changed the pipeline definition.

app.py

@@ -177,7 +177,7 @@ state_dict = torch.load(projector_path, map_location="cpu", weights_only=True)
 slider_projector.load_state_dict(state_dict, strict=True)
 
 _log("[worker] assembling pipeline (sharded/offloaded)…")
-pipe = FluxKontextSliderPipeline.from_pretrained(
+PIPELINE = FluxKontextSliderPipeline.from_pretrained(
     pretrained,
     token=token,
     trust_remote_code=True,
@@ -196,12 +196,11 @@ _log(f"[worker] loading LoRA from: {trained_models_path}")
 pipe.load_lora_weights(trained_models_path)
 _log("[worker] LoRA loaded.")
 
-# DO NOT pipe.to("cuda") here; keep auto device_map to avoid OOM
-PIPELINE = pipe
-if cuda_ok:
-    free, total = torch.cuda.mem_get_info()
-    _log(f"[worker] VRAM free/total: {free/1e9:.2f}/{total/1e9:.2f} GB")
-_log("[worker] PIPELINE ready.")
+# if cuda_ok:
+#     free, total = torch.cuda.mem_get_info()
+#     _log(f"[worker] VRAM free/total: {free/1e9:.2f}/{total/1e9:.2f} GB")
+# _log("[worker] PIPELINE ready.")
+
 
 # moving the pipeline to GPU
 PIPELINE.to('cuda')