openpangu_best_practice.yaml

apiversion: modelslim_v1
spec:
  process:
  - type: flex_smooth_quant
    proxy_quant_dtype: mxfp4
    proxy_quant_axes: -1
    enable_subgraph_type:
    - up-down
    include:
    - '*'
    exclude:
    - model.layers.50*
    - model.layers.1.*
    - model.layers.2.*
    - model.layers.3.*
    - model.layers.0.*
    - '*mlp.shared_experts*'
  - type: linear_quant
    qconfig:
      act:
        scope: per_block
        dtype: mxfp4
        symmetric: true
        method: minmax
      weight:
        scope: per_block
        dtype: mxfp4
        symmetric: true
        method: minmax
    include:
    - '*'
    exclude:
    - model.layers.1.*
    - model.layers.2.*
    - model.layers.3.*
    - model.layers.0.*
    - '*attn*'
    - '*mlp.shared_experts*'
    - '*mlp.gate'
    - model.layers.50*
  dataset: mix_calib.jsonl
  save:
  - type: ascendv1_saver
    part_file_size: 4
metadata:
  config_id: openpangu_mxfp4_e2m1
  score: 50.0
  label:
    w_bit: 4
    a_bit: 4
    is_sparse: false
    kv_cache: false
  verified_model_types: []
  verified_tags: {}
default_mxfp4_e2m1:
  act:
    scope: per_block
    dtype: mxfp4
    symmetric: true
    method: minmax
  weight:
    scope: per_block
    dtype: mxfp4
    symmetric: true
    method: minmax