ACE Step 1.5 produces garbage with --use-sage-attention #12270

Activity

1. rkfg

   added

   Potential BugUser is reporting a bug. This should be tested.User is reporting a bug. This should be tested.

   on Feb 4, 2026
2. iamrohitanshu commented on Feb 4, 2026

   

   iamrohitanshu

   on Feb 4, 2026

   Last edited by iamrohitanshu

   More actions

   I have an RTX 3060 12GB. SageAttn worked great with Wan workflows.
   But Qwen, Anima and AceStep, and also an IPAdapter for SDXL arch, have problems or generate black images with SageAttn.
   Not finding any solutions, I now keep 2 bat start files, one with SageAttn and one with --force-upcast-attention, which uses pytorch attn by default.
3. rkfg commented on Feb 4, 2026

   

   rkfg

   on Feb 4, 2026

   ContributorAuthor

   More actions

   Temporary hack:

   diff --git a/comfy/ldm/ace/ace_step15.py b/comfy/ldm/ace/ace_step15.py
   index d9054965..6a391b66 100644
   --- a/comfy/ldm/ace/ace_step15.py
   +++ b/comfy/ldm/ace/ace_step15.py
   @@ -3,7 +3,7 @@ import torch
    import torch.nn as nn
    import torch.nn.functional as F
    import itertools
   -from comfy.ldm.modules.attention import optimized_attention
   +from comfy.ldm.modules.attention import attention_pytorch
    import comfy.model_management
    from comfy.ldm.flux.layers import timestep_embedding
    
   @@ -183,7 +183,7 @@ class AceStepAttention(nn.Module):
                else:
                    attn_bias = window_bias
    
   -        attn_output = optimized_attention(query_states, key_states, value_states, self.num_heads, attn_bias, skip_reshape=True)
   +        attn_output = attention_pytorch(query_states, key_states, value_states, self.num_heads, attn_bias, skip_reshape=True)
            attn_output = self.o_proj(attn_output)
    
            return attn_output

   This simply replaces the currently selected attention method with PyTorch attention. No checks are performed so if you don't have it working it will fail. But I assume it should work on any modern Torch. Sage Attention doesn't bring any benefits to ACE Step 1.5 anyway since the model is already lightning fast. With this simple patch you can keep running ComfyUI with --use-sage-attention flag and it will still work.

   👍React with 👍4Reacted by festerz, tuolaku, rohitanshu and NonaBuhtig

   ❤️React with ❤️2Reacted by festerz and bjoernfal
4. comfyanonymous

   closed this as completedon Feb 5, 2026