Repository navigation
ACE-Step 1.5 CLIPLoader: Error(s) in loading state_dict for Qwen3_4B: size mismatch for model.embed_tokens.weight #12278
Description
Custom Node Testing
- I have tried disabling custom nodes and the issue persists (see how to disable custom nodes if you need help)
Expected Behavior
ACE-Step 1.5 AiO works fine with the CheckpointLoaderSimple.
However, trying to load other ACE-Step 1.5 LMs with a CLIPLoader fails (see error messages below).
Settings:
Load CLIP (CLIPLoader)
- type: ace
- device: default
Tested with v0.12.2 and HEAD (ef73070)
Error Messages
Error(s) in loading state_dict for Qwen3_06B:
size mismatch for model.embed_tokens.weight: copying a param with shape torch.Size([151669, 1024]) from checkpoint, the shape in current model is torch.Size([151936, 1024]).
Error(s) in loading state_dict for Ovis25_2B:
size mismatch for model.embed_tokens.weight: copying a param with shape torch.Size([217204, 2048]) from checkpoint, the shape in current model is torch.Size([151936, 2048]).
Error(s) in loading state_dict for Qwen3_4B:
size mismatch for model.embed_tokens.weight: copying a param with shape torch.Size([217204, 2560]) from checkpoint, the shape in current model is torch.Size([151936, 2560]).
Actual Behavior
At least qwen_1.7b_ace15 should load without problems. All of them? Even better!
Steps to Reproduce
I'm using this workflow.
Debug Logs
I forgot to get them before shutting down the server. Sorry!Other
For qwen_1.7b_ace15, the detected type Ovis25_2B is strange, too.
Activity
wodsoe commentedon Feb 4, 2026
Why comfyui not use the official model.safensors, which takes up more SSD space?
hum-ma commentedon Feb 4, 2026
The models have different architectures which might be making them more difficult to support. The 0.6B and 4.0B LMs are
Qwen3ForCausalLMwhile the 1.7B LM isQwen3Model, same as the 0.6B embedding model. The former two can even be quantized to GGUF but the latter are unknown by llama.cpphttps://huggingface.co/ACE-Step/acestep-5Hz-lm-0.6B/blob/main/config.json
https://huggingface.co/ACE-Step/acestep-5Hz-lm-4B/blob/main/config.json
https://huggingface.co/ACE-Step/Ace-Step1.5/blob/main/acestep-5Hz-lm-1.7B/config.json
https://huggingface.co/ACE-Step/Ace-Step1.5/blob/main/Qwen3-Embedding-0.6B/config.json
poluramus commentedon Feb 4, 2026
I'm having the same issue.
The commit fe25114 message says "Support the 4B ace step 1.5 lm model" was not supposed to fix it ?
iamrohitanshu commentedon Feb 4, 2026
You're supposed to use DualCLIPLoader. The 0.6b TE stays as is, the 1.7b TE can be replaced with 4b TE.
See the default comfyUI template (other than the aio one).
hum-ma commentedon Feb 4, 2026
It's a bit confusing because ACE-Step has this in the model guide table:
So it should be possible to use without the second model, loading just the 0.6B embedding model, but not one of the 0.6B, 1.7B or 4.0B LMs which cannot be used alone if I understand correctly.
But currently in ComfyUI we can neither load the embedding model alone nor the 0.6B + 0.6B models.
newsletternewsletter commentedon Feb 5, 2026
Yes, that was the solution! Thank you!