diffsynth-engine 0.7.0__tar.gz → 0.7.1.dev2__tar.gz

{diffsynth_engine-0.7.0 → diffsynth_engine-0.7.1.dev2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: diffsynth_engine
-Version: 0.7.0
+Version: 0.7.1.dev2
 Author: MuseAI x ModelScope
 Classifier: Programming Language :: Python :: 3
 Classifier: Operating System :: OS Independent

{diffsynth_engine-0.7.0 → diffsynth_engine-0.7.1.dev2}/diffsynth_engine/__init__.py

@@ -7,12 +7,14 @@ from .configs import (
     QwenImagePipelineConfig,
     HunyuanPipelineConfig,
     ZImagePipelineConfig,
+    Flux2KleinPipelineConfig,
     SDStateDicts,
     SDXLStateDicts,
     FluxStateDicts,
     WanStateDicts,
     QwenImageStateDicts,
     ZImageStateDicts,
+    Flux2StateDicts,
     AttnImpl,
     SpargeAttentionParams,
     VideoSparseAttentionParams,
@@ -26,6 +28,7 @@ from .pipelines import (
     SDImagePipeline,
     SDXLImagePipeline,
     FluxImagePipeline,
+    Flux2KleinPipeline,
     WanVideoPipeline,
     WanDMDPipeline,
     QwenImagePipeline,
@@ -59,12 +62,14 @@ __all__ = [
     "QwenImagePipelineConfig",
     "HunyuanPipelineConfig",
     "ZImagePipelineConfig",
+    "Flux2KleinPipelineConfig",
     "SDStateDicts",
     "SDXLStateDicts",
     "FluxStateDicts",
     "WanStateDicts",
     "QwenImageStateDicts",
     "ZImageStateDicts",
+    "Flux2StateDicts",
     "AttnImpl",
     "SpargeAttentionParams",
     "VideoSparseAttentionParams",
@@ -78,6 +83,7 @@ __all__ = [
     "SDXLImagePipeline",
     "SDXLControlNetUnion",
     "FluxImagePipeline",
+    "Flux2KleinPipeline",
     "FluxControlNet",
     "FluxIPAdapter",
     "FluxRedux",

diffsynth_engine-0.7.1.dev2/diffsynth_engine/conf/models/flux2/qwen3_8B_config.json

@@ -0,0 +1,68 @@
+{
+  "architectures": [
+    "Qwen3ForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 151643,
+  "dtype": "bfloat16",
+  "eos_token_id": 151645,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "initializer_range": 0.02,
+  "intermediate_size": 12288,
+  "layer_types": [
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention"
+  ],
+  "max_position_embeddings": 40960,
+  "max_window_layers": 36,
+  "model_type": "qwen3",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 36,
+  "num_key_value_heads": 8,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000,
+  "sliding_window": null,
+  "tie_word_embeddings": false,
+  "transformers_version": "4.56.1",
+  "use_cache": true,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

{diffsynth_engine-0.7.0 → diffsynth_engine-0.7.1.dev2}/diffsynth_engine/configs/__init__.py

@@ -11,6 +11,7 @@ from .pipeline import (
     QwenImagePipelineConfig,
     HunyuanPipelineConfig,
     ZImagePipelineConfig,
+    Flux2KleinPipelineConfig,
     BaseStateDicts,
     SDStateDicts,
     SDXLStateDicts,
@@ -19,6 +20,7 @@ from .pipeline import (
     WanS2VStateDicts,
     QwenImageStateDicts,
     ZImageStateDicts,
+    Flux2StateDicts,
     AttnImpl,
     SpargeAttentionParams,
     VideoSparseAttentionParams,
@@ -44,6 +46,7 @@ __all__ = [
     "QwenImagePipelineConfig",
     "HunyuanPipelineConfig",
     "ZImagePipelineConfig",
+    "Flux2KleinPipelineConfig",
     "BaseStateDicts",
     "SDStateDicts",
     "SDXLStateDicts",
@@ -52,6 +55,7 @@ __all__ = [
     "WanS2VStateDicts",
     "QwenImageStateDicts",
     "ZImageStateDicts",
+    "Flux2StateDicts",
     "AttnImpl",
     "SpargeAttentionParams",
     "VideoSparseAttentionParams",

{diffsynth_engine-0.7.0 → diffsynth_engine-0.7.1.dev2}/diffsynth_engine/configs/pipeline.py

@@ -3,6 +3,7 @@ import torch
 from enum import Enum
 from dataclasses import dataclass, field
 from typing import List, Dict, Tuple, Optional
+from typing_extensions import Literal
 
 from diffsynth_engine.configs.controlnet import ControlType
 
@@ -339,6 +340,47 @@ class ZImagePipelineConfig(AttentionConfig, OptimizationConfig, ParallelConfig,
         init_parallel_config(self)
 
 
+@dataclass
+class Flux2KleinPipelineConfig(AttentionConfig, OptimizationConfig, ParallelConfig, BaseConfig):
+    model_path: str | os.PathLike | List[str | os.PathLike]
+    model_dtype: torch.dtype = torch.bfloat16
+    vae_path: Optional[str | os.PathLike | List[str | os.PathLike]] = None
+    vae_dtype: torch.dtype = torch.bfloat16
+    encoder_path: Optional[str | os.PathLike | List[str | os.PathLike]] = None
+    encoder_dtype: torch.dtype = torch.bfloat16
+    image_encoder_path: Optional[str | os.PathLike | List[str | os.PathLike]] = None
+    image_encoder_dtype: torch.dtype = torch.bfloat16
+    model_size: Literal["4B", "9B"] = "4B"
+
+    @classmethod
+    def basic_config(
+        cls,
+        model_path: str | os.PathLike | List[str | os.PathLike],
+        encoder_path: Optional[str | os.PathLike | List[str | os.PathLike]] = None,
+        vae_path: Optional[str | os.PathLike | List[str | os.PathLike]] = None,
+        image_encoder_path: Optional[str | os.PathLike | List[str | os.PathLike]] = None,
+        device: str = "cuda",
+        parallelism: int = 1,
+        offload_mode: Optional[str] = None,
+        offload_to_disk: bool = False,
+    ) -> "Flux2KleinPipelineConfig":
+        return cls(
+            model_path=model_path,
+            device=device,
+            encoder_path=encoder_path,
+            vae_path=vae_path,
+            image_encoder_path=image_encoder_path,
+            parallelism=parallelism,
+            use_cfg_parallel=True if parallelism > 1 else False,
+            use_fsdp=True if parallelism > 1 else False,
+            offload_mode=offload_mode,
+            offload_to_disk=offload_to_disk,
+        )
+
+    def __post_init__(self):
+        init_parallel_config(self)
+
+
 @dataclass
 class BaseStateDicts:
     pass
@@ -398,7 +440,14 @@ class ZImageStateDicts:
     image_encoder: Optional[Dict[str, torch.Tensor]] = None
 
 
-def init_parallel_config(config: FluxPipelineConfig | QwenImagePipelineConfig | WanPipelineConfig | ZImagePipelineConfig):
+@dataclass
+class Flux2StateDicts:
+    model: Dict[str, torch.Tensor]
+    vae: Dict[str, torch.Tensor]
+    encoder: Dict[str, torch.Tensor]
+
+
+def init_parallel_config(config: FluxPipelineConfig | QwenImagePipelineConfig | WanPipelineConfig | ZImagePipelineConfig | Flux2KleinPipelineConfig):
     assert config.parallelism in (1, 2, 4, 8), "parallelism must be 1, 2, 4 or 8"
     config.batch_cfg = True if config.parallelism > 1 and config.use_cfg_parallel else config.batch_cfg
 

diffsynth_engine-0.7.1.dev2/diffsynth_engine/models/flux2/__init__.py

@@ -0,0 +1,7 @@
+from .flux2_dit import Flux2DiT
+from .flux2_vae import Flux2VAE
+
+__all__ = [
+    "Flux2DiT",
+    "Flux2VAE",
+]