Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-20 01:31:13 +02:00
Code Issues Packages Projects Releases Wiki Activity

update flux lora code

Browse Source 
Signed-off-by: Vladimir Mandic <mandic00@live.com>
This commit is contained in:
Vladimir Mandic
2024-09-30 10:16:02 -04:00
parent e6ff17d8c0
commit ac58b21d06
6 changed files with 19 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files
CHANGELOG.md
+3 -1
View File
@@ -44,7 +44,9 @@ And other goodies like XYZ grid improvements, additional Flux controlnets, addit
- full prompt parser will auto-select `xhinker` for flux models
- controlnet support for img2img and inpaint (in addition to previous txt2img controlnet)
- allow separate vae load
- add additional controlnets: [JasperAI](https://huggingface.co/collections/jasperai/flux1-dev-controlnets-66f27f9459d760dcafa32e08) **Depth**, **Upscaler**, **Surface**, thanks @EnragedAntelope
- support for both kohya and onetrainer loras in native load mode for fp16/nf4/fp4, thanks @AI-Casanova
- added native load mode for qint8/qint4 models
- add additional controlnets: [JasperAI](https://huggingface.co/collections/jasperai/flux1-dev-controlnets-66f27f9459d760dcafa32e08) **Depth**, **Upscaler**, **Surface**, thanks @EnragedAntelope
- **dtype**
- previously `cuda_dtype` in settings defaulted to `fp16` if available
- now `cuda_type` defaults to **Auto** which executes `bf16` and `fp16` tests on startup and selects best available dtype
extensions-builtin/Lora/lora_convert.py
+1 -1
View File
@@ -1,8 +1,8 @@
import os
import re
import bisect
import torch
from typing import Dict
import torch
from modules import shared
extensions-builtin/Lora/lora_patches.py
+10 -5
View File
@@ -16,6 +16,7 @@ class LoraPatches:
self.LayerNorm_load_state_dict = None
self.MultiheadAttention_forward = None
self.MultiheadAttention_load_state_dict = None
self.Linear4bit_forward = None
def apply(self):
if self.active or shared.opts.lora_force_diffusers:
@@ -23,12 +24,14 @@ class LoraPatches:
try:
import bitsandbytes
self.Linear4bit_forward = patches.patch(__name__, bitsandbytes.nn.Linear4bit, 'forward', networks.network_Linear4bit_forward)
except:
except Exception:
pass
if "Model" in shared.opts.optimum_quanto_weights or "Text Encoder" in shared.opts.optimum_quanto_weights:
try:
from optimum import quanto # pylint: disable=no-name-in-module
self.QLinear_forward = patches.patch(__name__, quanto.nn.QLinear, 'forward', networks.network_QLinear_forward) # pylint: disable=attribute-defined-outside-init
self.QConv2d_forward = patches.patch(__name__, quanto.nn.QConv2d, 'forward', networks.network_QConv2d_forward) # pylint: disable=attribute-defined-outside-init
except Exception:
pass
self.Linear_forward = patches.patch(__name__, torch.nn.Linear, 'forward', networks.network_Linear_forward)
self.Linear_load_state_dict = patches.patch(__name__, torch.nn.Linear, '_load_from_state_dict', networks.network_Linear_load_state_dict)
self.Conv2d_forward = patches.patch(__name__, torch.nn.Conv2d, 'forward', networks.network_Conv2d_forward)
@@ -49,13 +52,15 @@ class LoraPatches:
return
try:
import bitsandbytes
self.Linear4bit_forward = patches.undo(__name__, bitsandbytes.nn.Linear4bit, 'forward')
except:
self.Linear4bit_forward = patches.undo(__name__, bitsandbytes.nn.Linear4bit, 'forward') # pylint: disable=E1128
except Exception:
pass
if "Model" in shared.opts.optimum_quanto_weights or "Text Encoder" in shared.opts.optimum_quanto_weights:
try:
from optimum import quanto # pylint: disable=no-name-in-module
self.QLinear_forward = patches.undo(__name__, quanto.nn.QLinear, 'forward') # pylint: disable=E1128, attribute-defined-outside-init
self.QConv2d_forward = patches.undo(__name__, quanto.nn.QConv2d, 'forward') # pylint: disable=E1128, attribute-defined-outside-init
except Exception:
pass
self.Linear_forward = patches.undo(__name__, torch.nn.Linear, 'forward') # pylint: disable=E1128
self.Linear_load_state_dict = patches.undo(__name__, torch.nn.Linear, '_load_from_state_dict') # pylint: disable=E1128
self.Conv2d_forward = patches.undo(__name__, torch.nn.Conv2d, 'forward') # pylint: disable=E1128
extensions-builtin/Lora/networks.py
+3 -3
View File
@@ -296,7 +296,7 @@ def network_restore_weights_from_backup(self: Union[torch.nn.Conv2d, torch.nn.Li
elif hasattr(self, "qweight") and hasattr(self, "freeze"):
self.weight = torch.nn.Parameter(weights_backup.to(self.weight.device, copy=True))
self.freeze()
elif getattr(self, "quant_type", None) is not None:
elif getattr(self, "quant_type", None) in ['nf4', 'fp4']:
import bitsandbytes
device = self.weight.device
self.weight = bitsandbytes.nn.Params4bit(weights_backup, quant_state=self.quant_state,
@@ -336,7 +336,7 @@ def network_apply_weights(self: Union[torch.nn.Conv2d, torch.nn.Linear, torch.nn
raise RuntimeError("no backup weights found and current weights are not unchanged")
if isinstance(self, torch.nn.MultiheadAttention):
weights_backup = (self.in_proj_weight.clone().to(devices.cpu), self.out_proj.weight.clone().to(devices.cpu))
elif getattr(self.weight, "quant_type", None) == "nf4" or getattr(self.weight, "quant_type", None) == "nf4":
elif getattr(self.weight, "quant_type", None) in ['nf4', 'fp4']:
import bitsandbytes
with devices.inference_context():
weights_backup = bitsandbytes.functional.dequantize_4bit(self.weight,
@@ -373,7 +373,7 @@ def network_apply_weights(self: Union[torch.nn.Conv2d, torch.nn.Linear, torch.nn
if len(weight.shape) == 4 and weight.shape[1] == 9:
# inpainting model. zero pad updown to make channel[1] 4 to 9
updown = torch.nn.functional.pad(updown, (0, 0, 0, 0, 0, 5)) # pylint: disable=not-callable
if getattr(self.weight, "quant_type", None) == "nf4" or self.weight.numel() != updown.numel():
if getattr(self.weight, "quant_type", None) in ['nf4', 'fp4'] or self.weight.numel() != updown.numel():
import bitsandbytes
device = self.weight.device
weight = bitsandbytes.functional.dequantize_4bit(self.weight,
installer.py
+1 -1
View File
@@ -432,7 +432,7 @@ def check_python(supported_minors=[9, 10, 11, 12], reason=None):
# check diffusers version
def check_diffusers():
sha = 'aa73072f1f7014635e3de916cbcf47858f4c37a0'
sha = '8e7d6c03a366fdb0f551ce7b92f0871c863d4e08'
pkg = pkg_resources.working_set.by_key.get('diffusers', None)
minor = int(pkg.version.split('.')[1] if pkg is not None else 0)
cur = opts.get('diffusers_version', '') if minor > 0 else ''
modules/shared.py
+1 -1
View File
@@ -861,12 +861,12 @@ options_templates.update(options_section(('extra_networks', "Networks"), {
"extra_networks_styles": OptionInfo(True, "Show built-in styles"),
"lora_preferred_name": OptionInfo("filename", "LoRA preferred name", gr.Radio, {"choices": ["filename", "alias"]}),
"lora_add_hashes_to_infotext": OptionInfo(False, "LoRA add hash info"),
"lora_quant": OptionInfo("FP4","LoRA precision for merged layers in quantized models", gr.Radio, {"choices": ["FP4", "NF4"]}),
"lora_force_diffusers": OptionInfo(False if not cmd_opts.use_openvino else True, "LoRA force loading of all models using Diffusers"),
"lora_maybe_diffusers": OptionInfo(False, "LoRA force loading of specific models using Diffusers"),
"lora_fuse_diffusers": OptionInfo(False if not cmd_opts.use_openvino else True, "LoRA use merge when using alternative method"),
"lora_apply_tags": OptionInfo(0, "LoRA auto-apply tags", gr.Slider, {"minimum": -1, "maximum": 32, "step": 1}),
"lora_in_memory_limit": OptionInfo(0, "LoRA memory cache", gr.Slider, {"minimum": 0, "maximum": 24, "step": 1}),
"lora_quant": OptionInfo("NF4","LoRA precision in quantized models", gr.Radio, {"choices": ["NF4", "FP4"]}),
"lora_functional": OptionInfo(False, "Use Kohya method for handling multiple LoRA", gr.Checkbox, { "visible": False }),
"lora_load_gpu": OptionInfo(True if not cmd_opts.lowvram else False, "Load LoRA directly to GPU"),
"hypernetwork_enabled": OptionInfo(False, "Enable Hypernetwork support"),