4B SFT long context

config.yaml

# Model arguments
model_name_or_path: Qwen/Qwen3-4B-Instruct-2507
model_revision: main
attn_implementation: kernels-community/vllm-flash-attn3

# Data training arguments
dataset_mixture:
  datasets:
    - id: hf-carbon/seqqa-sft-v1_gemma-4-31B-it
      config: graded_correct
      split: train            
      columns:
        - messages
      weight: 1.0
  seed: 0
dataset_num_proc: 12

# SFT trainer config
assistant_only_loss: false
bf16: true
eval_strategy: 'no'
gradient_accumulation_steps: 2
gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
learning_rate: 1.0e-06
log_level: info
logging_steps: 1
logging_strategy: steps
lr_scheduler_type: cosine_with_min_lr
lr_scheduler_kwargs:
  min_lr_rate: 0.1
max_grad_norm: 0.2
max_length: 36864
num_train_epochs: 5
output_dir: data/Qwen3-4B-Instruct-2507-SFT
packing: true
per_device_train_batch_size: 2
per_device_eval_batch_size: 2
push_to_hub: true
report_to:
- wandb
save_strategy: epoch
save_total_limit: 1
seed: 42
use_liger_kernel: true
warmup_ratio: 0.03
wandb_entity: huggingface

"""
accelerate launch --config_file examples/accelerate_configs/deepspeed_zero3.yaml dev/scripts/sft.py --config dev/recipes/carbon-4b/sft/config_dummy.yaml
"""

import logging
import os
import sys

import datasets
import transformers
from transformers import set_seed
from transformers.trainer_utils import get_last_checkpoint

from trl import ModelConfig, SFTTrainer, TrlParser, get_peft_config
from trl.internal.callbacks import LogTrainingSamplesCallback, get_callbacks
from trl.internal.configs import ScriptArguments, SFTConfig
from trl.internal.dataset_utils import get_dataset
from trl.internal.hub import check_hub_revision_exists
from trl.internal.logging import init_trackio_training, init_wandb_training
from trl.internal.model_utils import get_model, get_tokenizer


logger = logging.getLogger(__name__)


def main(script_args, training_args, model_args):
    check_hub_revision_exists(training_args)
    # Set seed for reproducibility
    set_seed(training_args.seed)

    ###############
    # Setup logging
    ###############
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
        datefmt="%Y-%m-%d %H:%M:%S",
        handlers=[logging.StreamHandler(sys.stdout)],
    )
    log_level = training_args.get_process_log_level()
    logger.setLevel(log_level)
    datasets.utils.logging.set_verbosity(log_level)
    transformers.utils.logging.set_verbosity(log_level)
    transformers.utils.logging.enable_default_handler()
    transformers.utils.logging.enable_explicit_format()

    logger.info(f"Model parameters {model_args}")
    logger.info(f"Script parameters {script_args}")
    logger.info(f"Training parameters {training_args}")

    # Check for last checkpoint
    last_checkpoint = None
    if os.path.isdir(training_args.output_dir):
        last_checkpoint = get_last_checkpoint(training_args.output_dir)
    if last_checkpoint is not None and training_args.resume_from_checkpoint is None:
        logger.info(f"Checkpoint detected, resuming training at {last_checkpoint=}.")

    if "wandb" in training_args.report_to:
        init_wandb_training(training_args)
    if "trackio" in training_args.report_to:
        init_trackio_training(training_args)

    ################
    # Load datasets
    ################
    dataset = get_dataset(script_args)
    ################
    # Load tokenizer
    ################
    tokenizer = get_tokenizer(model_args, training_args)

    ###################
    # Load model
    ###################
    logger.info("*** Loading model ***")
    model = get_model(model_args, training_args)

    ############################
    # Initialize the SFT Trainer
    ############################
    trainer = SFTTrainer(
        model=model,
        args=training_args,
        train_dataset=dataset[script_args.dataset_train_split],
        eval_dataset=(dataset[script_args.dataset_test_split] if training_args.eval_strategy != "no" else None),
        processing_class=tokenizer,
        peft_config=get_peft_config(model_args),
        callbacks=get_callbacks(training_args, model_args),
    )

    # Add LogTrainingSamplesCallback to log training data samples to wandb
    if "wandb" in training_args.report_to or "trackio" in training_args.report_to:
        log_samples_callback = LogTrainingSamplesCallback(trainer=trainer, n_samples=50, decode_columns=["input_ids"])
        trainer.add_callback(log_samples_callback)

    ###############
    # Training loop
    ###############
    logger.info("*** Train ***")
    checkpoint = None
    if training_args.resume_from_checkpoint is not None:
        checkpoint = training_args.resume_from_checkpoint
    elif last_checkpoint is not None:
        checkpoint = last_checkpoint
    train_result = trainer.train(resume_from_checkpoint=checkpoint)
    metrics = train_result.metrics
    metrics["train_samples"] = len(dataset[script_args.dataset_train_split])
    trainer.log_metrics("train", metrics)
    trainer.save_metrics("train", metrics)
    trainer.save_state()

    ##################################
    # Save model and create model card
    ##################################
    logger.info("*** Save model ***")
    # Align the model's generation config with the tokenizer's eos token
    # to avoid unbounded generation in the transformers `pipeline()` function
    trainer.model.generation_config.eos_token_id = tokenizer.eos_token_id
    trainer.save_model(training_args.output_dir)
    logger.info(f"Model saved to {training_args.output_dir}")

    # Save everything else on main process
    kwargs = {
        "dataset_name": script_args.dataset_name,
        "tags": ["trl-internal"],
    }
    if trainer.accelerator.is_main_process:
        trainer.create_model_card(**kwargs)
        # Restore k,v cache for fast inference
        trainer.model.config.use_cache = True
        trainer.model.config.save_pretrained(training_args.output_dir)

    ##########
    # Evaluate
    ##########
    if training_args.do_eval:
        logger.info("*** Evaluate ***")
        metrics = trainer.evaluate()
        metrics["eval_samples"] = len(dataset[script_args.dataset_test_split])
        trainer.log_metrics("eval", metrics)
        trainer.save_metrics("eval", metrics)

    #############
    # push to hub
    #############
    if training_args.push_to_hub:
        logger.info("Pushing to hub...")
        trainer.push_to_hub(**kwargs)


if __name__ == "__main__":
    parser = TrlParser((ScriptArguments, SFTConfig, ModelConfig))
    script_args, training_args, model_args = parser.parse_args_and_config()
    main(script_args, training_args, model_args)

DeepSpeed config

compute_environment: LOCAL_MACHINE
debug: false
deepspeed_config:
  deepspeed_multinode_launcher: standard
  offload_optimizer_device: none
  offload_param_device: none
  zero3_init_flag: true
  zero3_save_16bit_model: true
  zero_stage: 3
distributed_type: DEEPSPEED
downcast_bf16: 'no'
machine_rank: 0
main_training_function: main
mixed_precision: bf16
num_machines: 1
num_processes: 8
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false