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.

Fix some OOM issues with split and sub quad attention.

comfyanonymousa373367b0cf3
diff --git a/comfy/ldm/modules/attention.py b/comfy/ldm/modules/attention.py
index dcf467489..4f10bbc35 100644
--- a/comfy/ldm/modules/attention.py
+++ b/comfy/ldm/modules/attention.py
@@ -222,9 +222,14 @@ def attention_split(q, k, v, heads, mask=None):
mem_free_total = model_management.get_free_memory(q.device)
+ if _ATTN_PRECISION =="fp32":
+ element_size = 4
+ else:
+ element_size = q.element_size()
+
gb = 1024 ** 3
- tensor_size = q.shape[0] * q.shape[1] * k.shape[1] * q.element_size()
- modifier = 3 if q.element_size() == 2 else 2.5
+ tensor_size = q.shape[0] * q.shape[1] * k.shape[1] * element_size
+ modifier = 3 if element_size == 2 else 2.5
mem_required = tensor_size * modifier
steps = 1
diff --git a/comfy/ldm/modules/sub_quadratic_attention.py b/comfy/ldm/modules/sub_quadratic_attention.py
index 4d42059b5..8e8e8054d 100644
--- a/comfy/ldm/modules/sub_quadratic_attention.py
+++ b/comfy/ldm/modules/sub_quadratic_attention.py
@@ -83,7 +83,8 @@ def _summarize_chunk(
)
max_score, _ = torch.max(attn_weights, -1, keepdim=True)
max_score = max_score.detach()
- torch.exp(attn_weights - max_score, out=attn_weights)
+ attn_weights -= max_score
+ torch.exp(attn_weights, out=attn_weights)
exp_weights = attn_weights.to(value.dtype)
exp_values = torch.bmm(exp_weights, value)
max_score = max_score.squeeze(-1)