Functions — github.com/2U1/Llama3.2-Vision-Finetune

hub / github.com/2U1/Llama3.2-Vision-Finetune / functions

Functions51 in github.com/2U1/Llama3.2-Vision-Finetune

⨍ Functions 51 ◇ Types & classes 10

↓ 5 callersFunctionpad_sequence

Pad a list of sequences to the same length. sequences: list of tensors in [seq_len, *] shape

src/train/data.py:84

↓ 4 callersFunctionget_peft_state_non_lora_maybe_zero_3

(named_params, require_grad_only=True)

src/train/train_utils.py:46

↓ 3 callersFunctionset_requires_grad

(parameters, requires_grad)

src/train/train_sft.py:37

↓ 3 callersFunctionset_requires_grad

(parameters, requires_grad)

src/train/train_dpo.py:37

↓ 2 callersFunction_pad_nm

(tensors, fill, is_bool=False)

src/train/data.py:44

↓ 2 callersFunctionencode_video

(video_path, max_num_frames=10)

src/train/data.py:69

↓ 2 callersFunctionget_peft_state_maybe_zero_3

(named_params, bias)

src/train/train_utils.py:21

↓ 2 callersFunctionmaybe_zero_3

(param, ignore_status=False, name=None)

src/train/train_utils.py:7

↓ 2 callersFunctionpad_cross_attention_mask

(cross_attention_masks)

src/train/data.py:103

↓ 2 callersFunctionpad_visual_batch

( batch_pixel_values, batch_aspect_ratio_ids, batch_aspect_ratio_mask, max_num_images=None,

src/train/data.py:16

↓ 2 callersFunctionrank0_print

(*args)

src/train/train_sft.py:16

↓ 2 callersFunctionrank0_print

(*args)

src/train/train_dpo.py:16

↓ 2 callersFunctionreplace_image_tokens

(input_string, start_count=0)

src/train/data.py:524

↓ 2 callersFunctionsafe_save_model_for_hf_trainer

Collects the state dict and dump to disk.

src/train/train_utils.py:53

↓ 2 callersFunctionvideo_to_image_tokens

(input_string, num_frames)

src/train/data.py:537

↓ 1 callersMethod__init__

( self, data_path: str | list, processor: transformers.ProcessorMixin, data_ar

src/train/data.py:327

↓ 1 callersMethodconcatenated_inputs

( batch: dict[str, Union[list, torch.LongTensor]], padding_value: int )

src/train/dpo_trainer.py:56

↓ 1 callersFunctionconfigure_llm

(model, training_args)

src/train/train_sft.py:54

↓ 1 callersFunctionconfigure_llm

(model, training_args)

src/train/train_dpo.py:54

↓ 1 callersFunctionconfigure_vision_tower

(model, training_args, compute_dtype, device)

src/train/train_sft.py:41

↓ 1 callersFunctionconfigure_vision_tower

(model, training_args, compute_dtype, device)

src/train/train_dpo.py:41

↓ 1 callersFunctionfind_target_linear_names

(model, num_lora_modules=-1, lora_namespan_exclude=[], verbose=True)

src/train/train_sft.py:20

↓ 1 callersFunctionfind_target_linear_names

(model, num_lora_modules=-1, lora_namespan_exclude=[], verbose=True)

src/train/train_dpo.py:20

↓ 1 callersFunctionget_model_name_from_path

(model_path)

src/utils.py:77

↓ 1 callersFunctionllava_to_openai

(conversations, is_video=False, num_frames=None)

src/train/data.py:542

↓ 1 callersFunctionload_pretrained_model

(model_path, model_base, model_name, load_8bit=False, load_4bit=False, device_map="

src/utils.py:16

↓ 1 callersFunctionmake_dpo_data_module

Make dataset and collator for DPO fine-tuning.

src/train/data.py:577

↓ 1 callersFunctionmake_supervised_data_module

Make dataset and collator for supervised fine-tuning.

src/train/data.py:566

↓ 1 callersFunctionmerge_lora

(args)

src/merge_lora_weights.py:3

↓ 1 callersFunctiontrain

()

src/train/train_sft.py:58

↓ 1 callersFunctiontrain

()

src/train/train_dpo.py:58

↓ 1 callersFunctionuniform_sample

src/train/data.py:270

Method__call__

(self, examples)

src/train/data.py:460

Method__getitem__

(self, i)

src/train/data.py:146

Method__getitem__

(self, i)

src/train/data.py:349

Method__init__

(self, *args, **kwargs)

src/train/dpo_trainer.py:43

Method__init__

( self, data_path: str | list, processor: transformers.ProcessorMixin, data_ar

src/train/data.py:124

Method__init__

(self, pad_token_id: int, processor: transformers.ProcessorMixin)

src/train/data.py:266

Method__init__

(self, pad_token_id: int, processor: transformers.ProcessorMixin)

src/train/data.py:456

Method__init__

(self, *args, **kwargs)

src/train/trainer.py:39

Method__len__

(self)

src/train/data.py:143

Method__len__

(self)

src/train/data.py:346

Method_prepare_dataset

( self, dataset, processing_class, args, dataset_name )

src/train/dpo_trainer.py:46

Method_save_checkpoint

(self, model, trial)

src/train/dpo_trainer.py:178

Method_save_checkpoint

(self, model, trial)

src/train/trainer.py:120

Methodconcatenated_forward

(self, model: nn.Module, batch: dict[str, Union[list, torch.LongTensor]])

src/train/dpo_trainer.py:87

Methodcreate_optimizer

Setup the optimizer. We provide a reasonable default that works well. If you want to use something else, you can pass a tuple in the

src/train/trainer.py:42

Functiondisable_torch_init

Disable the redundant torch default initialization to accelerate model creation.

src/utils.py:8

Functionmaybe_zero_3

(param, ignore_status=False, name=None)

src/train/dpo_trainer.py:27

Functionmaybe_zero_3

(param, ignore_status=False, name=None)

src/train/trainer.py:23