Force eager attention on Z-Image text encoder (vmap conflict with ZeroGPU); add Citrus theme + dark-mode color css

app.py

@@ -53,8 +53,26 @@ SR_SCALE = 4
 PID_INFERENCE_STEPS = 4
 
 print("[pid] loading Z-Image pipeline...", flush=True)
-pipeline, pipe_cfg = load_pipeline(BACKBONE, dtype=DTYPE)
+# transformers 4.57's SDPA causal-mask uses torch.vmap, which clashes with
+# ZeroGPU's __torch_function__ hijack during fake tensor allocation. Force
+# eager attention on the text encoder to skip the vmap codepath.
+from diffusers import ZImagePipeline
+from transformers import Qwen3Model
+
+_text_encoder = Qwen3Model.from_pretrained(
+    "Tongyi-MAI/Z-Image",
+    subfolder="text_encoder",
+    torch_dtype=DTYPE,
+    attn_implementation="eager",
+)
+pipeline = ZImagePipeline.from_pretrained(
+    "Tongyi-MAI/Z-Image",
+    torch_dtype=DTYPE,
+    text_encoder=_text_encoder,
+)
 pipeline.to("cuda")
+from pid._src.inference.pipeline_registry import get_config as _get_pipe_cfg
+pipe_cfg = _get_pipe_cfg(BACKBONE)
 
 print("[pid] loading PiD decoder...", flush=True)
 pid_meta = get_pid_checkpoint(BACKBONE, CKPT_TYPE)
@@ -183,7 +201,9 @@ PiD upsamples 4× during decode, so a 512² Z-Image latent track becomes a
 2048² super-resolved image.
 """
 
-with gr.Blocks() as demo:
+CSS = " .dark .gradio-container { color: var(--body-text-color);"
+
+with gr.Blocks(theme=gr.themes.Citrus(), css=CSS) as demo:
     gr.Markdown(DESCRIPTION)
     with gr.Row():
         with gr.Column(scale=1):