base_model: Fizzarolli/phi3-4x4b-uninitialized
model_type: MixtralForCausalLM
tokenizer_type: AutoTokenizer  # PreTrainedTokenizerFast

load_in_8bit: false
load_in_4bit: true
strict: false

datasets:
  - path: NeelNanda/pile-10k
    type: completion
    field: text
  - path: BEE-spoke-data/gutenberg-en-v1-clean
    type: completion
    field: text
dataset_prepared_path: last_run_prepared
val_set_size: 0.05
output_dir: ./lora-out

adapter: qlora
lora_model_dir:

sequence_len: 4072
sample_packing: true
pad_to_sequence_len: true

lora_r: 32
lora_alpha: 64
lora_dropout: 0.05
lora_target_modules:
    - k_proj
    - q_proj
    - v_proj
    - o_proj
lora_target_linear: true
lora_fan_in_fan_out:

wandb_project: phixtral3
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 6
num_epochs: 1
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 0.000001

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 10
evals_per_epoch: 0
eval_table_size:
saves_per_epoch: 20
debug:
fsdp:
  - full_shard
  - auto_wrap
fsdp_config:
  fsdp_limit_all_gathers: true
  fsdp_sync_module_states: true
  fsdp_offload_params: true
  fsdp_use_orig_params: false
  fsdp_cpu_ram_efficient_loading: true
  fsdp_transformer_layer_cls_to_wrap: MixtralSparseMoeBlock
  fsdp_state_dict_type: FULL_STATE_DICT
  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
  fsdp_sharding_strategy: FULL_SHARD
  fsdp_forward_prefetch: false
  fsdp_backward_prefetch: BACKWARD_PRE
weight_decay: 0.0
special_tokens: