Foreign-Mirrors/Fooocus
1
0
Fork 0
mirror of https://github.com/lllyasviel/Fooocus.git synced 2026-08-16 13:13:16 +02:00
Code Issues Packages Projects Releases Wiki Activity

feat: advanced params refactoring + prevent users from skipping/stopping other users tasks in queue (#981)

Browse Source 
* only make stop_button and skip_button interactive when rendering process starts

fix inconsistency in behaviour of stop_button and skip_button as it was possible to skip or stop other users processes while still being in queue

* use AsyncTask for last_stop handling instead of shared

* Revert "only make stop_button and skip_button interactive when rendering process starts"

This reverts commit d3f9156854.

* introduce state for task skipping/stopping

* fix return parameters of stop_clicked

* code cleanup, do not disable skip/stop on stop_clicked

* reset last_stop when skipping for further processing

* fix: replace fcbh with ldm_patched

* fix: use currentTask instead of ctrls after merging upstream

* feat: extract attribute disable_preview

* feat: extract attribute adm_scaler_positive

* feat: extract attribute adm_scaler_negative

* feat: extract attribute adm_scaler_end

* feat: extract attribute adaptive_cfg

* feat: extract attribute sampler_name

* feat: extract attribute scheduler_name

* feat: extract attribute generate_image_grid

* feat: extract attribute overwrite_step

* feat: extract attribute overwrite_switch

* feat: extract attribute overwrite_width

* feat: extract attribute overwrite_height

* feat: extract attribute overwrite_vary_strength

* feat: extract attribute overwrite_upscale_strength

* feat: extract attribute mixing_image_prompt_and_vary_upscale

* feat: extract attribute mixing_image_prompt_and_inpaint

* feat: extract attribute debugging_cn_preprocessor

* feat: extract attribute skipping_cn_preprocessor

* feat: extract attribute canny_low_threshold

* feat: extract attribute canny_high_threshold

* feat: extract attribute refiner_swap_method

* feat: extract freeu_ctrls attributes

freeu_enabled, freeu_b1, freeu_b2, freeu_s1, freeu_s2

* feat: extract inpaint_ctrls attributes

debugging_inpaint_preprocessor, inpaint_disable_initial_latent, inpaint_engine, inpaint_strength, inpaint_respective_field, inpaint_mask_upload_checkbox, invert_mask_checkbox, inpaint_erode_or_dilate

* wip: add TODOs

* chore: cleanup code

* feat: extract attribute controlnet_softness

* feat: extract remaining attributes, do not use globals in patch

* fix: resolve circular import, patch_all now in async_worker

* chore: cleanup pid code
This commit is contained in:
Manuel Schmid
2024-02-24 19:01:06 +01:00
committed by Manuel Schmid
parent 0ed01da4e4
commit 5b7ddf8b22
8 changed files with 218 additions and 196 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/patch.py
+38 -32
View File
@@ -17,7 +17,6 @@ import ldm_patched.controlnet.cldm
import ldm_patched.modules.model_patcher
import ldm_patched.modules.samplers
import ldm_patched.modules.args_parser
import modules.advanced_parameters as advanced_parameters
import warnings
import safetensors.torch
import modules.constants as constants
@@ -29,15 +28,25 @@ from modules.patch_precision import patch_all_precision
from modules.patch_clip import patch_all_clip
sharpness = 2.0
class PatchSettings:
def __init__(self,
sharpness=2.0,
adm_scaler_end=0.3,
positive_adm_scale=1.5,
negative_adm_scale=0.8,
controlnet_softness=0.25,
adaptive_cfg=7.0):
self.sharpness = sharpness
self.adm_scaler_end = adm_scaler_end
self.positive_adm_scale = positive_adm_scale
self.negative_adm_scale = negative_adm_scale
self.controlnet_softness = controlnet_softness
self.adaptive_cfg = adaptive_cfg
self.global_diffusion_progress = 0
self.eps_record = None
adm_scaler_end = 0.3
positive_adm_scale = 1.5
negative_adm_scale = 0.8
adaptive_cfg = 7.0
global_diffusion_progress = 0
eps_record = None
patch_settings = {}
def calculate_weight_patched(self, patches, weight, key):
@@ -201,14 +210,13 @@ class BrownianTreeNoiseSamplerPatched:
def compute_cfg(uncond, cond, cfg_scale, t):
global adaptive_cfg
mimic_cfg = float(adaptive_cfg)
pid = os.getpid()
mimic_cfg = float(patch_settings[pid].adaptive_cfg)
real_cfg = float(cfg_scale)
real_eps = uncond + real_cfg * (cond - uncond)
if cfg_scale > adaptive_cfg:
if cfg_scale > patch_settings[pid].adaptive_cfg:
mimicked_eps = uncond + mimic_cfg * (cond - uncond)
return real_eps * t + mimicked_eps * (1 - t)
else:
@@ -216,13 +224,13 @@ def compute_cfg(uncond, cond, cfg_scale, t):
def patched_sampling_function(model, x, timestep, uncond, cond, cond_scale, model_options=None, seed=None):
global eps_record
pid = os.getpid()
if math.isclose(cond_scale, 1.0) and not model_options.get("disable_cfg1_optimization", False):
final_x0 = calc_cond_uncond_batch(model, cond, None, x, timestep, model_options)[0]
if eps_record is not None:
eps_record = ((x - final_x0) / timestep).cpu()
if patch_settings[pid].eps_record is not None:
patch_settings[pid].eps_record = ((x - final_x0) / timestep).cpu()
return final_x0
@@ -231,16 +239,16 @@ def patched_sampling_function(model, x, timestep, uncond, cond, cond_scale, mode
positive_eps = x - positive_x0
negative_eps = x - negative_x0
alpha = 0.001 * sharpness * global_diffusion_progress
alpha = 0.001 * patch_settings[pid].sharpness * patch_settings[pid].global_diffusion_progress
positive_eps_degraded = anisotropic.adaptive_anisotropic_filter(x=positive_eps, g=positive_x0)
positive_eps_degraded_weighted = positive_eps_degraded * alpha + positive_eps * (1.0 - alpha)
final_eps = compute_cfg(uncond=negative_eps, cond=positive_eps_degraded_weighted,
cfg_scale=cond_scale, t=global_diffusion_progress)
cfg_scale=cond_scale, t=patch_settings[pid].global_diffusion_progress)
if eps_record is not None:
eps_record = (final_eps / timestep).cpu()
if patch_settings[pid].eps_record is not None:
patch_settings[pid].eps_record = (final_eps / timestep).cpu()
return x - final_eps
@@ -255,20 +263,19 @@ def round_to_64(x):
def sdxl_encode_adm_patched(self, **kwargs):
global positive_adm_scale, negative_adm_scale
clip_pooled = ldm_patched.modules.model_base.sdxl_pooled(kwargs, self.noise_augmentor)
width = kwargs.get("width", 1024)
height = kwargs.get("height", 1024)
target_width = width
target_height = height
pid = os.getpid()
if kwargs.get("prompt_type", "") == "negative":
width = float(width) * negative_adm_scale
height = float(height) * negative_adm_scale
width = float(width) * patch_settings[pid].negative_adm_scale
height = float(height) * patch_settings[pid].negative_adm_scale
elif kwargs.get("prompt_type", "") == "positive":
width = float(width) * positive_adm_scale
height = float(height) * positive_adm_scale
width = float(width) * patch_settings[pid].positive_adm_scale
height = float(height) * patch_settings[pid].positive_adm_scale
def embedder(number_list):
h = self.embedder(torch.tensor(number_list, dtype=torch.float32))
@@ -322,7 +329,7 @@ def patched_KSamplerX0Inpaint_forward(self, x, sigma, uncond, cond, cond_scale,
def timed_adm(y, timesteps):
if isinstance(y, torch.Tensor) and int(y.dim()) == 2 and int(y.shape[1]) == 5632:
y_mask = (timesteps > 999.0 * (1.0 - float(adm_scaler_end))).to(y)[..., None]
y_mask = (timesteps > 999.0 * (1.0 - float(patch_settings[os.getpid()].adm_scaler_end))).to(y)[..., None]
y_with_adm = y[..., :2816].clone()
y_without_adm = y[..., 2816:].clone()
return y_with_adm * y_mask + y_without_adm * (1.0 - y_mask)
@@ -332,6 +339,7 @@ def timed_adm(y, timesteps):
def patched_cldm_forward(self, x, hint, timesteps, context, y=None, **kwargs):
t_emb = ldm_patched.ldm.modules.diffusionmodules.openaimodel.timestep_embedding(timesteps, self.model_channels, repeat_only=False).to(x.dtype)
emb = self.time_embed(t_emb)
pid = os.getpid()
guided_hint = self.input_hint_block(hint, emb, context)
@@ -357,19 +365,17 @@ def patched_cldm_forward(self, x, hint, timesteps, context, y=None, **kwargs):
h = self.middle_block(h, emb, context)
outs.append(self.middle_block_out(h, emb, context))
if advanced_parameters.controlnet_softness > 0:
if patch_settings[pid].controlnet_softness > 0:
for i in range(10):
k = 1.0 - float(i) / 9.0
outs[i] = outs[i] * (1.0 - advanced_parameters.controlnet_softness * k)
outs[i] = outs[i] * (1.0 - patch_settings[pid].controlnet_softness * k)
return outs
def patched_unet_forward(self, x, timesteps=None, context=None, y=None, control=None, transformer_options={}, **kwargs):
global global_diffusion_progress
self.current_step = 1.0 - timesteps.to(x) / 999.0
global_diffusion_progress = float(self.current_step.detach().cpu().numpy().tolist()[0])
patch_settings[os.getpid()].global_diffusion_progress = float(self.current_step.detach().cpu().numpy().tolist()[0])
y = timed_adm(y, timesteps)
@@ -483,7 +489,7 @@ def patch_all():
if ldm_patched.modules.model_management.directml_enabled:
ldm_patched.modules.model_management.lowvram_available = True
ldm_patched.modules.model_management.OOM_EXCEPTION = Exception
patch_all_precision()
patch_all_clip()