| default_stage: | |
| default_modifiers: | |
| AWQModifier: | |
| targets: [Linear] | |
| ignore: [lm_head, embed_tokens] | |
| scheme: W4A16 | |
| bypass_divisibility_checks: false | |
| mappings: | |
| - smooth_layer: re:.*input_layernorm$ | |
| balance_layers: ['re:.*qkv_proj$'] | |
| activation_hook_target: null | |
| - smooth_layer: re:.*qkv_proj$ | |
| balance_layers: ['re:.*o_proj$'] | |
| activation_hook_target: null | |
| - smooth_layer: re:.*post_attention_layernorm$ | |
| balance_layers: ['re:.*gate_up_proj$'] | |
| activation_hook_target: null | |
| - smooth_layer: re:.*gate_up_proj$ | |
| balance_layers: ['re:.*down_proj$'] | |
| activation_hook_target: null | |
| duo_scaling: true | |
| n_grid: 20 | |