feat: add extensible Ascend A5 / Ascend 950 runtime support

README.md

@@ -75,7 +75,8 @@ image = pipe(prompt="a girl, qipao")
 image.save("image.png")
 ```
 
-For more details, please refer to our tutorials ([English](./docs/tutorial.md), [中文](./docs/tutorial_zh.md)).
+For more details, please refer to our tutorials ([English](./docs/tutorial.md), [中文](./docs/tutorial_zh.md))
+and the [Ascend A5 / Ascend 950 guide](./docs/ascend.md).
 
 ## Showcase
 

diffsynth_engine/__init__.py

@@ -18,6 +18,7 @@
     AttnImpl,
     SpargeAttentionParams,
     VideoSparseAttentionParams,
+    QuantizationConfig,
     LoraConfig,
     ControlNetParams,
     ControlType,
@@ -73,6 +74,7 @@
     "AttnImpl",
     "SpargeAttentionParams",
     "VideoSparseAttentionParams",
+    "QuantizationConfig",
     "LoraConfig",
     "ControlNetParams",
     "ControlType",

diffsynth_engine/configs/__init__.py

@@ -24,6 +24,7 @@
     AttnImpl,
     SpargeAttentionParams,
     VideoSparseAttentionParams,
+    QuantizationConfig,
     LoraConfig,
 )
 from .controlnet import (
@@ -59,6 +60,7 @@
     "AttnImpl",
     "SpargeAttentionParams",
     "VideoSparseAttentionParams",
+    "QuantizationConfig",
     "LoraConfig",
     "ControlType",
     "ControlNetParams",

diffsynth_engine/configs/pipeline.py

@@ -35,6 +35,7 @@ class AttnImpl(Enum):
     SAGE = "sage"  # Sage Attention
     SPARGE = "sparge"  # Sparge Attention
     VSA = "vsa"  # Video Sparse Attention
+    MINDIE = "mindie"  # MindIE-SD attention for Ascend NPU
 
 
 @dataclass
@@ -56,12 +57,28 @@ class AttentionConfig:
     attn_params: Optional[SpargeAttentionParams | VideoSparseAttentionParams] = None
 
 
+@dataclass
+class QuantizationConfig:
+    backend: Literal["auto", "native", "nunchaku", "mindie"] = "auto"
+    linear: Literal["none", "fp8", "int4"] = "none"
+    attention: Literal["none", "fp8"] = "none"
+
+    def __post_init__(self):
+        if self.backend not in {"auto", "native", "nunchaku", "mindie"}:
+            raise ValueError(f"Unsupported quantization backend: {self.backend}")
+        if self.linear not in {"none", "fp8", "int4"}:
+            raise ValueError(f"Unsupported linear quantization mode: {self.linear}")
+        if self.attention not in {"none", "fp8"}:
+            raise ValueError(f"Unsupported attention quantization mode: {self.attention}")
+
+
 @dataclass
 class OptimizationConfig:
     use_fp8_linear: bool = False
     use_fbcache: bool = False
     fbcache_relative_l1_threshold: float = 0.05
     use_torch_compile: bool = False
+    quantization: Optional[QuantizationConfig] = None
 
 
 @dataclass