Merge branch 'main' into ja

app.py

@@ -190,6 +190,7 @@ pipe = QwenImageEditPlusPipeline.from_pretrained(
     torch_dtype=dtype,
 )
 _hb.set()
+pipe.vae.enable_tiling()
 print(f"[startup] pipeline loaded in {time.perf_counter()-_t1_load:.1f}s", flush=True)
 
 print("[startup] using default attention processor.", flush=True)
@@ -531,6 +532,13 @@ def _make_step_callback(steps, timer, t0):
         if step_idx == 0:
             timer.mark("first_step")
         timer.mark("last_step")  # overwritten each step; final value = end of last step
+        # Text encoder is done after prompt encoding, before the denoising loop starts.
+        # Drop it (~15GB) ahead of VAE decode's fp32-upcast memory spike. Skipped when
+        # accelerate hooks are managing placement (offload-fallback path) to avoid
+        # fighting their own device bookkeeping — see the finally block below.
+        if step_idx == steps - 1 and getattr(pipeline.text_encoder, "_hf_hook", None) is None:
+            pipeline.text_encoder.to("cpu")
+            torch.cuda.empty_cache()
         delta_ms = (now - (step_times[-2] if len(step_times) > 1 else t0)) * 1000
         tag = " ← includes cold-start (offload hook install + first weight transfer)" if step_idx == 0 else ""
         print(f"[infer] step {step_idx+1}/{steps} done — {delta_ms:.0f}ms{tag} | t={now-t0:.1f}s")