default_stage: default_modifiers: GPTQModifier: targets: ['re:.*block_sparse_moe\.experts\.\d+\.w[1-3]$', 're:.*mlp\.experts\.\d+\.(gate|up|gate_up|down)_proj$'] ignore: ['re:.*self_attn.*', lm_head] scheme: W4A16 weight_observer: mse bypass_divisibility_checks: false block_size: 128 dampening_frac: 0.01 actorder: static offload_hessians: false