Update app.py

app.py

@@ -7,7 +7,7 @@ import time
 from diffusers import DiffusionPipeline, AutoencoderTiny
 from diffusers.models.attention_processor import AttnProcessor2_0
 from custom_pipeline import FluxWithCFGPipeline
-from huggingface_hub import hf_hub_download
+from huggingface_hub import login
 
 torch.backends.cuda.matmul.allow_tf32 = True
 
@@ -20,123 +20,51 @@ DEFAULT_INFERENCE_STEPS = 1
 
 # Device and model setup
 dtype = torch.float16
-device = "cuda"  # Explicitly set device to CUDA
-
-# Download the LoRA weights
-lora_weights_path = hf_hub_download(
-    repo_id="hugovntr/flux-schnell-realism",
-    filename="schnell-realism_v2.3.safetensors",
-)
-
 pipe = FluxWithCFGPipeline.from_pretrained(
-    "black-forest-labs/FLUX.1-schnell", torch_dtype=dtype
+    "black-forest-labs/FLUX.1-schnell", torch_dtype=dtype, use_safetensors=True, variant="fp16"
 )
-pipe.vae = AutoencoderTiny.from_pretrained("madebyollin/taef1", torch_dtype=dtype)
-pipe.to(device)  # Move the pipeline to CUDA
-
-# Load the LoRA weights
-pipe.load_lora_weights(lora_weights_path, adapter_name="better")
+pipe.vae = AutoencoderTiny.from_pretrained("madebyollin/taef1", torch_dtype=dtype, use_safetensors=True, variant="fp16")
+pipe.to("cuda")
+pipe.load_lora_weights('hugovntr/flux-schnell-realism', weight_name='schnell-realism_v2.3.safetensors', adapter_name="better")
 pipe.set_adapters(["better"], adapter_weights=[1.0])
 pipe.fuse_lora(adapter_name=["better"], lora_scale=1.0)
 pipe.unload_lora_weights()
 
-# Memory optimizations
-pipe.transformer.to(memory_format=torch.channels_last)
+# Enable xformers
 pipe.enable_xformers_memory_efficient_attention()
 
-# CUDA Graph setup
-static_inputs = None
-static_model = None
-graph = None
-
-def setup_cuda_graph(prompt, height, width, num_inference_steps):
-    global static_inputs, static_model, graph
-
-    batch_size = 1 if isinstance(prompt, str) else len(prompt)
-    num_images_per_prompt = 1
-
-    prompt_embeds, pooled_prompt_embeds, text_ids = pipe.encode_prompt(
-        prompt=prompt,
-        prompt_2=None,
-        prompt_embeds=None,
-        pooled_prompt_embeds=None,
-        device=device,
-        num_images_per_prompt=num_images_per_prompt,
-        max_sequence_length=300,
-        lora_scale=None,
-    )
-
-    latents, latent_image_ids = pipe.prepare_latents(
-        batch_size * num_images_per_prompt,
-        pipe.transformer.config.in_channels // 4,
-        height,
-        width,
-        prompt_embeds.dtype,
-        device,
-        None,
-        None,
-    )
-    sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
-    image_seq_len = latents.shape[1]
-    mu = calculate_timestep_shift(image_seq_len)
-
-    timesteps, num_inference_steps = prepare_timesteps(
-        pipe.scheduler,
-        num_inference_steps,
-        device,
-        None,
-        sigmas,
-        mu=mu,
-    )
-
-    guidance = torch.full([1], 3.5, device=device, dtype=torch.float16).expand(latents.shape[0]) if pipe.transformer.config.guidance_embeds else None
-
-    static_inputs = {
-        "hidden_states": latents.to(device),
-        "timestep": timesteps.to(device),
-        "guidance": guidance.to(device) if guidance is not None else None,
-        "pooled_projections": pooled_prompt_embeds.to(device),
-        "encoder_hidden_states": prompt_embeds.to(device),
-        "txt_ids": text_ids,
-        "img_ids": latent_image_ids,
-        "joint_attention_kwargs": None,
-    }
+# Compile the model (Optional, needs further testing for stability)
+# pipe.transformer = torch.compile(pipe.transformer, mode="reduce-overhead", fullgraph=True)
 
-    static_model = torch.cuda.make_graphed_callables(pipe.transformer, (static_inputs,))
-    graph = torch.cuda.CUDAGraph()
+# Capture CUDA Graph (Warm-up)
+static_inputs = {
+    "prompt": "warmup",
+    "width": DEFAULT_WIDTH,
+    "height": DEFAULT_HEIGHT,
+    "num_inference_steps": DEFAULT_INFERENCE_STEPS,
+    "generator": torch.Generator().manual_seed(0),
+}
 
-    with torch.cuda.graph(graph):
-        static_output = static_model(**static_inputs)
+pipe.capture_cuda_graph(**static_inputs)
+torch.cuda.empty_cache()
 
 # Inference function
-# @spaces.GPU(duration=25) # Remove decorator
+@spaces.GPU(duration=25)
 def generate_image(prompt, seed=24, width=DEFAULT_WIDTH, height=DEFAULT_HEIGHT, randomize_seed=False, num_inference_steps=2, progress=gr.Progress(track_tqdm=True)):
-    global static_inputs, graph
-
     if randomize_seed:
         seed = random.randint(0, MAX_SEED)
     generator = torch.Generator().manual_seed(int(float(seed)))
 
     start_time = time.time()
-    
-    if static_inputs is None:
-        setup_cuda_graph(prompt, height, width, num_inference_steps)
-
-    static_inputs["hidden_states"].copy_(pipe.prepare_latents(
-        1,
-        pipe.transformer.config.in_channels // 4,
-        height,
-        width,
-        static_inputs["encoder_hidden_states"].dtype,
-        "cuda",
-        generator,
-        None,
-    )[0])
 
-    graph.replay()
-    latents = static_inputs["hidden_states"]
-
-    img = pipe._decode_latents_to_image(latents, height, width, "pil")
+    # Only generate the last image in the sequence
+    img = pipe.generate_images( 
+            prompt=prompt,
+            width=width,
+            height=height,
+            num_inference_steps=num_inference_steps,
+            generator=generator
+        )
     latency = f"Latency: {(time.time()-start_time):.2f} seconds"    
     return img, seed, latency
 
@@ -159,7 +87,7 @@ with gr.Blocks() as demo:
         gr.Markdown("<span style='color: red;'>Note: Sometimes it stucks or stops generating images (I don't know why). In that situation just refresh the site.</span>")
 
         with gr.Row():
-            with gr.Column(scale=2): # Changed scale to 2
+            with gr.Column(scale=2.5):
                 result = gr.Image(label="Generated Image", show_label=False, interactive=False)
             with gr.Column(scale=1):
                 prompt = gr.Text(
@@ -192,8 +120,7 @@ with gr.Blocks() as demo:
                 fn=generate_image,
                 inputs=[prompt],
                 outputs=[result, seed, latency],
-                cache_examples=True, # Changed cache_examples
-                cache_mode="lazy" # Added cache_mode
+                cache_examples="lazy" 
             )
 
     enhanceBtn.click(
@@ -253,4 +180,4 @@ with gr.Blocks() as demo:
         )
 
 # Launch the app
-demo.queue(max_size=5).launch() # Removed concurrency_count
+demo.launch()