mirror of
https://github.com/invoke-ai/InvokeAI
synced 2024-08-30 20:32:17 +00:00
fix: Slow loading of Loras
Co-Authored-By: StAlKeR7779 <7768370+StAlKeR7779@users.noreply.github.com>
This commit is contained in:
blessedcoolant
committed by
psychedelicious
psychedelicious
parent
0f0336b6ef
commit
c0501ed5c2
@ -4,18 +4,17 @@ from contextlib import ExitStack
|
||||
from typing import List, Literal, Optional, Union
|
||||
|
||||
import einops
|
||||
|
||||
from pydantic import BaseModel, Field, validator
|
||||
import torch
|
||||
from diffusers import ControlNetModel, DPMSolverMultistepScheduler
|
||||
from diffusers.image_processor import VaeImageProcessor
|
||||
from diffusers.schedulers import SchedulerMixin as Scheduler
|
||||
from pydantic import BaseModel, Field, validator
|
||||
|
||||
from invokeai.app.util.misc import SEED_MAX, get_random_seed
|
||||
from invokeai.app.util.step_callback import stable_diffusion_step_callback
|
||||
|
||||
from ..models.image import ImageCategory, ImageField, ResourceOrigin
|
||||
from ...backend.image_util.seamless import configure_model_padding
|
||||
from ...backend.model_management.lora import ModelPatcher
|
||||
from ...backend.stable_diffusion import PipelineIntermediateState
|
||||
from ...backend.stable_diffusion.diffusers_pipeline import (
|
||||
ConditioningData, ControlNetData, StableDiffusionGeneratorPipeline,
|
||||
@ -24,7 +23,7 @@ from ...backend.stable_diffusion.diffusion.shared_invokeai_diffusion import \
|
||||
PostprocessingSettings
|
||||
from ...backend.stable_diffusion.schedulers import SCHEDULER_MAP
|
||||
from ...backend.util.devices import torch_dtype
|
||||
from ...backend.model_management.lora import ModelPatcher
|
||||
from ..models.image import ImageCategory, ImageField, ResourceOrigin
|
||||
from .baseinvocation import (BaseInvocation, BaseInvocationOutput,
|
||||
InvocationConfig, InvocationContext)
|
||||
from .compel import ConditioningField
|
||||
@ -32,14 +31,17 @@ from .controlnet_image_processors import ControlField
|
||||
from .image import ImageOutput
|
||||
from .model import ModelInfo, UNetField, VaeField
|
||||
|
||||
|
||||
class LatentsField(BaseModel):
|
||||
"""A latents field used for passing latents between invocations"""
|
||||
|
||||
latents_name: Optional[str] = Field(default=None, description="The name of the latents")
|
||||
latents_name: Optional[str] = Field(
|
||||
default=None, description="The name of the latents")
|
||||
|
||||
class Config:
|
||||
schema_extra = {"required": ["latents_name"]}
|
||||
|
||||
|
||||
class LatentsOutput(BaseInvocationOutput):
|
||||
"""Base class for invocations that output latents"""
|
||||
#fmt: off
|
||||
@ -53,11 +55,11 @@ class LatentsOutput(BaseInvocationOutput):
|
||||
|
||||
|
||||
def build_latents_output(latents_name: str, latents: torch.Tensor):
|
||||
return LatentsOutput(
|
||||
latents=LatentsField(latents_name=latents_name),
|
||||
width=latents.size()[3] * 8,
|
||||
height=latents.size()[2] * 8,
|
||||
)
|
||||
return LatentsOutput(
|
||||
latents=LatentsField(latents_name=latents_name),
|
||||
width=latents.size()[3] * 8,
|
||||
height=latents.size()[2] * 8,
|
||||
)
|
||||
|
||||
|
||||
SAMPLER_NAME_VALUES = Literal[
|
||||
@ -70,16 +72,19 @@ def get_scheduler(
|
||||
scheduler_info: ModelInfo,
|
||||
scheduler_name: str,
|
||||
) -> Scheduler:
|
||||
scheduler_class, scheduler_extra_config = SCHEDULER_MAP.get(scheduler_name, SCHEDULER_MAP['ddim'])
|
||||
orig_scheduler_info = context.services.model_manager.get_model(**scheduler_info.dict())
|
||||
scheduler_class, scheduler_extra_config = SCHEDULER_MAP.get(
|
||||
scheduler_name, SCHEDULER_MAP['ddim'])
|
||||
orig_scheduler_info = context.services.model_manager.get_model(
|
||||
**scheduler_info.dict())
|
||||
with orig_scheduler_info as orig_scheduler:
|
||||
scheduler_config = orig_scheduler.config
|
||||
|
||||
|
||||
if "_backup" in scheduler_config:
|
||||
scheduler_config = scheduler_config["_backup"]
|
||||
scheduler_config = {**scheduler_config, **scheduler_extra_config, "_backup": scheduler_config}
|
||||
scheduler_config = {**scheduler_config, **
|
||||
scheduler_extra_config, "_backup": scheduler_config}
|
||||
scheduler = scheduler_class.from_config(scheduler_config)
|
||||
|
||||
|
||||
# hack copied over from generate.py
|
||||
if not hasattr(scheduler, 'uses_inpainting_model'):
|
||||
scheduler.uses_inpainting_model = lambda: False
|
||||
@ -124,18 +129,18 @@ class TextToLatentsInvocation(BaseInvocation):
|
||||
"ui": {
|
||||
"tags": ["latents"],
|
||||
"type_hints": {
|
||||
"model": "model",
|
||||
"control": "control",
|
||||
# "cfg_scale": "float",
|
||||
"cfg_scale": "number"
|
||||
"model": "model",
|
||||
"control": "control",
|
||||
# "cfg_scale": "float",
|
||||
"cfg_scale": "number"
|
||||
}
|
||||
},
|
||||
}
|
||||
|
||||
# TODO: pass this an emitter method or something? or a session for dispatching?
|
||||
def dispatch_progress(
|
||||
self, context: InvocationContext, source_node_id: str, intermediate_state: PipelineIntermediateState
|
||||
) -> None:
|
||||
self, context: InvocationContext, source_node_id: str,
|
||||
intermediate_state: PipelineIntermediateState) -> None:
|
||||
stable_diffusion_step_callback(
|
||||
context=context,
|
||||
intermediate_state=intermediate_state,
|
||||
@ -143,9 +148,12 @@ class TextToLatentsInvocation(BaseInvocation):
|
||||
source_node_id=source_node_id,
|
||||
)
|
||||
|
||||
def get_conditioning_data(self, context: InvocationContext, scheduler) -> ConditioningData:
|
||||
c, extra_conditioning_info = context.services.latents.get(self.positive_conditioning.conditioning_name)
|
||||
uc, _ = context.services.latents.get(self.negative_conditioning.conditioning_name)
|
||||
def get_conditioning_data(
|
||||
self, context: InvocationContext, scheduler) -> ConditioningData:
|
||||
c, extra_conditioning_info = context.services.latents.get(
|
||||
self.positive_conditioning.conditioning_name)
|
||||
uc, _ = context.services.latents.get(
|
||||
self.negative_conditioning.conditioning_name)
|
||||
|
||||
conditioning_data = ConditioningData(
|
||||
unconditioned_embeddings=uc,
|
||||
@ -153,10 +161,10 @@ class TextToLatentsInvocation(BaseInvocation):
|
||||
guidance_scale=self.cfg_scale,
|
||||
extra=extra_conditioning_info,
|
||||
postprocessing_settings=PostprocessingSettings(
|
||||
threshold=0.0,#threshold,
|
||||
warmup=0.2,#warmup,
|
||||
h_symmetry_time_pct=None,#h_symmetry_time_pct,
|
||||
v_symmetry_time_pct=None#v_symmetry_time_pct,
|
||||
threshold=0.0, # threshold,
|
||||
warmup=0.2, # warmup,
|
||||
h_symmetry_time_pct=None, # h_symmetry_time_pct,
|
||||
v_symmetry_time_pct=None # v_symmetry_time_pct,
|
||||
),
|
||||
)
|
||||
|
||||
@ -164,31 +172,32 @@ class TextToLatentsInvocation(BaseInvocation):
|
||||
scheduler,
|
||||
|
||||
# for ddim scheduler
|
||||
eta=0.0, #ddim_eta
|
||||
eta=0.0, # ddim_eta
|
||||
|
||||
# for ancestral and sde schedulers
|
||||
generator=torch.Generator(device=uc.device).manual_seed(0),
|
||||
)
|
||||
return conditioning_data
|
||||
|
||||
def create_pipeline(self, unet, scheduler) -> StableDiffusionGeneratorPipeline:
|
||||
def create_pipeline(
|
||||
self, unet, scheduler) -> StableDiffusionGeneratorPipeline:
|
||||
# TODO:
|
||||
#configure_model_padding(
|
||||
# configure_model_padding(
|
||||
# unet,
|
||||
# self.seamless,
|
||||
# self.seamless_axes,
|
||||
#)
|
||||
# )
|
||||
|
||||
class FakeVae:
|
||||
class FakeVaeConfig:
|
||||
def __init__(self):
|
||||
self.block_out_channels = [0]
|
||||
|
||||
|
||||
def __init__(self):
|
||||
self.config = FakeVae.FakeVaeConfig()
|
||||
|
||||
return StableDiffusionGeneratorPipeline(
|
||||
vae=FakeVae(), # TODO: oh...
|
||||
vae=FakeVae(), # TODO: oh...
|
||||
text_encoder=None,
|
||||
tokenizer=None,
|
||||
unet=unet,
|
||||
@ -198,11 +207,12 @@ class TextToLatentsInvocation(BaseInvocation):
|
||||
requires_safety_checker=False,
|
||||
precision="float16" if unet.dtype == torch.float16 else "float32",
|
||||
)
|
||||
|
||||
|
||||
def prep_control_data(
|
||||
self,
|
||||
context: InvocationContext,
|
||||
model: StableDiffusionGeneratorPipeline, # really only need model for dtype and device
|
||||
# really only need model for dtype and device
|
||||
model: StableDiffusionGeneratorPipeline,
|
||||
control_input: List[ControlField],
|
||||
latents_shape: List[int],
|
||||
do_classifier_free_guidance: bool = True,
|
||||
@ -238,15 +248,17 @@ class TextToLatentsInvocation(BaseInvocation):
|
||||
print("Using HF model subfolders")
|
||||
print(" control_name: ", control_name)
|
||||
print(" control_subfolder: ", control_subfolder)
|
||||
control_model = ControlNetModel.from_pretrained(control_name,
|
||||
subfolder=control_subfolder,
|
||||
torch_dtype=model.unet.dtype).to(model.device)
|
||||
control_model = ControlNetModel.from_pretrained(
|
||||
control_name, subfolder=control_subfolder,
|
||||
torch_dtype=model.unet.dtype).to(
|
||||
model.device)
|
||||
else:
|
||||
control_model = ControlNetModel.from_pretrained(control_info.control_model,
|
||||
torch_dtype=model.unet.dtype).to(model.device)
|
||||
control_model = ControlNetModel.from_pretrained(
|
||||
control_info.control_model, torch_dtype=model.unet.dtype).to(model.device)
|
||||
control_models.append(control_model)
|
||||
control_image_field = control_info.image
|
||||
input_image = context.services.images.get_pil_image(control_image_field.image_name)
|
||||
input_image = context.services.images.get_pil_image(
|
||||
control_image_field.image_name)
|
||||
# self.image.image_type, self.image.image_name
|
||||
# FIXME: still need to test with different widths, heights, devices, dtypes
|
||||
# and add in batch_size, num_images_per_prompt?
|
||||
@ -263,41 +275,50 @@ class TextToLatentsInvocation(BaseInvocation):
|
||||
dtype=control_model.dtype,
|
||||
control_mode=control_info.control_mode,
|
||||
)
|
||||
control_item = ControlNetData(model=control_model,
|
||||
image_tensor=control_image,
|
||||
weight=control_info.control_weight,
|
||||
begin_step_percent=control_info.begin_step_percent,
|
||||
end_step_percent=control_info.end_step_percent,
|
||||
control_mode=control_info.control_mode,
|
||||
)
|
||||
control_item = ControlNetData(
|
||||
model=control_model, image_tensor=control_image,
|
||||
weight=control_info.control_weight,
|
||||
begin_step_percent=control_info.begin_step_percent,
|
||||
end_step_percent=control_info.end_step_percent,
|
||||
control_mode=control_info.control_mode,)
|
||||
control_data.append(control_item)
|
||||
# MultiControlNetModel has been refactored out, just need list[ControlNetData]
|
||||
return control_data
|
||||
|
||||
@torch.inference_mode()
|
||||
def invoke(self, context: InvocationContext) -> LatentsOutput:
|
||||
noise = context.services.latents.get(self.noise.latents_name)
|
||||
|
||||
# Get the source node id (we are invoking the prepared node)
|
||||
graph_execution_state = context.services.graph_execution_manager.get(context.graph_execution_state_id)
|
||||
graph_execution_state = context.services.graph_execution_manager.get(
|
||||
context.graph_execution_state_id)
|
||||
source_node_id = graph_execution_state.prepared_source_mapping[self.id]
|
||||
|
||||
def step_callback(state: PipelineIntermediateState):
|
||||
self.dispatch_progress(context, source_node_id, state)
|
||||
|
||||
unet_info = context.services.model_manager.get_model(**self.unet.unet.dict())
|
||||
with unet_info as unet:
|
||||
def _lora_loader():
|
||||
for lora in self.unet.loras:
|
||||
lora_info = context.services.model_manager.get_model(
|
||||
**lora.dict(exclude={"weight"}))
|
||||
yield (lora_info.context.model, lora.weight)
|
||||
del lora_info
|
||||
return
|
||||
|
||||
unet_info = context.services.model_manager.get_model(
|
||||
**self.unet.unet.dict())
|
||||
with ModelPatcher.apply_lora_unet(unet_info.context.model, _lora_loader()),\
|
||||
unet_info as unet:
|
||||
|
||||
scheduler = get_scheduler(
|
||||
context=context,
|
||||
scheduler_info=self.unet.scheduler,
|
||||
scheduler_name=self.scheduler,
|
||||
)
|
||||
|
||||
|
||||
pipeline = self.create_pipeline(unet, scheduler)
|
||||
conditioning_data = self.get_conditioning_data(context, scheduler)
|
||||
|
||||
loras = [(context.services.model_manager.get_model(**lora.dict(exclude={"weight"})).context.model, lora.weight) for lora in self.unet.loras]
|
||||
|
||||
control_data = self.prep_control_data(
|
||||
model=pipeline, context=context, control_input=self.control,
|
||||
latents_shape=noise.shape,
|
||||
@ -305,16 +326,15 @@ class TextToLatentsInvocation(BaseInvocation):
|
||||
do_classifier_free_guidance=True,
|
||||
)
|
||||
|
||||
with ModelPatcher.apply_lora_unet(pipeline.unet, loras):
|
||||
# TODO: Verify the noise is the right size
|
||||
result_latents, result_attention_map_saver = pipeline.latents_from_embeddings(
|
||||
latents=torch.zeros_like(noise, dtype=torch_dtype(unet.device)),
|
||||
noise=noise,
|
||||
num_inference_steps=self.steps,
|
||||
conditioning_data=conditioning_data,
|
||||
control_data=control_data, # list[ControlNetData]
|
||||
callback=step_callback,
|
||||
)
|
||||
# TODO: Verify the noise is the right size
|
||||
result_latents, result_attention_map_saver = pipeline.latents_from_embeddings(
|
||||
latents=torch.zeros_like(noise, dtype=torch_dtype(unet.device)),
|
||||
noise=noise,
|
||||
num_inference_steps=self.steps,
|
||||
conditioning_data=conditioning_data,
|
||||
control_data=control_data, # list[ControlNetData]
|
||||
callback=step_callback,
|
||||
)
|
||||
|
||||
# https://discuss.huggingface.co/t/memory-usage-by-later-pipeline-stages/23699
|
||||
torch.cuda.empty_cache()
|
||||
@ -323,14 +343,18 @@ class TextToLatentsInvocation(BaseInvocation):
|
||||
context.services.latents.save(name, result_latents)
|
||||
return build_latents_output(latents_name=name, latents=result_latents)
|
||||
|
||||
|
||||
class LatentsToLatentsInvocation(TextToLatentsInvocation):
|
||||
"""Generates latents using latents as base image."""
|
||||
|
||||
type: Literal["l2l"] = "l2l"
|
||||
|
||||
# Inputs
|
||||
latents: Optional[LatentsField] = Field(description="The latents to use as a base image")
|
||||
strength: float = Field(default=0.7, ge=0, le=1, description="The strength of the latents to use")
|
||||
latents: Optional[LatentsField] = Field(
|
||||
description="The latents to use as a base image")
|
||||
strength: float = Field(
|
||||
default=0.7, ge=0, le=1,
|
||||
description="The strength of the latents to use")
|
||||
|
||||
# Schema customisation
|
||||
class Config(InvocationConfig):
|
||||
@ -345,22 +369,31 @@ class LatentsToLatentsInvocation(TextToLatentsInvocation):
|
||||
},
|
||||
}
|
||||
|
||||
@torch.inference_mode()
|
||||
def invoke(self, context: InvocationContext) -> LatentsOutput:
|
||||
noise = context.services.latents.get(self.noise.latents_name)
|
||||
latent = context.services.latents.get(self.latents.latents_name)
|
||||
|
||||
# Get the source node id (we are invoking the prepared node)
|
||||
graph_execution_state = context.services.graph_execution_manager.get(context.graph_execution_state_id)
|
||||
graph_execution_state = context.services.graph_execution_manager.get(
|
||||
context.graph_execution_state_id)
|
||||
source_node_id = graph_execution_state.prepared_source_mapping[self.id]
|
||||
|
||||
def step_callback(state: PipelineIntermediateState):
|
||||
self.dispatch_progress(context, source_node_id, state)
|
||||
|
||||
unet_info = context.services.model_manager.get_model(
|
||||
**self.unet.unet.dict(),
|
||||
)
|
||||
def _lora_loader():
|
||||
for lora in self.unet.loras:
|
||||
lora_info = context.services.model_manager.get_model(
|
||||
**lora.dict(exclude={"weight"}))
|
||||
yield (lora_info.context.model, lora.weight)
|
||||
del lora_info
|
||||
return
|
||||
|
||||
with unet_info as unet:
|
||||
unet_info = context.services.model_manager.get_model(
|
||||
**self.unet.unet.dict())
|
||||
with ModelPatcher.apply_lora_unet(unet_info.context.model, _lora_loader()),\
|
||||
unet_info as unet:
|
||||
|
||||
scheduler = get_scheduler(
|
||||
context=context,
|
||||
@ -370,7 +403,7 @@ class LatentsToLatentsInvocation(TextToLatentsInvocation):
|
||||
|
||||
pipeline = self.create_pipeline(unet, scheduler)
|
||||
conditioning_data = self.get_conditioning_data(context, scheduler)
|
||||
|
||||
|
||||
control_data = self.prep_control_data(
|
||||
model=pipeline, context=context, control_input=self.control,
|
||||
latents_shape=noise.shape,
|
||||
@ -380,8 +413,7 @@ class LatentsToLatentsInvocation(TextToLatentsInvocation):
|
||||
|
||||
# TODO: Verify the noise is the right size
|
||||
initial_latents = latent if self.strength < 1.0 else torch.zeros_like(
|
||||
latent, device=unet.device, dtype=latent.dtype
|
||||
)
|
||||
latent, device=unet.device, dtype=latent.dtype)
|
||||
|
||||
timesteps, _ = pipeline.get_img2img_timesteps(
|
||||
self.steps,
|
||||
@ -389,18 +421,15 @@ class LatentsToLatentsInvocation(TextToLatentsInvocation):
|
||||
device=unet.device,
|
||||
)
|
||||
|
||||
loras = [(context.services.model_manager.get_model(**lora.dict(exclude={"weight"})).context.model, lora.weight) for lora in self.unet.loras]
|
||||
|
||||
with ModelPatcher.apply_lora_unet(pipeline.unet, loras):
|
||||
result_latents, result_attention_map_saver = pipeline.latents_from_embeddings(
|
||||
latents=initial_latents,
|
||||
timesteps=timesteps,
|
||||
noise=noise,
|
||||
num_inference_steps=self.steps,
|
||||
conditioning_data=conditioning_data,
|
||||
control_data=control_data, # list[ControlNetData]
|
||||
callback=step_callback
|
||||
)
|
||||
result_latents, result_attention_map_saver = pipeline.latents_from_embeddings(
|
||||
latents=initial_latents,
|
||||
timesteps=timesteps,
|
||||
noise=noise,
|
||||
num_inference_steps=self.steps,
|
||||
conditioning_data=conditioning_data,
|
||||
control_data=control_data, # list[ControlNetData]
|
||||
callback=step_callback
|
||||
)
|
||||
|
||||
# https://discuss.huggingface.co/t/memory-usage-by-later-pipeline-stages/23699
|
||||
torch.cuda.empty_cache()
|
||||
@ -417,9 +446,12 @@ class LatentsToImageInvocation(BaseInvocation):
|
||||
type: Literal["l2i"] = "l2i"
|
||||
|
||||
# Inputs
|
||||
latents: Optional[LatentsField] = Field(description="The latents to generate an image from")
|
||||
latents: Optional[LatentsField] = Field(
|
||||
description="The latents to generate an image from")
|
||||
vae: VaeField = Field(default=None, description="Vae submodel")
|
||||
tiled: bool = Field(default=False, description="Decode latents by overlaping tiles(less memory consumption)")
|
||||
tiled: bool = Field(
|
||||
default=False,
|
||||
description="Decode latents by overlaping tiles(less memory consumption)")
|
||||
|
||||
# Schema customisation
|
||||
class Config(InvocationConfig):
|
||||
@ -429,7 +461,7 @@ class LatentsToImageInvocation(BaseInvocation):
|
||||
},
|
||||
}
|
||||
|
||||
@torch.no_grad()
|
||||
@torch.inference_mode()
|
||||
def invoke(self, context: InvocationContext) -> ImageOutput:
|
||||
latents = context.services.latents.get(self.latents.latents_name)
|
||||
|
||||
@ -450,7 +482,7 @@ class LatentsToImageInvocation(BaseInvocation):
|
||||
# copied from diffusers pipeline
|
||||
latents = latents / vae.config.scaling_factor
|
||||
image = vae.decode(latents, return_dict=False)[0]
|
||||
image = (image / 2 + 0.5).clamp(0, 1) # denormalize
|
||||
image = (image / 2 + 0.5).clamp(0, 1) # denormalize
|
||||
# we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
|
||||
np_image = image.cpu().permute(0, 2, 3, 1).float().numpy()
|
||||
|
||||
@ -473,9 +505,9 @@ class LatentsToImageInvocation(BaseInvocation):
|
||||
height=image_dto.height,
|
||||
)
|
||||
|
||||
LATENTS_INTERPOLATION_MODE = Literal[
|
||||
"nearest", "linear", "bilinear", "bicubic", "trilinear", "area", "nearest-exact"
|
||||
]
|
||||
|
||||
LATENTS_INTERPOLATION_MODE = Literal["nearest", "linear",
|
||||
"bilinear", "bicubic", "trilinear", "area", "nearest-exact"]
|
||||
|
||||
|
||||
class ResizeLatentsInvocation(BaseInvocation):
|
||||
@ -484,21 +516,25 @@ class ResizeLatentsInvocation(BaseInvocation):
|
||||
type: Literal["lresize"] = "lresize"
|
||||
|
||||
# Inputs
|
||||
latents: Optional[LatentsField] = Field(description="The latents to resize")
|
||||
width: int = Field(ge=64, multiple_of=8, description="The width to resize to (px)")
|
||||
height: int = Field(ge=64, multiple_of=8, description="The height to resize to (px)")
|
||||
mode: LATENTS_INTERPOLATION_MODE = Field(default="bilinear", description="The interpolation mode")
|
||||
antialias: bool = Field(default=False, description="Whether or not to antialias (applied in bilinear and bicubic modes only)")
|
||||
latents: Optional[LatentsField] = Field(
|
||||
description="The latents to resize")
|
||||
width: int = Field(
|
||||
ge=64, multiple_of=8, description="The width to resize to (px)")
|
||||
height: int = Field(
|
||||
ge=64, multiple_of=8, description="The height to resize to (px)")
|
||||
mode: LATENTS_INTERPOLATION_MODE = Field(
|
||||
default="bilinear", description="The interpolation mode")
|
||||
antialias: bool = Field(
|
||||
default=False,
|
||||
description="Whether or not to antialias (applied in bilinear and bicubic modes only)")
|
||||
|
||||
def invoke(self, context: InvocationContext) -> LatentsOutput:
|
||||
latents = context.services.latents.get(self.latents.latents_name)
|
||||
|
||||
resized_latents = torch.nn.functional.interpolate(
|
||||
latents,
|
||||
size=(self.height // 8, self.width // 8),
|
||||
mode=self.mode,
|
||||
antialias=self.antialias if self.mode in ["bilinear", "bicubic"] else False,
|
||||
)
|
||||
latents, size=(self.height // 8, self.width // 8),
|
||||
mode=self.mode, antialias=self.antialias
|
||||
if self.mode in ["bilinear", "bicubic"] else False,)
|
||||
|
||||
# https://discuss.huggingface.co/t/memory-usage-by-later-pipeline-stages/23699
|
||||
torch.cuda.empty_cache()
|
||||
@ -515,21 +551,24 @@ class ScaleLatentsInvocation(BaseInvocation):
|
||||
type: Literal["lscale"] = "lscale"
|
||||
|
||||
# Inputs
|
||||
latents: Optional[LatentsField] = Field(description="The latents to scale")
|
||||
scale_factor: float = Field(gt=0, description="The factor by which to scale the latents")
|
||||
mode: LATENTS_INTERPOLATION_MODE = Field(default="bilinear", description="The interpolation mode")
|
||||
antialias: bool = Field(default=False, description="Whether or not to antialias (applied in bilinear and bicubic modes only)")
|
||||
latents: Optional[LatentsField] = Field(
|
||||
description="The latents to scale")
|
||||
scale_factor: float = Field(
|
||||
gt=0, description="The factor by which to scale the latents")
|
||||
mode: LATENTS_INTERPOLATION_MODE = Field(
|
||||
default="bilinear", description="The interpolation mode")
|
||||
antialias: bool = Field(
|
||||
default=False,
|
||||
description="Whether or not to antialias (applied in bilinear and bicubic modes only)")
|
||||
|
||||
def invoke(self, context: InvocationContext) -> LatentsOutput:
|
||||
latents = context.services.latents.get(self.latents.latents_name)
|
||||
|
||||
# resizing
|
||||
resized_latents = torch.nn.functional.interpolate(
|
||||
latents,
|
||||
scale_factor=self.scale_factor,
|
||||
mode=self.mode,
|
||||
antialias=self.antialias if self.mode in ["bilinear", "bicubic"] else False,
|
||||
)
|
||||
latents, scale_factor=self.scale_factor, mode=self.mode,
|
||||
antialias=self.antialias
|
||||
if self.mode in ["bilinear", "bicubic"] else False,)
|
||||
|
||||
# https://discuss.huggingface.co/t/memory-usage-by-later-pipeline-stages/23699
|
||||
torch.cuda.empty_cache()
|
||||
@ -548,7 +587,9 @@ class ImageToLatentsInvocation(BaseInvocation):
|
||||
# Inputs
|
||||
image: Union[ImageField, None] = Field(description="The image to encode")
|
||||
vae: VaeField = Field(default=None, description="Vae submodel")
|
||||
tiled: bool = Field(default=False, description="Encode latents by overlaping tiles(less memory consumption)")
|
||||
tiled: bool = Field(
|
||||
default=False,
|
||||
description="Encode latents by overlaping tiles(less memory consumption)")
|
||||
|
||||
# Schema customisation
|
||||
class Config(InvocationConfig):
|
||||
@ -558,7 +599,7 @@ class ImageToLatentsInvocation(BaseInvocation):
|
||||
},
|
||||
}
|
||||
|
||||
@torch.no_grad()
|
||||
@torch.inference_mode()
|
||||
def invoke(self, context: InvocationContext) -> LatentsOutput:
|
||||
# image = context.services.images.get(
|
||||
# self.image.image_type, self.image.image_name
|
||||
|
||||
Reference in New Issue
Block a user