Update app.py

app.py

@@ -128,6 +128,7 @@ def _build_pipeline_cpu() -> DiffusionPipeline:
     """
     log.info(f"Loading model backend: {MODEL_BACKEND}")
     if MODEL_BACKEND == "sdxl_lcm_unet":
+        # SDXL base with LCM UNet (no LoRA required)
         unet = UNet2DConditionModel.from_pretrained(
             "latent-consistency/lcm-sdxl",
             torch_dtype=torch.float32,
@@ -140,12 +141,17 @@ def _build_pipeline_cpu() -> DiffusionPipeline:
             cache_dir=CACHE_DIR,
         )
     elif MODEL_BACKEND == "ssd1b_lcm_lora":
+        # SSD-1B with LCM-LoRA (Diffusers backend; no PEFT required)
         _p = AutoPipelineForText2Image.from_pretrained(
             "segmind/SSD-1B",
             torch_dtype=torch.float32,
             cache_dir=CACHE_DIR,
         )
-        _p.load_lora_weights("latent-consistency/lcm-lora-ssd-1b")
+        _p.load_lora_weights(
+            "latent-consistency/lcm-lora-ssd-1b",
+            adapter_name="lcm",
+            use_peft_backend=False,   # <-- avoid PEFT requirement
+        )
         _p.fuse_lora()
     else:
         # Default: SDXL + LCM-LoRA (smaller download, great speed/quality)
@@ -154,7 +160,11 @@ def _build_pipeline_cpu() -> DiffusionPipeline:
             torch_dtype=torch.float32,
             cache_dir=CACHE_DIR,
         )
-        _p.load_lora_weights("latent-consistency/lcm-lora-sdxl")
+        _p.load_lora_weights(
+            "latent-consistency/lcm-lora-sdxl",
+            adapter_name="lcm",
+            use_peft_backend=False,   # <-- avoid PEFT requirement
+        )
         _p.fuse_lora()
 
     _p.scheduler = LCMScheduler.from_config(_p.scheduler.config)