From 8ae3b5bb43b8118c8f5e990dbc0f84f55d3c0360 Mon Sep 17 00:00:00 2001
From: Seunghoon Lee <lshqqytiger@naver.com>
Date: Sat, 17 Feb 2024 20:56:10 +0900
Subject: [PATCH] remove zluda force sync

---
 modules/shared.py |  2 +-
 modules/zluda.py  | 18 ++++--------------
 2 files changed, 5 insertions(+), 15 deletions(-)

diff --git a/modules/shared.py b/modules/shared.py
index 412ccfdbe..a192d9b9b 100644
--- a/modules/shared.py
+++ b/modules/shared.py
@@ -404,7 +404,7 @@ options_templates.update(options_section(('cuda', "Compute Settings"), {
     "olive_cache_optimized": OptionInfo(True, 'Olive cache optimized models'),
 
     "zluda_sep": OptionInfo("<h2>ZLUDA</h2>(experimental)", "", gr.HTML, {"visible": devices.backend == "cuda"}),
-    "zluda_force_sync": OptionInfo(False, 'ZLUDA force synchronization (experimental, restart required)', gr.Checkbox, {"visible": devices.backend == "cuda"}),
+    "zluda_enable_cudnn": OptionInfo(False, 'ZLUDA enable cuDNN (restart required)', gr.Checkbox, {"visible": devices.backend == "cuda"}),
 }))
 
 options_templates.update(options_section(('advanced', "Inference Settings"), {
diff --git a/modules/zluda.py b/modules/zluda.py
index 48f5dee8a..e2d7d15bb 100644
--- a/modules/zluda.py
+++ b/modules/zluda.py
@@ -1,23 +1,13 @@
 import platform
 import torch
-from diffusers.models.attention_processor import AttnProcessor
 from modules import shared, devices
 
 
 def initialize_zluda():
     if platform.system() == "Windows" and devices.cuda_ok and torch.cuda.get_device_name(devices.get_optimal_device()).endswith("[ZLUDA]"):
         shared.log.warning("Detected ZLUDA device. Currently, ZLUDA support is experimental and unstable.")
-        torch.backends.cudnn.enabled = False
+        torch.backends.cudnn.enabled = shared.opts.zluda_enable_cudnn
+        if torch.backends.cudnn.enabled:
+            shared.log.warning("cuDNN with ZLUDA won't work at this moment. Please wait for future update.")
         if shared.opts.cross_attention_optimization == "Scaled-Dot-Product":
-            shared.opts.cross_attention_optimization = "Batch matrix-matrix"
-        if shared.opts.zluda_force_sync:
-            patch_attention_processor(AttnProcessor)
-
-
-def patch_attention_processor(cls):
-    forward = cls.__call__
-    def patched(self, *args, **kwargs):
-        R = forward(self, *args, **kwargs)
-        torch.cuda.synchronize()
-        return R
-    cls.__call__ = patched
+            shared.log.warning("ZLUDA does not support Scaled-Dot-Product attention. Please consider changing it to Batch matrix-matrix.")