Merge_models.py doesn't merge SLA's proj_l layers?

After merging models using merge_models.py, I found that I cannot use --attention_type sla or --attention_type sagesla when inferencing with the merged model.

After debugging, I determined that the merged model does not satisfy has_sagesla_weights = any('proj_l' in k for k in state_dict.keys()), meaning the SLA proj_l layers were not merged into the final model.

Looking at the merge logic in merge_models.py, I see that only layers present in both diff_base_sd and diff_target_sd are merged:

if key in diff_base_sd and key in diff_target_sd:
    d_base_tensor = diff_base_sd[key] # pretrained
    d_target_tensor = diff_target_sd[key] # SLA

My Setup

• base: wan2pt1_1pt3B_res480p_t2v_rCM_3000.pth
• diff_base: Wan2.1-T2V-1.3B.pth (pretrained Wan2.1 model)
• diff_target: wan2pt1_1pt3B_res480p_t2v_SLA_4000.pth (SLA fine-tuned model)

Since the pretrained Wan2.1 model (diff_base_sd) doesn't have SLA layers, they cannot be merged with the SLA fine-tuned model.

Questions

1. Am I using the wrong diff_base_sd?
2. Is merge_models.py designed this way, meaning merged models can only use --attention_type original for inference?

Any guidance on how to properly merge models to retain SLA functionality would be greatly appreciated.

Start in merge_models.py and inspect the branch that merges keys shared by diff_base_sd and diff_target_sd. Reproduce the merge with the listed Wan2.1 and SLA checkpoints, then inspect the resulting state_dict for proj_l keys and test inference with --attention_type sla or --attention_type sagesla. Done means the merge behavior is clarified or corrected so SLA functionality is retained when the inputs support it.