Enable FA3 attention and torch.compile for faster inference

Replaces the ZeroGPU-era no-ops with real optimizations suitable for
dedicated GPU deployment: FA3 processor on all 60 transformer blocks
(with graceful fallback), and torch.compile(mode="reduce-overhead") on
the transformer for kernel fusion and CUDA graph capture.

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

app.py

@@ -86,9 +86,18 @@ pipe = QwenImageEditPlusPipeline.from_pretrained(
 _hb.set()
 print(f"[startup] pipeline loaded in {time.perf_counter()-_t1_load:.1f}s", flush=True)
 
-print("Using default attention processor (FA3 skipped for ZeroGPU GPU-arch compatibility).", flush=True)
-
-print("torch.compile skipped: lazy Triton kernel compilation inside @spaces.GPU always exceeds ZeroGPU's task timeout.")
+try:
+    from qwenimage.qwen_fa3_processor import QwenDoubleStreamAttnProcessorFA3
+    _fa3_processor = QwenDoubleStreamAttnProcessorFA3()
+    for _block in pipe.transformer.transformer_blocks:
+        _block.attn.processor = _fa3_processor
+    print("[startup] FA3 attention processor enabled.", flush=True)
+except Exception as _fa3_err:
+    print(f"[startup] FA3 not available ({_fa3_err}), using default attention processor.", flush=True)
+
+print("[startup] applying torch.compile to transformer (mode=reduce-overhead)...", flush=True)
+pipe.transformer = torch.compile(pipe.transformer, mode="reduce-overhead")
+print("[startup] torch.compile done.", flush=True)
 
 with open("examples.json") as _f:
     EXAMPLES_CONFIG = json.load(_f)