Release v0.1.5 · arcee-ai/mergekit

What's Changed

Fix auto-architecture for MoE models by @cg123 in #647
Add AFMoE support by @cg123 in #648
Add ApertusForCausalLM by @cg123 in #655
Allow tokensurgeon to attempt auto architecture by @cg123 in #656
Bug fix: Filter tokens not in the embedding (tokensurgeon) by @xz-liu in #657
Add RAM and tensor-local RAM+ variant by @cg123 in #662
Update tokenizer merging documentation by @cg123 in #663
Add GLM4 MOE models support by @pengdurice in #658
Add Qwen3-VL support by @pengdurice in #665
Add Qwen3-VL dense models by @pengdurice in #670
Add IBM Granite architecture support by @Bhavyashah20 in #683
Fix TypeError when loading model in MoE router without bitsandbytes by @umran666 in #690
Update to Transformers v5 by @cg123 in #692
Bump copyright date by @cg123 in #694
Add arXiv link for Karcher Mean paper by @neoheartbeats in #673
Add missing norm.weight parameters to cohere.json by @jukofyork in #674
Relax dependencies + switch to ruff by @cg123 in #705
Preserve existing tokenizer metadata by @NeonBohdan in #703
fix(moe/deepseek): create the output directory before writing config.json (#697) by @Anai-Guo in #701
Adding missing q_norm/k_norm weights to olmo2.json by @armsp in #699
Fix ENAMETOOLONG crash when chat_template is an inline Jinja string by @javierdejesusda in #698
New merge method API by @cg123 in #706
Remove vllm extra by @cg123 in #717
Fix magnitude pruning for small tensors in TIES merges by @tomaarsen in #716

Full Changelog: v0.1.4...v0.1.5