feat: Add LoRA and ControlNet support for Qwen Image models

nunchaku/lora/qwenimage/__init__.py

@@ -0,0 +1,38 @@
+"""
+Qwen Image LoRA module for Nunchaku.
+
+This module provides utilities for loading, converting, and composing LoRA weights
+for Qwen Image models in Nunchaku's quantized inference framework.
+
+Main Functions
+--------------
+- :func:`to_diffusers` : Convert LoRA to Diffusers format
+- :func:`to_nunchaku` : Convert LoRA to Nunchaku format
+- :func:`compose_lora` : Compose multiple LoRAs
+- :func:`is_nunchaku_format` : Check if LoRA is in Nunchaku format
+- :func:`convert_to_nunchaku_qwenimage_lowrank_dict` : Convert full model LoRA
+
+Example Usage
+-------------
+>>> from nunchaku.lora.qwenimage import to_nunchaku, compose_lora
+>>> # Convert a single LoRA
+>>> lora_nunchaku = to_nunchaku("lora.safetensors", output_path="lora_nunchaku.safetensors")
+>>> # Compose multiple LoRAs
+>>> composed = compose_lora([("lora1.safetensors", 0.8), ("lora2.safetensors", 0.6)])
+>>> # Convert the composed LoRA to Nunchaku format
+>>> lora_nunchaku = to_nunchaku(composed, output_path="composed_nunchaku.safetensors")
+"""
+
+from .compose import compose_lora
+from .diffusers_converter import to_diffusers
+from .nunchaku_converter import convert_to_nunchaku_qwenimage_lowrank_dict, fuse_vectors, to_nunchaku
+from .utils import is_nunchaku_format
+
+__all__ = [
+    "to_diffusers",
+    "to_nunchaku",
+    "compose_lora",
+    "convert_to_nunchaku_qwenimage_lowrank_dict",
+    "fuse_vectors",
+    "is_nunchaku_format",
+]