| default_stage: | |
| default_modifiers: | |
| AWQModifier: | |
| requires_calibration_data: true | |
| mappings: | |
| - smooth_layer: re:.*mlp\.experts\.\d+\.up_proj$ | |
| balance_layers: ['re:.*mlp\.experts\.\d+\.down_proj$'] | |
| activation_hook_target: null | |
| duo_scaling: true | |
| n_grid: 10 | |
| QuantizationModifier: | |
| targets: ['re:.*mlp\.experts\.\d+\.(gate_proj|up_proj|down_proj)$'] | |
| ignore: [lm_head, 're:.*embed_tokens.*', 're:.*mtp\..*', 're:.*\.ple\..*', 're:.*visual\..*', | |
| 're:.*\.gate$', 're:.*hyper_connection.*', 're:.*indexer.*', 're:.*\.linear_attn\..*', | |
| 're:.*self_attn\..*', 're:.*shared_expert.*'] | |
| scheme: W4A16 | |
| bypass_divisibility_checks: false | |
| requires_calibration_data: false | |