File size: 489 Bytes
d2f0beb |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 |
quant_stage:
quant_modifiers:
SmoothQuantModifier:
smoothing_strength: 0.7
mappings:
- - ['re:.*q_proj', 're:.*k_proj', 're:.*v_proj']
- re:.*input_layernorm
- - ['re:.*gate_proj', 're:.*up_proj']
- re:.*post_attention_layernorm
- - ['re:.*down_proj']
- re:.*up_proj
GPTQModifier:
sequential_update: true
dampening_frac: 0.01
ignore: [lm_head]
scheme: W8A8
targets: [Linear]
observer: mse
|