overrides.yaml for optimizing glm-4.5 models in exl3

gistfile1.txt

sources:
  - id: 4
    model_dir: /mnt/llms/models/bullerwins/GLM-4.5-exl3-4.0bpw
  - id: 5
    model_dir: /mnt/llms/models/bullerwins/GLM-4.5-exl3-5.0bpw

overrides:
  #   Attention & router tensors – cheap, big gain on MoE models
  - key: "*.self_attn.*"
    source: 5          # +2 bpw
  - key: "*.shared_experts.*"
    source: 5          # +2 bpw (router / shared experts)

  #  Highest-error dense layers – +1 bpw each
  - key: "model.layers.84.*"  # top offenders from rfn_err ranking
    source: 4
  - key: "model.layers.82.*"
    source: 4
  - key: "model.layers.85.*"
    source: 4
  - key: "model.layers.83.*"
    source: 4
  - key: "model.layers.81.*"
    source: 4
  - key: "model.layers.80.*"
    source: 4
  - key: "model.layers.86.*"
    source: 4
  - key: "model.layers.79.*"
    source: 4
  - key: "model.layers.78.*"
    source: 4
  - key: "model.layers.91.*"
    source: 4