default_stage: default_modifiers: QuantizationModifier: targets: [Linear] ignore: [lm_head, 're:.*embed_tokens$', 're:.*norm$', 're:.*vpm.*', 're:.*resampler.*', 're:.*apm.*', 're:.*audio_projection_layer.*', 're:.*tts.*'] scheme: W8A16 bypass_divisibility_checks: false