Update app.py

app.py

@@ -47,23 +47,7 @@ from safetensors.torch import load_file
 from ltx_core.loader.primitives import LoraPathStrengthAndSDOps
 from ltx_core.loader.sd_ops import LTXV_LORA_COMFY_RENAMING_MAP
 
-try:
-    from ltx_core.loader.fuse_loras import apply_loras
-except ImportError:
-    from ltx_core.loader.fuse_loras import fuse_lora_weights
-
-    def apply_loras(model_sd, loras, dtype=None):
-        # fuse_lora_weights is the lower-level helper the repo uses internally;
-        # this wrapper turns its output into a regular state_dict.
-        return {
-            k: v
-            for k, v in fuse_lora_weights(
-                model_sd,
-                loras,
-                dtype=dtype,
-                preserve_input_device=False,
-            )
-        }
+from ltx_core.loader.fuse_loras import apply_loras
 
 from ltx_core.components.diffusion_steps import EulerDiffusionStep
 from ltx_core.components.noisers import GaussianNoiser
@@ -459,11 +443,17 @@ def apply_current_loras_to_transformer(
             )
         ]
 
-        fused_state = apply_loras(
-            BASE_TRANSFORMER_STATE,
+        base_model_sd = _StateDictModel(
+            {k: v.clone() for k, v in BASE_TRANSFORMER_STATE.items()}
+        )
+        
+        fused_model_sd = apply_loras(
+            base_model_sd,
             loras,
             dtype=pipeline.model_ledger.dtype,
         )
+        
+        fused_state = fused_model_sd.sd if hasattr(fused_model_sd, "sd") else fused_model_sd
         LORA_STATE_CACHE[key] = fused_state
 
     with torch.no_grad():
@@ -499,7 +489,9 @@ BASE_TRANSFORMER_STATE = {
     k: v.detach().cpu().contiguous()
     for k, v in _transformer.state_dict().items()
 }
-
+class _StateDictModel:
+    def __init__(self, sd: dict[str, torch.Tensor]):
+        self.sd = sd
 ACTIVE_LORA_KEY: str | None = None
 LORA_STATE_CACHE: dict[str, dict[str, torch.Tensor]] = {}
 _video_encoder = _orig_video_encoder_factory()