The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
Allow zeroing out of embeds with unused attention mask.
comfyanonymousce649d61c05a
diff --git a/comfy/sd1_clip.py b/comfy/sd1_clip.pyindex 78e556b56..3b812b44c 100644--- a/comfy/sd1_clip.py+++ b/comfy/sd1_clip.py@@ -169,7 +169,7 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):tokens = torch.LongTensor(tokens).to(device)attention_mask = None- if self.enable_attention_masks:+ if self.enable_attention_masks or self.zero_out_masked:attention_mask = torch.zeros_like(tokens)end_token = self.special_tokens.get("end", -1)for x in range(attention_mask.shape[0]):@@ -178,7 +178,11 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):if tokens[x, y] == end_token:break- outputs = self.transformer(tokens, attention_mask, intermediate_output=self.layer_idx, final_layer_norm_intermediate=self.layer_norm_hidden_state)+ attention_mask_model = None+ if self.enable_attention_masks:+ attention_mask_model = attention_mask++ outputs = self.transformer(tokens, attention_mask_model, intermediate_output=self.layer_idx, final_layer_norm_intermediate=self.layer_norm_hidden_state)self.transformer.set_input_embeddings(backup_embeds)if self.layer == "last":@@ -186,7 +190,7 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):else:z = outputs[1].float()- if self.zero_out_masked and attention_mask is not None:+ if self.zero_out_masked:z *= attention_mask.unsqueeze(-1).float()pooled_output = None