Update app.py

app.py

@@ -154,9 +154,6 @@ def load_models() -> str | None:
         pipe.load_components(dtype=torch.bfloat16)
         pipe.transformer.set_attention_backend(ATTENTION)
 
-        pipe.load_components(dtype=torch.bfloat16)
-        pipe.transformer.set_attention_backend(ATTENTION)
-
         # --- Diagnostic: dump the LoRA files' key names/shapes and the transformer's own shapes to the Space
         # logs, so the exact rename map can be worked out without a notebook or shell. Set H3_LORA_DEBUG=0 in
         # the Space's env vars to silence this once you're done, or just delete this block later.
@@ -192,14 +189,6 @@ def load_models() -> str | None:
             for k, shape in sorted(norm_out.items()):
                 print(f"[lora-debug]   {k}  {shape}", flush=True)
 
-        if LORA_REPO.lower() not in ("", "off", "none"):
-            from peft.tuners.tuners_utils import BaseTunerLayer
-
-            try:
-                counts = {}
-                for name, filename in LORA_FILES.items():
-                    ...
-
         # Approach B: attach both LoRA adapters as PEFT layers on the transformer, inactive (weight 0) until a
         # request asks for them. `load_lora_adapter` is the model-level loader (`PeftAdapterMixin`), used because
         # `MiniMaxH3ModularPipeline` has no pipeline-level `load_lora_weights` of its own.