diff --git a/comfy/model_management.py b/comfy/model_management.py
index 212ce9af2..dd8a2a28f 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -236,6 +236,19 @@ try:
 except:
     pass
 
+
+try:
+    if is_amd():
+        arch = torch.cuda.get_device_properties(get_torch_device()).gcnArchName
+        logging.info("AMD arch: {}".format(arch))
+        if args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
+            if int(torch_version[0]) >= 2 and int(torch_version[2]) >= 7:  # works on 2.6 but doesn't actually seem to improve much
+                if arch in ["gfx1100"]: #TODO: more arches
+                    ENABLE_PYTORCH_ATTENTION = True
+except:
+    pass
+
+
 if ENABLE_PYTORCH_ATTENTION:
     torch.backends.cuda.enable_math_sdp(True)
     torch.backends.cuda.enable_flash_sdp(True)