[WIP] [LoRA] support omi hidream lora.

src/diffusers/loaders/lora_conversion_utils.py

@@ -1789,12 +1789,58 @@ def get_alpha_scales(down_weight, key):
     return converted_state_dict
 
 
-def _convert_non_diffusers_hidream_lora_to_diffusers(state_dict, non_diffusers_prefix="diffusion_model"):
-    if not all(k.startswith(non_diffusers_prefix) for k in state_dict):
-        raise ValueError("Invalid LoRA state dict for HiDream.")
-    converted_state_dict = {k.removeprefix(f"{non_diffusers_prefix}."): v for k, v in state_dict.items()}
-    converted_state_dict = {f"transformer.{k}": v for k, v in converted_state_dict.items()}
-    return converted_state_dict
+def _convert_non_diffusers_hidream_lora_to_diffusers(state_dict):
+    non_diffusers_prefix = "diffusion_model"
+    is_kohya = all(k.startswith(f"{non_diffusers_prefix}.") for k in state_dict)
+
+    def _convert_kohya(state_dict):
+        converted_state_dict = {k.removeprefix(f"{non_diffusers_prefix}."): v for k, v in state_dict.items()}
+        converted_state_dict = {f"transformer.{k}": v for k, v in converted_state_dict.items()}
+        return converted_state_dict
+
+    if is_kohya:
+        return _convert_kohya(state_dict)
+
+    else:
+        assert any(k.startswith(("clip_g.", "clip_l.", "t5.", "llama.", "transformer.")) for k in state_dict)
+        converted_state_dict = {}
+        component = "transformer"
+        compoent_sd = {k: v for k, v in state_dict.items() if k.startswith(f"{component}.")}
+
+        def _convert_omi(key, state_dict, component):
+            down_key = f"{key}.lora_down.weight"
+            down_weight = state_dict.pop(down_key)
+            lora_rank = down_weight.shape[0]
+
+            up_weight_key = f"{key}.lora_up.weight"
+            up_weight = state_dict.pop(up_weight_key)
+
+            alpha_key = f"{key}.alpha"
+            alpha = state_dict.pop(alpha_key)
+
+            # scale weight by alpha and dim
+            scale = alpha / lora_rank
+            # calculate scale_down and scale_up
+            scale_down = scale
+            scale_up = 1.0
+            while scale_down * 2 < scale_up:
+                scale_down *= 2
+                scale_up /= 2
+            down_weight = down_weight * scale_down
+            up_weight = up_weight * scale_up
+
+            diffusers_down_key = f"{key}.lora_A.weight"
+            converted_state_dict[f"{component}.{diffusers_down_key}"] = down_weight
+            converted_state_dict[f"{component}.{diffusers_down_key.replace('.lora_A.', '.lora_B.')}"] = up_weight
+
+        all_unique_keys = {
+            k.replace(".lora_down.weight", "").replace(".lora_up.weight", "").replace(".alpha", "")
+            for k in compoent_sd
+        }
+        for k in all_unique_keys:
+            _convert_omi(k, compoent_sd, component=component)
+
+        return converted_state_dict
 
 
 def _convert_non_diffusers_ltxv_lora_to_diffusers(state_dict, non_diffusers_prefix="diffusion_model"):

src/diffusers/loaders/lora_pipeline.py

@@ -5489,7 +5489,9 @@ def lora_state_dict(
             logger.warning(warn_msg)
             state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}
 
-        is_non_diffusers_format = any("diffusion_model" in k for k in state_dict)
+        kohya_format = any("diffusion_model" in k for k in state_dict)
+        is_omi_format = any(k.startswith(("clip_g.", "clip_l.", "t5.", "llama.", "transformer.")) for k in state_dict)
+        is_non_diffusers_format = kohya_format or is_omi_format
         if is_non_diffusers_format:
             state_dict = _convert_non_diffusers_hidream_lora_to_diffusers(state_dict)