Merge fad68161c9 into 51435f1718

Fix sd 2.0/2.1 lora fine-tuning text encoder params mismatch issue.
With the original text encoder and its corresponding parser, stable diffusion 2.0/2.1/2.1-unclip all cannot be properly loaded due to the text encoder transformer version difference and checkpoints' state_dict key name difference.
After this fix, all these three versions are tested to work well when executing lora fine tuning.

README-ja.md

@@ -50,9 +50,6 @@ Stable Diffusion等の画像生成モデルの学習、モデルによる画像
 
 ### 更新履歴
 
-- 次のリリースに含まれる予定の主な変更点は以下の通りです。リリース前の変更点は予告なく変更される可能性があります。
-    - Intel GPUの互換性を向上しました。[PR #2307](https://github.com/kohya-ss/sd-scripts/pull/2307) WhitePr氏に感謝します。
-
 - **Version 0.10.3 (2026-04-02):**
     - Animaでfp16で学習する際の安定性をさらに改善しました。[PR #2302](https://github.com/kohya-ss/sd-scripts/pull/2302) 問題をご報告いただいた方々に深く感謝します。
 

README.md

@@ -47,9 +47,6 @@ If you find this project helpful, please consider supporting its development via
 
 ### Change History
 
-- The following are the main changes planned for the next release. Please note that these changes may be subject to change without notice before the release.
-    - Improved compatibility with Intel GPUs. Thanks to WhitePr for [PR #2307](https://github.com/kohya-ss/sd-scripts/pull/2307).
-
 - **Version 0.10.3 (2026-04-02):**
     - Stability when training with fp16 on Anima has been further improved. See [PR #2302](https://github.com/kohya-ss/sd-scripts/pull/2302) for details. We deeply appreciate those who reported the issue.
 

library/ipex/__init__.py

@@ -1,7 +1,6 @@
 import os
 import sys
 import torch
-from packaging import version
 try:
     import intel_extension_for_pytorch as ipex # pylint: disable=import-error, unused-import
     has_ipex = True
@@ -9,7 +8,7 @@ except Exception:
     has_ipex = False
 from .hijacks import ipex_hijacks
 
-torch_version = version.parse(torch.__version__)
+torch_version = float(torch.__version__[:3])
 
 # pylint: disable=protected-access, missing-function-docstring, line-too-long
 
@@ -57,6 +56,7 @@ def ipex_init(): # pylint: disable=too-many-statements
             torch.cuda.__path__ = torch.xpu.__path__
             torch.cuda.set_stream = torch.xpu.set_stream
             torch.cuda.torch = torch.xpu.torch
+            torch.cuda.Union = torch.xpu.Union
             torch.cuda.__annotations__ = torch.xpu.__annotations__
             torch.cuda.__package__ = torch.xpu.__package__
             torch.cuda.__builtins__ = torch.xpu.__builtins__
@@ -64,12 +64,14 @@ def ipex_init(): # pylint: disable=too-many-statements
             torch.cuda.StreamContext = torch.xpu.StreamContext
             torch.cuda._lazy_call = torch.xpu._lazy_call
             torch.cuda.random = torch.xpu.random
+            torch.cuda._device = torch.xpu._device
             torch.cuda.__name__ = torch.xpu.__name__
+            torch.cuda._device_t = torch.xpu._device_t
             torch.cuda.__spec__ = torch.xpu.__spec__
             torch.cuda.__file__ = torch.xpu.__file__
             # torch.cuda.is_current_stream_capturing = torch.xpu.is_current_stream_capturing
 
-            if torch_version < version.parse("2.3"):
+            if torch_version < 2.3:
                 torch.cuda._initialization_lock = torch.xpu.lazy_init._initialization_lock
                 torch.cuda._initialized = torch.xpu.lazy_init._initialized
                 torch.cuda._is_in_bad_fork = torch.xpu.lazy_init._is_in_bad_fork
@@ -112,22 +114,17 @@ def ipex_init(): # pylint: disable=too-many-statements
                 torch.cuda.threading = torch.xpu.threading
                 torch.cuda.traceback = torch.xpu.traceback
 
-            if torch_version < version.parse("2.5"):
+            if torch_version < 2.5:
                 torch.cuda.os = torch.xpu.os
                 torch.cuda.Device = torch.xpu.Device
                 torch.cuda.warnings = torch.xpu.warnings
                 torch.cuda.classproperty = torch.xpu.classproperty
                 torch.UntypedStorage.cuda = torch.UntypedStorage.xpu
 
-            if torch_version < version.parse("2.7"):
+            if torch_version < 2.7:
                 torch.cuda.Tuple = torch.xpu.Tuple
                 torch.cuda.List = torch.xpu.List
 
-            if torch_version < version.parse("2.11"):
-                torch.cuda._device_t = torch.xpu._device_t
-                torch.cuda._device = torch.xpu._device
-                torch.cuda.Union = torch.xpu.Union
-
 
             # Memory:
             if 'linux' in sys.platform and "WSL2" in os.popen("uname -a").read():
@@ -163,7 +160,7 @@ def ipex_init(): # pylint: disable=too-many-statements
             torch.cuda.initial_seed = torch.xpu.initial_seed
 
             # C
-            if torch_version < version.parse("2.3"):
+            if torch_version < 2.3:
                 torch._C._cuda_getCurrentRawStream = ipex._C._getCurrentRawStream
                 ipex._C._DeviceProperties.multi_processor_count = ipex._C._DeviceProperties.gpu_subslice_count
                 ipex._C._DeviceProperties.major = 12

library/model_util.py

@@ -11,7 +11,7 @@ init_ipex()
 
 import diffusers
 from transformers import CLIPTextModel, CLIPTokenizer, CLIPTextConfig, logging
-from diffusers import AutoencoderKL, DDIMScheduler, StableDiffusionPipeline  # , UNet2DConditionModel
+from diffusers import AutoencoderKL, DDIMScheduler, StableDiffusionPipeline, StableUnCLIPImg2ImgPipeline  # , UNet2DConditionModel
 from safetensors.torch import load_file, save_file
 from library.original_unet import UNet2DConditionModel
 from library.utils import setup_logging
@@ -658,6 +658,77 @@ def convert_ldm_clip_checkpoint_v2(checkpoint, max_length):
     return new_sd
 
 
+def convert_ldm_clip_checkpoint_v2_fix(checkpoint, max_length):
+    # 嫌になるくらい違うぞ！
+    def convert_key(key):
+        if not key.startswith("cond_stage_model"):
+            return None
+
+        # common conversion
+        key = key.replace("cond_stage_model.model.transformer.", "text_model.encoder.")
+        key = key.replace("cond_stage_model.model.", "text_model.")
+
+        if "resblocks" in key:
+            # resblocks conversion
+            key = key.replace(".resblocks.", ".layers.")
+            if ".ln_" in key:
+                key = key.replace(".ln_", ".layer_norm")
+            elif ".mlp." in key:
+                key = key.replace(".c_fc.", ".fc1.")
+                key = key.replace(".c_proj.", ".fc2.")
+            elif ".attn.out_proj" in key:
+                key = key.replace(".attn.out_proj.", ".self_attn.out_proj.")
+            elif ".attn.in_proj" in key:
+                key = None  # 特殊なので後で処理する
+            else:
+                raise ValueError(f"unexpected key in SD: {key}")
+        elif ".positional_embedding" in key:
+            key = key.replace(".positional_embedding", ".embeddings.position_embedding.weight")
+        elif ".text_projection" in key:
+            key = None  # 使われない???
+        elif ".logit_scale" in key:
+            key = None  # 使われない???
+        elif ".token_embedding" in key:
+            key = key.replace(".token_embedding.weight", ".embeddings.token_embedding.weight")
+        elif ".ln_final" in key:
+            key = key.replace(".ln_final", ".final_layer_norm")
+        return key
+
+    keys = list(checkpoint.keys())
+    new_sd = {}
+    for key in keys:
+        # remove resblocks 23
+        if ".resblocks.23." in key:
+            continue
+        if 'embedder.model' in key:
+            continue
+        new_key = convert_key(key)
+        if new_key is None:
+            continue
+        new_sd[new_key] = checkpoint[key]
+
+    # attnの変換
+    for key in keys:
+        if ".resblocks.23." in key:
+            continue
+        if 'embedder.model' in key:
+            continue
+        if ".resblocks" in key and ".attn.in_proj_" in key:
+            # 三つに分割
+            values = torch.chunk(checkpoint[key], 3)
+
+            key_suffix = ".weight" if "weight" in key else ".bias"
+            key_pfx = key.replace("cond_stage_model.model.transformer.resblocks.", "text_model.encoder.layers.")
+            key_pfx = key_pfx.replace("_weight", "")
+            key_pfx = key_pfx.replace("_bias", "")
+            key_pfx = key_pfx.replace(".attn.in_proj", ".self_attn.")
+            new_sd[key_pfx + "q_proj" + key_suffix] = values[0]
+            new_sd[key_pfx + "k_proj" + key_suffix] = values[1]
+            new_sd[key_pfx + "v_proj" + key_suffix] = values[2]
+
+    return new_sd
+
+
 # endregion
 
 
@@ -1017,33 +1088,58 @@ def load_models_from_stable_diffusion_checkpoint(v2, ckpt_path, device="cpu", dt
     vae = AutoencoderKL(**vae_config).to(device)
     info = vae.load_state_dict(converted_vae_checkpoint)
     logger.info(f"loading vae: {info}")
-
+    
-    # convert text_model
     if v2:
-        converted_text_encoder_checkpoint = convert_ldm_clip_checkpoint_v2(state_dict, 77)
+        try:
-        cfg = CLIPTextConfig(
+            converted_text_encoder_checkpoint = convert_ldm_clip_checkpoint_v2_fix(state_dict, 77)
-            vocab_size=49408,
+            cfg = CLIPTextConfig(
-            hidden_size=1024,
+                attention_dropout = 0.0,
-            intermediate_size=4096,
+                bos_token_id = 0,
-            num_hidden_layers=23,
+                dropout = 0.0,
-            num_attention_heads=16,
+                eos_token_id = 2,
-            max_position_embeddings=77,
+                hidden_act = "gelu",
-            hidden_act="gelu",
+                hidden_size = 1024,
-            layer_norm_eps=1e-05,
+                initializer_factor = 1.0,
-            dropout=0.0,
+                initializer_range = 0.02,
-            attention_dropout=0.0,
+                intermediate_size = 4096,
-            initializer_range=0.02,
+                layer_norm_eps = 1e-05,
-            initializer_factor=1.0,
+                max_position_embeddings = 77,
-            pad_token_id=1,
+                model_type = "clip_text_model",
-            bos_token_id=0,
+                num_attention_heads = 16,
-            eos_token_id=2,
+                num_hidden_layers = 23,
-            model_type="clip_text_model",
+                pad_token_id = 1,
-            projection_dim=512,
+                projection_dim = 512,
-            torch_dtype="float32",
+                torch_dtype = "float16",
-            transformers_version="4.25.0.dev0",
+                transformers_version = "4.28.0.dev0",
-        )
+                vocab_size = 49408
-        text_model = CLIPTextModel._from_config(cfg)
+            )
-        info = text_model.load_state_dict(converted_text_encoder_checkpoint)
+            text_model = CLIPTextModel._from_config(cfg)
+            info = text_model.load_state_dict(converted_text_encoder_checkpoint)
+        except Exception as e:
+            converted_text_encoder_checkpoint = convert_ldm_clip_checkpoint_v2(state_dict, 77)
+            cfg = CLIPTextConfig(
+                vocab_size=49408,
+                hidden_size=1024,
+                intermediate_size=4096,
+                num_hidden_layers=23,
+                num_attention_heads=16,
+                max_position_embeddings=77,
+                hidden_act="gelu",
+                layer_norm_eps=1e-05,
+                dropout=0.0,
+                attention_dropout=0.0,
+                initializer_range=0.02,
+                initializer_factor=1.0,
+                pad_token_id=1,
+                bos_token_id=0,
+                eos_token_id=2,
+                model_type="clip_text_model",
+                projection_dim=512,
+                torch_dtype="float32",
+                transformers_version="4.25.0.dev0",
+            )
+            text_model = CLIPTextModel._from_config(cfg)
+            info = text_model.load_state_dict(converted_text_encoder_checkpoint)     
     else:
         converted_text_encoder_checkpoint = convert_ldm_clip_checkpoint_v1(state_dict)
 
@@ -1077,6 +1173,25 @@ def load_models_from_stable_diffusion_checkpoint(v2, ckpt_path, device="cpu", dt
 
     return text_model, vae, unet
 
+# def load_models_from_stable_diffusion_checkpoint(v2, ckpt_path, device="cpu", dtype=torch.float32):
+#     pipe = StableUnCLIPImg2ImgPipeline.from_pretrained(ckpt_path, torch_dtype=torch.float32).to(device)
+    
+#     # Load the UNet model
+#     unet = pipe.unet.to(device)
+
+#     # Load the VAE model
+#     vae = pipe.vae.to(device)
+
+#     # Load the text model
+#     text_encoder = pipe.text_encoder.to(device)
+
+#     # Log information
+#     logger.info(f"Loaded UNet: {unet}")
+#     logger.info(f"Loaded VAE: {vae}")
+#     logger.info(f"Loaded Text Encoder: {text_encoder}")
+
+#     return text_encoder, vae, unet
+
 
 def get_model_version_str_for_sd1_sd2(v2, v_parameterization):
     # only for reference