Create app.py

app.py

@@ -0,0 +1,111 @@
+import os
+import gc
+import time
+import random
+import torch
+import gradio as gr
+
+from diffusers import DiffusionPipeline
+
+# =========================
+# HARD CPU MODE
+# =========================
+os.environ["CUDA_VISIBLE_DEVICES"] = ""
+os.environ["HF_HUB_DISABLE_TELEMETRY"] = "1"
+os.environ["TOKENIZERS_PARALLELISM"] = "false"
+
+cpu_cores = os.cpu_count() or 1
+torch.set_num_threads(cpu_cores)
+torch.set_num_interop_threads(cpu_cores)
+
+os.environ["OMP_NUM_THREADS"] = str(cpu_cores)
+os.environ["MKL_NUM_THREADS"] = str(cpu_cores)
+
+torch.backends.mkldnn.enabled = True
+
+device = torch.device("cpu")
+dtype = torch.bfloat16 if torch.cpu.is_bf16_supported() else torch.float32
+
+MODEL_ID = "tensorart/stable-diffusion-3.5-medium-turbo"
+CACHE_DIR = "models"
+
+# =========================
+# LOAD PIPELINE
+# =========================
+def load_pipeline():
+    pipe = DiffusionPipeline.from_pretrained(
+        MODEL_ID,
+        torch_dtype=dtype,
+        cache_dir=CACHE_DIR,
+        low_cpu_mem_usage=True
+    )
+
+    pipe.enable_attention_slicing()
+    pipe.enable_vae_slicing()
+    pipe.enable_sequential_cpu_offload()
+
+    pipe = pipe.to(device)
+
+    return pipe
+
+pipe = load_pipeline()
+
+# =========================
+# GENERATION
+# =========================
+def generate(prompt, seed, progress=gr.Progress()):
+    if not prompt:
+        raise gr.Error("Prompt required")
+
+    if seed < 0:
+        seed = random.randint(0, 2**31 - 1)
+
+    generator = torch.Generator(device=device).manual_seed(seed)
+
+    steps = 6
+    width = 512
+    height = 512
+
+    start = time.time()
+
+    def callback(step, timestep, latents):
+        done = step + 1
+        elapsed = time.time() - start
+        eta = (elapsed / done) * (steps - done)
+        progress(done / steps, desc=f"Step {done}/{steps} | ETA {eta:.1f}s")
+
+    with torch.inference_mode():
+        gc.collect()
+        image = pipe(
+            prompt=prompt,
+            width=width,
+            height=height,
+            num_inference_steps=steps,
+            guidance_scale=0.0,
+            generator=generator,
+            callback=callback,
+            callback_steps=1
+        ).images[0]
+        gc.collect()
+
+    return image, seed
+
+# =========================
+# UI
+# =========================
+with gr.Blocks(title="SD 3.5 Medium Turbo CPU Ultra Lean") as demo:
+    gr.Markdown("# Stable Diffusion 3.5 Medium Turbo — 16GB CPU Mode")
+
+    prompt = gr.Textbox(label="Prompt", lines=3)
+    seed = gr.Number(label="Seed (-1 random)", value=-1, precision=0)
+    btn = gr.Button("Generate")
+
+    image_out = gr.Image()
+    seed_out = gr.Number(interactive=False)
+
+    btn.click(generate, inputs=[prompt, seed], outputs=[image_out, seed_out])
+
+demo.queue(max_size=5, concurrency_count=1)
+
+if __name__ == "__main__":
+    demo.launch(server_name="0.0.0.0", server_port=7860)