Update app.py

app.py

@@ -2,7 +2,6 @@ import os
 import subprocess
 import sys
 import io
-from kernels import get_kernel
 import gradio as gr
 import numpy as np
 import random
@@ -38,14 +37,11 @@ def remote_text_encoder(prompts):
     return prompt_embeds
 
 # Load model
-fa3_kernel = get_kernel("kernels-community/flash-attn3", revision="fake-ops-return-probs")
-
 repo_id = "black-forest-labs/FLUX.2-dev"
 
 dit = Flux2Transformer2DModel.from_pretrained(
     repo_id,
     subfolder="transformer",
-    attn_implementation=fa3_kernel,
     torch_dtype=torch.bfloat16
 )
 
@@ -56,16 +52,18 @@ pipe = Flux2Pipeline.from_pretrained(
     torch_dtype=torch.bfloat16
 )
 pipe.to("cuda")
-# pipe.transformer.compile_repeated_blocks(dynamic=True)
-
-#optimize_pipeline_(
-#    pipe,
-#    prompt_embeds=remote_text_encoder("prompt").to("cuda"),
-#    guidance_scale=2.5,
-#    width=1024,
-#    height=1024,
-#    num_inference_steps=1,
-#)
+
+pipe.transformer.set_attention_backend("_flash_3_hub")
+
+optimize_pipeline_(
+    pipe,
+    image=[Image.new("RGB", (1024, 1024))],
+    prompt_embeds = remote_text_encoder("prompt").to("cuda"),
+    guidance_scale=2.5,
+    width=1024,
+    height=1024,
+    num_inference_steps=1
+)
 
 @spaces.GPU(duration=180)
 def infer(prompt, input_images, seed=42, randomize_seed=False, width=1024, height=1024, num_inference_steps=50, guidance_scale=2.5, progress=gr.Progress(track_tqdm=True)):