default_stage: default_modifiers: QuantizationModifier: targets: ['re:.*block_sparse_moe\.experts\.\d+\.w1$', 're:.*block_sparse_moe\.experts\.\d+\.w2$', 're:.*block_sparse_moe\.experts\.\d+\.w3$'] ignore: [lm_head, 're:.*block_sparse_moe.gate$'] scheme: NVFP4 weight_observer: mse bypass_divisibility_checks: false