fix bug

app.py

@@ -160,6 +160,7 @@ attention_backend_error: str | None = None
 aoti_error: str | None = None
 transformer_compiled: bool = False
 transformer_compile_attempted: bool = False
+compile_error: str | None = None
 inductor_configured: bool = False
 
 SCHEDULERS = {"FlowMatch Euler": FlowMatchEulerDiscreteScheduler}
@@ -413,14 +414,19 @@ def configure_inductor_for_compile() -> None:
 
 
 def maybe_compile_transformer() -> None:
-    global transformer_compiled, transformer_compile_attempted
+    global transformer_compiled, transformer_compile_attempted, compile_error
     if not ENABLE_COMPILE or transformer_compile_attempted:
         return
     if pipe is None or getattr(pipe, "transformer", None) is None:
         return
 
     transformer_compile_attempted = True
+    compile_error = None
     configure_inductor_for_compile()
+    try:
+        torch._dynamo.config.suppress_errors = True
+    except Exception:  # noqa: BLE001
+        pass
 
     try:
         if getattr(pipe, "vae", None) is not None and hasattr(pipe.vae, "disable_tiling"):
@@ -434,6 +440,7 @@ def maybe_compile_transformer() -> None:
         transformer_compiled = True
     except Exception as exc:  # noqa: BLE001
         transformer_compiled = False
+        compile_error = str(exc)
         print(f"torch.compile failed (continuing without compile): {exc}")
 
 
@@ -513,7 +520,6 @@ def generate_image(
     max_shift: float,
 ) -> Tuple[torch.Tensor, int]:
     width, height = parse_resolution(resolution)
-    generator = torch.Generator("cuda").manual_seed(seed)
     set_scheduler(
         pipeline,
         str(scheduler_name),
@@ -533,8 +539,9 @@ def generate_image(
         except Exception as exc:  # noqa: BLE001
             print(f"LoRA scale update failed (continuing without changing LoRA state): {exc}")
 
-    with torch.inference_mode():
-        image = pipeline(
+    def run_pipeline() -> torch.Tensor:
+        generator = torch.Generator("cuda").manual_seed(seed)
+        return pipeline(
             prompt=prompt,
             height=height,
             width=width,
@@ -543,6 +550,25 @@ def generate_image(
             generator=generator,
             max_sequence_length=int(max_sequence_length),
         ).images[0]
+
+    try:
+        with torch.inference_mode():
+            image = run_pipeline()
+    except Exception as exc:  # noqa: BLE001
+        transformer = getattr(pipeline, "transformer", None)
+        message = str(exc)
+        is_dynamo_error = isinstance(exc, AssertionError) or "torch._dynamo" in message or "ConstantVariable" in message
+        if transformer is not None and hasattr(transformer, "_orig_mod") and is_dynamo_error:
+            global transformer_compiled, transformer_compile_attempted, compile_error
+            compile_error = message
+            transformer_compiled = False
+            transformer_compile_attempted = True
+            pipeline.transformer = transformer._orig_mod
+            print(f"torch.compile runtime failed; falling back to eager: {exc}")
+            with torch.inference_mode():
+                image = run_pipeline()
+        else:
+            raise
     return image, seed
 
 
@@ -709,7 +735,7 @@ with gr.Blocks(title="Z-Image + LoRA") as demo:
     gr.Markdown(
         """<div align="center">
 
-# Z-Image Generation (No SD fallback)
+# Z-Image with LoRA
 
 </div>"""
     )
@@ -728,6 +754,8 @@ with gr.Blocks(title="Z-Image + LoRA") as demo:
         details_md_blocks.append(f"**Attention backend error**\n```\n{attention_backend_error}\n```")
     if aoti_error:
         details_md_blocks.append(f"**AoTI error**\n```\n{aoti_error}\n```")
+    if compile_error:
+        details_md_blocks.append(f"**torch.compile error**\n```\n{compile_error}\n```")
 
     with gr.Column(elem_id="floating_status_panel"):
         with gr.Accordion("Status / Debug", open=False):