| default_stage: | |
| default_modifiers: | |
| AWQModifier: | |
| requires_calibration_data: true | |
| mappings: | |
| - smooth_layer: re:.*input_layernorm$ | |
| balance_layers: ['re:.*q_proj$', 're:.*k_proj$', 're:.*v_proj$'] | |
| activation_hook_target: null | |
| - smooth_layer: re:.*v_proj$ | |
| balance_layers: ['re:.*o_proj$'] | |
| activation_hook_target: null | |
| - smooth_layer: re:.*post_attention_layernorm$ | |
| balance_layers: ['re:.*gate_proj$', 're:.*up_proj$'] | |
| activation_hook_target: null | |
| - smooth_layer: re:.*up_proj$ | |
| balance_layers: ['re:.*down_proj$'] | |
| activation_hook_target: null | |
| offload_device: !!python/object/apply:torch.device [cpu] | |
| duo_scaling: both | |
| n_grid: 40 | |
| QuantizationModifier: | |
| targets: [Linear] | |
| ignore: [lm_head] | |
| scheme: W4A16_ASYM | |
| bypass_divisibility_checks: false | |
| requires_calibration_data: false | |