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.

Don't crash when using flux kv cache with split batches. (#14422)

comfyanonymous10d466b0e3a8
diff --git a/comfy_extras/nodes_flux.py b/comfy_extras/nodes_flux.py
index afc663b22..ef1757ae5 100644
--- a/comfy_extras/nodes_flux.py
+++ b/comfy_extras/nodes_flux.py
@@ -245,6 +245,11 @@ class KV_Attn_Input:
cache_key = "{}_{}".format(extra_options["block_type"], extra_options["block_index"])
if cache_key in self.cache:
kk, vv = self.cache[cache_key]
+
+ # Fix batch size changing.
+ kk = comfy.utils.repeat_to_batch_size(kk, k.shape[0])
+ vv = comfy.utils.repeat_to_batch_size(vv, v.shape[0])
+
self.set_cache = False
return {"q": q, "k": torch.cat((k, kk), dim=2), "v": torch.cat((v, vv), dim=2)}