scripts/vlm/llama4/llama4_finetune.py

# 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.

"""
Mock Data Example:
  torchrun --nproc_per_node=2 scripts/vlm/llama4/llama4_finetune.py \
  --devices=2 --tp=2 --data_type=mock --mbs=1 --gbs=4 --use_toy_model
"""

import argparse

import torch
from lightning.pytorch.loggers import WandbLogger
from megatron.core.optimizer import OptimizerConfig

from nemo import lightning as nl
from nemo.collections import llm, vlm
from nemo.collections.common.tokenizers import AutoTokenizer
from nemo.collections.vlm.data.data_module import EnergonDataModule
from nemo.collections.vlm.llama4.data.task_encoder import TaskEncoder as Llama4TaskEncoder
from nemo.collections.vlm.llama4.data.task_encoder import TaskEncoderConfig as Llama4TaskEncoderConfig
from nemo.collections.vlm.llama4.model.base import Llama4OmniModel
from nemo.lightning.pytorch.callbacks.megatron_comm_overlap import MegatronCommOverlapCallback
from nemo.lightning.pytorch.optim import CosineAnnealingScheduler
from nemo.lightning.pytorch.optim.megatron import MegatronOptimizerModule
from nemo.utils.exp_manager import TimingCallback


def main(args):
    # pylint: disable=C0115,C0116

    # Global and micro batch sizes
    gbs = args.gbs
    mbs = args.mbs
    max_steps = args.max_steps
    num_workers = args.num_workers
    val_check_interval = 500

    decoder_seq_length = args.decoder_seq_length

    # Submodules configurations
    # switch to 128E with  vlm.Llama4MaverickExperts128Config()
    llama4_config = vlm.Llama4ScoutExperts16Config()
    if args.use_toy_model:
        decoder_seq_length = 4096
        val_check_interval = 50
        llama4_config.vision_transformer_config.num_layers = 2
        llama4_config.language_transformer_config.num_layers = 2
        llama4_config.language_transformer_config.num_moe_experts = 2
        num_workers = 0

    if args.data_type == "llava":
        raise NotImplementedError
    elif args.data_type == "energon":

        task_encoder = Llama4TaskEncoder(
            config=Llama4TaskEncoderConfig(
                hf_path='meta-llama/Llama-4-Scout-17B-16E-Instruct',
            )
        )
        data = EnergonDataModule(
            path=args.data_path,
            train_encoder=task_encoder,
            seq_length=decoder_seq_length,
            global_batch_size=gbs,
            micro_batch_size=mbs,
            num_workers=num_workers,
        )
    elif args.data_type == "mock":
        llama_tokenizer = AutoTokenizer('meta-llama/Llama-4-Scout-17B-16E-Instruct')

        data = vlm.Llama4MockDataModule(
            seq_length=decoder_seq_length,
            global_batch_size=gbs,
            micro_batch_size=mbs,
            tokenizer=llama_tokenizer,
            image_processor=None,
            num_workers=num_workers,
            packed_sequence=args.use_packed_sequence,
        )
    else:
        raise ValueError(f"Data type {args.data_type} not supported")

    from megatron.core.distributed import DistributedDataParallelConfig

    # Training strategy setup
    strategy = nl.MegatronStrategy(
        tensor_model_parallel_size=args.tp_size,
        expert_tensor_parallel_size=args.tp_size,
        expert_model_parallel_size=args.ep_size,
        pipeline_model_parallel_size=args.pp_size,
        encoder_pipeline_model_parallel_size=args.encoder_pp_size,
        context_parallel_size=args.cp_size,
        pipeline_dtype=torch.bfloat16,
        sequence_parallel=True,
        ddp=DistributedDataParallelConfig(
            check_for_nan_in_grad=True,
            grad_reduce_in_fp32=True,
            overlap_grad_reduce=True,
            overlap_param_gather=True,
            average_in_collective=True,
        ),
        ckpt_load_strictness="log_all",
    )

    model = Llama4OmniModel(llama4_config, tokenizer=data.tokenizer)

    # Checkpoint callback setup
    checkpoint_callback = nl.ModelCheckpoint(
        save_last=True,
        monitor="reduced_train_loss",
        save_top_k=2,
        every_n_train_steps=1000,
        dirpath=args.log_dir,
    )

    from nemo.lightning.pytorch.callbacks import NsysCallback

    # Trainer setup
    trainer = nl.Trainer(
        num_nodes=args.num_nodes,
        devices=args.devices,
        max_steps=max_steps,
        accelerator="gpu",
        strategy=strategy,
        plugins=nl.MegatronMixedPrecision(precision="bf16-mixed"),
        callbacks=[
            checkpoint_callback,
            TimingCallback(),
            MegatronCommOverlapCallback(tp_comm_overlap=False),
            NsysCallback(start_step=10, end_step=12, ranks=[0], gen_shape=True),
        ],
        val_check_interval=val_check_interval,
        limit_val_batches=gbs,
        log_every_n_steps=1,
        num_sanity_val_steps=0,
    )

    # Logger setup
    nemo_logger = nl.NeMoLogger(
        log_dir=args.log_dir,
        name=args.name,
        wandb=WandbLogger(project=args.wandb_project, name=args.name) if args.wandb_project is not None else None,
    )

    # Auto resume setup
    resume = nl.AutoResume(
        resume_if_exists=True,
        resume_ignore_no_checkpoint=True,
        resume_from_directory=args.log_dir,
        restore_config=nl.RestoreConfig(path=args.restore_path) if args.restore_path is not None else None,
    )

    # Optimizer and scheduler setup
    opt_config = OptimizerConfig(
        optimizer='adam',
        lr=args.lr,
        adam_beta1=0.9,
        adam_beta2=0.95,
        use_distributed_optimizer=True,
        bf16=True,
    )
    sched = CosineAnnealingScheduler(
        max_steps=trainer.max_steps,
        warmup_steps=150,
        constant_steps=0,
        min_lr=1.0e-07,
    )
    opt = MegatronOptimizerModule(opt_config, sched)

    # PEFT setup
    if args.peft == 'lora':
        peft = vlm.peft.LoRA(
            target_modules=[
                "linear_qkv",
                "linear_proj",
                "linear_fc1",
                "linear_fc2",
            ]
        )
    else:
        peft = None

    llm.finetune(
        model=model,
        data=data,
        trainer=trainer,
        peft=peft,
        log=nemo_logger,
        optim=opt,
        resume=resume,
    )


if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="Llama4 Model Training Script")

    # Argument parsing
    parser.add_argument("--data_type", type=str, required=False, default="mock", help="mock | energon")
    parser.add_argument("--data_path", type=str, required=False, default=None, help="Path to the dataset JSON file")
    parser.add_argument(
        "--log_dir", type=str, required=False, default="/results", help="Directory for logging and checkpoints"
    )
    parser.add_argument(
        "--language_model_path", type=str, required=False, default=None, help="Path to the pretrained language model"
    )
    parser.add_argument(
        "--restore_path", type=str, required=False, default=None, help="Path to restore model from checkpoint"
    )
    parser.add_argument("--devices", type=int, required=False, default=1)
    parser.add_argument("--num_workers", type=int, required=False, default=4)
    parser.add_argument("--num_nodes", type=int, required=False, default=1)
    parser.add_argument("--max_steps", type=int, required=False, default=5190)
    parser.add_argument("--tp_size", type=int, required=False, default=1)
    parser.add_argument("--pp_size", type=int, required=False, default=1)
    parser.add_argument("--cp_size", type=int, required=False, default=1)
    parser.add_argument("--ep_size", type=int, required=False, default=1)
    parser.add_argument("--encoder_pp_size", type=int, required=False, default=0)
    parser.add_argument("--projector_type", type=str, required=False, default="mcore_mlp")
    parser.add_argument("--name", type=str, required=False, default="llama4_pretrain")
    parser.add_argument("--peft", type=str, default='none', help="none | lora")
    parser.add_argument("--wandb_project", type=str, required=False, default=None)
    parser.add_argument("--gbs", type=int, required=False, default=128, help="Global batch size")
    parser.add_argument("--mbs", type=int, required=False, default=1, help="Micro batch size")
    parser.add_argument("--lr", type=float, required=False, default=2.0e-06, help="Learning rate")
    parser.add_argument("--decoder_seq_length", type=int, required=False, default=8192, help="decoder sequence length")
    parser.add_argument(
        "--use_packed_sequence",
        action="store_true",
    )
    parser.add_argument(
        "--use_toy_model",
        action="store_true",
        help="Toy size model used for testing",
    )
    args = parser.parse_args()
    main(args)