Update app.py

app.py

@@ -1,19 +1,54 @@
-import gradio as gr
+import os
+
+# ===== CPU / Thread 最佳化（關鍵）=====
+CPU_THREADS = str(os.cpu_count() or 2)
+
+os.environ["OMP_NUM_THREADS"] = CPU_THREADS
+os.environ["OPENVINO_NUM_THREADS"] = CPU_THREADS
+os.environ["MKL_NUM_THREADS"] = CPU_THREADS
+os.environ["NUMEXPR_NUM_THREADS"] = CPU_THREADS
+
+# 避免 thread oversubscription
+os.environ["OMP_WAIT_POLICY"] = "PASSIVE"
+os.environ["KMP_BLOCKTIME"] = "0"
+
 import torch
+import gradio as gr
 from optimum.intel import OVZImagePipeline
 
-# ===== Load model (only once) =====
-pipe = OVZImagePipeline.from_pretrained(
-    "hsuwill000/Z-Image-Turbo-ov",
-    device="cpu"
-)
+# ===== Global objects（避免重複初始化）=====
+pipe = None
+generator = torch.Generator("cpu")
+
+
+# ===== Load model（lazy + only once）=====
+def load_model():
+    global pipe
 
-# ===== Inference function =====
+    if pipe is None:
+        pipe = OVZImagePipeline.from_pretrained(
+            "hsuwill000/Z-Image-Turbo-ov",
+            device="cpu"
+        )
+
+        # ===== Warmup（避免第一次慢）=====
+        pipe(
+            prompt="warmup",
+            height=512,
+            width=512,
+            num_inference_steps=1,
+            guidance_scale=0.0,
+        )
+
+    return pipe
+
+
+# ===== Inference =====
 def generate_image(prompt, height, width, steps, seed):
-    if seed == -1:
-        generator = torch.Generator("cpu")
-    else:
-        generator = torch.Generator("cpu").manual_seed(int(seed))
+    pipe = load_model()
+
+    if seed != -1:
+        generator.manual_seed(int(seed))
 
     image = pipe(
         prompt=prompt,
@@ -31,24 +66,28 @@ def generate_image(prompt, height, width, steps, seed):
 default_prompt = """Cinematic portrait of a Japanese girl as Shinobu Kocho, realistic facial features, gradient purple compound eyes, intricate butterfly hair ornament, wearing silk haori with butterfly wing patterns, soft moonlight, hyper-realistic, depth of field, purple butterfly particles, 8k resolution, ethereal lighting"""
 
 
-# ===== Gradio UI =====
+# ===== UI =====
 with gr.Blocks() as demo:
-    gr.Markdown("# 🚀 Z-Image Turbo (OpenVINO CPU)")
+    gr.Markdown("## 🚀 Z-Image Turbo (OpenVINO CPU Optimized)")
 
     with gr.Row():
         with gr.Column():
-            prompt = gr.Textbox(label="Prompt", value=default_prompt, lines=5)
+            prompt = gr.Textbox(
+                label="Prompt",
+                value=default_prompt,
+                lines=4
+            )
 
             height = gr.Slider(256, 1024, value=512, step=64, label="Height")
             width = gr.Slider(256, 1024, value=512, step=64, label="Width")
 
-            steps = gr.Slider(1, 20, value=9, step=1, label="Inference Steps")
+            steps = gr.Slider(1, 20, value=9, step=1, label="Steps")
             seed = gr.Number(value=-1, label="Seed (-1 = random)")
 
             run_btn = gr.Button("Generate")
 
         with gr.Column():
-            output = gr.Image(label="Generated Image")
+            output = gr.Image(label="Result")
 
     run_btn.click(
         fn=generate_image,
@@ -56,6 +95,10 @@ with gr.Blocks() as demo:
         outputs=output,
     )
 
+
+# ===== Queue（最小化延遲）=====
+demo.queue(concurrency_count=1, max_size=2)
+
 # ===== Launch =====
 if __name__ == "__main__":
     demo.launch()