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.

Add a way for nodes to add pre attn patches to flux model. (#12861)

comfyanonymousc4fb0271cd7f
diff --git a/comfy/ldm/flux/layers.py b/comfy/ldm/flux/layers.py
index 8b3f500d7..e20d498f8 100644
--- a/comfy/ldm/flux/layers.py
+++ b/comfy/ldm/flux/layers.py
@@ -223,12 +223,19 @@ class DoubleStreamBlock(nn.Module):
del txt_k, img_k
v = torch.cat((txt_v, img_v), dim=2)
del txt_v, img_v
+
+ extra_options["img_slice"] = [txt.shape[1], q.shape[2]]
+ if "attn1_patch" in transformer_patches:
+ patch = transformer_patches["attn1_patch"]
+ for p in patch:
+ out = p(q, k, v, pe=pe, attn_mask=attn_mask, extra_options=extra_options)
+ q, k, v, pe, attn_mask = out.get("q", q), out.get("k", k), out.get("v", v), out.get("pe", pe), out.get("attn_mask", attn_mask)
+
# run actual attention
attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options)
del q, k, v
if "attn1_output_patch" in transformer_patches:
- extra_options["img_slice"] = [txt.shape[1], attn.shape[1]]
patch = transformer_patches["attn1_output_patch"]
for p in patch:
attn = p(attn, extra_options)
@@ -321,6 +328,12 @@ class SingleStreamBlock(nn.Module):
del qkv
q, k = self.norm(q, k, v)
+ if "attn1_patch" in transformer_patches:
+ patch = transformer_patches["attn1_patch"]
+ for p in patch:
+ out = p(q, k, v, pe=pe, attn_mask=attn_mask, extra_options=extra_options)
+ q, k, v, pe, attn_mask = out.get("q", q), out.get("k", k), out.get("v", v), out.get("pe", pe), out.get("attn_mask", attn_mask)
+
# compute attention
attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options)
del q, k, v
diff --git a/comfy/ldm/flux/math.py b/comfy/ldm/flux/math.py
index 5e764bb46..824daf5e6 100644
--- a/comfy/ldm/flux/math.py
+++ b/comfy/ldm/flux/math.py
@@ -31,6 +31,8 @@ def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
def _apply_rope1(x: Tensor, freqs_cis: Tensor):
x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2)
+ if x_.shape[2] != 1 and freqs_cis.shape[2] != 1 and x_.shape[2] != freqs_cis.shape[2]:
+ freqs_cis = freqs_cis[:, :, :x_.shape[2]]
x_out = freqs_cis[..., 0] * x_[..., 0]
x_out.addcmul_(freqs_cis[..., 1], x_[..., 1])
diff --git a/comfy/ldm/flux/model.py b/comfy/ldm/flux/model.py
index ef4dcf7c5..00f12c031 100644
--- a/comfy/ldm/flux/model.py
+++ b/comfy/ldm/flux/model.py
@@ -170,7 +170,7 @@ class Flux(nn.Module):
if "post_input" in patches:
for p in patches["post_input"]:
- out = p({"img": img, "txt": txt, "img_ids": img_ids, "txt_ids": txt_ids})
+ out = p({"img": img, "txt": txt, "img_ids": img_ids, "txt_ids": txt_ids, "transformer_options": transformer_options})
img = out["img"]
txt = out["txt"]
img_ids = out["img_ids"]