qkv, (0, hidden_size * 3, hidden_size * 4))
+
+        block_map = {
+                        "norm.linear.weight": "modulation.lin.weight",
+                        "norm.linear.bias": "modulation.lin.bias",
+                        "proj_out.weight": "linear2.weight",
+                        "proj_out.bias": "linear2.bias",
+                        "attn.norm_q.weight": "norm.query_norm.scale",
+                        "attn.norm_k.weight": "norm.key_norm.scale",
+                    }
+
+        for k in block_map:
+            key_map["{}.{}".format(prefix_from, k)] = "{}.{}".format(prefix_to, block_map[k])
+
+    MAP_BASIC = {
+        ("final_layer.linear.bias", "proj_out.bias"),
+        ("final_layer.linear.weight", "proj_out.weight"),
+        ("img_in.bias", "x_embedder.bias"),
+        ("img_in.weight", "x_embedder.weight"),
+        ("time_in.in_layer.bias", "time_text_embed.timestep_embedder.linear_1.bias"),
+        ("time_in.in_layer.weight", "time_text_embed.timestep_embedder.linear_1.weight"),
+        ("time_in.out_layer.bias", "time_text_embed.timestep_embedder.linear_2.bias"),
+        ("time_in.out_layer.weight", "time_text_embed.timestep_embedder.linear_2.weight"),
+        ("txt_in.bias", "context_embedder.bias"),
+        ("txt_in.weight", "context_embedder.weight"),
+        ("vector_in.in_layer.bias", "time_text_embed.text_embedder.linear_1.bias"),
+        ("vector_in.in_layer.weight", "time_text_embed.text_embedder.linear_1.weight"),
+        ("vector_in.out_layer.bias", "time_text_embed.text_embedder.linear_2.bias"),
+        ("vector_in.out_layer.weight", "time_text_embed.text_embedder.linear_2.weight"),
+        ("guidance_in.in_layer.bias", "time_text_embed.guidance_embedder.linear_1.bias"),
+        ("guidance_in.in_layer.weight", "time_text_embed.guidance_embedder.linear_1.weight"),
+        ("guidance_in.out_layer.bias", "time_text_embed.guidance_embedder.linear_2.bias"),
+        ("guidance_in.out_layer.weight", "time_text_embed.guidance_embedder.linear_2.weight"),
+        ("final_layer.adaLN_modulation.1.bias", "norm_out.linear.bias", swap_scale_shift),
+        ("final_layer.adaLN_modulation.1.weight", "norm_out.linear.weight", swap_scale_shift),
+    }
+
+    for k in MAP_BASIC:
+        if len(k) > 2:
+            key_map[k[1]] = ("{}{}".format(output_prefix, k[0]), None, k[2])
+        else:
+            key_map[k[1]] = "{}{}".format(output_prefix, k[0])
+
+    return key_map
+
 def repeat_to_batch_size(tensor, batch_size, dim=0):
     if tensor.shape[dim] > batch_size:
         return tensor.narrow(dim, 0, batch_size)