default_stage: default_modifiers: QuantizationModifier: targets: [Linear] ignore: [lm_head, model.embed_tokens, 're:.*block_sparse_moe\.gate$'] scheme: NVFP4 bypass_divisibility_checks: false