mirrors/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2025-09-13 13:05:07 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'master' into attention-select

Browse Source
This commit is contained in:
Jedrzej Kosinski
2025-08-29 23:35:38 -07:00
parent cb959f9669 885015eecf
commit d9bb4530b0
15 changed files with 253 additions and 105 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
comfy/k_diffusion/sampling.py
View File

@@ -853,6 +853,11 @@ def sample_dpmpp_2m_sde(model, x, sigmas, extra_args=None, callback=None, disabl
return x
@torch.no_grad()
def sample_dpmpp_2m_sde_heun(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type='heun'):
return sample_dpmpp_2m_sde(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler, solver_type=solver_type)
@torch.no_grad()
def sample_dpmpp_3m_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
"""DPM-Solver++(3M) SDE."""
@@ -925,6 +930,16 @@ def sample_dpmpp_3m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, di
return sample_dpmpp_3m_sde(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler)
@torch.no_grad()
def sample_dpmpp_2m_sde_heun_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type='heun'):
if len(sigmas) <= 1:
return x
extra_args = {} if extra_args is None else extra_args
sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=extra_args.get("seed", None), cpu=False) if noise_sampler is None else noise_sampler
return sample_dpmpp_2m_sde_heun(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler, solver_type=solver_type)
@torch.no_grad()
def sample_dpmpp_2m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type='midpoint'):
if len(sigmas) <= 1:

4
comfy/ldm/flux/model.py
View File

@@ -161,7 +161,7 @@ class Flux(nn.Module):
if i < len(control_i):
add = control_i[i]
if add is not None:
img += add
img[:, :add.shape[1]] += add
if img.dtype == torch.float16:
img = torch.nan_to_num(img, nan=0.0, posinf=65504, neginf=-65504)
@@ -194,7 +194,7 @@ class Flux(nn.Module):
if i < len(control_o):
add = control_o[i]
if add is not None:
img[:, txt.shape[1] :, ...] += add
img[:, txt.shape[1] : txt.shape[1] + add.shape[1], ...] += add
img = img[:, txt.shape[1] :, ...]

2
comfy/ldm/qwen_image/model.py
View File

@@ -463,7 +463,7 @@ class QwenImageTransformer2DModel(nn.Module):
if i < len(control_i):
add = control_i[i]
if add is not None:
hidden_states += add
hidden_states[:, :add.shape[1]] += add
hidden_states = self.norm_out(hidden_states, temb)
hidden_states = self.proj_out(hidden_states)

6
comfy/ldm/wan/model.py
View File

@@ -1257,6 +1257,7 @@ class WanModel_S2V(WanModel):
audio_emb = None
# embeddings
bs, _, time, height, width = x.shape
x = self.patch_embedding(x.float()).to(x.dtype)
if control_video is not None:
x = x + self.cond_encoder(control_video)
@@ -1274,11 +1275,12 @@ class WanModel_S2V(WanModel):
if reference_latent is not None:
ref = self.patch_embedding(reference_latent.float()).to(x.dtype)
ref = ref.flatten(2).transpose(1, 2)
freqs_ref = self.rope_encode(reference_latent.shape[-3], reference_latent.shape[-2], reference_latent.shape[-1], t_start=30, device=x.device, dtype=x.dtype)
freqs_ref = self.rope_encode(reference_latent.shape[-3], reference_latent.shape[-2], reference_latent.shape[-1], t_start=max(30, time + 9), device=x.device, dtype=x.dtype)
ref = ref + cond_mask_weight[1]
x = torch.cat([x, ref], dim=1)
freqs = torch.cat([freqs, freqs_ref], dim=1)
t = torch.cat([t, torch.zeros((t.shape[0], reference_latent.shape[-3]), device=t.device, dtype=t.dtype)], dim=1)
del ref, freqs_ref
if reference_motion is not None:
motion_encoded, freqs_motion = self.frame_packer(reference_motion, self)
@@ -1288,6 +1290,7 @@ class WanModel_S2V(WanModel):
t = torch.repeat_interleave(t, 2, dim=1)
t = torch.cat([t, torch.zeros((t.shape[0], 3), device=t.device, dtype=t.dtype)], dim=1)
del motion_encoded, freqs_motion
# time embeddings
e = self.time_embedding(
@@ -1298,7 +1301,6 @@ class WanModel_S2V(WanModel):
# context
context = self.text_embedding(context)
patches_replace = transformer_options.get("patches_replace", {})
blocks_replace = patches_replace.get("dit", {})
for i, block in enumerate(self.blocks):

40
comfy/model_base.py
View File

@@ -150,6 +150,7 @@ class BaseModel(torch.nn.Module):
logging.debug("adm {}".format(self.adm_channels))
self.memory_usage_factor = model_config.memory_usage_factor
self.memory_usage_factor_conds = ()
self.memory_usage_shape_process = {}
def apply_model(self, x, t, c_concat=None, c_crossattn=None, control=None, transformer_options={}, **kwargs):
return comfy.patcher_extension.WrapperExecutor.new_class_executor(
@@ -350,8 +351,15 @@ class BaseModel(torch.nn.Module):
input_shapes = [input_shape]
for c in self.memory_usage_factor_conds:
shape = cond_shapes.get(c, None)
if shape is not None and len(shape) > 0:
input_shapes += shape
if shape is not None:
if c in self.memory_usage_shape_process:
out = []
for s in shape:
out.append(self.memory_usage_shape_process[c](s))
shape = out
if len(shape) > 0:
input_shapes += shape
if comfy.model_management.xformers_enabled() or comfy.model_management.pytorch_attention_flash_attention():
dtype = self.get_dtype()
@@ -1102,9 +1110,10 @@ class WAN21(BaseModel):
shape_image[1] = extra_channels
image = torch.zeros(shape_image, dtype=noise.dtype, layout=noise.layout, device=noise.device)
else:
latent_dim = self.latent_format.latent_channels
image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
for i in range(0, image.shape[1], 16):
image[:, i: i + 16] = self.process_latent_in(image[:, i: i + 16])
for i in range(0, image.shape[1], latent_dim):
image[:, i: i + latent_dim] = self.process_latent_in(image[:, i: i + latent_dim])
image = utils.resize_to_batch_size(image, noise.shape[0])
if extra_channels != image.shape[1] + 4:
@@ -1204,6 +1213,8 @@ class WAN21_Camera(WAN21):
class WAN22_S2V(WAN21):
def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.WanModel_S2V)
self.memory_usage_factor_conds = ("reference_latent", "reference_motion")
self.memory_usage_shape_process = {"reference_motion": lambda shape: [shape[0], shape[1], 1.5, shape[-2], shape[-1]]}
def extra_conds(self, **kwargs):
out = super().extra_conds(**kwargs)
@@ -1224,18 +1235,25 @@ class WAN22_S2V(WAN21):
out['control_video'] = comfy.conds.CONDRegular(self.process_latent_in(control_video))
return out
class WAN22(BaseModel):
def extra_conds_shapes(self, **kwargs):
out = {}
ref_latents = kwargs.get("reference_latents", None)
if ref_latents is not None:
out['reference_latent'] = list([1, 16, sum(map(lambda a: math.prod(a.size()), ref_latents)) // 16])
reference_motion = kwargs.get("reference_motion", None)
if reference_motion is not None:
out['reference_motion'] = reference_motion.shape
return out
class WAN22(WAN21):
def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.WanModel)
super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.WanModel)
self.image_to_video = image_to_video
def extra_conds(self, **kwargs):
out = super().extra_conds(**kwargs)
cross_attn = kwargs.get("cross_attn", None)
if cross_attn is not None:
out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
denoise_mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None))
denoise_mask = kwargs.get("denoise_mask", None)
if denoise_mask is not None:
out["denoise_mask"] = comfy.conds.CONDRegular(denoise_mask)
return out

2
comfy/samplers.py Normal file → Executable file
View File

@@ -729,7 +729,7 @@ class Sampler:
KSAMPLER_NAMES = ["euler", "euler_cfg_pp", "euler_ancestral", "euler_ancestral_cfg_pp", "heun", "heunpp2","dpm_2", "dpm_2_ancestral",
"lms", "dpm_fast", "dpm_adaptive", "dpmpp_2s_ancestral", "dpmpp_2s_ancestral_cfg_pp", "dpmpp_sde", "dpmpp_sde_gpu",
"dpmpp_2m", "dpmpp_2m_cfg_pp", "dpmpp_2m_sde", "dpmpp_2m_sde_gpu", "dpmpp_3m_sde", "dpmpp_3m_sde_gpu", "ddpm", "lcm",
"dpmpp_2m", "dpmpp_2m_cfg_pp", "dpmpp_2m_sde", "dpmpp_2m_sde_gpu", "dpmpp_2m_sde_heun", "dpmpp_2m_sde_heun_gpu", "dpmpp_3m_sde", "dpmpp_3m_sde_gpu", "ddpm", "lcm",
"ipndm", "ipndm_v", "deis", "res_multistep", "res_multistep_cfg_pp", "res_multistep_ancestral", "res_multistep_ancestral_cfg_pp",
"gradient_estimation", "gradient_estimation_cfg_pp", "er_sde", "seeds_2", "seeds_3", "sa_solver", "sa_solver_pece"]

2
comfy/supported_models.py
View File

@@ -700,7 +700,7 @@ class Flux(supported_models_base.BASE):
unet_extra_config = {}
latent_format = latent_formats.Flux
memory_usage_factor = 2.8
memory_usage_factor = 3.1 # TODO: debug why flux mem usage is so weird on windows.
supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32]