Revert FA3 and torch.compile: incompatible with this GPU environment

torch.compile falls back everywhere due to a torch.device ConstantVariable
bug in the installed PyTorch version. FA3 fails at runtime with
"no kernel image available" — the vllm-flash-attn3 kernel is not compiled
for this GPU's compute capability. Both add overhead with no benefit here.

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

app.py

@@ -86,19 +86,7 @@ pipe = QwenImageEditPlusPipeline.from_pretrained(
 _hb.set()
 print(f"[startup] pipeline loaded in {time.perf_counter()-_t1_load:.1f}s", flush=True)
 
-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)
-
-torch._dynamo.config.suppress_errors = 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)
+print("[startup] using default attention processor.", flush=True)
 
 with open("examples.json") as _f:
     EXAMPLES_CONFIG = json.load(_f)