abdelrahmane01's picture
Upload AWQ quantized model
03609fa verified
default_stage:
default_modifiers:
AWQModifier:
targets: [Linear]
ignore: [lm_head]
scheme: W4A16
bypass_divisibility_checks: false
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
duo_scaling: true
n_grid: 20