nvfp4_fp8_at_5p4bits.yaml

# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

# AutoQuantize: per-layer search over {NVFP4 (W4A4), FP8 (W8A8)} at 5.4 effective bits.

# modelopt-schema: modelopt.recipe.config.ModelOptAutoQuantizeRecipe
imports:
  base_disabled_layers: configs/auto_quantize/units/base_disabled_layers
  base_cost_excluded_layers: configs/auto_quantize/units/base_cost_excluded_layers
  nvfp4: configs/ptq/presets/model/nvfp4
  fp8: configs/ptq/presets/model/fp8

metadata:
  recipe_type: auto_quantize
  description: Mixed NVFP4 + FP8 per-layer search at 5.4 effective bits.

auto_quantize:
  constraints:
    effective_bits: 5.4

  candidate_formats:
    - $import: nvfp4
    - $import: fp8

  auto_quantize_method: gradient
  score_size: 128

  # Base (model-agnostic) non-quantizable layers, spliced from the shared unit. Arch-specific
  # models use a recipe under huggingface/<model>/auto_quantize/ that appends to this set.
  disabled_layers:
    - $import: base_disabled_layers

  # Base (model-agnostic) cost-excluded layers, spliced from the shared unit: VL vision
  # towers / MTP are kept out of the effective-bits denominator. No-op on non-VL models.
  cost_excluded_layers:
    - $import: base_cost_excluded_layers