mixed precision

app.py

@@ -5,181 +5,63 @@ import spaces
 import torch
 from diffusers import FluxPipeline
 
-# Check for CUDA and set device
+# Enable cuDNN benchmarking for potential performance improvement
+torch.backends.cudnn.benchmark = True
+
+# Set up device and data types
 device = "cuda" if torch.cuda.is_available() else "cpu"
-print(f"Using device: {device}")
+DTYPE = torch.float16
 
 # Load the model
-pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell", torch_dtype=torch.float16)
-pipe = pipe.to(device)
+pipe = FluxPipeline.from_pretrained(
+    "black-forest-labs/FLUX.1-schnell",
+    torch_dtype=torch.bfloat16,
+)
 
-# Convert text encoders to full precision
-pipe.text_encoder = pipe.text_encoder.to(torch.float32)
-if hasattr(pipe, 'text_encoder_2'):
-    pipe.text_encoder_2 = pipe.text_encoder_2.to(torch.float32)
+# Configure the pipeline
+pipe.enable_sequential_cpu_offload()
+pipe.vae.enable_tiling()
+pipe = pipe.to(DTYPE)
 
 MAX_SEED = np.iinfo(np.int32).max
 MAX_IMAGE_SIZE = 2048
 
 @spaces.GPU()
 def infer(prompt, seed=42, randomize_seed=False, width=1024, height=1024, num_inference_steps=4, progress=gr.Progress(track_tqdm=True)):
-    try:
-        if randomize_seed:
-            seed = random.randint(0, MAX_SEED)
-        generator = torch.Generator(device=device).manual_seed(seed)
-        
-        # Use full precision for text encoding
-        with torch.no_grad():
-            text_inputs = pipe.tokenizer(prompt, return_tensors="pt").to(device)
-            text_embeddings = pipe.text_encoder(text_inputs.input_ids)[0]
-        
-        # Use mixed precision for the rest of the pipeline
-        with torch.autocast(device_type=device, dtype=torch.float16):
-            image = pipe(
-                prompt_embeds=text_embeddings,
-                width=width,
-                height=height,
-                num_inference_steps=num_inference_steps,
-                generator=generator,
-                guidance_scale=0.0
-            ).images[0]
-        
-        return image, seed
-    except Exception as e:
-        print(f"Error during inference: {e}")
-        return None, seed
- 
-examples = [
-    "a tiny astronaut hatching from an egg on the moon",
-    "a cat holding a sign that says hello world",
-    "an anime illustration of a wiener schnitzel",
-]
-
-css = """
-#col-container {
-    margin: 0 auto;
-    max-width: 720px;
-}
-.container {
-    margin: 0 auto;
-    padding: 20px;
-    border-radius: 10px;
-    background-color: #f0f0f0;
-}
-.title {
-    text-align: center;
-    color: #2c3e50;
-    margin-bottom: 20px;
-}
-.subtitle {
-    text-align: center;
-    color: #34495e;
-    margin-bottom: 30px;
-}
-.speed-info {
-    background-color: #e74c3c;
-    color: white;
-    padding: 10px;
-    border-radius: 5px;
-    text-align: center;
-    margin-bottom: 20px;
-}
-.prompt-container {
-    display: flex;
-    gap: 10px;
-    margin-bottom: 20px;
-}
-.advanced-settings {
-    background-color: #ecf0f1;
-    padding: 15px;
-    border-radius: 5px;
-    margin-top: 20px;
-}
-"""
+    if randomize_seed:
+        seed = random.randint(0, MAX_SEED)
+    generator = torch.Generator(device=device).manual_seed(seed)
+    
+    image = pipe(
+        prompt,
+        num_inference_steps=num_inference_steps,
+        num_images_per_prompt=1,
+        guidance_scale=0.0,
+        height=height,
+        width=width,
+        generator=generator,
+    ).images[0]
+    
+    return image, seed
 
-with gr.Blocks(css=css) as demo:
-    with gr.Column(elem_id="col-container"):
-        gr.HTML(
-            """
-            <div class="container">
-                <h1 class="title">FLUX.1 [schnell] - Mixed Precision Edition</h1>
-                <h3 class="subtitle">12B param rectified flow transformer optimized for maximum inference speed</h3>
-                <div class="speed-info">
-                    <strong>Mixed Precision Pipeline:</strong> FP32 Text Encoders + FP16 Core for optimal speed and quality
-                </div>
-            </div>
-            """
-        )
-        
-        with gr.Column(elem_id="prompt-container"):
-            prompt = gr.Text(
-                label="Enter your prompt",
-                placeholder="A futuristic cityscape with flying cars",
-                lines=2
-            )
-            run_button = gr.Button("Generate Image", variant="primary")
-        
-        result = gr.Image(label="Generated Image")
-        
-        with gr.Accordion("Advanced Settings", open=False):
-            with gr.Column(elem_id="advanced-settings"):
-                seed = gr.Slider(
-                    label="Seed",
-                    minimum=0,
-                    maximum=MAX_SEED,
-                    step=1,
-                    value=0,
-                    info="Set to 0 for random seed"
-                )
-                randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
-                
-                with gr.Row():
-                    width = gr.Slider(
-                        label="Width",
-                        minimum=256,
-                        maximum=MAX_IMAGE_SIZE,
-                        step=32,
-                        value=1024,
-                    )
-                    height = gr.Slider(
-                        label="Height",
-                        minimum=256,
-                        maximum=MAX_IMAGE_SIZE,
-                        step=32,
-                        value=1024,
-                    )
-                
-                num_inference_steps = gr.Slider(
-                    label="Number of inference steps",
-                    minimum=1,
-                    maximum=50,
-                    step=1,
-                    value=4,
-                    info="Lower values = faster generation, higher values = potentially better quality"
-                )
-        
-        gr.Markdown(
-            """
-            ### About FLUX.1 [schnell]
-            - Distilled from [FLUX.1 [pro]](https://blackforestlabs.ai/)
-            - Optimized for 4-step generation
-            - Mixed precision pipeline for maximum speed
-            
-            [[Blog]](https://blackforestlabs.ai/announcing-black-forest-labs/) | [[Model]](https://huggingface.co/black-forest-labs/FLUX.1-schnell)
-            """
-        )
-        
-        gr.Examples(
-            examples=examples,
-            fn=infer,
-            inputs=[prompt],
-            outputs=[result, seed],
-            cache_examples="lazy"
-        )
+# Gradio interface
+with gr.Blocks() as demo:
+    gr.Markdown("# FLUX.1 [schnell] Image Generator")
+    with gr.Row():
+        with gr.Column():
+            prompt = gr.Textbox(label="Prompt")
+            run_button = gr.Button("Generate")
+        with gr.Column():
+            result = gr.Image(label="Generated Image")
+    with gr.Accordion("Advanced Settings", open=False):
+        seed = gr.Slider(minimum=0, maximum=MAX_SEED, step=1, label="Seed", randomize=True)
+        randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
+        width = gr.Slider(minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=1024, label="Width")
+        height = gr.Slider(minimum=256, maximum=MAX_IMAGE_SIZE, step=32, value=1024, label="Height")
+        num_inference_steps = gr.Slider(minimum=1, maximum=50, step=1, value=4, label="Number of inference steps")
     
-    gr.on(
-        triggers=[run_button.click, prompt.submit],
-        fn=infer,
+    run_button.click(
+        infer,
         inputs=[prompt, seed, randomize_seed, width, height, num_inference_steps],
         outputs=[result, seed]
     )