# the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loading
# FSDP cpu_ram_efficient_loading is used to reduce the initial CPU memory usage when loading the model
base_model: axolotl-ai-co/gpt-oss-120b-dequantized

use_kernels: false

dp_shard_size: 16  # requires 2x8xH100 nodes

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding

datasets:
  - path: HuggingFaceH4/Multilingual-Thinking
    type: chat_template
    field_thinking: thinking
    template_thinking_key: thinking

dataset_prepared_path: last_run_prepared
val_set_size: 0
output_dir: ./outputs/gpt-oss-out/
save_total_limit: 2  # the 120B model can use up to 720GB of disk space per checkpoint, so let's only keep the last 2

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

trackio_project_name:
trackio_run_name:
trackio_space_id:

gradient_accumulation_steps: 2
micro_batch_size: 1
num_epochs: 1

optimizer: adamw_torch_fused  # 8bit optimizers do not work with FSDP2 offload
lr_scheduler: constant_with_warmup
learning_rate: 2e-5

bf16: true
tf32: true

flash_attention: true
attn_implementation: kernels-community/vllm-flash-attn3  # this is not needed if using flash_attn >= 2.8.3

gradient_checkpointing: true
activation_offloading: true

logging_steps: 1
saves_per_epoch: 1

warmup_ratio: 0.03

special_tokens:
eot_tokens:
  - "<|end|>"

fsdp_version: 2
fsdp_config:
  offload_params: true
  state_dict_type: SHARDED_STATE_DICT
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: GptOssDecoderLayer
  reshard_after_forward: true
  cpu_ram_efficient_loading: true