Skip to main content
Version: Next

inline::huggingface-cpu

Description​

HuggingFace-based post-training provider for fine-tuning models using the HuggingFace ecosystem.

Configuration​

FieldTypeRequiredDefaultDescription
device<class 'str'>Nocuda
distributed_backendLiteral['fsdp', 'deepspeed'No
checkpoint_formatLiteral['full_state', 'huggingface'Nohuggingface
chat_template<class 'str'>No`<user
model_specific_config<class 'dict'>No&#123;'trust_remote_code': True, 'attn_implementation': 'sdpa'&#125;
max_seq_length<class 'int'>No2048
gradient_checkpointing<class 'bool'>NoFalse
save_total_limit<class 'int'>No3
logging_steps<class 'int'>No10
warmup_ratio<class 'float'>No0.1
weight_decay<class 'float'>No0.01
dataloader_num_workers<class 'int'>No4
dataloader_pin_memory<class 'bool'>NoTrue
dpo_beta<class 'float'>No0.1
use_reference_model<class 'bool'>NoTrue
dpo_loss_typeLiteral['sigmoid', 'hinge', 'ipo', 'kto_pair'Nosigmoid
dpo_output_dir<class 'str'>No

Sample Configuration​

checkpoint_format: huggingface
distributed_backend: null
device: cpu
dpo_output_dir: ~/.llama/dummy/dpo_output