diffsynth 2.1.3__tar.gz → 2.1.4__tar.gz

{diffsynth-2.1.3 → diffsynth-2.1.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: diffsynth
-Version: 2.1.3
+Version: 2.1.4
 Summary: Enjoy the magic of Diffusion models!
 Author: ModelScope Team
 License-Expression: Apache-2.0
@@ -107,6 +107,8 @@ See also:
 
 > Currently, the development personnel of this project are limited, with most of the work handled by [Artiprocher](https://github.com/Artiprocher) and [mi804](https://github.com/mi804). Therefore, the progress of new feature development will be relatively slow, and the speed of responding to and resolving issues is limited. We apologize for this and ask developers to understand.
 
+- **August 25, 2026** We have open-sourced [DiffSynth-WebUI](https://github.com/modelscope/DiffSynth-WebUI), enabling one-click private deployment of LoRA training services. Combined with the model quantization feature, you can train large models even with consumer-grade GPUs.
+
 - **August 19, 2026** We have released the model quantization feature. It provides a unified `QuantizeConfig` entry point supporting multiple quantization backends including bitsandbytes, torchao, and comfy-kitchen, with capabilities such as online quantization, loading pre-quantized weights, mixed quantization, saving quantized models, and quantization + LoRA training. For details, please refer to the [documentation](/docs/en/Pipeline_Usage/Quantization.md).
 
 - **August 17, 2026** MiniMax-Music3 open-sourced, welcome a new member to the audio model family! Support includes text-to-music generation and low VRAM inference. For details, please refer to the [documentation](/docs/en/Model_Details/MiniMax-Music3.md) and [example code](/examples/minimax_music3/).

{diffsynth-2.1.3 → diffsynth-2.1.4}/README.md

@@ -40,6 +40,8 @@ See also:
 
 > Currently, the development personnel of this project are limited, with most of the work handled by [Artiprocher](https://github.com/Artiprocher) and [mi804](https://github.com/mi804). Therefore, the progress of new feature development will be relatively slow, and the speed of responding to and resolving issues is limited. We apologize for this and ask developers to understand.
 
+- **August 25, 2026** We have open-sourced [DiffSynth-WebUI](https://github.com/modelscope/DiffSynth-WebUI), enabling one-click private deployment of LoRA training services. Combined with the model quantization feature, you can train large models even with consumer-grade GPUs.
+
 - **August 19, 2026** We have released the model quantization feature. It provides a unified `QuantizeConfig` entry point supporting multiple quantization backends including bitsandbytes, torchao, and comfy-kitchen, with capabilities such as online quantization, loading pre-quantized weights, mixed quantization, saving quantized models, and quantization + LoRA training. For details, please refer to the [documentation](/docs/en/Pipeline_Usage/Quantization.md).
 
 - **August 17, 2026** MiniMax-Music3 open-sourced, welcome a new member to the audio model family! Support includes text-to-music generation and low VRAM inference. For details, please refer to the [documentation](/docs/en/Model_Details/MiniMax-Music3.md) and [example code](/examples/minimax_music3/).

{diffsynth-2.1.3 → diffsynth-2.1.4}/diffsynth/core/data/operators.py

@@ -307,8 +307,8 @@ class LoadAudioWithTorchaudio(DataProcessingOperator, FrameSamplerByRateMixin):
                 padding = target_samples - current_samples
                 waveform = torch.nn.functional.pad(waveform, (0, padding))
             return waveform, sample_rate
-        except:
-            warnings.warn(f"Cannot load audio in {data}. The audio will be `None`.")
+        except Exception as e:
+            warnings.warn(f"Cannot load audio in {data} due to {e}. The audio will be `None`.")
             return None
 
 

{diffsynth-2.1.3 → diffsynth-2.1.4}/diffsynth/core/loader/file.py

@@ -1,5 +1,5 @@
 from safetensors import safe_open
-import torch, hashlib
+import torch, hashlib, json
 
 
 def load_state_dict(file_path, torch_dtype=None, device="cpu", pin_memory=False, verbose=0):
@@ -34,10 +34,26 @@ def load_state_dict_from_safetensors(file_path, torch_dtype=None, device="cpu"):
 
 
 def load_metadata_from_safetensors(file_path):
+    if isinstance(file_path, list):
+        metadata = {}
+        for file_path_ in file_path:
+            for key, value in load_metadata_from_safetensors(file_path_).items():
+                metadata[key] = merge_metadata_value(metadata[key], value) if key in metadata else value
+        return metadata
     with safe_open(file_path, framework="pt", device="cpu") as f:
         return f.metadata() or {}
 
 
+def merge_metadata_value(old_value, new_value):
+    try:
+        old_items, new_items = json.loads(old_value), json.loads(new_value)
+    except (json.JSONDecodeError, TypeError):
+        return new_value
+    if isinstance(old_items, list) and isinstance(new_items, list):
+        return json.dumps(old_items + [i for i in new_items if i not in old_items])
+    return new_value
+
+
 def load_state_dict_from_bin(file_path, torch_dtype=None, device="cpu"):
     state_dict = torch.load(file_path, map_location=device, weights_only=True)
     if len(state_dict) == 1:

{diffsynth-2.1.3 → diffsynth-2.1.4}/diffsynth/core/loader/model.py

@@ -37,7 +37,7 @@ def load_model(model_class, path, config=None, torch_dtype=torch.bfloat16, devic
             devices = [vram_config[k] for k in ("offload_device", "onload_device", "preparing_device", "computation_device")]
             load_device = [d for d in devices if d != "disk"][0]
             disk_map = DiskMap(path, load_device, torch_dtype=None, state_dict_converter=state_dict_converter)
-            metadata = load_metadata_from_safetensors(path[0] if isinstance(path, list) else path)
+            metadata = load_metadata_from_safetensors(path)
             model = quantize.prepare_for_prequantized_load(model, compute_dtype=vram_config["computation_dtype"])
             model = enable_vram_management(model, module_map, vram_config=vram_config, disk_map=disk_map, vram_limit=vram_limit, quantize=quantize, metadata=metadata)
         else:

{diffsynth-2.1.3 → diffsynth-2.1.4}/diffsynth/diffusion/flow_match.py

@@ -302,7 +302,7 @@ class FlowMatchScheduler():
         num_train_timesteps = 1000
         if special_case == "stage2":
             sigmas = torch.Tensor([0.909375, 0.725, 0.421875])
-        elif special_case == "ditilled_stage1":
+        elif special_case == "distilled_stage1":
             sigmas = torch.Tensor([1.0, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875])
         else:
             dynamic_shift_len = dynamic_shift_len or 4096

{diffsynth-2.1.3 → diffsynth-2.1.4}/diffsynth/diffusion/training_module.py

@@ -193,9 +193,25 @@ class DiffusionTrainingModule(torch.nn.Module):
             quant_config = QuantizeConfig(method=method, exclude_modules=exclude_modules)
             if not quant_config.backend.capabilities().get("is_differentiable", True):
                 raise ValueError(f"Quantization method `{method}` is not differentiable, so it cannot be used for training (frozen quantized layers must pass gradients through to LoRA branches). Choose a method whose backend declares `is_differentiable=True`.")
-            quant_map[model_string] = quant_config
+            quant_map[self.normalize_quant_key(model_string)] = quant_config
         return quant_map
 
+    def normalize_quant_key(self, model_string):
+        if isinstance(model_string, str) and model_string.startswith("[") and model_string.endswith("]"):
+            try:
+                model_string = json.loads(model_string)
+            except json.JSONDecodeError:
+                raise ValueError(f"Failed to parse quant option model string: `{model_string}`. A multi-file model must be written as a JSON list of files, matching its `--model_paths` entry.")
+        if isinstance(model_string, list):
+            return tuple(model_string)
+        return model_string
+
+    def get_quant_config(self, quant_map, model_string):
+        quant_config = quant_map.get(self.normalize_quant_key(model_string))
+        if quant_config is None and len(quant_map) > 0:
+            print(f"No quant option matches `{model_string}`. This model is loaded without quantization. Parsed quant options: {({key: config.method for key, config in quant_map.items()})}.")
+        return quant_config
+
     def parse_model_configs(self, model_paths, model_id_with_origin_paths, fp8_models=None, offload_models=None, quant_options=None, device="cpu"):
         fp8_models = [] if fp8_models is None else fp8_models.split(",")
         offload_models = [] if offload_models is None else offload_models.split(",")
@@ -209,7 +225,7 @@ class DiffusionTrainingModule(torch.nn.Module):
                     offload=path in offload_models,
                     device=device
                 )
-                model_configs.append(ModelConfig(path=path, quantize=quant_map.get(path), **vram_config))
+                model_configs.append(ModelConfig(path=path, quantize=self.get_quant_config(quant_map, path), **vram_config))
         if model_id_with_origin_paths is not None:
             model_id_with_origin_paths = model_id_with_origin_paths.split(",")
             for model_id_with_origin_path in model_id_with_origin_paths:
@@ -219,7 +235,7 @@ class DiffusionTrainingModule(torch.nn.Module):
                     device=device
                 )
                 config = self.parse_path_or_model_id(model_id_with_origin_path)
-                model_configs.append(ModelConfig(model_id=config.model_id, origin_file_pattern=config.origin_file_pattern, quantize=quant_map.get(model_id_with_origin_path), **vram_config))
+                model_configs.append(ModelConfig(model_id=config.model_id, origin_file_pattern=config.origin_file_pattern, quantize=self.get_quant_config(quant_map, model_id_with_origin_path), **vram_config))
         return model_configs
     
 

{diffsynth-2.1.3 → diffsynth-2.1.4}/diffsynth/pipelines/ltx2_audio_video.py

@@ -213,7 +213,7 @@ class LTX2AudioVideoPipeline(BasePipeline):
         progress_bar_cmd=tqdm,
     ):
         # Scheduler
-        self.scheduler.set_timesteps(num_inference_steps, denoising_strength=denoising_strength, special_case="ditilled_stage1" if use_distilled_pipeline else None)
+        self.scheduler.set_timesteps(num_inference_steps, denoising_strength=denoising_strength, special_case="distilled_stage1" if use_distilled_pipeline else None)
         # Inputs
         inputs_posi = {
             "prompt": prompt,

{diffsynth-2.1.3 → diffsynth-2.1.4}/diffsynth/version.py

@@ -1,5 +1,5 @@
 # Make sure to modify __release_datetime__ to release time when making official release.
-__version__ = '2.1.3'
+__version__ = '2.1.4'
 # default release datetime for branches under active development is set
 # to be a time far-far-away-into-the-future
 __release_datetime__ = '2099-10-13 08:56:12'

{diffsynth-2.1.3 → diffsynth-2.1.4}/diffsynth.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: diffsynth
-Version: 2.1.3
+Version: 2.1.4
 Summary: Enjoy the magic of Diffusion models!
 Author: ModelScope Team
 License-Expression: Apache-2.0
@@ -107,6 +107,8 @@ See also:
 
 > Currently, the development personnel of this project are limited, with most of the work handled by [Artiprocher](https://github.com/Artiprocher) and [mi804](https://github.com/mi804). Therefore, the progress of new feature development will be relatively slow, and the speed of responding to and resolving issues is limited. We apologize for this and ask developers to understand.
 
+- **August 25, 2026** We have open-sourced [DiffSynth-WebUI](https://github.com/modelscope/DiffSynth-WebUI), enabling one-click private deployment of LoRA training services. Combined with the model quantization feature, you can train large models even with consumer-grade GPUs.
+
 - **August 19, 2026** We have released the model quantization feature. It provides a unified `QuantizeConfig` entry point supporting multiple quantization backends including bitsandbytes, torchao, and comfy-kitchen, with capabilities such as online quantization, loading pre-quantized weights, mixed quantization, saving quantized models, and quantization + LoRA training. For details, please refer to the [documentation](/docs/en/Pipeline_Usage/Quantization.md).
 
 - **August 17, 2026** MiniMax-Music3 open-sourced, welcome a new member to the audio model family! Support includes text-to-music generation and low VRAM inference. For details, please refer to the [documentation](/docs/en/Model_Details/MiniMax-Music3.md) and [example code](/examples/minimax_music3/).