Add VAE decode logging: pre-decode mem snapshot, tiling prediction, peak reset

Adds three log points to support issue #6 investigation:
- [startup] VAE tile threshold confirmation after enable_tiling()
- [infer] whether tiling will activate for this image size
- [infer] pre-VAE-decode memory snapshot + reset_peak_memory_stats() so
the existing "VAE decode + postprocess done" peak reflects only VAE decode

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

app.py

@@ -192,6 +192,7 @@ pipe = QwenImageEditPlusPipeline.from_pretrained(
 )
 _hb.set()
 pipe.vae.enable_tiling(tile_sample_min_height=Mode.HIGH_DETAIL.max_dim, tile_sample_min_width=Mode.HIGH_DETAIL.max_dim)
+print(f"[startup] VAE tiling: threshold={pipe.vae.tile_sample_min_height}x{pipe.vae.tile_sample_min_width}px use_tiling={pipe.vae.use_tiling}", flush=True)
 print(f"[startup] pipeline loaded in {time.perf_counter()-_t1_load:.1f}s", flush=True)
 
 print("[startup] setting cuDNN SDPA attention processor...", flush=True)
@@ -565,7 +566,7 @@ def _instrument_first_touch(modules_with_names, t0):
         handle_box["h"] = module.register_forward_pre_hook(_make_hook(name, handle_box))
 
 
-def _make_step_callback(steps, timer, t0, mode: Mode):
+def _make_step_callback(steps, timer, t0, mode: Mode, cuda_ok: bool = False):
     """Build the diffusers step callback that logs per-step timing and marks timer checkpoints."""
     step_times = []
     def _step_cb(pipeline, step_idx, timestep, cb_kwargs):
@@ -592,6 +593,9 @@ def _make_step_callback(steps, timer, t0, mode: Mode):
                 print("[infer] skipping text_encoder offload (int8, .to() unsupported / smaller footprint)")
             else:
                 print(f"[infer] skipping text_encoder offload for mode={mode.value} (ample headroom at this resolution)")
+        if step_idx == steps - 1 and cuda_ok:
+            print(f"[infer] pre-VAE-decode — {_gpu_mem_str(True, sync=True)} | t={time.perf_counter()-t0:.1f}s")
+            torch.cuda.reset_peak_memory_stats()
         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")
@@ -629,9 +633,15 @@ def _infer_gpu(pil_images, prompt, seed, guidance_scale, steps, width, height, m
         )
 
     print(f"[infer] {len(pil_images)} image(s) pre-decoded, output={width}x{height}, seed={seed}")
+    if _cuda_ok:
+        _will_tile = pipe.vae.use_tiling and (
+            width > pipe.vae.tile_sample_min_width or height > pipe.vae.tile_sample_min_height
+        )
+        print(f"[infer] VAE tiling will {'activate' if _will_tile else 'NOT activate'} "
+              f"(threshold={pipe.vae.tile_sample_min_height}x{pipe.vae.tile_sample_min_width}px)")
 
     generator = torch.Generator(device=device).manual_seed(seed)
-    step_cb = _make_step_callback(steps, timer, t0, mode)
+    step_cb = _make_step_callback(steps, timer, t0, mode, _cuda_ok)
 
     timer.mark("pipe_start")
     print(f"[infer] calling pipe... t={time.perf_counter()-t0:.1f}s")