Skip to content

(#1808) add options for validation adapters and fix bug preventing validation page from rendering #1843

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Oct 26, 2025
Merged

(#1808) add options for validation adapters and fix bug preventing validation page from rendering #1843

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
1d6d76e
(#1808) add options for validation adapters and fix bug preventing va…
Oct 26, 2025
8ec63ff
fix offload check in newer Diffusers returning is_group_offload using…
Oct 26, 2025
cd10093
expand validation_adapter* settings to cover more use cases
Oct 26, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
41 changes: 41 additions & 0 deletions simpletuner/helpers/configuration/cmd_args.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -837,6 +837,47 @@ def _set_tf32(enabled: bool) -> None:
elif args.sana_complex_human_instruction == "None":
args.sana_complex_human_instruction = None

if isinstance(getattr(args, "validation_adapter_path", None), str):
candidate = args.validation_adapter_path.strip()
args.validation_adapter_path = candidate or None

if getattr(args, "validation_adapter_config", None):
args.validation_adapter_config = _parse_json_like_option(
args.validation_adapter_config,
"--validation_adapter_config",
)

if args.validation_adapter_path and args.validation_adapter_config:
raise ValueError("Provide either --validation_adapter_path or --validation_adapter_config, not both.")

if isinstance(getattr(args, "validation_adapter_name", None), str):
candidate = args.validation_adapter_name.strip()
args.validation_adapter_name = candidate or None

strength_value = getattr(args, "validation_adapter_strength", None)
if strength_value is None or strength_value in ("", "None"):
args.validation_adapter_strength = 1.0
else:
try:
strength = float(strength_value)
except (TypeError, ValueError):
raise ValueError(f"Invalid --validation_adapter_strength value: {strength_value}") from None
if strength <= 0:
raise ValueError("--validation_adapter_strength must be greater than 0.")
args.validation_adapter_strength = strength

mode_value = getattr(args, "validation_adapter_mode", None)
if mode_value in (None, "", "None"):
args.validation_adapter_mode = "adapter_only"
else:
normalized_mode = str(mode_value).strip().lower()
valid_modes = {"adapter_only", "comparison", "none"}
if normalized_mode not in valid_modes:
raise ValueError(
f"Invalid --validation_adapter_mode '{mode_value}'. Expected one of: {', '.join(sorted(valid_modes))}."
)
args.validation_adapter_mode = normalized_mode

if args.attention_mechanism != "diffusers" and not torch.cuda.is_available():
warning_log("For non-CUDA systems, only Diffusers attention mechanism is officially supported.")

Expand Down
21 changes: 17 additions & 4 deletions simpletuner/helpers/models/chroma/pipeline.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -45,6 +45,7 @@
from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection, T5EncoderModel, T5TokenizerFast

from simpletuner.helpers.models.chroma.transformer import ChromaTransformer2DModel
from simpletuner.helpers.utils.offloading import restore_offload_state, unpack_offload_state

if is_torch_xla_available():
import torch_xla.core.xla_model as xm
Expand Down Expand Up @@ -423,7 +424,12 @@ def load_lora_into_controlnet(

# In case the pipeline has been already offloaded to CPU - temporarily remove the hooks
# otherwise loading LoRA weights will lead to an error
is_model_cpu_offload, is_sequential_cpu_offload = cls._optionally_disable_offloading(_pipeline)
offload_state = cls._optionally_disable_offloading(_pipeline)
(
is_model_cpu_offload,
is_sequential_cpu_offload,
is_group_offload,
) = unpack_offload_state(offload_state)

peft_kwargs = {}
if is_peft_version(">=", "0.13.1"):
Expand All @@ -447,7 +453,8 @@ def load_lora_into_controlnet(
if warn_msg:
logger.warning(warn_msg)

cls._optionally_enable_offloading(is_model_cpu_offload, is_sequential_cpu_offload, _pipeline)
# Offload back.
restore_offload_state(_pipeline, is_model_cpu_offload, is_sequential_cpu_offload, is_group_offload)

@classmethod
def load_lora_into_transformer(
Expand Down Expand Up @@ -527,7 +534,12 @@ def load_lora_into_transformer(

# In case the pipeline has been already offloaded to CPU - temporarily remove the hooks
# otherwise loading LoRA weights will lead to an error
is_model_cpu_offload, is_sequential_cpu_offload = cls._optionally_disable_offloading(_pipeline)
offload_state = cls._optionally_disable_offloading(_pipeline)
(
is_model_cpu_offload,
is_sequential_cpu_offload,
is_group_offload,
) = unpack_offload_state(offload_state)

peft_kwargs = {}
if is_peft_version(">=", "0.13.1"):
Expand All @@ -551,7 +563,8 @@ def load_lora_into_transformer(
if warn_msg:
logger.warning(warn_msg)

cls._optionally_enable_offloading(is_model_cpu_offload, is_sequential_cpu_offload, _pipeline)
# Offload back.
restore_offload_state(_pipeline, is_model_cpu_offload, is_sequential_cpu_offload, is_group_offload)

# Copied from diffusers.loaders.lora_pipeline.StableDiffusionLoraLoaderMixin.load_lora_into_text_encoder
def load_lora_into_text_encoder(
Expand Down
39 changes: 24 additions & 15 deletions simpletuner/helpers/models/flux/pipeline.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -66,6 +66,8 @@
T5TokenizerFast,
)

from simpletuner.helpers.utils.offloading import restore_offload_state, unpack_offload_state

if is_torch_xla_available():
import torch_xla.core.xla_model as xm

Expand Down Expand Up @@ -521,7 +523,12 @@ def load_lora_into_controlnet(
adapter_name = get_adapter_name(controlnet)

# In case the pipeline has been already offloaded to CPU - temporarily remove the hooks
is_model_cpu_offload, is_sequential_cpu_offload = cls._optionally_disable_offloading(_pipeline)
offload_state = cls._optionally_disable_offloading(_pipeline)
(
is_model_cpu_offload,
is_sequential_cpu_offload,
is_group_offload,
) = unpack_offload_state(offload_state)

peft_kwargs = {}
if is_peft_version(">=", "0.13.1"):
Expand All @@ -534,10 +541,7 @@ def load_lora_into_controlnet(
logger.info(f"Loaded ControlNet LoRA with incompatible keys: {incompatible_keys}")

# Offload back.
if is_model_cpu_offload:
_pipeline.enable_model_cpu_offload()
elif is_sequential_cpu_offload:
_pipeline.enable_sequential_cpu_offload()
restore_offload_state(_pipeline, is_model_cpu_offload, is_sequential_cpu_offload, is_group_offload)

@classmethod
def load_lora_into_transformer(
Expand Down Expand Up @@ -617,7 +621,12 @@ def load_lora_into_transformer(

# In case the pipeline has been already offloaded to CPU - temporarily remove the hooks
# otherwise loading LoRA weights will lead to an error
is_model_cpu_offload, is_sequential_cpu_offload = cls._optionally_disable_offloading(_pipeline)
offload_state = cls._optionally_disable_offloading(_pipeline)
(
is_model_cpu_offload,
is_sequential_cpu_offload,
is_group_offload,
) = unpack_offload_state(offload_state)

peft_kwargs = {}
if is_peft_version(">=", "0.13.1"):
Expand Down Expand Up @@ -652,10 +661,7 @@ def load_lora_into_transformer(
logger.warning(warn_msg)

# Offload back.
if is_model_cpu_offload:
_pipeline.enable_model_cpu_offload()
elif is_sequential_cpu_offload:
_pipeline.enable_sequential_cpu_offload()
restore_offload_state(_pipeline, is_model_cpu_offload, is_sequential_cpu_offload, is_group_offload)
# Unsafe code />

@classmethod
Expand Down Expand Up @@ -769,7 +775,12 @@ def load_lora_into_text_encoder(
if adapter_name is None:
adapter_name = get_adapter_name(text_encoder)

is_model_cpu_offload, is_sequential_cpu_offload = cls._optionally_disable_offloading(_pipeline)
offload_state = cls._optionally_disable_offloading(_pipeline)
(
is_model_cpu_offload,
is_sequential_cpu_offload,
is_group_offload,
) = unpack_offload_state(offload_state)

# inject LoRA layers and load the state dict
# in transformers we automatically check whether the adapter name is already in use or not
Expand All @@ -786,10 +797,8 @@ def load_lora_into_text_encoder(
text_encoder.to(device=text_encoder.device, dtype=text_encoder.dtype)

# Offload back.
if is_model_cpu_offload:
_pipeline.enable_model_cpu_offload()
elif is_sequential_cpu_offload:
_pipeline.enable_sequential_cpu_offload()
restore_offload_state(_pipeline, is_model_cpu_offload, is_sequential_cpu_offload, is_group_offload)

# Unsafe code />

@classmethod
Expand Down
25 changes: 15 additions & 10 deletions simpletuner/helpers/models/hidream/pipeline.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -46,6 +46,7 @@
)

from simpletuner.helpers.models.hidream.schedule import FlowUniPCMultistepScheduler
from simpletuner.helpers.utils.offloading import restore_offload_state, unpack_offload_state

if is_torch_xla_available():
import torch_xla.core.xla_model as xm
Expand Down Expand Up @@ -486,7 +487,12 @@ def load_lora_into_controlnet(
adapter_name = get_adapter_name(controlnet)

# In case the pipeline has been already offloaded to CPU - temporarily remove the hooks
is_model_cpu_offload, is_sequential_cpu_offload = cls._optionally_disable_offloading(_pipeline)
offload_state = cls._optionally_disable_offloading(_pipeline)
(
is_model_cpu_offload,
is_sequential_cpu_offload,
is_group_offload,
) = unpack_offload_state(offload_state)

peft_kwargs = {}
if is_peft_version(">=", "0.13.1"):
Expand All @@ -499,10 +505,7 @@ def load_lora_into_controlnet(
logger.info(f"Loaded ControlNet LoRA with incompatible keys: {incompatible_keys}")

# Offload back.
if is_model_cpu_offload:
_pipeline.enable_model_cpu_offload()
elif is_sequential_cpu_offload:
_pipeline.enable_sequential_cpu_offload()
restore_offload_state(_pipeline, is_model_cpu_offload, is_sequential_cpu_offload, is_group_offload)

@classmethod
def load_lora_into_transformer(
Expand Down Expand Up @@ -638,7 +641,12 @@ def load_lora_into_text_encoder(
if adapter_name is None:
adapter_name = get_adapter_name(text_encoder)

is_model_cpu_offload, is_sequential_cpu_offload = cls._optionally_disable_offloading(_pipeline)
offload_state = cls._optionally_disable_offloading(_pipeline)
(
is_model_cpu_offload,
is_sequential_cpu_offload,
is_group_offload,
) = unpack_offload_state(offload_state)

# inject LoRA layers and load the state dict
# in transformers we automatically check whether the adapter name is already in use or not
Expand All @@ -655,10 +663,7 @@ def load_lora_into_text_encoder(
text_encoder.to(device=text_encoder.device, dtype=text_encoder.dtype)

# Offload back.
if is_model_cpu_offload:
_pipeline.enable_model_cpu_offload()
elif is_sequential_cpu_offload:
_pipeline.enable_sequential_cpu_offload()
restore_offload_state(_pipeline, is_model_cpu_offload, is_sequential_cpu_offload, is_group_offload)

@classmethod
def save_lora_weights(
Expand Down
38 changes: 23 additions & 15 deletions simpletuner/helpers/models/sd3/pipeline.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -53,6 +53,8 @@
from huggingface_hub.utils import validate_hf_hub_args
from transformers import CLIPTextModelWithProjection, CLIPTokenizer, T5EncoderModel, T5TokenizerFast

from simpletuner.helpers.utils.offloading import restore_offload_state, unpack_offload_state

if is_torch_xla_available():
import torch_xla.core.xla_model as xm

Expand Down Expand Up @@ -498,7 +500,12 @@ def load_lora_into_controlnet(
adapter_name = get_adapter_name(controlnet)

# In case the pipeline has been already offloaded to CPU - temporarily remove the hooks
is_model_cpu_offload, is_sequential_cpu_offload = cls._optionally_disable_offloading(_pipeline)
offload_state = cls._optionally_disable_offloading(_pipeline)
(
is_model_cpu_offload,
is_sequential_cpu_offload,
is_group_offload,
) = unpack_offload_state(offload_state)

peft_kwargs = {}
if is_peft_version(">=", "0.13.1"):
Expand All @@ -511,10 +518,7 @@ def load_lora_into_controlnet(
logger.info(f"Loaded ControlNet LoRA with incompatible keys: {incompatible_keys}")

# Offload back.
if is_model_cpu_offload:
_pipeline.enable_model_cpu_offload()
elif is_sequential_cpu_offload:
_pipeline.enable_sequential_cpu_offload()
restore_offload_state(_pipeline, is_model_cpu_offload, is_sequential_cpu_offload, is_group_offload)

@classmethod
def load_lora_into_transformer(
Expand Down Expand Up @@ -584,7 +588,12 @@ def load_lora_into_transformer(

# In case the pipeline has been already offloaded to CPU - temporarily remove the hooks
# otherwise loading LoRA weights will lead to an error
is_model_cpu_offload, is_sequential_cpu_offload = cls._optionally_disable_offloading(_pipeline)
offload_state = cls._optionally_disable_offloading(_pipeline)
(
is_model_cpu_offload,
is_sequential_cpu_offload,
is_group_offload,
) = unpack_offload_state(offload_state)

peft_kwargs = {}
if is_peft_version(">=", "0.13.1"):
Expand Down Expand Up @@ -619,10 +628,7 @@ def load_lora_into_transformer(
logger.warning(warn_msg)

# Offload back.
if is_model_cpu_offload:
_pipeline.enable_model_cpu_offload()
elif is_sequential_cpu_offload:
_pipeline.enable_sequential_cpu_offload()
restore_offload_state(_pipeline, is_model_cpu_offload, is_sequential_cpu_offload, is_group_offload)
# Unsafe code />

@classmethod
Expand Down Expand Up @@ -736,7 +742,12 @@ def load_lora_into_text_encoder(
if adapter_name is None:
adapter_name = get_adapter_name(text_encoder)

is_model_cpu_offload, is_sequential_cpu_offload = cls._optionally_disable_offloading(_pipeline)
offload_state = cls._optionally_disable_offloading(_pipeline)
(
is_model_cpu_offload,
is_sequential_cpu_offload,
is_group_offload,
) = unpack_offload_state(offload_state)

# inject LoRA layers and load the state dict
# in transformers we automatically check whether the adapter name is already in use or not
Expand All @@ -753,10 +764,7 @@ def load_lora_into_text_encoder(
text_encoder.to(device=text_encoder.device, dtype=text_encoder.dtype)

# Offload back.
if is_model_cpu_offload:
_pipeline.enable_model_cpu_offload()
elif is_sequential_cpu_offload:
_pipeline.enable_sequential_cpu_offload()
restore_offload_state(_pipeline, is_model_cpu_offload, is_sequential_cpu_offload, is_group_offload)
# Unsafe code />

@classmethod
Expand Down
Loading