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.

Cleaner handling of attention mask in ltxv model code.

comfyanonymousff838657fac7
diff --git a/comfy/ldm/lightricks/model.py b/comfy/ldm/lightricks/model.py
index eeeeaea04..2a02acd65 100644
--- a/comfy/ldm/lightricks/model.py
+++ b/comfy/ldm/lightricks/model.py
@@ -456,9 +456,8 @@ class LTXVModel(torch.nn.Module):
x = self.patchify_proj(x)
timestep = timestep * 1000.0
- attention_mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1]))
- attention_mask = attention_mask.masked_fill(attention_mask.to(torch.bool), float("-inf")) # not sure about this
- # attention_mask = (context != 0).any(dim=2).to(dtype=x.dtype)
+ if attention_mask is not None and not torch.is_floating_point(attention_mask):
+ attention_mask = (attention_mask - 1).to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])) * torch.finfo(x.dtype).max
pe = precompute_freqs_cis(indices_grid, dim=self.inner_dim, out_dtype=x.dtype)