CoolFace
Modelpublic

nm-testing/multiple_modifiers_gptq_awq-e2e

sourceHugging Faceupdated 6h agoView on Hugging Face
0likes284downloads
recipe.yaml26 linesDownload Raw Back to root
1quant_stage:2  quant_modifiers:3    AWQModifier:4      requires_calibration_data: true5      mappings:6      - smooth_layer: re:.*input_layernorm$7        balance_layers: ['re:.*q_proj$', 're:.*k_proj$', 're:.*v_proj$']8        activation_hook_target: null9      duo_scaling: true10      n_grid: 2011    QuantizationModifier:12      targets: ['re:.*self_attn\.(k|q|o|v)_proj$']13      ignore: [lm_head]14      scheme: W4A1615      bypass_divisibility_checks: false16      requires_calibration_data: false17    GPTQModifier:18      targets: ['re:.*mlp\.(down|gate|up)_proj$']19      ignore: [lm_head]20      scheme: W8A821      bypass_divisibility_checks: false22      requires_calibration_data: true23      block_size: 12824      dampening_frac: 0.00125      batched_quantization: auto26