Make buggy xformers fall back on pytorch attention.

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)