diffsynth-engine 0.6.1.dev20__py3-none-any.whl → 0.6.1.dev22__py3-none-any.whl

diffsynth_engine/configs/pipeline.py

@@ -242,6 +242,8 @@ class QwenImagePipelineConfig(AttentionConfig, OptimizationConfig, ParallelConfi
     vae_tile_size: Tuple[int, int] = (34, 34)
     vae_tile_stride: Tuple[int, int] = (18, 16)
 
+    load_encoder: bool = True
+
     @classmethod
     def basic_config(
         cls,

diffsynth_engine/pipelines/flux_image.py

@@ -830,7 +830,7 @@ class FluxImagePipeline(BasePipeline):
                 masked_image = image.clone()
                 masked_image[(mask > 0.5).repeat(1, 3, 1, 1)] = -1
                 latent = self.encode_image(masked_image)
-                mask = torch.nn.functional.interpolate(mask, size=(latent.shape[2], latent.shape[3]))
+                mask = torch.nn.functional.interpolate(mask, size=(latent.shape[2], latent.shape[3])).to(latent.dtype)
                 mask = 1 - mask
                 latent = torch.cat([latent, mask], dim=1)
             elif self.config.control_type == ControlType.bfl_fill:

diffsynth_engine/pipelines/qwen_image.py

@@ -186,6 +186,7 @@ class QwenImagePipeline(BasePipeline):
         logger.info(f"loading state dict from {config.vae_path} ...")
         vae_state_dict = cls.load_model_checkpoint(config.vae_path, device="cpu", dtype=config.vae_dtype)
 
+        encoder_state_dict = None
         if config.encoder_path is None:
             config.encoder_path = fetch_model(
                 "MusePublic/Qwen-image",
@@ -197,8 +198,9 @@ class QwenImagePipeline(BasePipeline):
                     "text_encoder/model-00004-of-00004.safetensors",
                 ],
             )
-        logger.info(f"loading state dict from {config.encoder_path} ...")
-        encoder_state_dict = cls.load_model_checkpoint(config.encoder_path, device="cpu", dtype=config.encoder_dtype)
+        if config.load_encoder:
+            logger.info(f"loading state dict from {config.encoder_path} ...")
+            encoder_state_dict = cls.load_model_checkpoint(config.encoder_path, device="cpu", dtype=config.encoder_dtype)
 
         state_dicts = QwenImageStateDicts(
             model=model_state_dict,
@@ -225,22 +227,25 @@ class QwenImagePipeline(BasePipeline):
     @classmethod
     def _from_state_dict(cls, state_dicts: QwenImageStateDicts, config: QwenImagePipelineConfig) -> "QwenImagePipeline":
         init_device = "cpu" if config.offload_mode is not None else config.device
-        tokenizer = Qwen2TokenizerFast.from_pretrained(QWEN_IMAGE_TOKENIZER_CONF_PATH)
-        processor = Qwen2VLProcessor.from_pretrained(
-            tokenizer_config_path=QWEN_IMAGE_TOKENIZER_CONF_PATH,
-            image_processor_config_path=QWEN_IMAGE_PROCESSOR_CONFIG_FILE,
-        )
-        with open(QWEN_IMAGE_VISION_CONFIG_FILE, "r", encoding="utf-8") as f:
-            vision_config = Qwen2_5_VLVisionConfig(**json.load(f))
-        with open(QWEN_IMAGE_CONFIG_FILE, "r", encoding="utf-8") as f:
-            text_config = Qwen2_5_VLConfig(**json.load(f))
-        encoder = Qwen2_5_VLForConditionalGeneration.from_state_dict(
-            state_dicts.encoder,
-            vision_config=vision_config,
-            config=text_config,
-            device=("cpu" if config.use_fsdp else init_device),
-            dtype=config.encoder_dtype,
-        )
+        tokenizer, processor, encoder = None, None, None
+        if config.load_encoder:
+            tokenizer = Qwen2TokenizerFast.from_pretrained(QWEN_IMAGE_TOKENIZER_CONF_PATH)
+            processor = Qwen2VLProcessor.from_pretrained(
+                tokenizer_config_path=QWEN_IMAGE_TOKENIZER_CONF_PATH,
+                image_processor_config_path=QWEN_IMAGE_PROCESSOR_CONFIG_FILE,
+            )
+            with open(QWEN_IMAGE_VISION_CONFIG_FILE, "r", encoding="utf-8") as f:
+                vision_config = Qwen2_5_VLVisionConfig(**json.load(f))
+            with open(QWEN_IMAGE_CONFIG_FILE, "r", encoding="utf-8") as f:
+                text_config = Qwen2_5_VLConfig(**json.load(f))
+            encoder = Qwen2_5_VLForConditionalGeneration.from_state_dict(
+                state_dicts.encoder,
+                vision_config=vision_config,
+                config=text_config,
+                device=("cpu" if config.use_fsdp else init_device),
+                dtype=config.encoder_dtype,
+            )
+
         with open(QWEN_IMAGE_VAE_CONFIG_FILE, "r", encoding="utf-8") as f:
             vae_config = json.load(f)
         vae = QwenImageVAE.from_state_dict(

{diffsynth_engine-0.6.1.dev20.dist-info → diffsynth_engine-0.6.1.dev22.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: diffsynth_engine
-Version: 0.6.1.dev20
+Version: 0.6.1.dev22
 Author: MuseAI x ModelScope
 Classifier: Programming Language :: Python :: 3
 Classifier: Operating System :: OS Independent

{diffsynth_engine-0.6.1.dev20.dist-info → diffsynth_engine-0.6.1.dev22.dist-info}/RECORD

@@ -80,7 +80,7 @@ diffsynth_engine/conf/tokenizers/wan/umt5-xxl/tokenizer.json,sha256=bhl7TT29cdoU
 diffsynth_engine/conf/tokenizers/wan/umt5-xxl/tokenizer_config.json,sha256=7Zo6iw-qcacKMoR-BDX-A25uES1N9O23u0ipIeNE3AU,61728
 diffsynth_engine/configs/__init__.py,sha256=f6Y-j_ZQs7bM4Lr7Mh9CXFEBrSNLc9k5GJyJqjLAGiY,1187
 diffsynth_engine/configs/controlnet.py,sha256=f3vclyP3lcAjxDGD9C1vevhqqQ7W2LL_c6Wye0uxk3Q,1180
-diffsynth_engine/configs/pipeline.py,sha256=u4P0JnzSsvS_tfbTYyUARdT88k7TEGRYNqjaAPZlY40,13223
+diffsynth_engine/configs/pipeline.py,sha256=FwHIvj2VdxtmiHxOUYoAzs5YVBprxobhV9AJ2CFrV4c,13254
 diffsynth_engine/kernels/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 diffsynth_engine/models/__init__.py,sha256=8Ze7cSE8InetgXWTNb0neVA2Q44K7WlE-h7O-02m2sY,119
 diffsynth_engine/models/base.py,sha256=BA5vgMqfy_cjuL2OtXbrFD-Qg5xQnaumHpj5TabwSy8,2559
@@ -140,9 +140,9 @@ diffsynth_engine/models/wan/wan_text_encoder.py,sha256=OERlmwOqthAFPNnnT2sXJ4Ojy
 diffsynth_engine/models/wan/wan_vae.py,sha256=dC7MoUFeXRL7SIY0LG1OOUiZW-pp9IbXCghutMxpXr4,38889
 diffsynth_engine/pipelines/__init__.py,sha256=jh-4LSJ0vqlXiT8BgFgRIQxuAr2atEPyHrxXWj-Ud1U,604
 diffsynth_engine/pipelines/base.py,sha256=BWW7LW0E2qwu8G-6bP3nmeO7VCQxC8srOo8tE4aKA4o,14993
-diffsynth_engine/pipelines/flux_image.py,sha256=Dpy8AkwywuLAhvJ6cjg5TgzhSUgFQtv6p2JTTkzUHbo,50919
+diffsynth_engine/pipelines/flux_image.py,sha256=vJKvnYmeeQVX2O1Zjtm4NLrltBp66VSZ-KjAUqJ8zJ8,50936
 diffsynth_engine/pipelines/hunyuan3d_shape.py,sha256=TNV0Wr09Dj2bzzlpua9WioCClOj3YiLfE6utI9aWL8A,8164
-diffsynth_engine/pipelines/qwen_image.py,sha256=jt4rg-U5qWsFD0kUeDwKzgIiTAC80Cj8aq1YQOR1_-k,33052
+diffsynth_engine/pipelines/qwen_image.py,sha256=rksB8tiAEp9TIcLLca269dNFQRPIDxffThKRMuR06A0,33280
 diffsynth_engine/pipelines/sd_image.py,sha256=nr-Nhsnomq8CsUqhTM3i2l2zG01YjwXdfRXgr_bC3F0,17891
 diffsynth_engine/pipelines/sdxl_image.py,sha256=v7ZACGPb6EcBunL6e5E9jynSQjE7GQx8etEV-ZLP91g,21704
 diffsynth_engine/pipelines/utils.py,sha256=lk7sFGEk-fGjgadLpwwppHKG-yZ0RC-4ZmHW7pRRe8A,473
@@ -185,8 +185,8 @@ diffsynth_engine/utils/video.py,sha256=8FCaeqIdUsWMgWI_6SO9SPynsToGcLCQAVYFTc4CD
 diffsynth_engine/utils/memory/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 diffsynth_engine/utils/memory/linear_regression.py,sha256=oW_EQEw13oPoyUrxiL8A7Ksa5AuJ2ynI2qhCbfAuZbg,3930
 diffsynth_engine/utils/memory/memory_predcit_model.py,sha256=EXprSl_zlVjgfMWNXP-iw83Ot3hyMcgYaRPv-dvyL84,3943
-diffsynth_engine-0.6.1.dev20.dist-info/licenses/LICENSE,sha256=x7aBqQuVI0IYnftgoTPI_A0I_rjdjPPQkjnU6N2nikM,11346
-diffsynth_engine-0.6.1.dev20.dist-info/METADATA,sha256=JZJRwz1ckJI0aMe_StTIj8LGkSYg8rRMQvqUsvqgx3s,1164
-diffsynth_engine-0.6.1.dev20.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-diffsynth_engine-0.6.1.dev20.dist-info/top_level.txt,sha256=6zgbiIzEHLbhgDKRyX0uBJOV3F6VnGGBRIQvSiYYn6w,17
-diffsynth_engine-0.6.1.dev20.dist-info/RECORD,,
+diffsynth_engine-0.6.1.dev22.dist-info/licenses/LICENSE,sha256=x7aBqQuVI0IYnftgoTPI_A0I_rjdjPPQkjnU6N2nikM,11346
+diffsynth_engine-0.6.1.dev22.dist-info/METADATA,sha256=K5yUarSjYpbjDvqrG1i7rKrY2r1ILPkcuNupcMTDsvY,1164
+diffsynth_engine-0.6.1.dev22.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+diffsynth_engine-0.6.1.dev22.dist-info/top_level.txt,sha256=6zgbiIzEHLbhgDKRyX0uBJOV3F6VnGGBRIQvSiYYn6w,17
+diffsynth_engine-0.6.1.dev22.dist-info/RECORD,,