The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
Fix some OOM issues with split attention.
comfyanonymous125b03eeadd2
diff --git a/comfy/ldm/modules/attention.py b/comfy/ldm/modules/attention.pyindex 4f10bbc35..9840cc7f5 100644--- a/comfy/ldm/modules/attention.py+++ b/comfy/ldm/modules/attention.py@@ -229,7 +229,7 @@ def attention_split(q, k, v, heads, mask=None):gb = 1024 ** 3tensor_size = q.shape[0] * q.shape[1] * k.shape[1] * element_size- modifier = 3 if element_size == 2 else 2.5+ modifier = 3mem_required = tensor_size * modifiersteps = 1@@ -257,10 +257,10 @@ def attention_split(q, k, v, heads, mask=None):s1 = einsum('b i d, b j d -> b i j', q[:, i:end].float(), k.float()) * scaleelse:s1 = einsum('b i d, b j d -> b i j', q[:, i:end], k) * scale- first_op_done = Trues2 = s1.softmax(dim=-1).to(v.dtype)del s1+ first_op_done = Truer1[:, i:end] = einsum('b i j, b j d -> b i d', s2, v)del s2