hcpdiff 2.1__tar.gz → 2.2.1__tar.gz

{hcpdiff-2.1 → hcpdiff-2.2.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hcpdiff
-Version: 2.1
+Version: 2.2.1
 Summary: A universal Diffusion toolbox
 Home-page: https://github.com/IrisRainbowNeko/HCP-Diffusion
 Author: Ziyi Dong
@@ -17,7 +17,7 @@ Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
 Requires-Python: >=3.8
 Description-Content-Type: text/markdown
 License-File: LICENSE
-Requires-Dist: rainbowneko
+Requires-Dist: rainbowneko==1.6
 Requires-Dist: diffusers
 Requires-Dist: matplotlib
 Requires-Dist: pyarrow
@@ -65,6 +65,8 @@ Compared to the original DreamArtist, it offers better stability, image quality,
 
 ## Installation
 
+Install [pytorch](https://pytorch.org/)
+
 Install via pip:
 
 ```bash
@@ -205,6 +207,18 @@ After parsing, the framework will instantiate the components accordingly. This m
 | CCIP Score       | 🚧 In Development                        |
 | Corrupt Score    | 🚧 In Development                        |
 
+---
+
+### ⚡️ Image Generation
+
+| 功能                           | 描述/支持情况                            |
+|------------------------------|------------------------------------|
+| Batch Generation             | ✅ Supported                   |
+| Generate from Prompt Dataset | ✅ Supported                               |
+| Image to Image               | ✅ Supported                               |
+| Inpaint                      | ✅ Supported                               |
+| Token Weight                 | ✅ Supported |
+
 </details>
 
 ---
@@ -248,9 +262,13 @@ hcp_run --cfg cfgs/workflow/text2img_cli.py \
     seed=42
 ```
 
-### Tutorials
+### 📚 Tutorials
 
-🚧 In Development
++ 🧠 [Model Training Guide](https://hcpdiff.readthedocs.io/en/latest/user_guides/train.html)
++ 🔧 [LoRA Training Tutorial](https://hcpdiff.readthedocs.io/enlatest/tutorial/lora.html)
++ 🎨 [Image Generation Guide](https://hcpdiff.readthedocs.io/en/latest/user_guides/workflow.html)
++ ⚙️ [Configuration File Explanation](https://hcpdiff.readthedocs.io/en/latest/user_guides/cfg.html)
++ 🧩 [Model Format Explanation](https://hcpdiff.readthedocs.io/en/latest/user_guides/model_format.html)
 
 ---
 

{hcpdiff-2.1 → hcpdiff-2.2.1}/README.md

@@ -28,6 +28,8 @@ Compared to the original DreamArtist, it offers better stability, image quality,
 
 ## Installation
 
+Install [pytorch](https://pytorch.org/)
+
 Install via pip:
 
 ```bash
@@ -168,6 +170,18 @@ After parsing, the framework will instantiate the components accordingly. This m
 | CCIP Score       | 🚧 In Development                        |
 | Corrupt Score    | 🚧 In Development                        |
 
+---
+
+### ⚡️ Image Generation
+
+| 功能                           | 描述/支持情况                            |
+|------------------------------|------------------------------------|
+| Batch Generation             | ✅ Supported                   |
+| Generate from Prompt Dataset | ✅ Supported                               |
+| Image to Image               | ✅ Supported                               |
+| Inpaint                      | ✅ Supported                               |
+| Token Weight                 | ✅ Supported |
+
 </details>
 
 ---
@@ -211,9 +225,13 @@ hcp_run --cfg cfgs/workflow/text2img_cli.py \
     seed=42
 ```
 
-### Tutorials
+### 📚 Tutorials
 
-🚧 In Development
++ 🧠 [Model Training Guide](https://hcpdiff.readthedocs.io/en/latest/user_guides/train.html)
++ 🔧 [LoRA Training Tutorial](https://hcpdiff.readthedocs.io/enlatest/tutorial/lora.html)
++ 🎨 [Image Generation Guide](https://hcpdiff.readthedocs.io/en/latest/user_guides/workflow.html)
++ ⚙️ [Configuration File Explanation](https://hcpdiff.readthedocs.io/en/latest/user_guides/cfg.html)
++ 🧩 [Model Format Explanation](https://hcpdiff.readthedocs.io/en/latest/user_guides/model_format.html)
 
 ---
 

{hcpdiff-2.1 → hcpdiff-2.2.1}/hcpdiff/ckpt_manager/__init__.py

@@ -1,4 +1,4 @@
 from .format import EmbFormat, DiffusersSD15Format, DiffusersModelFormat, DiffusersSDXLFormat, DiffusersPixArtFormat, OfficialSDXLFormat, \
-    OfficialSD15Format
+    OfficialSD15Format, LoraWebuiFormat
 from .ckpt import EmbSaver, easy_emb_saver
 from .loader import HCPLoraLoader

{hcpdiff-2.1 → hcpdiff-2.2.1}/hcpdiff/ckpt_manager/format/lora_webui.py

@@ -2,7 +2,7 @@ import math
 import re
 from typing import List, Dict, Any
 
-from rainbowneko.ckpt_manager.format import CkptFormat
+from rainbowneko.ckpt_manager.format import CkptFormat, SafeTensorFormat
 from torch.serialization import FILE_LIKE
 
 class LoraConverter:
@@ -36,7 +36,12 @@ class LoraConverter:
         if auto_scale_alpha:
             sd_unet = self.alpha_scale_from_webui(sd_unet)
             sd_TE = self.alpha_scale_from_webui(sd_TE)
-        return {'plugin':sd_TE}, {'plugin':sd_unet}
+
+        sd = {
+            **{f'denoiser.{k}':v for k,v in sd_unet.items()},
+            **{f'TE.{k}':v for k,v in sd_TE.items()},
+        }
+        return {'base': sd}
 
     def convert_to_webui(self, sd_unet, sd_TE, auto_scale_alpha=False, sdxl=False):
         sd_unet = self.convert_to_webui_(sd_unet, prefix=self.prefix_unet)
@@ -207,9 +212,12 @@ class LoraConverter:
         return state
 
 class LoraWebuiFormat(CkptFormat):
-    def __init__(self, format, auto_scale_alpha=False):
+    def __init__(self, format=None, auto_scale_alpha=False):
         self.converter = LoraConverter()
         self.auto_scale_alpha = auto_scale_alpha
+
+        if format is None:
+            format = SafeTensorFormat()
         self.format = format
 
     def save_ckpt(self, sd_model: Dict[str, Any], save_f: FILE_LIKE):
@@ -240,5 +248,5 @@ class LoraWebuiFormat(CkptFormat):
                 sdxl = True
                 break
 
-        sd_TE, sd_unet = self.converter.convert_from_webui(sd_webui, auto_scale_alpha=self.auto_scale_alpha, sdxl=sdxl)
-        return sd_TE, sd_unet
+        sd_all = self.converter.convert_from_webui(sd_webui, auto_scale_alpha=self.auto_scale_alpha, sdxl=sdxl)
+        return sd_all

{hcpdiff-2.1 → hcpdiff-2.2.1}/hcpdiff/data/__init__.py

@@ -1,4 +1,4 @@
 from .dataset import TextImagePairDataset
-from .source import Text2ImageSource, Text2ImageLossMapSource, Text2ImageCondSource, T2IFolderClassSource
-from .handler import StableDiffusionHandler, LossMapHandler, DiffusionImageHandler
+from .source import Text2ImageSource, Text2ImageLossMapSource, Text2ImageCondSource, T2IFolderClassSource, TextSource
+from .handler import StableDiffusionHandler, LossMapHandler, DiffusionImageHandler, DiffusionTextHandler
 from .cache import VaeCache

{hcpdiff-2.1 → hcpdiff-2.2.1}/hcpdiff/data/handler/__init__.py

@@ -1,3 +1,3 @@
-from .diffusion import StableDiffusionHandler, DiffusionImageHandler, LossMapHandler
+from .diffusion import StableDiffusionHandler, DiffusionImageHandler, LossMapHandler, DiffusionTextHandler
 from .text import TokenizeHandler, TagEraseHandler, TagDropoutHandler, TagShuffleHandler, TemplateFillHandler
 from .controlnet import ControlNetHandler

{hcpdiff-2.1 → hcpdiff-2.2.1}/hcpdiff/data/handler/diffusion.py

@@ -49,14 +49,11 @@ class DiffusionImageHandler(DataHandler):
         else:
             return self.handlers(dict(image=image, image_size=image_size))
 
-class StableDiffusionHandler(DataHandler):
-    def __init__(self, bucket, encoder_attention_mask=False, key_map_in=('image -> image', 'image_size -> image_size', 'prompt -> prompt'),
-                 key_map_out=('image -> image', 'coord -> coord', 'prompt -> prompt'),
-                 erase=0.15, dropout=0.0, shuffle=0.0, word_names={}, tokenize=True):
+class DiffusionTextHandler(DataHandler):
+    def __init__(self, encoder_attention_mask=False, erase=0.0, dropout=0.0, shuffle=0.0, word_names={}, tokenize=True,
+                 key_map_in=('prompt -> prompt', ), key_map_out=('prompt -> prompt', )):
         super().__init__(key_map_in, key_map_out)
 
-        self.image_handlers = DiffusionImageHandler(bucket)
-
         text_handlers = {}
         if dropout>0:
             text_handlers['dropout'] = TagDropoutHandler(p=dropout)
@@ -67,7 +64,20 @@ class StableDiffusionHandler(DataHandler):
         text_handlers['fill'] = TemplateFillHandler(word_names)
         if tokenize:
             text_handlers['tokenize'] = TokenizeHandler(encoder_attention_mask)
-        self.text_handlers = HandlerChain(**text_handlers)
+        self.handlers = HandlerChain(**text_handlers)
+
+    def handle(self, prompt: Union[str, Dict[str, str]]):
+        return self.handlers(dict(prompt=prompt))
+
+class StableDiffusionHandler(DataHandler):
+    def __init__(self, bucket, encoder_attention_mask=False, key_map_in=('image -> image', 'image_size -> image_size', 'prompt -> prompt'),
+                 key_map_out=('image -> image', 'coord -> coord', 'prompt -> prompt'),
+                 erase=0.0, dropout=0.0, shuffle=0.0, word_names={}, tokenize=True):
+        super().__init__(key_map_in, key_map_out)
+
+        self.image_handlers = DiffusionImageHandler(bucket)
+        self.text_handlers = DiffusionTextHandler(encoder_attention_mask=encoder_attention_mask, erase=erase, dropout=dropout, shuffle=shuffle,
+                                                  word_names=word_names, tokenize=tokenize)
 
     def handle(self, image: Image.Image, image_size: np.ndarray[int], prompt: str):
         return dict(**self.image_handlers(dict(image=image, image_size=image_size)), **self.text_handlers(dict(prompt=prompt)))

{hcpdiff-2.1 → hcpdiff-2.2.1}/hcpdiff/data/source/__init__.py

@@ -1,3 +1,4 @@
 from .text2img import Text2ImageSource, Text2ImageLossMapSource
 from .text2img_cond import Text2ImageCondSource
-from .folder_class import T2IFolderClassSource
+from .folder_class import T2IFolderClassSource
+from .text import TextSource

hcpdiff-2.2.1/hcpdiff/data/source/text.py

@@ -0,0 +1,40 @@
+from rainbowneko.data import UnLabelSource, DataSource
+from rainbowneko.data.label_loader import BaseLabelLoader, auto_label_loader
+from typing import Union, Dict, Any
+import random
+
+class TextSource(DataSource):
+    def __init__(self, label_file, prompt_template=None, repeat=1, **kwargs):
+        super().__init__(repeat=repeat)
+        self.label_file = label_file
+        self.label_dict = self._load_label_data(label_file)
+        self.img_ids = self._load_img_ids(self.label_dict)
+        self.prompt_template = self.load_template(prompt_template)
+
+    def _load_img_ids(self, label_dict):
+        return list(label_dict.keys()) * self.repeat
+
+    def _load_label_data(self, label_file: Union[str, BaseLabelLoader]):
+        if label_file is None:
+            return {}
+        elif isinstance(label_file, str):
+            return auto_label_loader(label_file).load()
+        else:
+            return label_file.load()
+
+    def load_template(self, template_file):
+        if template_file is None:
+            return ['{caption}']
+        else:
+            with open(template_file, 'r', encoding='utf-8') as f:
+                return f.read().strip().split('\n')
+
+    def __getitem__(self, index) -> Dict[str, Any]:
+        img_name = self.img_ids[index]
+        return {
+            'id':img_name,
+            'prompt':{
+                'template':random.choice(self.prompt_template),
+                'caption':self.label_dict[img_name],
+            }
+        }

{hcpdiff-2.1 → hcpdiff-2.2.1}/hcpdiff/data/source/text2img.py

@@ -25,7 +25,7 @@ class Text2ImageSource(ImageLabelSource):
 
     def __getitem__(self, index) -> Dict[str, Any]:
         img_name = self.img_ids[index]
-        path = os.path.join(self.img_root, img_name)
+        path = self.img_root / img_name
 
         return {
             'id':img_name,

{hcpdiff-2.1 → hcpdiff-2.2.1}/hcpdiff/easy/cfg/__init__.py

@@ -1,3 +1,3 @@
 from .sd15_train import SD15_lora_train, cfg_data_SD_ARB, cfg_data_SD_resize_crop, SD15_finetuning
 from .sdxl_train import SDXL_lora_train, SDXL_finetuning
-from .t2i import SD15_t2i, SDXL_t2i, SDXL_t2i_lora, SD15_t2i_lora
+from .t2i import SD15_t2i, SDXL_t2i, SDXL_t2i_lora, SD15_t2i_lora, SDXL_t2i_parts, SD15_t2i_parts

{hcpdiff-2.1 → hcpdiff-2.2.1}/hcpdiff/easy/cfg/sd15_train.py

@@ -1,9 +1,10 @@
 import torch
-from rainbowneko.ckpt_manager import ckpt_saver, LAYERS_TRAINABLE, plugin_saver
+from rainbowneko.ckpt_manager import ckpt_saver, LAYERS_TRAINABLE, NekoPluginSaver, SafeTensorFormat
 from rainbowneko.data import RatioBucket, FixedBucket
 from rainbowneko.parser import CfgWDPluginParser, neko_cfg, CfgWDModelParser, disable_neko_cfg
 from rainbowneko.utils import ConstantLR, Path_Like
 
+from hcpdiff.ckpt_manager import LoraWebuiFormat
 from hcpdiff.data import TextImagePairDataset, Text2ImageSource, StableDiffusionHandler
 from hcpdiff.data import VaeCache
 from hcpdiff.easy import SD15_auto_loader
@@ -46,7 +47,7 @@ def SD15_finetuning(base_model: str, train_steps: int, dataset, save_step: int =
 
             optimizer=optimizer,
 
-            scheduler=ConstantLR(
+            lr_scheduler=ConstantLR(
                 _partial_=True,
                 warmup_steps=warmup_steps,
             ),
@@ -69,7 +70,7 @@ def SD15_finetuning(base_model: str, train_steps: int, dataset, save_step: int =
 @neko_cfg
 def SD15_lora_train(base_model: str, train_steps: int, dataset, save_step: int = 200, lr: float = 1e-4, rank: int = 4, alpha: float = None,
                     clip_skip: int = 0, with_conv: bool = False, dtype: str = 'fp16', low_vram: bool = False, warmup_steps: int = 0,
-                    name: str = 'SD15'):
+                    name: str = 'SD15', save_webui_format=False):
     with disable_neko_cfg:
         if alpha is None:
             alpha = rank
@@ -95,6 +96,11 @@ def SD15_lora_train(base_model: str, train_steps: int, dataset, save_step: int =
     else:
         optimizer = torch.optim.AdamW(_partial_=True, betas=(0.9, 0.99))
 
+    if save_webui_format:
+        lora_format = LoraWebuiFormat()
+    else:
+        lora_format = SafeTensorFormat()
+
     from cfgs.train.py.examples import SD_FT
 
     return dict(
@@ -114,8 +120,8 @@ def SD15_lora_train(base_model: str, train_steps: int, dataset, save_step: int =
 
         ckpt_saver=dict(
             _replace_ = True,
-            lora_unet=plugin_saver(
-                ckpt_type='safetensors',
+            lora_unet=NekoPluginSaver(
+                format=lora_format,
                 target_plugin='lora1',
             )
         ),
@@ -126,7 +132,7 @@ def SD15_lora_train(base_model: str, train_steps: int, dataset, save_step: int =
 
             optimizer=optimizer,
 
-            scheduler=ConstantLR(
+            lr_scheduler=ConstantLR(
                 _partial_=True,
                 warmup_steps=warmup_steps,
             ),

{hcpdiff-2.1 → hcpdiff-2.2.1}/hcpdiff/easy/cfg/sdxl_train.py

@@ -1,11 +1,12 @@
 import torch
-from rainbowneko.ckpt_manager import ckpt_saver, plugin_saver, LAYERS_TRAINABLE
+from rainbowneko.ckpt_manager import ckpt_saver, NekoPluginSaver, LAYERS_TRAINABLE, SafeTensorFormat
 from rainbowneko.parser import CfgWDPluginParser, neko_cfg, CfgWDModelParser, disable_neko_cfg
 from rainbowneko.utils import ConstantLR
 
 from hcpdiff.easy import SDXL_auto_loader
 from hcpdiff.models import SDXLWrapper
 from hcpdiff.models.lora_layers_patch import LoraLayer
+from hcpdiff.ckpt_manager import LoraWebuiFormat
 
 @neko_cfg
 def SDXL_finetuning(base_model: str, train_steps: int, dataset, save_step: int = 500, lr: float = 1e-5,
@@ -43,7 +44,7 @@ def SDXL_finetuning(base_model: str, train_steps: int, dataset, save_step: int =
 
             optimizer=optimizer,
 
-            scheduler=ConstantLR(
+            lr_scheduler=ConstantLR(
                 _partial_=True,
                 warmup_steps=warmup_steps,
             ),
@@ -64,7 +65,8 @@ def SDXL_finetuning(base_model: str, train_steps: int, dataset, save_step: int =
 
 @neko_cfg
 def SDXL_lora_train(base_model: str, train_steps: int, dataset, save_step: int = 200, lr: float = 1e-4, rank: int = 4, alpha: float = None,
-                    with_conv: bool = False, dtype: str = 'fp16', low_vram: bool = False, warmup_steps: int = 0, name: str = 'SD15'):
+                    with_conv: bool = False, dtype: str = 'fp16', low_vram: bool = False, warmup_steps: int = 0, name: str = 'SDXL',
+                    save_webui_format=False):
     with disable_neko_cfg:
         if alpha is None:
             alpha = rank
@@ -90,6 +92,11 @@ def SDXL_lora_train(base_model: str, train_steps: int, dataset, save_step: int =
     else:
         optimizer = torch.optim.AdamW(_partial_=True, betas=(0.9, 0.99))
 
+    if save_webui_format:
+        lora_format = LoraWebuiFormat()
+    else:
+        lora_format = SafeTensorFormat()
+
     from cfgs.train.py.examples import SD_FT
 
     return dict(
@@ -109,8 +116,8 @@ def SDXL_lora_train(base_model: str, train_steps: int, dataset, save_step: int =
 
         ckpt_saver=dict(
             _replace_ = True,
-            lora_unet=plugin_saver(
-                ckpt_type='safetensors',
+            lora_unet=NekoPluginSaver(
+                format=lora_format,
                 target_plugin='lora1',
             )
         ),
@@ -121,7 +128,7 @@ def SDXL_lora_train(base_model: str, train_steps: int, dataset, save_step: int =
 
             optimizer=optimizer,
 
-            scheduler=ConstantLR(
+            lr_scheduler=ConstantLR(
                 _partial_=True,
                 warmup_steps=warmup_steps,
             ),

{hcpdiff-2.1 → hcpdiff-2.2.1}/hcpdiff/easy/cfg/t2i.py

@@ -1,5 +1,6 @@
 import torch
 from rainbowneko.infer.workflow import (Actions, PrepareAction, LoopAction, LoadModelAction)
+from rainbowneko.ckpt_manager import NekoModelLoader
 from rainbowneko.parser import neko_cfg, disable_neko_cfg
 from typing import Union, List
 
@@ -25,6 +26,29 @@ def build_model(pretrained_model='ckpts/any5', noise_sampler=Diffusers_SD.dpmpp_
         ),
     ])
 
+@neko_cfg
+def load_parts(info: List[str]) -> Actions:
+    acts = []
+    for i, path in enumerate(info):
+        part_unet = LoadModelAction(cfg={
+            f'part_unet_{i}':NekoModelLoader(
+                path=path,
+                state_prefix='denoiser.'
+            )
+        }, key_map_in=('denoiser -> model', 'in_preview -> in_preview'))
+        part_TE = LoadModelAction(cfg={
+            f'part_TE_{i}':NekoModelLoader(
+                path=path,
+                state_prefix='TE.',
+            )
+        }, key_map_in=('TE -> model', 'in_preview -> in_preview'))
+
+        with disable_neko_cfg:
+            acts.append(part_unet)
+            acts.append(part_TE)
+
+    return Actions(acts)
+
 @neko_cfg
 def load_lora(info: List[List]) -> Actions:
     lora_acts = []
@@ -37,7 +61,7 @@ def load_lora(info: List[List]) -> Actions:
             )
         }, key_map_in=('denoiser -> model', 'in_preview -> in_preview'))
         lora_TE = LoadModelAction(cfg={
-            f'lora_unet_{i}':HCPLoraLoader(
+            f'lora_TE_{i}':HCPLoraLoader(
                 path=item[0],
                 state_prefix='TE.',
                 alpha=item[1],
@@ -59,9 +83,9 @@ def optimize_model() -> Actions:
     ])
 
 @neko_cfg
-def text(prompt, negative_prompt=negative_prompt, bs=4) -> Actions:
+def text(prompt, negative_prompt=negative_prompt, bs=4, N_repeats=1, layer_skip=1) -> Actions:
     return Actions([
-        TextHookAction(N_repeats=1, layer_skip=1),
+        TextHookAction(N_repeats=N_repeats, layer_skip=layer_skip),
         AttnMultTextEncodeAction(
             prompt=prompt,
             negative_prompt=negative_prompt,
@@ -84,9 +108,9 @@ def build_model_SDXL(pretrained_model='ckpts/any5', noise_sampler=Diffusers_SD.d
     ])
 
 @neko_cfg
-def text_SDXL(prompt, negative_prompt=negative_prompt, bs=4) -> Actions:
+def text_SDXL(prompt, negative_prompt=negative_prompt, bs=4, N_repeats=1, layer_skip=1) -> Actions:
     return Actions([
-        TextHookAction(N_repeats=1, layer_skip=1, TE_final_norm=False),
+        TextHookAction(N_repeats=N_repeats, layer_skip=layer_skip, TE_final_norm=False),
         AttnMultTextEncodeAction(
             prompt=prompt,
             negative_prompt=negative_prompt,
@@ -128,11 +152,24 @@ def resize(width=1024, height=1024):
 
 @neko_cfg
 def SD15_t2i(pretrained_model, prompt, negative_prompt=negative_prompt, noise_sampler=Diffusers_SD.dpmpp_2m_karras, bs=4, width=512, height=512,
-             seed=None, N_steps=20, guidance_scale=7.0, save_root='output_pipe/'):
+             seed=None, N_steps=20, guidance_scale=7.0, save_root='output_pipe/', N_repeats=1, layer_skip=1):
+    return dict(workflow=Actions(actions=[
+        build_model(pretrained_model=pretrained_model, noise_sampler=noise_sampler),
+        optimize_model(),
+        text(prompt=prompt, negative_prompt=negative_prompt, bs=bs, N_repeats=N_repeats, layer_skip=layer_skip),
+        config_diffusion(width=width, height=height, seed=seed, N_steps=N_steps),
+        diffusion(guidance_scale=guidance_scale),
+        decode(save_root=save_root)
+    ]))
+
+@neko_cfg
+def SD15_t2i_parts(pretrained_model, parts, prompt, negative_prompt=negative_prompt, noise_sampler=Diffusers_SD.dpmpp_2m_karras, bs=4, width=512, height=512,
+             seed=None, N_steps=20, guidance_scale=7.0, save_root='output_pipe/', N_repeats=1, layer_skip=1):
     return dict(workflow=Actions(actions=[
         build_model(pretrained_model=pretrained_model, noise_sampler=noise_sampler),
+        load_parts(parts),
         optimize_model(),
-        text(prompt=prompt, negative_prompt=negative_prompt, bs=bs),
+        text(prompt=prompt, negative_prompt=negative_prompt, bs=bs, N_repeats=N_repeats, layer_skip=layer_skip),
         config_diffusion(width=width, height=height, seed=seed, N_steps=N_steps),
         diffusion(guidance_scale=guidance_scale),
         decode(save_root=save_root)
@@ -140,12 +177,12 @@ def SD15_t2i(pretrained_model, prompt, negative_prompt=negative_prompt, noise_sa
 
 @neko_cfg
 def SD15_t2i_lora(pretrained_model, lora_info, prompt, negative_prompt=negative_prompt, noise_sampler=Diffusers_SD.dpmpp_2m_karras, bs=4,
-                  width=512, height=512, seed=None, N_steps=20, guidance_scale=7.0, save_root='output_pipe/'):
+                  width=512, height=512, seed=None, N_steps=20, guidance_scale=7.0, save_root='output_pipe/', N_repeats=1, layer_skip=1):
     return dict(workflow=Actions(actions=[
         build_model(pretrained_model=pretrained_model, noise_sampler=noise_sampler),
         load_lora(info=lora_info),
         optimize_model(),
-        text(prompt=prompt, negative_prompt=negative_prompt, bs=bs),
+        text(prompt=prompt, negative_prompt=negative_prompt, bs=bs, N_repeats=N_repeats, layer_skip=layer_skip),
         config_diffusion(width=width, height=height, seed=seed, N_steps=N_steps),
         diffusion(guidance_scale=guidance_scale),
         decode(save_root=save_root)
@@ -153,24 +190,38 @@ def SD15_t2i_lora(pretrained_model, lora_info, prompt, negative_prompt=negative_
 
 @neko_cfg
 def SDXL_t2i(pretrained_model, prompt, negative_prompt=negative_prompt, noise_sampler=Diffusers_SD.dpmpp_2m_karras, bs=4, width=1024, height=1024,
-             seed=None, N_steps=20, guidance_scale=7.0, save_root='output_pipe/'):
+             seed=None, N_steps=20, guidance_scale=7.0, save_root='output_pipe/', N_repeats=1, layer_skip=1):
     return dict(workflow=Actions(actions=[
         build_model_SDXL(pretrained_model=pretrained_model, noise_sampler=noise_sampler),
         optimize_model(),
-        text_SDXL(prompt=prompt, negative_prompt=negative_prompt, bs=bs),
+        text_SDXL(prompt=prompt, negative_prompt=negative_prompt, bs=bs, N_repeats=N_repeats, layer_skip=layer_skip),
         config_diffusion(width=width, height=height, seed=seed, N_steps=N_steps),
         diffusion(guidance_scale=guidance_scale),
         decode(save_root=save_root)
     ]))
 
+@neko_cfg
+def SDXL_t2i_parts(pretrained_model, parts, prompt, negative_prompt=negative_prompt, noise_sampler=Diffusers_SD.dpmpp_2m_karras, bs=4, width=1024, height=1024,
+             seed=None, N_steps=20, guidance_scale=7.0, save_root='output_pipe/', N_repeats=1, layer_skip=1):
+    return dict(workflow=Actions(actions=[
+        build_model_SDXL(pretrained_model=pretrained_model, noise_sampler=noise_sampler),
+        load_parts(parts),
+        optimize_model(),
+        text_SDXL(prompt=prompt, negative_prompt=negative_prompt, bs=bs, N_repeats=N_repeats, layer_skip=layer_skip),
+        config_diffusion(width=width, height=height, seed=seed, N_steps=N_steps),
+        diffusion(guidance_scale=guidance_scale),
+        decode(save_root=save_root)
+    ]))
+
+
 @neko_cfg
 def SDXL_t2i_lora(pretrained_model, lora_info, prompt, negative_prompt=negative_prompt, noise_sampler=Diffusers_SD.dpmpp_2m_karras, bs=4,
-                  width=1024, height=1024, seed=None, N_steps=20, guidance_scale=7.0, save_root='output_pipe/'):
+                  width=1024, height=1024, seed=None, N_steps=20, guidance_scale=7.0, save_root='output_pipe/', N_repeats=1, layer_skip=1):
     return dict(workflow=Actions(actions=[
         build_model_SDXL(pretrained_model=pretrained_model, noise_sampler=noise_sampler),
         load_lora(info=lora_info),
         optimize_model(),
-        text_SDXL(prompt=prompt, negative_prompt=negative_prompt, bs=bs),
+        text_SDXL(prompt=prompt, negative_prompt=negative_prompt, bs=bs, N_repeats=N_repeats, layer_skip=layer_skip),
         config_diffusion(width=width, height=height, seed=seed, N_steps=N_steps),
         diffusion(guidance_scale=guidance_scale),
         decode(save_root=save_root)

{hcpdiff-2.1 → hcpdiff-2.2.1}/hcpdiff/models/text_emb_ex.py

@@ -126,6 +126,10 @@ class EmbeddingPTInterpHook(SinglePluginBlock):
         BOS = repeat(inputs_embeds[0,0,:], 'e -> r 1 e', r=self.N_repeats)
         EOS = repeat(inputs_embeds[0,-1,:], 'e -> r 1 e', r=self.N_repeats)
 
+        # make DDP happy
+        if len(self.emb_train) > 0:
+            BOS = BOS + sum(emb.mean()*0 for emb in self.emb_train if emb.requires_grad)
+
         replaced_embeds = []
         for item, rep_idxs, ids_raw in zip(inputs_embeds, rep_idxs_B, self.input_ids):
             # insert pt to embeddings

{hcpdiff-2.1 → hcpdiff-2.2.1}/hcpdiff/trainer_ac.py

@@ -42,13 +42,6 @@ class HCPTrainer(Trainer):
     def pt_trainable(self):
         return self.cfgs.emb_pt is not None
 
-    def get_loss(self, ds_name, model_pred, inputs):
-        loss = super().get_loss(ds_name, model_pred, inputs)
-        # make DDP happy
-        if len(self.train_pts)>0:
-            loss = loss+0*sum([emb.mean() for emb in self.train_pts.values()])
-        return loss
-
     def save_model(self, from_raw=False):
         NekoSaver.save_all(
             self.model_raw,

hcpdiff-2.2.1/hcpdiff/trainer_deepspeed.py

@@ -0,0 +1,47 @@
+import argparse
+import warnings
+
+import torch
+from rainbowneko.ckpt_manager import NekoPluginSaver
+from rainbowneko.train.trainer import TrainerDeepspeed
+from rainbowneko.utils import xformers_available
+
+from hcpdiff.trainer_ac import HCPTrainer, load_config_with_cli
+
+class HCPTrainerDeepspeed(TrainerDeepspeed, HCPTrainer):
+    def config_model(self):
+        if self.cfgs.model.enable_xformers:
+            if xformers_available:
+                self.model_wrapper.enable_xformers()
+            else:
+                warnings.warn("xformers is not available. Make sure it is installed correctly")
+
+        if self.model_wrapper.vae is not None:
+            self.vae_dtype = self.weight_dtype_map.get(self.cfgs.model.get('vae_dtype', None), torch.float32)
+            self.model_wrapper.set_dtype(self.weight_dtype, self.vae_dtype)
+
+        if self.cfgs.model.gradient_checkpointing:
+            self.model_wrapper.enable_gradient_checkpointing()
+
+        if self.is_local_main_process:
+            for saver in self.ckpt_saver.values():
+                if isinstance(saver, NekoPluginSaver):
+                    saver.plugin_from_raw = True
+
+def hcp_train():
+    import subprocess
+    parser = argparse.ArgumentParser(description='HCP-Diffusion Launcher')
+    parser.add_argument('--launch_cfg', type=str, default='cfgs/launcher/deepspeed.yaml')
+    args, train_args = parser.parse_known_args()
+
+    subprocess.run(["accelerate", "launch", '--config_file', args.launch_cfg, "-m",
+                       "hcpdiff.trainer_deepspeed"]+train_args, check=True)
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='HCP Diffusion Trainer for DeepSpeed')
+    parser.add_argument("--cfg", type=str, default=None, required=True)
+    args, cfg_args = parser.parse_known_args()
+
+    parser, conf = load_config_with_cli(args.cfg, args_list=cfg_args)  # skip --cfg
+    trainer = HCPTrainerDeepspeed(parser, conf)
+    trainer.train()

{hcpdiff-2.1 → hcpdiff-2.2.1}/hcpdiff/workflow/diffusion.py

@@ -32,14 +32,15 @@ class SeedAction(BasicAction):
         self.seed = seed
         self.bs = bs
 
-    def forward(self, device, gen_step=0, **states):
+    def forward(self, device, seed=None, **states):
         bs = states['prompt_embeds'].shape[0]//2 if 'prompt_embeds' in states else self.bs
-        if self.seed is None:
+        seed = seed or self.seed
+        if seed is None:
             seeds = [None]*bs
-        elif isinstance(self.seed, int):
-            seeds = list(range(self.seed+gen_step*bs, self.seed+(gen_step+1)*bs))
+        elif isinstance(seed, int):
+            seeds = list(range(seed, seed+bs))
         else:
-            seeds = self.seed
+            seeds = seed
         seeds = [s or random.randint(0, 1 << 30) for s in seeds]
 
         G = prepare_seed(seeds, device=device)