Llama-3-70B config (works on one GPU no deepspeed; OOMs on multiple GPUs during merge)

khk-llama-3-70B.yml

base_model: meta-llama/Meta-Llama-3-70B
# deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json

load_in_8bit: false
load_in_4bit: true
strict: false

datasets:
  # This will be the path used for the data when it is saved to the Volume in the cloud.
  - path: data.jsonl
    ds_type: json
    type: input_output

dataset_prepared_path: last_run_prepared
val_set_size: 0.05
output_dir: ./lora-out

sequence_len: 1024
sample_packing: false
eval_sample_packing: false
pad_to_sequence_len: false

adapter: qlora
lora_model_dir:
# lora_r: 128
lora_r: 16
lora_alpha: 16
# lora_modules_to_save: [embed_tokens, lm_head]
lora_dropout: 0.05
# lora_target_linear: true
lora_fan_in_fan_out:
lora_target_modules:
  - q_proj
  - v_proj
#  - k_proj
#  - o_proj
#  - gate_proj
#  - down_proj
#  - up_proj

wandb_project: khk-llama-3-70b
wandb_entity:
wandb_watch:
wandb_run_id:

gradient_accumulation_steps: 1
micro_batch_size: 1
num_epochs: 1
# optimizer: adamw_torch
optimizer: adamw_8bit
lr_scheduler: constant
learning_rate: 1e-5

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

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

warmup_steps: 100
# evals_per_epoch: 5
eval_steps: 1.0
eval_batch_size: 1
saves_per_epoch: 5
save_total_limit: 10
eval_table_size:
save_steps:
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  pad_token: <|end_of_text|>