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.

Make buggy xformers fall back on pytorch attention.

comfyanonymous3e5ea74ad356
diff --git a/comfy/ldm/modules/attention.py b/comfy/ldm/modules/attention.py
index 947e2008c..d511dda16 100644
--- a/comfy/ldm/modules/attention.py
+++ b/comfy/ldm/modules/attention.py
@@ -278,9 +278,20 @@ def attention_split(q, k, v, heads, mask=None):
)
return r1
+BROKEN_XFORMERS = False
+try:
+ x_vers = xformers.__version__
+ #I think 0.0.23 is also broken (q with bs bigger than 65535 gives CUDA error)
+ BROKEN_XFORMERS = x_vers.startswith("0.0.21") or x_vers.startswith("0.0.22") or x_vers.startswith("0.0.23")
+except:
+ pass
+
def attention_xformers(q, k, v, heads, mask=None):
b, _, dim_head = q.shape
dim_head //= heads
+ if BROKEN_XFORMERS:
+ if b * heads > 65535:
+ return attention_pytorch(q, k, v, heads, mask)
q, k, v = map(
lambda t: t.unsqueeze(3)