Update to Transformers v5

[cg123]

Note

High Risk
Core merge/load paths and MoE output formats changed for a major Transformers bump; incorrect conversion or optional-weight logic could silently drop or mis-merge expert weights.

Overview
Upgrades mergekit to Transformers 5 (transformers>=5.0,<6.0, peft>=0.18.0) and adds a conversion layer that uses Hugging Face’s checkpoint conversion registry to map legacy checkpoint keys to the v5 model layout (e.g. packed gate_up_proj / down_proj for MoE experts).

Tensor loading and architecture inference now fall back to conversion when a weight name is missing on disk; auto-inference builds weight templates from the base model’s v5 state_dict and treats weights as present only if conversion (or a direct key) works across all merge inputs and every layer.

Mixtral and Qwen3 MoE move from Python ModuleArchitecture classes to JSON definitions with v5 tensor names; MoE merge writers assemble per-expert sources and save via conversion instead of manual block_sparse_moe / per-expert MLP remapping. Glm4 MoE layer templates align with the same packed expert layout.

Smaller changes: no_init_weights replaces custom NoInit in evolve; MoE router uses BitsAndBytesConfig for 4/8-bit loads; mergekit-evolve is deprecated with a runtime warning.

^{Reviewed by Cursor Bugbot for commit 0d59494. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

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

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

^{Reviewed by Cursor Bugbot for commit 17f7785. Configure here.}