nemotron_nano_v3_hellaswag_peft.yaml

# Copyright (c) 2025, NVIDIA CORPORATION.  All rights reserved.
#
# 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.



# To run this recipe:
#   automodel examples/llm_finetune/nemotron/nemotron_nano_v3_hellaswag_peft.yaml --nproc-per-node 8
# Adjust --nproc-per-node to the number of GPUs available on your machine.

recipe: TrainFinetuneRecipeForNextTokenPrediction

step_scheduler:
  global_batch_size: 8
  local_batch_size: 1
  ckpt_every_steps: 1000
  val_every_steps: 1000  # will run every x number of gradient steps
  max_steps: 100

dist_env:
  backend: nccl
  timeout_minutes: 1

rng:
  _target_: nemo_automodel.components.training.rng.StatefulRNG
  seed: 1111
  ranked: true

model:
  _target_: nemo_automodel.NeMoAutoModelForCausalLM.from_pretrained
  pretrained_model_name_or_path: nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
  backend:
    _target_: nemo_automodel.components.models.common.BackendConfig
    rms_norm: torch_fp32
    linear: torch

# torch.compile configuration
compile:
  enabled: false
  mode: "default"  # Options: "default", "reduce-overhead", "max-autotune"
  fullgraph: false
  dynamic: true  # Set to false for better performance with fixed shapes
  backend: null  # Use default backend (inductor)

checkpoint:
  enabled: true
  checkpoint_dir: checkpoints/
  model_save_format: safetensors
  save_consolidated: final

peft:
  _target_: nemo_automodel.components._peft.lora.PeftConfig
  exclude_modules: ["*.out_proj"]  # mamba layers use custom kernels that take in the out_proj.weight directly, thus lora doesn't work here.
  dim: 8
  alpha: 32
  use_triton: True

distributed:
  strategy: fsdp2
  dp_size: none
  tp_size: 1
  cp_size: 1
  ep_size: 4

  sequence_parallel: false

loss_fn:
  _target_: nemo_automodel.components.loss.masked_ce.MaskedCrossEntropy

dataset:
  _target_: nemo_automodel.components.datasets.llm.hellaswag.HellaSwag
  path_or_dataset: rowan/hellaswag
  split: train

packed_sequence:
  packed_sequence_size: 0

dataloader:
  _target_: torchdata.stateful_dataloader.StatefulDataLoader
  collate_fn: nemo_automodel.components.datasets.utils.default_collater
  shuffle: True

validation_dataset:
  _target_: nemo_automodel.components.datasets.llm.hellaswag.HellaSwag
  path_or_dataset: rowan/hellaswag
  split: validation
  num_samples_limit: 64

validation_dataloader:
  _target_: torchdata.stateful_dataloader.StatefulDataLoader
  collate_fn: nemo_automodel.components.datasets.utils.default_collater

optimizer:
  _target_: torch.optim.Adam
  betas: [0.9, 0.999]
  eps: 1e-8
  lr: 1.0e-5
  weight_decay: 0

lr_scheduler:
  lr_decay_style: cosine
  min_lr: 1.0e-6

ci:
  vllm_deploy: true
  vllm_smoke_test: true
  recipe_owner: adil-a
  time: "00:15:00"
  known_issue_id: AM-149
  checkpoint_robustness:
    hf_kl_threshold: 1e-1
    hf_device_map_auto: true
    experts_implementation: grouped_mm
    trust_remote_code: true
    tokenizer_name: nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
    distributed.ep_size: 8
    no_check_resume: true
    dataset.num_samples_limit: 500
    validation_dataset.num_samples_limit: 500

# wandb:
#   project: <your_wandb_project>
#   entity: <your_wandb_entity>
#   name: <your_wandb_exp_name>
#   save_dir: <your_wandb_save_dir>