mirror of
https://github.com/comfyanonymous/ComfyUI.git
synced 2025-09-11 03:58:22 +00:00
Disable autocast in unet for increased speed.
This commit is contained in:
comfyanonymous
@@ -2,7 +2,6 @@ from .k_diffusion import sampling as k_diffusion_sampling
|
||||
from .k_diffusion import external as k_diffusion_external
|
||||
from .extra_samplers import uni_pc
|
||||
import torch
|
||||
import contextlib
|
||||
from comfy import model_management
|
||||
from .ldm.models.diffusion.ddim import DDIMSampler
|
||||
from .ldm.modules.diffusionmodules.util import make_ddim_timesteps
|
||||
@@ -577,11 +576,6 @@ class KSampler:
|
||||
apply_empty_x_to_equal_area(positive, negative, 'control', lambda cond_cnets, x: cond_cnets[x])
|
||||
apply_empty_x_to_equal_area(positive, negative, 'gligen', lambda cond_cnets, x: cond_cnets[x])
|
||||
|
||||
if self.model.get_dtype() == torch.float16:
|
||||
precision_scope = torch.autocast
|
||||
else:
|
||||
precision_scope = contextlib.nullcontext
|
||||
|
||||
if self.model.is_adm():
|
||||
positive = encode_adm(self.model, positive, noise.shape[0], noise.shape[3], noise.shape[2], self.device, "positive")
|
||||
negative = encode_adm(self.model, negative, noise.shape[0], noise.shape[3], noise.shape[2], self.device, "negative")
|
||||
@@ -612,67 +606,67 @@ class KSampler:
|
||||
else:
|
||||
max_denoise = True
|
||||
|
||||
with precision_scope(model_management.get_autocast_device(self.device)):
|
||||
if self.sampler == "uni_pc":
|
||||
samples = uni_pc.sample_unipc(self.model_wrap, noise, latent_image, sigmas, sampling_function=sampling_function, max_denoise=max_denoise, extra_args=extra_args, noise_mask=denoise_mask, callback=callback, disable=disable_pbar)
|
||||
elif self.sampler == "uni_pc_bh2":
|
||||
samples = uni_pc.sample_unipc(self.model_wrap, noise, latent_image, sigmas, sampling_function=sampling_function, max_denoise=max_denoise, extra_args=extra_args, noise_mask=denoise_mask, callback=callback, variant='bh2', disable=disable_pbar)
|
||||
elif self.sampler == "ddim":
|
||||
timesteps = []
|
||||
for s in range(sigmas.shape[0]):
|
||||
timesteps.insert(0, self.model_wrap.sigma_to_t(sigmas[s]))
|
||||
noise_mask = None
|
||||
if denoise_mask is not None:
|
||||
noise_mask = 1.0 - denoise_mask
|
||||
|
||||
ddim_callback = None
|
||||
if callback is not None:
|
||||
total_steps = len(timesteps) - 1
|
||||
ddim_callback = lambda pred_x0, i: callback(i, pred_x0, None, total_steps)
|
||||
if self.sampler == "uni_pc":
|
||||
samples = uni_pc.sample_unipc(self.model_wrap, noise, latent_image, sigmas, sampling_function=sampling_function, max_denoise=max_denoise, extra_args=extra_args, noise_mask=denoise_mask, callback=callback, disable=disable_pbar)
|
||||
elif self.sampler == "uni_pc_bh2":
|
||||
samples = uni_pc.sample_unipc(self.model_wrap, noise, latent_image, sigmas, sampling_function=sampling_function, max_denoise=max_denoise, extra_args=extra_args, noise_mask=denoise_mask, callback=callback, variant='bh2', disable=disable_pbar)
|
||||
elif self.sampler == "ddim":
|
||||
timesteps = []
|
||||
for s in range(sigmas.shape[0]):
|
||||
timesteps.insert(0, self.model_wrap.sigma_to_t(sigmas[s]))
|
||||
noise_mask = None
|
||||
if denoise_mask is not None:
|
||||
noise_mask = 1.0 - denoise_mask
|
||||
|
||||
sampler = DDIMSampler(self.model, device=self.device)
|
||||
sampler.make_schedule_timesteps(ddim_timesteps=timesteps, verbose=False)
|
||||
z_enc = sampler.stochastic_encode(latent_image, torch.tensor([len(timesteps) - 1] * noise.shape[0]).to(self.device), noise=noise, max_denoise=max_denoise)
|
||||
samples, _ = sampler.sample_custom(ddim_timesteps=timesteps,
|
||||
conditioning=positive,
|
||||
batch_size=noise.shape[0],
|
||||
shape=noise.shape[1:],
|
||||
verbose=False,
|
||||
unconditional_guidance_scale=cfg,
|
||||
unconditional_conditioning=negative,
|
||||
eta=0.0,
|
||||
x_T=z_enc,
|
||||
x0=latent_image,
|
||||
img_callback=ddim_callback,
|
||||
denoise_function=sampling_function,
|
||||
extra_args=extra_args,
|
||||
mask=noise_mask,
|
||||
to_zero=sigmas[-1]==0,
|
||||
end_step=sigmas.shape[0] - 1,
|
||||
disable_pbar=disable_pbar)
|
||||
ddim_callback = None
|
||||
if callback is not None:
|
||||
total_steps = len(timesteps) - 1
|
||||
ddim_callback = lambda pred_x0, i: callback(i, pred_x0, None, total_steps)
|
||||
|
||||
sampler = DDIMSampler(self.model, device=self.device)
|
||||
sampler.make_schedule_timesteps(ddim_timesteps=timesteps, verbose=False)
|
||||
z_enc = sampler.stochastic_encode(latent_image, torch.tensor([len(timesteps) - 1] * noise.shape[0]).to(self.device), noise=noise, max_denoise=max_denoise)
|
||||
samples, _ = sampler.sample_custom(ddim_timesteps=timesteps,
|
||||
conditioning=positive,
|
||||
batch_size=noise.shape[0],
|
||||
shape=noise.shape[1:],
|
||||
verbose=False,
|
||||
unconditional_guidance_scale=cfg,
|
||||
unconditional_conditioning=negative,
|
||||
eta=0.0,
|
||||
x_T=z_enc,
|
||||
x0=latent_image,
|
||||
img_callback=ddim_callback,
|
||||
denoise_function=sampling_function,
|
||||
extra_args=extra_args,
|
||||
mask=noise_mask,
|
||||
to_zero=sigmas[-1]==0,
|
||||
end_step=sigmas.shape[0] - 1,
|
||||
disable_pbar=disable_pbar)
|
||||
|
||||
else:
|
||||
extra_args["denoise_mask"] = denoise_mask
|
||||
self.model_k.latent_image = latent_image
|
||||
self.model_k.noise = noise
|
||||
|
||||
if max_denoise:
|
||||
noise = noise * torch.sqrt(1.0 + sigmas[0] ** 2.0)
|
||||
else:
|
||||
extra_args["denoise_mask"] = denoise_mask
|
||||
self.model_k.latent_image = latent_image
|
||||
self.model_k.noise = noise
|
||||
noise = noise * sigmas[0]
|
||||
|
||||
if max_denoise:
|
||||
noise = noise * torch.sqrt(1.0 + sigmas[0] ** 2.0)
|
||||
else:
|
||||
noise = noise * sigmas[0]
|
||||
k_callback = None
|
||||
total_steps = len(sigmas) - 1
|
||||
if callback is not None:
|
||||
k_callback = lambda x: callback(x["i"], x["denoised"], x["x"], total_steps)
|
||||
|
||||
k_callback = None
|
||||
total_steps = len(sigmas) - 1
|
||||
if callback is not None:
|
||||
k_callback = lambda x: callback(x["i"], x["denoised"], x["x"], total_steps)
|
||||
|
||||
if latent_image is not None:
|
||||
noise += latent_image
|
||||
if self.sampler == "dpm_fast":
|
||||
samples = k_diffusion_sampling.sample_dpm_fast(self.model_k, noise, sigma_min, sigmas[0], total_steps, extra_args=extra_args, callback=k_callback, disable=disable_pbar)
|
||||
elif self.sampler == "dpm_adaptive":
|
||||
samples = k_diffusion_sampling.sample_dpm_adaptive(self.model_k, noise, sigma_min, sigmas[0], extra_args=extra_args, callback=k_callback, disable=disable_pbar)
|
||||
else:
|
||||
samples = getattr(k_diffusion_sampling, "sample_{}".format(self.sampler))(self.model_k, noise, sigmas, extra_args=extra_args, callback=k_callback, disable=disable_pbar)
|
||||
if latent_image is not None:
|
||||
noise += latent_image
|
||||
if self.sampler == "dpm_fast":
|
||||
samples = k_diffusion_sampling.sample_dpm_fast(self.model_k, noise, sigma_min, sigmas[0], total_steps, extra_args=extra_args, callback=k_callback, disable=disable_pbar)
|
||||
elif self.sampler == "dpm_adaptive":
|
||||
samples = k_diffusion_sampling.sample_dpm_adaptive(self.model_k, noise, sigma_min, sigmas[0], extra_args=extra_args, callback=k_callback, disable=disable_pbar)
|
||||
else:
|
||||
samples = getattr(k_diffusion_sampling, "sample_{}".format(self.sampler))(self.model_k, noise, sigmas, extra_args=extra_args, callback=k_callback, disable=disable_pbar)
|
||||
|
||||
return self.model.process_latent_out(samples.to(torch.float32))
|
||||
|
||||
Reference in New Issue
Block a user