recipe.yaml

default_stage:
  default_modifiers:
    SpinQuantModifier:
      rotations: [R1, R2, R4]
      transform_type: hadamard
      randomize: false
      learnable: false
      precision: torch.float64
      transform_block_size: 16
      transform_config:
        config_groups:
          R1:
            type: hadamard
            apply:
            - targets: ['re:.*embed_tokens$', 're:.*o_proj$', 're:.*down_proj$']
              location: weight_output
              inverse: false
              ignore: []
            - targets: ['re:.*q_proj$', 're:.*k_proj$', 're:.*v_proj$', 're:.*up_proj$', 're:.*gate_proj$',
                lm_head]
              location: weight_input
              inverse: true
              ignore: []
            randomize: false
            requires_grad: false
            head_dim: 16
            precision: torch.float64
          R2:
            type: hadamard
            apply:
            - targets: ['re:.*v_proj$']
              location: weight_output
              inverse: false
              ignore: []
            - targets: ['re:.*o_proj$']
              location: weight_input
              inverse: true
              ignore: []
            randomize: false
            requires_grad: false
            head_dim: 16
            precision: torch.float64
          R4:
            type: hadamard
            apply:
            - targets: ['re:.*down_proj$']
              location: input
              inverse: false
              ignore: []
            - targets: ['re:.*down_proj$']
              location: weight_input
              inverse: true
              ignore: []
            randomize: false
            requires_grad: false
            head_dim: 16
            precision: torch.float64
    QuantizationModifier:
      targets: [Linear]
      ignore: [lm_head]
      scheme: NVFP4A16