h9%FSSKrSSKJrJr SSKJr \"SS55rg)N) dataclassfield)Optionalc\rSrSr%Sr\"SSS0S9r\\\ S'\"SSS 0S9r \\ S '\"SS /S QS .S9r \\\ S'\"SSS0S9r \ \ S'\"SSS0S9r\\\ S'\"SSS0S9r\ \ S'\"SSS0S9r\\ S'\"SSS0S9r\\ S'\"SSS0S9r\\ S'\"SSS0S9r\\\\ S '\"SSS!0S9r\\\\ S"'\"SSS#0S9r\\\\ S$'\"S%SS&0S9r\\ S''\"SSS(0S9r\ \ S)'\"SSS*0S9r\ \ S+'\"SSS,0S9r\ \ S-'\"SSS.0S9r\ \ S/'\"S0S1S2S0/S .S9r\\ S3'\"SSS40S9r\ \ S5'\"SS /S QS .S9r \\\ S6'S7r!S8r"g)9 ModelConfigu Configuration class for the models. Using [`~transformers.HfArgumentParser`] we can turn this class into [argparse](https://docs.python.org/3/library/argparse#module-argparse) arguments that can be specified on the command line. Parameters: model_name_or_path (`str` or `None`, *optional*, defaults to `None`): Model checkpoint for weights initialization. model_revision (`str`, *optional*, defaults to `"main"`): Specific model version to use. It can be a branch name, a tag name, or a commit id. dtype (`Literal["auto", "bfloat16", "float16", "float32"]` or `None`, *optional*, defaults to `None`): Override the default `torch.dtype` and load the model under this dtype. Possible values are - `"bfloat16"`: `torch.bfloat16` - `"float16"`: `torch.float16` - `"float32"`: `torch.float32` - `"auto"`: Automatically derive the dtype from the model's weights. trust_remote_code (`bool`, *optional*, defaults to `False`): Whether to allow for custom models defined on the Hub in their own modeling files. This option should only be set to `True` for repositories you trust and in which you have read the code, as it will execute code present on the Hub on your local machine. attn_implementation (`str` or `None`, *optional*, defaults to `None`): Which attention implementation to use. You can run `--attn_implementation=flash_attention_2`, in which case you must install this manually by running `pip install flash-attn --no-build-isolation`. use_peft (`bool`, *optional*, defaults to `False`): Whether to use PEFT for training. lora_r (`int`, *optional*, defaults to `16`): LoRA R value. lora_alpha (`int`, *optional*, defaults to `32`): LoRA alpha. lora_dropout (`float`, *optional*, defaults to `0.05`): LoRA dropout. lora_target_modules (`Union[str, list[str]]` or `None`, *optional*, defaults to `None`): LoRA target modules. lora_target_parameters (`Union[str, list[str]]` or `None`, *optional*, defaults to `None`): List of target parameters for LoRA. lora_modules_to_save (`list[str]` or `None`, *optional*, defaults to `None`): Model layers to unfreeze & train. lora_task_type (`str`, *optional*, defaults to `"CAUSAL_LM"`): Task type to pass for LoRA (use `"SEQ_CLS"` for reward modeling). use_rslora (`bool`, *optional*, defaults to `False`): Whether to use Rank-Stabilized LoRA, which sets the adapter scaling factor to `lora_alpha/√r`, instead of the original default value of `lora_alpha/r`. use_dora (`bool`, *optional*, defaults to `False`): Enable [Weight-Decomposed Low-Rank Adaptation (DoRA)](https://huggingface.co/papers/2402.09353). This technique decomposes the updates of the weights into two parts, magnitude and direction. Direction is handled by normal LoRA, whereas the magnitude is handled by a separate learnable parameter. This can improve the performance of LoRA, especially at low ranks. Right now, DoRA only supports linear and Conv2D layers. DoRA introduces a bigger overhead than pure LoRA, so it is recommended to merge weights for inference. load_in_8bit (`bool`, *optional*, defaults to `False`): Whether to use 8 bit precision for the base model. Works only with LoRA. load_in_4bit (`bool`, *optional*, defaults to `False`): Whether to use 4 bit precision for the base model. Works only with LoRA. bnb_4bit_quant_type (`str`, *optional*, defaults to `"nf4"`): Quantization type (`"fp4"` or `"nf4"`). use_bnb_nested_quant (`bool`, *optional*, defaults to `False`): Whether to use nested quantization. Nhelpz,Model checkpoint for weights initialization.)defaultmetadatamodel_name_or_pathmainzSSpecific model version to use. It can be a branch name, a tag name, or a commit id.model_revisionzGOverride the default `torch.dtype` and load the model under this dtype.)autobfloat16float16float32)r choicesdtypeFzWhether to allow for custom models defined on the Hub in their own modeling files. This option should only be set to `True` for repositories you trust and in which you have read the code, as it will execute code present on the Hub on your local machine.trust_remote_codezWhich attention implementation to use. You can run `--attn_implementation=flash_attention_2`, in which case you must install this manually by running `pip install flash-attn --no-build-isolation`.attn_implementationz!Whether to use PEFT for training.use_peftz LoRA R value.lora_r z LoRA alpha. lora_alphag?z LoRA dropout. lora_dropoutzLoRA target modules.lora_target_modulesz#List of target parameters for LoRA.lora_target_parametersz!Model layers to unfreeze & train.lora_modules_to_save CAUSAL_LMz?Task type to pass for LoRA (use 'SEQ_CLS' for reward modeling).lora_task_typeuWhether to use Rank-Stabilized LoRA, which sets the adapter scaling factor to `lora_alpha/√r`, instead of the original default value of `lora_alpha/r`. use_rsloraaEnable Weight-Decomposed Low-Rank Adaptation (DoRA). This technique decomposes the updates of the weights into two parts, magnitude and direction. Direction is handled by normal LoRA, whereas the magnitude is handled by a separate learnable parameter. This can improve the performance of LoRA, especially at low ranks. Right now, DoRA only supports linear and Conv2D layers. DoRA introduces a bigger overhead than pure LoRA, so it is recommended to merge weights for inference.use_dorazHWhether to use 8 bit precision for the base model. Works only with LoRA. load_in_8bitzHWhether to use 4 bit precision for the base model. Works only with LoRA. load_in_4bitnf4zQuantization type.fp4bnb_4bit_quant_typez#Whether to use nested quantization.use_bnb_nested_quant torch_dtypecUR(aUR(a [S5eUR(a=UR(d,[ R "S[5 URUl[URS5(a/[UR5S:XaURSUl ggg)Nz8You can't use 8 bit and 4 bit precision at the same timez^`torch_dtype` is deprecated and will be removed in version 0.27.0, please use `dtype` instead.__len__r) r$r% ValueErrorr*rwarningswarnDeprecationWarninghasattrrlen)selfs R/home/james-whalen/.local/lib/python3.13/site-packages/trl/trainer/model_config.py __post_init__ModelConfig.__post_init__s   !2!2WX X   DJJ MMp" ))DJ 4++Y 7 7C@X@XHd/*FC-(J /*L%05010$s),38?@3HT#Y/16=>1(49- [\NC G J c  Hd deL$deL$ %.E5>J "'?@"$ "']A "K# Cr8r)r/ dataclassesrrtypingrrrEr8r5rHs/( tCtC tCr8