Unify Efficient Fine-Tuning of 100+ LLMs
new_special_tokens
argumentexport_device
argumentquantization_device_map
argumentexamples
folderdataset_info.json
by specifying --dataset_dir ONLINE
moe_aux_loss_coef
to control the coefficient of auxiliary loss in MoE models.addtional_target
in unsloth by @kno10 in #3201This patch mainly fixes #2983
In commit 9bec3c98a22c91b1c28fda757db51eb780291641, we built the optimizer and scheduler inside the trainers, which inadvertently introduced a bug: when DeepSpeed was enabled, the trainers in transformers would build an optimizer and scheduler before calling the create_optimizer_and_scheduler
method [1], then the optimizer created by our method would overwrite the original one, while the scheduler would not. Consequently, the scheduler would no longer affect the learning rate in the optimizer, leading to a regression in the training result. We have fixed this bug in 3bcd41b639899e72bcabc51d59bac8967af19899 and 8c77b1091296e204dc3c8c1f157c288ca5b236bd. Thank @HideLord for helping us identify this critical bug.
[1] https://github.com/huggingface/transformers/blob/v4.39.1/src/transformers/trainer.py#L1877-L1881
We have also fixed #2961 #2981 #2982 #2983 #2991 #3010
--infer_backend vllm
apply_chat_template
by adding a chat template to the tokenizer after fine-tuningtests/llama_pro.py
for usageuse_rslora
option for the LoRA methodtorch_dtype
check in export model by @fenglui in #2262use_cache
in export model by @yhyu13 in #2266test_toolcall.py
by @mini-tiger in #2435--dataset glaive_toolcall
for tool using #2226--use_unsloth
, see benchmarking here
checkpoint_dir
and use adapter_name_or_path
insteadresume_lora_training
with create_new_adapter
llmtuner.model.patcher
The above changes were made by @hiyouga in #1864
dpo_ftx
argument, suggested by @lylcst in https://github.com/hiyouga/LLaMA-Factory/issues/1347#issuecomment-1846943606
export_quantization_bit
and export_quantization_dataset
arguments--stage=rm
in api_demo.py
--reward_model_type api
export_size
argument