From f73e8cde88794bd9568474e92f4421aa5622ff1a Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 29 Jul 2026 18:10:28 -0700 Subject: [PATCH] Fallback to cudnn attention on linux if flash attention doesn't work. (#15146) --- comfy/ops.py | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-) diff --git a/comfy/ops.py b/comfy/ops.py index 5e1cce333c3..9d692dcc719 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -41,7 +41,7 @@ def scaled_dot_product_attention(q, k, v, *args, **kwargs): try: - if torch.cuda.is_available() and comfy.model_management.WINDOWS: + if torch.cuda.is_available(): from torch.nn.attention import SDPBackend, sdpa_kernel import inspect if "set_priority" in inspect.signature(sdpa_kernel).parameters: @@ -51,7 +51,10 @@ def scaled_dot_product_attention(q, k, v, *args, **kwargs): SDPBackend.MATH, ] - SDPA_BACKEND_PRIORITY.insert(0, SDPBackend.CUDNN_ATTENTION) + if comfy.model_management.WINDOWS: + SDPA_BACKEND_PRIORITY.insert(0, SDPBackend.CUDNN_ATTENTION) + else: + SDPA_BACKEND_PRIORITY.insert(1, SDPBackend.CUDNN_ATTENTION) def scaled_dot_product_attention(q, k, v, *args, **kwargs): if q.nelement() < 1024 * 128: # arbitrary number, for small inputs cudnn attention seems slower