# Config for multi-device full finetuning in full_finetune_distributed.py
# using a Llama3 8B Instruct model
#
# This config assumes that you've run the following command before launching
# this run:
#   tune download meta-llama/Meta-Llama-3-8B-Instruct --output-dir /tmp/Meta-Llama-3-8B-Instruct --hf-token <HF_TOKEN>
#
# To launch on 4 devices, run the following command from root:
#   tune run --nproc_per_node 4 full_finetune_distributed --config llama3/8B_full
#
# You can add specific overrides through the command line. For example
# to override the checkpointer directory while launching training
# you can run:
#   tune run --nproc_per_node 4 full_finetune_distributed --config llama3/8B_full checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
#
# This config works best when the model is being fine-tuned on 2+ GPUs.
# Single device full finetuning requires more memory optimizations. It's
# best to use 8B_full_single_device.yaml for those cases
# Tokenizer
tokenizer:
  _component_: torchtune.models.llama3.llama3_s_tokenizer
  path: ../model_zoo/tokenizer.model
  max_seq_len: 4096

# Dataset
dataset:
  _component_: torchtune.datasets.chat_dataset
  source: jan-hq/mixed-instruction-speech-multiturn-noise-clean
  conversation_style: openai
  max_seq_len: 4096
  split: train
  train_on_input: True

seed: 42
shuffle: False
# Model Arguments
model:
  _component_: torchtune.models.llama3_1.llama3_1_s_8b
  # path: model_zoo/Llama3.1_s_8b_init
checkpointer:
  _component_: torchtune.utils.FullModelHFCheckpointerSaveSteps
  checkpoint_dir: ../model_zoo/llama3.1-s-cp-7000
  checkpoint_files: [
    model-00001-of-00004.safetensors,
    model-00002-of-00004.safetensors,
    model-00003-of-00004.safetensors,
    model-00004-of-00004.safetensors,
  ]
  recipe_checkpoint: null
  output_dir: ../model_zoo/llama3-s-instruct-lr-3e-5
  model_type: LLAMA3
resume_from_checkpoint: False
save_every_n_steps: 200
max_checkpoints: 3
# Fine-tuning arguments
batch_size: 4
epochs: 1
max_steps_per_epoch: null
gradient_accumulation_steps: 8
compile: False
# Optimizer and Scheduler
optimizer:
  _component_: torch.optim.AdamW #change this to use adam_mini: torchtune.modules.optimizer.Adam_mini
  weight_decay: 0.005
  lr: 1.5e-5
  fused: True
lr_scheduler:
  _component_: torchtune.modules.get_cosine_schedule_with_warmup
  num_warmup_steps: 8

loss:
  _component_: torch.nn.CrossEntropyLoss

fsdp:
  cpu_offload: False

# Training env
device: cuda
dtype: bf16

# Memory management
enable_activation_checkpointing: True
memory_efficient_fsdp_wrap: True
ac_mode: 'selective'


# Logging
metric_logger:
  _component_: torchtune.utils.metric_logging.DiskLogger
  log_dir: ${output_dir}
output_dir: ../model_zoo/Llama3-instruct-log-lr-3e-5/
log_every_n_steps: 1
log_peak_memory_stats: False