Skip to content

Instantly share code, notes, and snippets.

@mtisz
Created May 15, 2024 16:47
Show Gist options
  • Save mtisz/2e9f7d8acb1a65f0b58f2427a402f387 to your computer and use it in GitHub Desktop.
Save mtisz/2e9f7d8acb1a65f0b58f2427a402f387 to your computer and use it in GitHub Desktop.
Axolotl Config for Llama-3-70B QLoRA
base_model: meta-llama/Meta-Llama-3-70B
model_type: LlamaForCausalLM
tokenizer_type: AutoTokenizer
load_in_8bit: false
load_in_4bit: true
strict: false
datasets:
- path: /home/migel/ai_datasets/tess-v1.5b-chatml.jsonl
type: sharegpt
conversation: llama3
chat_template: llama3
adapter: qlora
lora_r: 128
lora_alpha: 16
lora_modules_to_save: [embed_tokens, lm_head]
lora_dropout: 0.05
lora_target_linear: true
dataset_prepared_path: last_run_prepared
val_set_size: 0.05
output_dir: /home/migel/whiterabbitneo-llama3-70B
sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true
wandb_project: llama-3
wandb_watch:
wandb_run_id:
wandb_log_model:
gradient_accumulation_steps: 4
micro_batch_size: 3
num_epochs: 2
optimizer: adamw_8bit
lr_scheduler: constant
learning_rate: 1e-5
train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false
gradient_checkpointing: true
gradient_checkpointing_kwargs:
use_reentrant: false
early_stopping_patience:
resume_from_checkpoint:
logging_steps: 1
xformers_attention:
flash_attention: true
warmup_steps: 100
evals_per_epoch: 5
eval_table_size:
saves_per_epoch: 5
save_total_limit: 10
save_steps:
debug:
deepspeed: /home/migel/axolotl/deepspeed_configs/zero3_bf16.json
weight_decay: 0.00
fsdp:
fsdp_config:
special_tokens:
pad_token: "<|end_of_text|>"
@rumbleFTW
Copy link

How much VRAM is needed for fine-tuning using this config?

@mtisz
Copy link
Author

mtisz commented Jun 9, 2024

Hey, this is optimized for 320GB VRAM. You can play around with micro_batch_size, gradient_accumulation_steps and lora_r to suit your needs.

@rumbleFTW
Copy link

Sweet. This is really helpful. Thanks!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment