recipe.yaml

test_stage:
  obcq_modifiers:
    LogarithmicEqualizationModifier:
      mappings: [
        [["re:.*q_proj", "re:.*k_proj", "re:.*v_proj"], "re:.*input_layernorm"],
        [["re:.*gate_proj", "re:.*up_proj"], "re:.*post_attention_layernorm"],
      ] 
    QuantizationModifier:
      ignore:
        # These operations don't make sense to quantize
        - LlamaRotaryEmbedding
        - LlamaRMSNorm
        - SiLUActivation
        - MatMulOutput_QK
        - MatMulOutput_PV
        # Skip quantizing the layers with the most sensitive activations
        - model.layers.21.mlp.down_proj
        - model.layers.7.mlp.down_proj
        - model.layers.2.mlp.down_proj
        - model.layers.8.self_attn.q_proj
        - model.layers.8.self_attn.k_proj
      post_oneshot_calibration: true
      scheme_overrides:
        # Enable channelwise quantization for better accuracy
        Linear:
          weights:
            num_bits: 8
            symmetric: true
            strategy: channel
        MatMulLeftInput_QK:
          input_activations:
            num_bits: 8
            symmetric: true
        # For the embeddings, only weight-quantization makes sense
        Embedding:
          input_activations: null
          weights:
            num_bits: 8
            symmetric: false
    SparseGPTModifier:
      sparsity: 0.5
      block_size: 128
      sequential_update: true
      quantize: true
      percdamp: 0.01
      mask_structure: "0:0"
      targets: ["re:model.layers.\\d*$"]