# timeout /t 3 /nobreak && mergekit-yaml C:\mergekit-main\moe_karcher.yaml C:\mergekit-main\moe_karcher --copy-tokenizer --allow-crimes --out-shard-size 5B --trust-remote-code --lazy-unpickle --random-seed 420 --cuda

# MoE_Karcher
# MoE-aware Karcher merge that:
#    1. Identifies expert weights by pattern matching
#    2. Blends corresponding experts across MoE models
#    3. Handles router weights separately with optional strategies
merge_method: moe_karcher
base_model: B:\8B\Meme-Trix-MoE-14B-A8B-v1
models:
  - model: B:\8B\Meme-Trix-MoE-14B-A8B-v1
  - model: B:\8B\Babsie--CrossroadsLoki-MoE-2x8B
parameters:
  max_iter: 1000
  tol: 1e-9
  router_strategy: karcher  # Options: karcher, average, first, random_init
  blend_experts: true  # Blend corresponding experts (expert[0] + expert[0], etc.)
dtype: float32
out_dtype: bfloat16
tokenizer:
  source: union
# chat_template: auto
name: Trixster-MoE-Karcher-14B-A8B-v1