Update app.py

app.py

@@ -276,7 +276,14 @@ PENDING_LORA_KEY: str | None = None
 PENDING_LORA_STATE: dict[str, torch.Tensor] | None = None
 PENDING_LORA_STATUS: str = "No LoRA state prepared yet."
 
-checkpoint_path = hf_hub_download(repo_id=LTX_MODEL_REPO, filename="ltx-2.3-22b-distilled.safetensors")
+weights_dir = Path("weights")
+weights_dir.mkdir(exist_ok=True)
+checkpoint_path = hf_hub_download(
+    repo_id=LTX_MODEL_REPO,
+    filename="ltx-2.3-22b-distilled.safetensors",
+    local_dir=str(weights_dir),
+    local_dir_use_symlinks=False,
+)
 spatial_upsampler_path = hf_hub_download(repo_id=LTX_MODEL_REPO, filename="ltx-2.3-spatial-upscaler-x2-1.1.safetensors")
 gemma_root = snapshot_download(repo_id=GEMMA_REPO)
 
@@ -372,29 +379,16 @@ def prepare_lora_cache(
     new_transformer_cpu = None
     try:
         progress(0.35, desc="Building fused CPU transformer")
-        tmp_ledger = ledger.with_loras(tuple(loras_for_builder))
-
-        orig_tmp_target = getattr(tmp_ledger, "_target_device", None)
-        orig_tmp_device = getattr(tmp_ledger, "device", None)
-        try:
-            tmp_ledger._target_device = lambda: torch.device("cpu")
-            tmp_ledger.device = torch.device("cpu")
-            new_transformer_cpu = tmp_ledger.transformer()
-        finally:
-            if orig_tmp_target is not None:
-                tmp_ledger._target_device = orig_tmp_target
-            else:
-                try:
-                    delattr(tmp_ledger, "_target_device")
-                except Exception:
-                    pass
-            if orig_tmp_device is not None:
-                tmp_ledger.device = orig_tmp_device
-            else:
-                try:
-                    delattr(tmp_ledger, "device")
-                except Exception:
-                    pass
+        tmp_ledger = pipeline.model_ledger.__class__(
+            dtype=ledger.dtype,
+            device=torch.device("cpu"),
+            checkpoint_path=str(checkpoint_path),
+            spatial_upsampler_path=str(spatial_upsampler_path),
+            gemma_root_path=str(gemma_root),
+            loras=tuple(loras_for_builder),
+            quantization=getattr(ledger, "quantization", None),
+        )
+        new_transformer_cpu = tmp_ledger.transformer()
 
         progress(0.70, desc="Extracting fused state_dict")
         state = new_transformer_cpu.state_dict()