Add a command line option to disable upcasting in some cross attention ops.

3661e106 · comfyanonymous · 50db297c · 3661e106
Commit 3661e106 authored Jan 29, 2023 by comfyanonymous
Show whitespace changes
Inline Side-by-side

Showing with 4 additions and 0 deletions

main.py main.py +4 -0

No files found.
--- a/main.py
+++ b/main.py
@@ -6,6 +6,10 @@ import threading
 import queue
 import traceback
+if '--dont-upcast-attention' in sys.argv:
+    print("disabling upcasting of attention")
+    os.environ['ATTN_PRECISION'] = "fp16"
 import torch
 import nodes