| quant_stage: | |
| quant_modifiers: | |
| AWQModifier: | |
| mappings: | |
| - smooth_layer: re:.*input_layernorm$ | |
| balance_layers: ['re:.*q_proj$', 're:.*k_proj$', 're:.*v_proj$'] | |
| activation_hook_target: null | |
| duo_scaling: true | |
| n_grid: 20 | |
| QuantizationModifier: | |
| targets: ['re:.*self_attn\.(k|q|o|v)_proj$'] | |
| ignore: [lm_head] | |
| scheme: W4A16 | |
| bypass_divisibility_checks: false | |
| GPTQModifier: | |
| targets: ['re:.*mlp\.(down|gate|up)_proj$'] | |
| ignore: [lm_head] | |
| scheme: W8A8 | |
| bypass_divisibility_checks: false | |
| block_size: 128 | |
| dampening_frac: 0.001 | |
| offload_hessians: false | |