app.py

import spaces
import os
import sys
import site

if "--use-sage-attention" not in sys.argv:
    sys.argv.append("--use-sage-attention")
    print("🚀 [SageAttention] Injected '--use-sage-attention' into sys.argv.")

APP_DIR = os.path.dirname(os.path.abspath(__file__))
if APP_DIR not in sys.path:
    sys.path.insert(0, APP_DIR)
    print(f"✅ Added project root '{APP_DIR}' to sys.path.")

SAGE_PATCH_APPLIED = False

def apply_sage_attention_patch():
    global SAGE_PATCH_APPLIED
    if SAGE_PATCH_APPLIED:
        return "SageAttention patch already applied."

    try:
        from comfy import model_management
        import sageattention
        print("--- [Runtime Patch] sageattention package found. Applying patch... ---")
        # Disable SageAttention to avoid runtime errors on CPU‑only setups.
        if hasattr(model_management, "sage_attention_enabled"):
            model_management.sage_attention_enabled = lambda: False
        if hasattr(model_management, "pytorch_attention_enabled"):
            model_management.pytorch_attention_enabled = lambda: True
        SAGE_PATCH_APPLIED = True
        return "✅ Successfully disabled SageAttention (using standard PyTorch attention)."
    except Exception as e:
        SAGE_PATCH_APPLIED = False
        msg = f"--- [Runtime Patch] ⚠️ SageAttention patch could not be applied: {e} ---"
        print(msg)
        return msg
    except Exception as e:
        SAGE_PATCH_APPLIED = False
        msg = f"--- [Runtime Patch] ❌ An error occurred while applying SageAttention patch: {e} ---"
        print(msg)
        return msg

# @spaces.cpu (disabled for CPU)
def dummy_gpu_for_startup():
    try:
        print("--- [GPU Startup] Dummy function for startup check initiated. ---")
        patch_result = apply_sage_attention_patch()
        print(f"--- [GPU Startup] {patch_result} ---")
        print("--- [GPU Startup] Startup check passed. ---")
        return "Startup check passed."
    except BaseException as e:
        err_msg = str(e)
        if "uncorrectable ECC error" in err_msg or "cudaErrorECCUncorrectable" in err_msg:
            print("\n" + "="*80)
            print(f"🚨 [Fatal GPU Error] Captured uncorrectable ECC error during inference: {err_msg}")
            print("🚨 Terminating process to trigger an automatic container restart...")
            print("="*80 + "\n")
            os._exit(1)
        raise e


def main():
    from comfy_integration import setup as setup_comfyui
    from utils.app_utils import load_ipadapter_presets
    
    print("--- [Setup] Starting ComfyUI initialization ---")
    setup_comfyui.initialize_comfyui()
    
    print("--- [Setup] Applying SageAttention Runtime Patch ---")
    patch_result = apply_sage_attention_patch()
    print(f"--- [Setup] {patch_result} ---")

    print("--- [Setup] Reloading site-packages to detect newly installed packages... ---")
    try:
        site.main()
        print("--- [Setup] ✅ Site-packages reloaded. ---")
    except Exception as e:
        print(f"--- [Setup] ⚠️  Warning: Could not fully reload site-packages: {e} ---")

    print("--- Initiating GPU Startup Check & SageAttention Patch Verification ---")
    try:
        dummy_gpu_for_startup()
    except Exception as e:
        print(f"--- [GPU Startup] ⚠️ Warning: Startup check failed: {e} ---")

    print("--- Starting Application Setup ---")
    
    print("--- Loading IPAdapter presets ---")
    load_ipadapter_presets()
    print("--- ✅ IPAdapter setup complete. ---")


    print("--- Environment configured. Proceeding with module imports. ---")
    from ui.layout import build_ui
    from ui.events import attach_event_handlers

    print(f"✅ Working directory is stable: {os.getcwd()}")

    demo = build_ui(attach_event_handlers)
        
    print("--- Launching Gradio Interface ---")
    demo.queue().launch(server_name="0.0.0.0", server_port=7860)


if __name__ == "__main__":
    main()