Skip to content

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.

Fallback to pytorch attention if sage attention fails.

comfyanonymouse471c726e57b
diff --git a/comfy/ldm/modules/attention.py b/comfy/ldm/modules/attention.py
index 7908d1313..ede506463 100644
--- a/comfy/ldm/modules/attention.py
+++ b/comfy/ldm/modules/attention.py
@@ -471,7 +471,7 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha
def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
if skip_reshape:
b, _, _, dim_head = q.shape
- tensor_layout="HND"
+ tensor_layout = "HND"
else:
b, _, dim_head = q.shape
dim_head //= heads
@@ -479,7 +479,7 @@ def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=
lambda t: t.view(b, -1, heads, dim_head),
(q, k, v),
)
- tensor_layout="NHD"
+ tensor_layout = "NHD"
if mask is not None:
# add a batch dimension if there isn't already one
@@ -489,7 +489,17 @@ def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=
if mask.ndim == 3:
mask = mask.unsqueeze(1)
- out = sageattn(q, k, v, attn_mask=mask, is_causal=False, tensor_layout=tensor_layout)
+ try:
+ out = sageattn(q, k, v, attn_mask=mask, is_causal=False, tensor_layout=tensor_layout)
+ except Exception as e:
+ logging.error("Error running sage attention: {}, using pytorch attention instead.".format(e))
+ if tensor_layout == "NHD":
+ q, k, v = map(
+ lambda t: t.transpose(1, 2),
+ (q, k, v),
+ )
+ return attention_pytorch(q, k, v, heads, mask=mask, skip_reshape=True, skip_output_reshape=skip_output_reshape)
+
if tensor_layout == "HND":
if not skip_output_reshape:
out = (