From 5be3e216e57822394f2f6db842f759489c72e2fb Mon Sep 17 00:00:00 2001 From: realAsma Date: Thu, 20 Aug 2026 19:00:35 +0000 Subject: [PATCH] Set NVFP4 MSE effective bits to 4.5 Signed-off-by: realAsma --- .../ptq/presets/model/nvfp4_w4a4_weight_mse_fp8_sweep.yaml | 1 + tests/unit/recipe/test_presets.py | 4 ++++ 2 files changed, 5 insertions(+) diff --git a/modelopt_recipes/configs/ptq/presets/model/nvfp4_w4a4_weight_mse_fp8_sweep.yaml b/modelopt_recipes/configs/ptq/presets/model/nvfp4_w4a4_weight_mse_fp8_sweep.yaml index 3ae22dbc3a6..a503f8e86fd 100644 --- a/modelopt_recipes/configs/ptq/presets/model/nvfp4_w4a4_weight_mse_fp8_sweep.yaml +++ b/modelopt_recipes/configs/ptq/presets/model/nvfp4_w4a4_weight_mse_fp8_sweep.yaml @@ -25,6 +25,7 @@ imports: algorithm: method: mse fp8_scale_sweep: true +effective_bits: 4.5 quant_cfg: - $import: base_disable_all - quantizer_name: '*weight_quantizer' diff --git a/tests/unit/recipe/test_presets.py b/tests/unit/recipe/test_presets.py index 89645773f93..15b7e19061b 100644 --- a/tests/unit/recipe/test_presets.py +++ b/tests/unit/recipe/test_presets.py @@ -63,6 +63,10 @@ def test_kv_none_sentinel_is_not_a_discovered_preset(): assert presets.KV_CACHE_NONE not in presets.KV_QUANT_CFG_CHOICES +def test_nvfp4_mse_preset_uses_nvfp4_storage_cost(): + assert presets.QUANT_CFG_CHOICES["nvfp4_w4a4_weight_mse_fp8_sweep"]["effective_bits"] == 4.5 + + def test_w4a16_nvfp4_preset_disables_vllm_marlin_incompatible_projections(): disabled_quantizers = { entry["quantizer_name"]