github/ai-toolkit
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

Huge memory optimizations, many big fixes

Browse Source
This commit is contained in:
Jaret Burkett
2023-08-27 17:48:02 -06:00
parent cc49786ee9
commit c446f768ea
15 changed files with 86 additions and 78 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
jobs/BaseJob.py
View File

@@ -6,19 +6,16 @@ from jobs.process import BaseProcess
class BaseJob:
config: OrderedDict
job: str
name: str
meta: OrderedDict
process: List[BaseProcess]
def __init__(self, config: OrderedDict):
if not config:
raise ValueError('config is required')
self.process: List[BaseProcess]
self.config = config['config']
self.raw_config = config
self.job = config['job']
self.torch_profiler = self.get_conf('torch_profiler', False)
self.name = self.get_conf('name', required=True)
if 'meta' in config:
self.meta = config['meta']

3
jobs/ExtensionJob.py
View File

@@ -1,7 +1,8 @@
import os
from collections import OrderedDict
from jobs import BaseJob
from toolkit.extension import get_all_extensions_process_dict
from toolkit.paths import CONFIG_ROOT
class ExtensionJob(BaseJob):

1
jobs/GenerateJob.py
View File

@@ -14,7 +14,6 @@ process_dict = {
class GenerateJob(BaseJob):
process: List[GenerateProcess]
def __init__(self, config: OrderedDict):
super().__init__(config)

1
jobs/TrainJob.py
View File

@@ -26,7 +26,6 @@ process_dict = {
class TrainJob(BaseJob):
process: List[BaseExtractProcess]
def __init__(self, config: OrderedDict):
super().__init__(config)

7
jobs/process/BaseExtensionProcess.py
View File

@@ -4,10 +4,6 @@ from jobs.process.BaseProcess import BaseProcess
class BaseExtensionProcess(BaseProcess):
process_id: int
config: OrderedDict
progress_bar: ForwardRef('tqdm') = None
def __init__(
self,
process_id: int,
@@ -15,6 +11,9 @@ class BaseExtensionProcess(BaseProcess):
config: OrderedDict
):
super().__init__(process_id, job, config)
self.process_id: int
self.config: OrderedDict
self.progress_bar: ForwardRef('tqdm') = None
def run(self):
super().run()

9
jobs/process/BaseExtractProcess.py
View File

@@ -12,11 +12,6 @@ from toolkit.train_tools import get_torch_dtype
class BaseExtractProcess(BaseProcess):
process_id: int
config: OrderedDict
output_folder: str
output_filename: str
output_path: str
def __init__(
self,
@@ -25,6 +20,10 @@ class BaseExtractProcess(BaseProcess):
config: OrderedDict
):
super().__init__(process_id, job, config)
self.config: OrderedDict
self.output_folder: str
self.output_filename: str
self.output_path: str
self.process_id = process_id
self.job = job
self.config = config

4
jobs/process/BaseMergeProcess.py
View File

@@ -9,8 +9,6 @@ from toolkit.train_tools import get_torch_dtype
class BaseMergeProcess(BaseProcess):
process_id: int
config: OrderedDict
def __init__(
self,
@@ -19,6 +17,8 @@ class BaseMergeProcess(BaseProcess):
config: OrderedDict
):
super().__init__(process_id, job, config)
self.process_id: int
self.config: OrderedDict
self.output_path = self.get_conf('output_path', required=True)
self.dtype = self.get_conf('dtype', self.job.dtype)
self.torch_dtype = get_torch_dtype(self.dtype)

2
jobs/process/BaseProcess.py
View File

@@ -4,7 +4,6 @@ from collections import OrderedDict
class BaseProcess(object):
meta: OrderedDict
def __init__(
self,
@@ -13,6 +12,7 @@ class BaseProcess(object):
config: OrderedDict
):
self.process_id = process_id
self.meta: OrderedDict
self.job = job
self.config = config
self.raw_process_config = config

8
jobs/process/BaseSDTrainProcess.py
View File

@@ -9,6 +9,7 @@ from toolkit.data_loader import get_dataloader_from_datasets
from toolkit.embedding import Embedding
from toolkit.lora_special import LoRASpecialNetwork
from toolkit.optimizer import get_optimizer
from toolkit.paths import CONFIG_ROOT
from toolkit.scheduler import get_lr_scheduler
from toolkit.stable_diffusion_model import StableDiffusion
@@ -31,11 +32,12 @@ def flush():
class BaseSDTrainProcess(BaseTrainProcess):
sd: StableDiffusion
embedding: Union[Embedding, None] = None
def __init__(self, process_id: int, job, config: OrderedDict, custom_pipeline=None):
super().__init__(process_id, job, config)
self.sd: StableDiffusion
self.embedding: Union[Embedding, None] = None
self.custom_pipeline = custom_pipeline
self.step_num = 0
self.start_step = 0
@@ -344,7 +346,9 @@ class BaseSDTrainProcess(BaseTrainProcess):
# remove grads for these
noisy_latents.requires_grad = False
noisy_latents = noisy_latents.detach()
noise.requires_grad = False
noise = noise.detach()
return noisy_latents, noise, timesteps, conditioned_prompts, imgs

11
jobs/process/BaseTrainProcess.py
View File

@@ -14,11 +14,6 @@ if TYPE_CHECKING:
class BaseTrainProcess(BaseProcess):
process_id: int
config: OrderedDict
writer: 'SummaryWriter'
job: Union['TrainJob', 'BaseJob', 'ExtensionJob']
progress_bar: 'tqdm' = None
def __init__(
self,
@@ -27,6 +22,12 @@ class BaseTrainProcess(BaseProcess):
config: OrderedDict
):
super().__init__(process_id, job, config)
self.process_id: int
self.config: OrderedDict
self.writer: 'SummaryWriter'
self.job: Union['TrainJob', 'BaseJob', 'ExtensionJob']
self.progress_bar: 'tqdm' = None
self.progress_bar = None
self.writer = None
self.training_folder = self.get_conf('training_folder', self.job.training_folder if hasattr(self.job, 'training_folder') else None)

2
jobs/process/GenerateProcess.py
View File

@@ -16,9 +16,9 @@ import random
class GenerateConfig:
prompts: List[str]
def __init__(self, **kwargs):
self.prompts: List[str]
self.sampler = kwargs.get('sampler', 'ddpm')
self.width = kwargs.get('width', 512)
self.height = kwargs.get('height', 512)

3
jobs/process/ModRescaleLoraProcess.py
View File

@@ -24,6 +24,9 @@ class ModRescaleLoraProcess(BaseProcess):
config: OrderedDict
):
super().__init__(process_id, job, config)
self.process_id: int
self.config: OrderedDict
self.progress_bar: ForwardRef('tqdm') = None
self.input_path = self.get_conf('input_path', required=True)
self.output_path = self.get_conf('output_path', required=True)
self.replace_meta = self.get_conf('replace_meta', default=False)