torch with gpu

app.py

@@ -25,10 +25,7 @@ from illusion_style import css
 BASE_MODEL = "SG161222/Realistic_Vision_V5.1_noVAE"
 
 
-if torch.cuda.is_available():
-    device='gpu'
-else:
-    device='cpu'
+device='gpu'
 
 # Initialize both pipelines
 vae = AutoencoderKL.from_pretrained("stabilityai/sd-vae-ft-mse", torch_dtype=torch.float16)
@@ -42,6 +39,163 @@ main_pipe = StableDiffusionControlNetPipeline.from_pretrained(
     torch_dtype=torch.float16,
 ).to(device)
 
+#main_pipe.unet = torch.compile(main_pipe.unet, mode="reduce-overhead", fullgraph=True)
+#main_pipe.unet.to(memory_format=torch.channels_last)
+#main_pipe.unet = torch.compile(main_pipe.unet, mode="reduce-overhead", fullgraph=True)
+#model_id = "stabilityai/sd-x2-latent-upscaler"
+image_pipe = StableDiffusionControlNetImg2ImgPipeline(**main_pipe.components)
+
+
+#image_pipe.unet = torch.compile(image_pipe.unet, mode="reduce-overhead", fullgraph=True)
+#upscaler = StableDiffusionLatentUpscalePipeline.from_pretrained(model_id, torch_dtype=torch.float16)
+#upscaler.to("cuda")
+
+
+# Sampler map
+SAMPLER_MAP = {
+    "DPM++ Karras SDE": lambda config: DPMSolverMultistepScheduler.from_config(config, use_karras=True, algorithm_type="sde-dpmsolver++"),
+    "Euler": lambda config: EulerDiscreteScheduler.from_config(config),
+}
+
+def center_crop_resize(img, output_size=(512, 512)):
+    width, height = img.size
+
+    # Calculate dimensions to crop to the center
+    new_dimension = min(width, height)
+    left = (width - new_dimension)/2
+    top = (height - new_dimension)/2
+    right = (width + new_dimension)/2
+    bottom = (height + new_dimension)/2
+
+    # Crop and resize
+    img = img.crop((left, top, right, bottom))
+    img = img.resize(output_size)
+
+    return img
+
+def common_upscale(samples, width, height, upscale_method, crop=False):
+        if crop == "center":
+            old_width = samples.shape[3]
+            old_height = samples.shape[2]
+            old_aspect = old_width / old_height
+            new_aspect = width / height
+            x = 0
+            y = 0
+            if old_aspect > new_aspect:
+                x = round((old_width - old_width * (new_aspect / old_aspect)) / 2)
+            elif old_aspect < new_aspect:
+                y = round((old_height - old_height * (old_aspect / new_aspect)) / 2)
+            s = samples[:,:,y:old_height-y,x:old_width-x]
+        else:
+            s = samples
+
+        return torch.nn.functional.interpolate(s, size=(height, width), mode=upscale_method)
+
+def upscale(samples, upscale_method, scale_by):
+        #s = samples.copy()
+        width = round(samples["images"].shape[3] * scale_by)
+        height = round(samples["images"].shape[2] * scale_by)
+        s = common_upscale(samples["images"], width, height, upscale_method, "disabled")
+        return (s)
+
+def check_inputs(prompt: str, control_image: Image.Image):
+    if control_image is None:
+        raise gr.Error("Please select or upload an Input Illusion")
+    if prompt is None or prompt == "":
+        raise gr.Error("Prompt is required")
+
+def convert_to_pil(base64_image):
+    pil_image = Image.open(base64_image)
+    return pil_image
+
+def convert_to_base64(pil_image):
+    with tempfile.NamedTemporaryFile(suffix='.png', delete=False) as temp_file:
+        image.save(temp_file.name)
+    return temp_file.name
+
+# Inference function
+@spaces.GPU
+def inference(
+    control_image: Image.Image,
+    prompt: str,
+    negative_prompt: str,
+    guidance_scale: float = 8.0,
+    controlnet_conditioning_scale: float = 1,
+    control_guidance_start: float = 1,    
+    control_guidance_end: float = 1,
+    upscaler_strength: float = 0.5,
+    seed: int = -1,
+    sampler = "DPM++ Karras SDE",
+    progress = gr.Progress(track_tqdm=True),
+    profile: gr.OAuthProfile | None = None,
+):
+    start_time = time.time()
+    start_time_struct = time.localtime(start_time)
+    start_time_formatted = time.strftime("%H:%M:%S", start_time_struct)
+    print(f"Inference started at {start_time_formatted}")
+    
+    # Generate the initial image
+    #init_image = init_pipe(prompt).images[0]
+
+    # Rest of your existing code
+    control_image_small = center_crop_resize(control_image)
+    control_image_large = center_crop_resize(control_image, (1024, 1024))
+
+    main_pipe.scheduler = SAMPLER_MAP[sampler](main_pipe.scheduler.config)
+    my_seed = random.randint(0, 2**32 - 1) if seed == -1 else seed
+    generator = torch.Generator(device=device).manual_seed(my_seed)
+    
+    out = main_pipe(
+        prompt=prompt,
+        negative_prompt=negative_prompt,
+        image=control_image_small,
+        guidance_scale=float(guidance_scale),
+        controlnet_conditioning_scale=float(controlnet_conditioning_scale),
+        generator=generator,
+        control_guidance_start=float(control_guidance_start),
+        control_guidance_end=float(control_guidance_end),
+        num_inference_steps=15,
+        output_type="latent"
+    )
+    upscaled_latents = upscale(out, "nearest-exact", 2)
+    out_image = image_pipe(
+        prompt=prompt,
+        negative_prompt=negative_prompt,
+        control_image=control_image_large,        
+        image=upscaled_latents,
+        guidance_scale=float(guidance_scale),
+        generator=generator,
+        num_inference_steps=20,
+        strength=upscaler_strength,
+        control_guidance_start=float(control_guidance_start),
+        control_guidance_end=float(control_guidance_end),
+        controlnet_conditioning_scale=float(controlnet_conditioning_scale)
+    )
+    end_time = time.time()
+    end_time_struct = time.localtime(end_time)
+    end_time_formatted = time.strftime("%H:%M:%S", end_time_struct)
+    print(f"Inference ended at {end_time_formatted}, taking {end_time-start_time}s")
+
+    # Save image + metadata
+    user_history.save_image(
+        label=prompt,
+        image=out_image["images"][0],
+        profile=profile,
+        metadata={
+            "prompt": prompt,
+            "negative_prompt": negative_prompt,
+            "guidance_scale": guidance_scale,
+            "controlnet_conditioning_scale": controlnet_conditioning_scale,
+            "control_guidance_start": control_guidance_start,
+            "control_guidance_end": control_guidance_end,
+            "upscaler_strength": upscaler_strength,
+            "seed": seed,
+            "sampler": sampler,
+        },
+    )
+
+    return out_image["images"][0], gr.update(visible=True), gr.update(visible=True), my_seed
+
 def greet(name):
     return "Hello " + name + "!!"