examples/llm_pretrain/deepseekv3_pretrain.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, please use the following command:
# torchrun --nproc_per_node 8 
#     --nnodes $NNODES
#     --master_addr $MASTER_ADDR 
#     --master_port 29400
#     --node_rank $NODE_RANK 
#     examples/llm_pretrain/pretrain.py 
#     --config examples/llm_pretrain/deepseekv3_pretrain.yaml


step_scheduler:
  global_batch_size: 512
  local_batch_size: 8
  ckpt_every_steps: 500
  val_every_steps: 100  # will run every x number of gradient steps
  num_epochs: 1
  max_steps: 9500

dist_env:
  backend: nccl
  timeout_minutes: 10

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

parallelize_fn:
  _target_: nemo_automodel.components.models.deepseek_v3.parallelizer.parallelize_model
  activation_checkpointing: false

# Option 1 - shorthand builder w/ automatic optimized path
model:
  _target_: nemo_automodel.NeMoAutoModelForCausalLM.from_config
  config:
    _target_: transformers.AutoConfig.from_pretrained
    pretrained_model_name_or_path: deepseek-ai/DeepSeek-V3
  is_meta_device: true

# Option 2 - more fine-grained builder
# model:
#   _target_: nemo_automodel.NeMoAutoModelForCausalLM.from_config
#   config:
#     _target_: transformers.AutoConfig.from_pretrained
#     pretrained_model_name_or_path: deepseek-ai/DeepSeek-V3
#   is_meta_device: true
#   backend:
#     _target_: nemo_automodel.components.moe.utils.BackendConfig
#     attn: te
#     linear: te
#     rms_norm: te
#     enable_deepep: true
#     fake_balanced_gate: false
#     enable_hf_state_dict_adapter: true
#   moe_config:
#     _target_: nemo_automodel.components.moe.layers.MoEConfig
#     dim: 7168
#     ...

checkpoint:
  enabled: true
  checkpoint_dir: checkpoints/
  model_save_format: torch_save # torch_save or safetensors
  save_consolidated: false # saves the model in a consolidated safetensors format. Requires model_save_format to be safetensors.
  load_base_model: false

distributed:
  _target_: nemo_automodel.components.distributed.fsdp2.FSDP2Manager
  tp_size: 1
  cp_size: 1
  pp_size: 8
  dp_replicate_size: 1
  dp_size: 16
  ep_size: 16

loss_fn:
  _target_: nemo_automodel.components.loss.linear_ce.FusedLinearCrossEntropy

dataset:
  _target_: nemo_automodel.components.datasets.llm.megatron_dataset.MegatronPretraining
  paths: /path/to/fineweb_edu/processed_data_*_text_document*  # REPLACE THIS
  index_mapping_dir: /path/to/fineweb_edu/mapping_dir  # REPLACE THIS
  tokenizer:
    _target_: transformers.AutoTokenizer.from_pretrained
    pretrained_model_name_or_path: deepseek-ai/DeepSeek-V3
    trust_remote_code: true
  seq_length: 1024
  split: "0.99, 0.01, 0.00"  # train, validation, test
  splits_to_build: "train"  # has to be one of train, validation, test

dataloader:
  _target_: torchdata.stateful_dataloader.StatefulDataLoader
  collate_fn: torch.utils.data.default_collate
  dataloader_type: "single"  # or "cyclic"

validation_dataset:
  _target_: nemo_automodel.components.datasets.llm.megatron_dataset.MegatronPretraining
  paths: /path/to/fineweb_edu/processed_data_*_text_document*  # REPLACE THIS
  index_mapping_dir: /path/to/fineweb_edu/mapping_dir  # REPLACE THIS
  tokenizer:
    _target_: transformers.AutoTokenizer.from_pretrained
    pretrained_model_name_or_path: deepseek-ai/DeepSeek-V3
    trust_remote_code: true
  seq_length: 1024
  split: "0.99, 0.01, 0.00"  # train, validation, test
  splits_to_build: "validation"  # has to be one of train, validation, test
  num_val_samples: 1024

validation_dataloader:
  _target_: torchdata.stateful_dataloader.StatefulDataLoader
  collate_fn: torch.utils.data.default_collate
  dataloader_type: "single"

optimizer:
  _target_: torch.optim.AdamW
  betas: [0.9, 0.95]
  eps: 1e-5
  lr: 4.2e-4
  weight_decay: 0.1

lr_scheduler:
  lr_decay_style: cosine
  lr_warmup_steps: 500
  min_lr: 0.0

autopipeline:
  _target_: nemo_automodel.components.distributed.pipelining.autopipeline.AutoPipeline
  pp_microbatch_size: 1
  pp_schedule: interleaved1f1b
  scale_grads_in_schedule: False
  round_virtual_stages_to_pp_multiple: down
  patch_inner_model: False
  patch_causal_lm_model: False
  layers_per_stage: 2

# Uncomment and configure for W&B logging
# wandb:
#   project: <your_wandb_project>
#   entity: <your_wandb_entity>
#   name: <your_wandb_exp_name>
#   save_dir: <your_wandb_save_dir> 