Add Qwen3.5-MoE architecture support and fix Qwen3-VL-MoE linear attention weights

[alchernysh]

Summary

• Add linear attention (gated delta net) weights to the Qwen3-VL-MoE architecture — hybrid layers use linear_attn instead of self_attn on most layers, and these weights were missing from the definition, so mergekit silently dropped them during merges.
• Add architecture support for Qwen3.5-MoE (Qwen3_5MoeForConditionalGeneration), including a visual module for the vision tower + merger. Without it, this multimodal model's model.visual.* tensors were silently dropped from merge output entirely.

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Want higher recall? High effort reviews run extra passes and find more bugs. A team admin can switch effort levels in the Cursor dashboard.

^{Reviewed by Cursor Bugbot for commit fcdaa9a. Configure here.}

[cursor]

Missing deepstack merger weights

High Severity

The new visual module lists patch embed, vision blocks, and the main merger, but not model.visual.deepstack_merger_list.*. Qwen3.5 MoE multimodal vision uses DeepStack like Qwen3-VL; mergekit only writes planned weights, so those tensors are omitted from merge output.

Additional Locations (1)

• mergekit/_data/architectures/qwen3_vl.json#L149-L170

 

^{Reviewed by Cursor Bugbot for commit fcdaa9a. Configure here.}

[alchernysh]

I have read the CLA Document and I hereby sign the CLA