tune attn params

2024-08-02 04:18:47 -07:00
parent e5860a4999
commit 76e0d17af3
3 changed files with 15 additions and 30 deletions
@@ -22,7 +22,7 @@ if memory_management.xformers_enabled():
 FORCE_UPCAST_ATTENTION_DTYPE = memory_management.force_upcast_attention_dtype()


-def get_attn_precision(attn_precision):
+def get_attn_precision(attn_precision=torch.float32):
    if args.disable_attention_upcast:
        return None
    if FORCE_UPCAST_ATTENTION_DTYPE is not None: