Merge branch 'master' into worksplit-multigpu-wip

comfy/cli_args.py

@@ -49,7 +49,8 @@ parser.add_argument("--temp-directory", type=str, default=None, help="Set the Co
 parser.add_argument("--input-directory", type=str, default=None, help="Set the ComfyUI input directory. Overrides --base-directory.")
 parser.add_argument("--auto-launch", action="store_true", help="Automatically launch ComfyUI in the default browser.")
 parser.add_argument("--disable-auto-launch", action="store_true", help="Disable auto launching the browser.")
-parser.add_argument("--cuda-device", type=str, default=None, metavar="DEVICE_ID", help="Set the ids of cuda devices this instance will use.")
+parser.add_argument("--cuda-device", type=str, default=None, metavar="DEVICE_ID", help="Set the ids of cuda devices this instance will use. All other devices will not be visible.")
+parser.add_argument("--default-device", type=int, default=None, metavar="DEFAULT_DEVICE_ID", help="Set the id of the default device, all other devices will stay visible.")
 cm_group = parser.add_mutually_exclusive_group()
 cm_group.add_argument("--cuda-malloc", action="store_true", help="Enable cudaMallocAsync (enabled by default for torch 2.0 and up).")
 cm_group.add_argument("--disable-cuda-malloc", action="store_true", help="Disable cudaMallocAsync.")

comfy/k_diffusion/sampling.py

@@ -1210,39 +1210,21 @@ def sample_deis(model, x, sigmas, extra_args=None, callback=None, disable=None,
     return x_next
 
 
-@torch.no_grad()
-def sample_euler_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None):
-    extra_args = {} if extra_args is None else extra_args
-
-    temp = [0]
-    def post_cfg_function(args):
-        temp[0] = args["uncond_denoised"]
-        return args["denoised"]
-
-    model_options = extra_args.get("model_options", {}).copy()
-    extra_args["model_options"] = comfy.model_patcher.set_model_options_post_cfg_function(model_options, post_cfg_function, disable_cfg1_optimization=True)
-
-    s_in = x.new_ones([x.shape[0]])
-    for i in trange(len(sigmas) - 1, disable=disable):
-        sigma_hat = sigmas[i]
-        denoised = model(x, sigma_hat * s_in, **extra_args)
-        d = to_d(x, sigma_hat, temp[0])
-        if callback is not None:
-            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised})
-        # Euler method
-        x = denoised + d * sigmas[i + 1]
-    return x
-
 @torch.no_grad()
 def sample_euler_ancestral_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
-    """Ancestral sampling with Euler method steps."""
+    """Ancestral sampling with Euler method steps (CFG++)."""
     extra_args = {} if extra_args is None else extra_args
     seed = extra_args.get("seed", None)
     noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler
 
-    temp = [0]
+    model_sampling = model.inner_model.model_patcher.get_model_object("model_sampling")
+    lambda_fn = partial(sigma_to_half_log_snr, model_sampling=model_sampling)
+
+    uncond_denoised = None
+
     def post_cfg_function(args):
-        temp[0] = args["uncond_denoised"]
+        nonlocal uncond_denoised
+        uncond_denoised = args["uncond_denoised"]
         return args["denoised"]
 
     model_options = extra_args.get("model_options", {}).copy()
@@ -1251,15 +1233,33 @@ def sample_euler_ancestral_cfg_pp(model, x, sigmas, extra_args=None, callback=No
     s_in = x.new_ones([x.shape[0]])
     for i in trange(len(sigmas) - 1, disable=disable):
         denoised = model(x, sigmas[i] * s_in, **extra_args)
-        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})
-        d = to_d(x, sigmas[i], temp[0])
-        # Euler method
-        x = denoised + d * sigma_down
-        if sigmas[i + 1] > 0:
-            x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * sigma_up
+        if sigmas[i + 1] == 0:
+            # Denoising step
+            x = denoised
+        else:
+            alpha_s = sigmas[i] * lambda_fn(sigmas[i]).exp()
+            alpha_t = sigmas[i + 1] * lambda_fn(sigmas[i + 1]).exp()
+            d = to_d(x, sigmas[i], alpha_s * uncond_denoised)   # to noise
+
+            # DDIM stochastic sampling
+            sigma_down, sigma_up = get_ancestral_step(sigmas[i] / alpha_s, sigmas[i + 1] / alpha_t, eta=eta)
+            sigma_down = alpha_t * sigma_down
+
+            # Euler method
+            x = alpha_t * denoised + sigma_down * d
+            if eta > 0 and s_noise > 0:
+                x = x + alpha_t * noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * sigma_up
     return x
+
+
+@torch.no_grad()
+def sample_euler_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None):
+    """Euler method steps (CFG++)."""
+    return sample_euler_ancestral_cfg_pp(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=0.0, s_noise=0.0, noise_sampler=None)
+
+
 @torch.no_grad()
 def sample_dpmpp_2s_ancestral_cfg_pp(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."""

comfy/ldm/wan/vae.py

@@ -52,15 +52,6 @@ class RMS_norm(nn.Module):
             x, dim=(1 if self.channel_first else -1)) * self.scale * self.gamma.to(x) + (self.bias.to(x) if self.bias is not None else 0)
 
 
-class Upsample(nn.Upsample):
-
-    def forward(self, x):
-        """
-        Fix bfloat16 support for nearest neighbor interpolation.
-        """
-        return super().forward(x.float()).type_as(x)
-
-
 class Resample(nn.Module):
 
     def __init__(self, dim, mode):
@@ -73,11 +64,11 @@ class Resample(nn.Module):
         # layers
         if mode == 'upsample2d':
             self.resample = nn.Sequential(
-                Upsample(scale_factor=(2., 2.), mode='nearest-exact'),
+                nn.Upsample(scale_factor=(2., 2.), mode='nearest-exact'),
                 ops.Conv2d(dim, dim // 2, 3, padding=1))
         elif mode == 'upsample3d':
             self.resample = nn.Sequential(
-                Upsample(scale_factor=(2., 2.), mode='nearest-exact'),
+                nn.Upsample(scale_factor=(2., 2.), mode='nearest-exact'),
                 ops.Conv2d(dim, dim // 2, 3, padding=1))
             self.time_conv = CausalConv3d(
                 dim, dim * 2, (3, 1, 1), padding=(1, 0, 0))
@@ -157,29 +148,6 @@ class Resample(nn.Module):
                     feat_idx[0] += 1
         return x
 
-    def init_weight(self, conv):
-        conv_weight = conv.weight
-        nn.init.zeros_(conv_weight)
-        c1, c2, t, h, w = conv_weight.size()
-        one_matrix = torch.eye(c1, c2)
-        init_matrix = one_matrix
-        nn.init.zeros_(conv_weight)
-        #conv_weight.data[:,:,-1,1,1] = init_matrix * 0.5
-        conv_weight.data[:, :, 1, 0, 0] = init_matrix  #* 0.5
-        conv.weight.data.copy_(conv_weight)
-        nn.init.zeros_(conv.bias.data)
-
-    def init_weight2(self, conv):
-        conv_weight = conv.weight.data
-        nn.init.zeros_(conv_weight)
-        c1, c2, t, h, w = conv_weight.size()
-        init_matrix = torch.eye(c1 // 2, c2)
-        #init_matrix = repeat(init_matrix, 'o ... -> (o 2) ...').permute(1,0,2).contiguous().reshape(c1,c2)
-        conv_weight[:c1 // 2, :, -1, 0, 0] = init_matrix
-        conv_weight[c1 // 2:, :, -1, 0, 0] = init_matrix
-        conv.weight.data.copy_(conv_weight)
-        nn.init.zeros_(conv.bias.data)
-
 
 class ResidualBlock(nn.Module):
 
@@ -494,12 +462,6 @@ class WanVAE(nn.Module):
         self.decoder = Decoder3d(dim, z_dim, dim_mult, num_res_blocks,
                                  attn_scales, self.temperal_upsample, dropout)
 
-    def forward(self, x):
-        mu, log_var = self.encode(x)
-        z = self.reparameterize(mu, log_var)
-        x_recon = self.decode(z)
-        return x_recon, mu, log_var
-
     def encode(self, x):
         self.clear_cache()
         ## cache
@@ -545,18 +507,6 @@ class WanVAE(nn.Module):
         self.clear_cache()
         return out
 
-    def reparameterize(self, mu, log_var):
-        std = torch.exp(0.5 * log_var)
-        eps = torch.randn_like(std)
-        return eps * std + mu
-
-    def sample(self, imgs, deterministic=False):
-        mu, log_var = self.encode(imgs)
-        if deterministic:
-            return mu
-        std = torch.exp(0.5 * log_var.clamp(-30.0, 20.0))
-        return mu + std * torch.randn_like(std)
-
     def clear_cache(self):
         self._conv_num = count_conv3d(self.decoder)
         self._conv_idx = [0]

comfy/model_management.py

@@ -106,7 +106,7 @@ if args.directml is not None:
     lowvram_available = False #TODO: need to find a way to get free memory in directml before this can be enabled by default.
 
 try:
-    import intel_extension_for_pytorch as ipex
+    import intel_extension_for_pytorch as ipex  # noqa: F401
     _ = torch.xpu.device_count()
     xpu_available = xpu_available or torch.xpu.is_available()
 except:
@@ -133,6 +133,11 @@ try:
 except:
     mlu_available = False
 
+try:
+    ixuca_available = hasattr(torch, "corex")
+except:
+    ixuca_available = False
+
 if args.cpu:
     cpu_state = CPUState.CPU
 
@@ -156,6 +161,12 @@ def is_mlu():
         return True
     return False
 
+def is_ixuca():
+    global ixuca_available
+    if ixuca_available:
+        return True
+    return False
+
 def get_torch_device():
     global directml_enabled
     global cpu_state
@@ -210,8 +221,9 @@ def get_total_memory(dev=None, torch_total_too=False):
         elif is_intel_xpu():
             stats = torch.xpu.memory_stats(dev)
             mem_reserved = stats['reserved_bytes.all.current']
+            mem_total_xpu = torch.xpu.get_device_properties(dev).total_memory
             mem_total_torch = mem_reserved
-            mem_total = torch.xpu.get_device_properties(dev).total_memory
+            mem_total = mem_total_xpu
         elif is_ascend_npu():
             stats = torch.npu.memory_stats(dev)
             mem_reserved = stats['reserved_bytes.all.current']
@@ -312,7 +324,7 @@ try:
         if torch_version_numeric[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
-    if is_intel_xpu() or is_ascend_npu() or is_mlu():
+    if is_intel_xpu() or is_ascend_npu() or is_mlu() or is_ixuca():
         if args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
             ENABLE_PYTORCH_ATTENTION = True
 except:
@@ -331,7 +343,10 @@ try:
         logging.info("ROCm version: {}".format(rocm_version))
         if args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
             if torch_version_numeric >= (2, 7):  # works on 2.6 but doesn't actually seem to improve much
-                if any((a in arch) for a in ["gfx90a", "gfx942", "gfx1100", "gfx1101", "gfx1151"]):  # TODO: more arches, TODO: gfx1201 and gfx950
+                if any((a in arch) for a in ["gfx90a", "gfx942", "gfx1100", "gfx1101", "gfx1151"]):  # TODO: more arches, TODO: gfx950
+                    ENABLE_PYTORCH_ATTENTION = True
+            if torch_version_numeric >= (2, 8):
+                if any((a in arch) for a in ["gfx1201"]):
                     ENABLE_PYTORCH_ATTENTION = True
         if torch_version_numeric >= (2, 7) and rocm_version >= (6, 4):
             if any((a in arch) for a in ["gfx1201", "gfx942", "gfx950"]):  # TODO: more arches
@@ -401,6 +416,8 @@ def get_torch_device_name(device):
             except:
                 allocator_backend = ""
             return "{} {} : {}".format(device, torch.cuda.get_device_name(device), allocator_backend)
+        elif device.type == "xpu":
+            return "{} {}".format(device, torch.xpu.get_device_name(device))
         else:
             return "{}".format(device.type)
     elif is_intel_xpu():
@@ -904,6 +921,7 @@ def vae_dtype(device=None, allowed_dtypes=[]):
             return d
 
         # NOTE: bfloat16 seems to work on AMD for the VAE but is extremely slow in some cases compared to fp32
+        # slowness still a problem on pytorch nightly 2.9.0.dev20250720+rocm6.4 tested on RDNA3
         if d == torch.bfloat16 and (not is_amd()) and should_use_bf16(device):
             return d
 
@@ -957,7 +975,7 @@ def device_supports_non_blocking(device):
     if is_device_mps(device):
         return False #pytorch bug? mps doesn't support non blocking
     if is_intel_xpu():
-        return False
+        return True
     if args.deterministic: #TODO: figure out why deterministic breaks non blocking from gpu to cpu (previews)
         return False
     if directml_enabled:
@@ -996,6 +1014,8 @@ def get_offload_stream(device):
         stream_counter = (stream_counter + 1) % len(ss)
         if is_device_cuda(device):
             ss[stream_counter].wait_stream(torch.cuda.current_stream())
+        elif is_device_xpu(device):
+            ss[stream_counter].wait_stream(torch.xpu.current_stream())
         stream_counters[device] = stream_counter
         return s
     elif is_device_cuda(device):
@@ -1007,6 +1027,15 @@ def get_offload_stream(device):
         stream_counter = (stream_counter + 1) % len(ss)
         stream_counters[device] = stream_counter
         return s
+    elif is_device_xpu(device):
+        ss = []
+        for k in range(NUM_STREAMS):
+            ss.append(torch.xpu.Stream(device=device, priority=0))
+        STREAMS[device] = ss
+        s = ss[stream_counter]
+        stream_counter = (stream_counter + 1) % len(ss)
+        stream_counters[device] = stream_counter
+        return s
     return None
 
 def sync_stream(device, stream):
@@ -1014,6 +1043,8 @@ def sync_stream(device, stream):
         return
     if is_device_cuda(device):
         torch.cuda.current_stream().wait_stream(stream)
+    elif is_device_xpu(device):
+        torch.xpu.current_stream().wait_stream(stream)
 
 def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, stream=None):
     if device is None or weight.device == device:
@@ -1055,6 +1086,8 @@ def xformers_enabled():
         return False
     if is_mlu():
         return False
+    if is_ixuca():
+        return False
     if directml_enabled:
         return False
     return XFORMERS_IS_AVAILABLE
@@ -1090,6 +1123,8 @@ def pytorch_attention_flash_attention():
             return True
         if is_amd():
             return True #if you have pytorch attention enabled on AMD it probably supports at least mem efficient attention
+        if is_ixuca():
+            return True
     return False
 
 def force_upcast_attention_dtype():
@@ -1120,8 +1155,8 @@ def get_free_memory(dev=None, torch_free_too=False):
             stats = torch.xpu.memory_stats(dev)
             mem_active = stats['active_bytes.all.current']
             mem_reserved = stats['reserved_bytes.all.current']
-            mem_free_torch = mem_reserved - mem_active
             mem_free_xpu = torch.xpu.get_device_properties(dev).total_memory - mem_reserved
+            mem_free_torch = mem_reserved - mem_active
             mem_free_total = mem_free_xpu + mem_free_torch
         elif is_ascend_npu():
             stats = torch.npu.memory_stats(dev)
@@ -1170,6 +1205,9 @@ def is_device_cpu(device):
 def is_device_mps(device):
     return is_device_type(device, 'mps')
 
+def is_device_xpu(device):
+    return is_device_type(device, 'xpu')
+
 def is_device_cuda(device):
     return is_device_type(device, 'cuda')
 
@@ -1201,7 +1239,10 @@ def should_use_fp16(device=None, model_params=0, prioritize_performance=True, ma
         return False
 
     if is_intel_xpu():
-        return True
+        if torch_version_numeric < (2, 3):
+            return True
+        else:
+            return torch.xpu.get_device_properties(device).has_fp16
 
     if is_ascend_npu():
         return True
@@ -1209,6 +1250,9 @@ def should_use_fp16(device=None, model_params=0, prioritize_performance=True, ma
     if is_mlu():
         return True
 
+    if is_ixuca():
+        return True
+
     if torch.version.hip:
         return True
 
@@ -1264,11 +1308,17 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, ma
         return False
 
     if is_intel_xpu():
-        return True
+        if torch_version_numeric < (2, 6):
+            return True
+        else:
+            return torch.xpu.get_device_capability(device)['has_bfloat16_conversions']
 
     if is_ascend_npu():
         return True
 
+    if is_ixuca():
+        return True
+
     if is_amd():
         arch = torch.cuda.get_device_properties(device).gcnArchName
         if any((a in arch) for a in ["gfx1030", "gfx1031", "gfx1010", "gfx1011", "gfx1012", "gfx906", "gfx900", "gfx803"]):  # RDNA2 and older don't support bf16

comfy/utils.py

@@ -698,6 +698,26 @@ def resize_to_batch_size(tensor, batch_size):
 
     return output
 
+def resize_list_to_batch_size(l, batch_size):
+    in_batch_size = len(l)
+    if in_batch_size == batch_size or in_batch_size == 0:
+        return l
+
+    if batch_size <= 1:
+        return l[:batch_size]
+
+    output = []
+    if batch_size < in_batch_size:
+        scale = (in_batch_size - 1) / (batch_size - 1)
+        for i in range(batch_size):
+            output.append(l[min(round(i * scale), in_batch_size - 1)])
+    else:
+        scale = in_batch_size / batch_size
+        for i in range(batch_size):
+           output.append(l[min(math.floor((i + 0.5) * scale), in_batch_size - 1)])
+
+    return output
+
 def convert_sd_to(state_dict, dtype):
     keys = list(state_dict.keys())
     for k in keys:

comfy/weight_adapter/__init__.py

@@ -15,9 +15,20 @@ adapters: list[type[WeightAdapterBase]] = [
     OFTAdapter,
     BOFTAdapter,
 ]
+adapter_maps: dict[str, type[WeightAdapterBase]] = {
+    "LoRA": LoRAAdapter,
+    "LoHa": LoHaAdapter,
+    "LoKr": LoKrAdapter,
+    "OFT": OFTAdapter,
+    ## We disable not implemented algo for now
+    # "GLoRA": GLoRAAdapter,
+    # "BOFT": BOFTAdapter,
+}
+
 
 __all__ = [
     "WeightAdapterBase",
     "WeightAdapterTrainBase",
-    "adapters"
+    "adapters",
+    "adapter_maps",
 ] + [a.__name__ for a in adapters]

comfy/weight_adapter/base.py

@@ -133,3 +133,43 @@ def tucker_weight_from_conv(up, down, mid):
 def tucker_weight(wa, wb, t):
     temp = torch.einsum("i j ..., j r -> i r ...", t, wb)
     return torch.einsum("i j ..., i r -> r j ...", temp, wa)
+
+
+def factorization(dimension: int, factor: int = -1) -> tuple[int, int]:
+    """
+    return a tuple of two value of input dimension decomposed by the number closest to factor
+    second value is higher or equal than first value.
+
+    examples)
+    factor
+        -1               2                4               8               16               ...
+    127 -> 1, 127   127 -> 1, 127    127 -> 1, 127   127 -> 1, 127   127 -> 1, 127
+    128 -> 8, 16    128 -> 2, 64     128 -> 4, 32    128 -> 8, 16    128 -> 8, 16
+    250 -> 10, 25   250 -> 2, 125    250 -> 2, 125   250 -> 5, 50    250 -> 10, 25
+    360 -> 8, 45    360 -> 2, 180    360 -> 4, 90    360 -> 8, 45    360 -> 12, 30
+    512 -> 16, 32   512 -> 2, 256    512 -> 4, 128   512 -> 8, 64    512 -> 16, 32
+    1024 -> 32, 32  1024 -> 2, 512   1024 -> 4, 256  1024 -> 8, 128  1024 -> 16, 64
+    """
+
+    if factor > 0 and (dimension % factor) == 0 and dimension >= factor**2:
+        m = factor
+        n = dimension // factor
+        if m > n:
+            n, m = m, n
+        return m, n
+    if factor < 0:
+        factor = dimension
+    m, n = 1, dimension
+    length = m + n
+    while m < n:
+        new_m = m + 1
+        while dimension % new_m != 0:
+            new_m += 1
+        new_n = dimension // new_m
+        if new_m + new_n > length or new_m > factor:
+            break
+        else:
+            m, n = new_m, new_n
+    if m > n:
+        n, m = m, n
+    return m, n

comfy/weight_adapter/loha.py

@@ -3,7 +3,120 @@ from typing import Optional
 
 import torch
 import comfy.model_management
-from .base import WeightAdapterBase, weight_decompose
+from .base import WeightAdapterBase, WeightAdapterTrainBase, weight_decompose
+
+
+class HadaWeight(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, w1u, w1d, w2u, w2d, scale=torch.tensor(1)):
+        ctx.save_for_backward(w1d, w1u, w2d, w2u, scale)
+        diff_weight = ((w1u @ w1d) * (w2u @ w2d)) * scale
+        return diff_weight
+
+    @staticmethod
+    def backward(ctx, grad_out):
+        (w1d, w1u, w2d, w2u, scale) = ctx.saved_tensors
+        grad_out = grad_out * scale
+        temp = grad_out * (w2u @ w2d)
+        grad_w1u = temp @ w1d.T
+        grad_w1d = w1u.T @ temp
+
+        temp = grad_out * (w1u @ w1d)
+        grad_w2u = temp @ w2d.T
+        grad_w2d = w2u.T @ temp
+
+        del temp
+        return grad_w1u, grad_w1d, grad_w2u, grad_w2d, None
+
+
+class HadaWeightTucker(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, t1, w1u, w1d, t2, w2u, w2d, scale=torch.tensor(1)):
+        ctx.save_for_backward(t1, w1d, w1u, t2, w2d, w2u, scale)
+
+        rebuild1 = torch.einsum("i j ..., j r, i p -> p r ...", t1, w1d, w1u)
+        rebuild2 = torch.einsum("i j ..., j r, i p -> p r ...", t2, w2d, w2u)
+
+        return rebuild1 * rebuild2 * scale
+
+    @staticmethod
+    def backward(ctx, grad_out):
+        (t1, w1d, w1u, t2, w2d, w2u, scale) = ctx.saved_tensors
+        grad_out = grad_out * scale
+
+        temp = torch.einsum("i j ..., j r -> i r ...", t2, w2d)
+        rebuild = torch.einsum("i j ..., i r -> r j ...", temp, w2u)
+
+        grad_w = rebuild * grad_out
+        del rebuild
+
+        grad_w1u = torch.einsum("r j ..., i j ... -> r i", temp, grad_w)
+        grad_temp = torch.einsum("i j ..., i r -> r j ...", grad_w, w1u.T)
+        del grad_w, temp
+
+        grad_w1d = torch.einsum("i r ..., i j ... -> r j", t1, grad_temp)
+        grad_t1 = torch.einsum("i j ..., j r -> i r ...", grad_temp, w1d.T)
+        del grad_temp
+
+        temp = torch.einsum("i j ..., j r -> i r ...", t1, w1d)
+        rebuild = torch.einsum("i j ..., i r -> r j ...", temp, w1u)
+
+        grad_w = rebuild * grad_out
+        del rebuild
+
+        grad_w2u = torch.einsum("r j ..., i j ... -> r i", temp, grad_w)
+        grad_temp = torch.einsum("i j ..., i r -> r j ...", grad_w, w2u.T)
+        del grad_w, temp
+
+        grad_w2d = torch.einsum("i r ..., i j ... -> r j", t2, grad_temp)
+        grad_t2 = torch.einsum("i j ..., j r -> i r ...", grad_temp, w2d.T)
+        del grad_temp
+        return grad_t1, grad_w1u, grad_w1d, grad_t2, grad_w2u, grad_w2d, None
+
+
+class LohaDiff(WeightAdapterTrainBase):
+    def __init__(self, weights):
+        super().__init__()
+        # Unpack weights tuple from LoHaAdapter
+        w1a, w1b, alpha, w2a, w2b, t1, t2, _ = weights
+
+        # Create trainable parameters
+        self.hada_w1_a = torch.nn.Parameter(w1a)
+        self.hada_w1_b = torch.nn.Parameter(w1b)
+        self.hada_w2_a = torch.nn.Parameter(w2a)
+        self.hada_w2_b = torch.nn.Parameter(w2b)
+
+        self.use_tucker = False
+        if t1 is not None and t2 is not None:
+            self.use_tucker = True
+            self.hada_t1 = torch.nn.Parameter(t1)
+            self.hada_t2 = torch.nn.Parameter(t2)
+        else:
+            # Keep the attributes for consistent access
+            self.hada_t1 = None
+            self.hada_t2 = None
+
+        # Store rank and non-trainable alpha
+        self.rank = w1b.shape[0]
+        self.alpha = torch.nn.Parameter(torch.tensor(alpha), requires_grad=False)
+
+    def __call__(self, w):
+        org_dtype = w.dtype
+
+        scale = self.alpha / self.rank
+        if self.use_tucker:
+            diff_weight = HadaWeightTucker.apply(self.hada_t1, self.hada_w1_a, self.hada_w1_b, self.hada_t2, self.hada_w2_a, self.hada_w2_b, scale)
+        else:
+            diff_weight = HadaWeight.apply(self.hada_w1_a, self.hada_w1_b, self.hada_w2_a, self.hada_w2_b, scale)
+
+        # Add the scaled difference to the original weight
+        weight = w.to(diff_weight) + diff_weight.reshape(w.shape)
+
+        return weight.to(org_dtype)
+
+    def passive_memory_usage(self):
+        """Calculates memory usage of the trainable parameters."""
+        return sum(param.numel() * param.element_size() for param in self.parameters())
 
 
 class LoHaAdapter(WeightAdapterBase):
@@ -13,6 +126,25 @@ class LoHaAdapter(WeightAdapterBase):
         self.loaded_keys = loaded_keys
         self.weights = weights
 
+    @classmethod
+    def create_train(cls, weight, rank=1, alpha=1.0):
+        out_dim = weight.shape[0]
+        in_dim = weight.shape[1:].numel()
+        mat1 = torch.empty(out_dim, rank, device=weight.device, dtype=weight.dtype)
+        mat2 = torch.empty(rank, in_dim, device=weight.device, dtype=weight.dtype)
+        torch.nn.init.normal_(mat1, 0.1)
+        torch.nn.init.constant_(mat2, 0.0)
+        mat3 = torch.empty(out_dim, rank, device=weight.device, dtype=weight.dtype)
+        mat4 = torch.empty(rank, in_dim, device=weight.device, dtype=weight.dtype)
+        torch.nn.init.normal_(mat3, 0.1)
+        torch.nn.init.normal_(mat4, 0.01)
+        return LohaDiff(
+            (mat1, mat2, alpha, mat3, mat4, None, None, None)
+        )
+
+    def to_train(self):
+        return LohaDiff(self.weights)
+
     @classmethod
     def load(
         cls,

comfy/weight_adapter/lokr.py

@@ -3,7 +3,77 @@ from typing import Optional
 
 import torch
 import comfy.model_management
-from .base import WeightAdapterBase, weight_decompose
+from .base import (
+    WeightAdapterBase,
+    WeightAdapterTrainBase,
+    weight_decompose,
+    factorization,
+)
+
+
+class LokrDiff(WeightAdapterTrainBase):
+    def __init__(self, weights):
+        super().__init__()
+        (lokr_w1, lokr_w2, alpha, lokr_w1_a, lokr_w1_b, lokr_w2_a, lokr_w2_b, lokr_t2, dora_scale) = weights
+        self.use_tucker = False
+        if lokr_w1_a is not None:
+            _, rank_a = lokr_w1_a.shape[0], lokr_w1_a.shape[1]
+            rank_a, _ = lokr_w1_b.shape[0], lokr_w1_b.shape[1]
+            self.lokr_w1_a = torch.nn.Parameter(lokr_w1_a)
+            self.lokr_w1_b = torch.nn.Parameter(lokr_w1_b)
+            self.w1_rebuild = True
+            self.ranka = rank_a
+
+        if lokr_w2_a is not None:
+            _, rank_b = lokr_w2_a.shape[0], lokr_w2_a.shape[1]
+            rank_b, _ = lokr_w2_b.shape[0], lokr_w2_b.shape[1]
+            self.lokr_w2_a = torch.nn.Parameter(lokr_w2_a)
+            self.lokr_w2_b = torch.nn.Parameter(lokr_w2_b)
+            if lokr_t2 is not None:
+                self.use_tucker = True
+                self.lokr_t2 = torch.nn.Parameter(lokr_t2)
+            self.w2_rebuild = True
+            self.rankb = rank_b
+
+        if lokr_w1 is not None:
+            self.lokr_w1 = torch.nn.Parameter(lokr_w1)
+            self.w1_rebuild = False
+
+        if lokr_w2 is not None:
+            self.lokr_w2 = torch.nn.Parameter(lokr_w2)
+            self.w2_rebuild = False
+
+        self.alpha = torch.nn.Parameter(torch.tensor(alpha), requires_grad=False)
+
+    @property
+    def w1(self):
+        if self.w1_rebuild:
+            return (self.lokr_w1_a @ self.lokr_w1_b) * (self.alpha / self.ranka)
+        else:
+            return self.lokr_w1
+
+    @property
+    def w2(self):
+        if self.w2_rebuild:
+            if self.use_tucker:
+                w2 = torch.einsum(
+                    'i j k l, j r, i p -> p r k l',
+                    self.lokr_t2,
+                    self.lokr_w2_b,
+                    self.lokr_w2_a
+                )
+            else:
+                w2 = self.lokr_w2_a @ self.lokr_w2_b
+            return w2 * (self.alpha / self.rankb)
+        else:
+            return self.lokr_w2
+
+    def __call__(self, w):
+        diff = torch.kron(self.w1, self.w2)
+        return w + diff.reshape(w.shape).to(w)
+
+    def passive_memory_usage(self):
+        return sum(param.numel() * param.element_size() for param in self.parameters())
 
 
 class LoKrAdapter(WeightAdapterBase):
@@ -13,6 +83,20 @@ class LoKrAdapter(WeightAdapterBase):
         self.loaded_keys = loaded_keys
         self.weights = weights
 
+    @classmethod
+    def create_train(cls, weight, rank=1, alpha=1.0):
+        out_dim = weight.shape[0]
+        in_dim = weight.shape[1:].numel()
+        out1, out2 = factorization(out_dim, rank)
+        in1, in2 = factorization(in_dim, rank)
+        mat1 = torch.empty(out1, in1, device=weight.device, dtype=weight.dtype)
+        mat2 = torch.empty(out2, in2, device=weight.device, dtype=weight.dtype)
+        torch.nn.init.kaiming_uniform_(mat2, a=5**0.5)
+        torch.nn.init.constant_(mat1, 0.0)
+        return LokrDiff(
+            (mat1, mat2, alpha, None, None, None, None, None, None)
+        )
+
     @classmethod
     def load(
         cls,

comfy/weight_adapter/oft.py

@@ -3,7 +3,58 @@ from typing import Optional
 
 import torch
 import comfy.model_management
-from .base import WeightAdapterBase, weight_decompose
+from .base import WeightAdapterBase, WeightAdapterTrainBase, weight_decompose, factorization
+
+
+class OFTDiff(WeightAdapterTrainBase):
+    def __init__(self, weights):
+        super().__init__()
+        # Unpack weights tuple from LoHaAdapter
+        blocks, rescale, alpha, _ = weights
+
+        # Create trainable parameters
+        self.oft_blocks = torch.nn.Parameter(blocks)
+        if rescale is not None:
+            self.rescale = torch.nn.Parameter(rescale)
+            self.rescaled = True
+        else:
+            self.rescaled = False
+        self.block_num, self.block_size, _ = blocks.shape
+        self.constraint = float(alpha)
+        self.alpha = torch.nn.Parameter(torch.tensor(alpha), requires_grad=False)
+
+    def __call__(self, w):
+        org_dtype = w.dtype
+        I = torch.eye(self.block_size, device=self.oft_blocks.device)
+
+        ## generate r
+        # for Q = -Q^T
+        q = self.oft_blocks - self.oft_blocks.transpose(1, 2)
+        normed_q = q
+        if self.constraint:
+            q_norm = torch.norm(q) + 1e-8
+            if q_norm > self.constraint:
+                normed_q = q * self.constraint / q_norm
+        # use float() to prevent unsupported type
+        r = (I + normed_q) @ (I - normed_q).float().inverse()
+
+        ## Apply chunked matmul on weight
+        _, *shape = w.shape
+        org_weight = w.to(dtype=r.dtype)
+        org_weight = org_weight.unflatten(0, (self.block_num, self.block_size))
+        # Init R=0, so add I on it to ensure the output of step0 is original model output
+        weight = torch.einsum(
+            "k n m, k n ... -> k m ...",
+            r,
+            org_weight,
+        ).flatten(0, 1)
+        if self.rescaled:
+            weight = self.rescale * weight
+        return weight.to(org_dtype)
+
+    def passive_memory_usage(self):
+        """Calculates memory usage of the trainable parameters."""
+        return sum(param.numel() * param.element_size() for param in self.parameters())
 
 
 class OFTAdapter(WeightAdapterBase):
@@ -13,6 +64,18 @@ class OFTAdapter(WeightAdapterBase):
         self.loaded_keys = loaded_keys
         self.weights = weights
 
+    @classmethod
+    def create_train(cls, weight, rank=1, alpha=1.0):
+        out_dim = weight.shape[0]
+        block_size, block_num = factorization(out_dim, rank)
+        block = torch.zeros(block_num, block_size, block_size, device=weight.device, dtype=weight.dtype)
+        return OFTDiff(
+            (block, None, alpha, None)
+        )
+
+    def to_train(self):
+        return OFTDiff(self.weights)
+
     @classmethod
     def load(
         cls,
@@ -60,6 +123,8 @@ class OFTAdapter(WeightAdapterBase):
         blocks = v[0]
         rescale = v[1]
         alpha = v[2]
+        if alpha is None:
+            alpha = 0
         dora_scale = v[3]
 
         blocks = comfy.model_management.cast_to_device(blocks, weight.device, intermediate_dtype)