Test ZeroGPU

app.py

@@ -1,35 +1,16 @@
-import gc, random
+import gc, random, os
 import gradio as gr
 import torch, spaces
 from diffusers import StableDiffusionPipeline, DPMSolverMultistepScheduler
 
-# --- gradio_client bool-schema hotfix (safe no-op if not needed) ---
-try:
-    import gradio_client.utils as _gcu
-    _orig_get_type = _gcu.get_type
-    def _get_type_safe(schema):
-        # Handle JSON Schema booleans (e.g., additionalProperties: True/False)
-        if isinstance(schema, bool):
-            return "any"
-        return _orig_get_type(schema)
-    _gcu.get_type = _get_type_safe
-except Exception:
-    pass
-# -------------------------------------------------------------------
-
-# sanity: show CPU at startup (no GPU yet)
-zero = torch.tensor([0.0])
-print("startup device:", zero.device)  # should print: cpu
-
-# ====== SD 1.5 minimal loader (lazy) ======
 MODEL_ID = "runwayml/stable-diffusion-v1-5"
 DTYPE = torch.float16
 _PIPE = None
 
 def get_pipe():
-    """Create the pipeline on first use (safe for ZeroGPU)."""
     global _PIPE
     if _PIPE is None:
+        # Build on CPU
         pipe = StableDiffusionPipeline.from_pretrained(
             MODEL_ID,
             torch_dtype=DTYPE,
@@ -37,11 +18,9 @@ def get_pipe():
             use_safetensors=True,
             low_cpu_mem_usage=True,
         )
-        # fast/stable scheduler
         pipe.scheduler = DPMSolverMultistepScheduler.from_config(
             pipe.scheduler.config, use_karras_sigmas=True, algorithm_type="dpmsolver++"
         )
-        # memory savers
         pipe.enable_attention_slicing()
         pipe.enable_vae_slicing()
         pipe.enable_model_cpu_offload()
@@ -52,20 +31,11 @@ def snap8(x: int) -> int:
     x = max(256, min(1024, int(x)))
     return x - (x % 8)
 
-@spaces.GPU  # tiny demo proving CUDA is present inside GPU block
-def greet(n: float):
-    print("inside greet, cuda available:", torch.cuda.is_available())
-    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-    t = zero.to(device) + torch.tensor([float(n)], device=device)
-    return f"Hello {t.item():.3f} Tensor (device: {t.device})"
-
-# SD 1.5 text -> image
 @spaces.GPU(duration=120)
 def generate(prompt: str, negative: str, steps: int, cfg: float, width: int, height: int, seed: int):
-    pipe = get_pipe()
+    pipe = get_pipe()           # stays CPU/offloaded until now
     w, h = snap8(width), snap8(height)
 
-    # seed
     if int(seed) < 0:
         seed = random.randint(0, 2**31 - 1)
     gen = torch.Generator(device="cuda").manual_seed(int(seed))
@@ -85,16 +55,10 @@ def generate(prompt: str, negative: str, steps: int, cfg: float, width: int, hei
         )
     return out.images[0]
 
-# ====== UI ======
 with gr.Blocks() as demo:
-    gr.Markdown("# ZeroGPU demo + Stable Diffusion 1.5 (minimal)")
+    gr.Markdown("# ZeroGPU + SD1.5 (minimal)")
 
-    with gr.Tab("ZeroGPU sanity"):
-        n = gr.Number(label="Add to zero", value=1.0)
-        hello = gr.Textbox(label="Result")
-        gr.Button("Greet").click(greet, n, hello)
-
-    with gr.Tab("SD 1.5: Text → Image"):
+    with gr.Tab("Text → Image"):
         prompt  = gr.Textbox(label="Prompt", value="a cozy reading nook, warm sunlight, cinematic lighting, highly detailed")
         negative = gr.Textbox(label="Negative (optional)", value="lowres, blurry, watermark, text")
         steps   = gr.Slider(8, 40, value=28, step=1, label="Steps")
@@ -103,10 +67,9 @@ with gr.Blocks() as demo:
         height  = gr.Slider(256, 1024, value=640, step=16, label="Height")
         seed    = gr.Number(value=-1, precision=0, label="Seed (-1 random)")
         out_img = gr.Image(label="Image", interactive=False)
-        gr.Button("Generate").click(
-            generate, [prompt, negative, steps, cfg, width, height, seed], out_img
-        )
+        gr.Button("Generate").click(generate, [prompt, negative, steps, cfg, width, height, seed], out_img)
 
 if __name__ == "__main__":
-    # On Spaces, expose a shareable URL & bind to 0.0.0.0
-    demo.queue(max_size=12).launch(share=False, show_error=True, mcp_server=True)
+    # On Spaces: keep it simple; don’t pass odd kwargs.
+    # If you see “localhost is not accessible”, add share=True.
+    demo.queue(max_size=12).launch(share=True)

requirements.txt

@@ -6,3 +6,5 @@ safetensors
 gradio==4.44.1
 huggingface-hub
 spaces
+qrcode[pil]
+opencv-python