Functions — github.com/apple/ml-mgie

hub / github.com/apple/ml-mgie / functions

Functions40 in github.com/apple/ml-mgie

⨍ Functions 40 ◇ Types & classes 10

↓ 7 callersMethodget_model

(self)

mgie_llava.py:230

↓ 3 callersFunctionb2f

(b)

mgie_train.py:43

↓ 3 callersMethodget_vision_tower

(self)

mgie_llava.py:46

↓ 3 callersFunctionresize

(f)

mgie_train.py:44

↓ 2 callersMethod__init__

(self)

mgie_llava.py:188

↓ 2 callersFunction_tokenize_fn

Tokenize a list of strings.

mgie_train.py:219

↓ 2 callersFunctionimg2npy

(f)

mgie_train.py:54

↓ 2 callersFunctionmaybe_zero_3

(param, ignore_status=False, name=None)

mgie_train.py:114

↓ 2 callersFunctionpreprocess

Given a list of sources, each is a conversation list. This transform: 1. Add signal '### ' at the beginning each sentence, with end signal '\

mgie_train.py:445

↓ 1 callersMethod__init__

(self, data_path: str, tokenizer: transformers.PreTrainedTokenizer)

mgie_train.py:482

↓ 1 callersFunction_add_speaker_and_signal

Add speaker and start/end signal on each round.

mgie_train.py:257

↓ 1 callersFunction_mask_targets

(target, tokenized_lens, speakers)

mgie_train.py:246

↓ 1 callersFunctionfind_all_linear_names

(model)

mgie_train.py:162

↓ 1 callersMethodforward

(self, llm, emb)

mgie_llava.py:198

↓ 1 callersFunctionget_peft_state_maybe_zero_3

(named_params, bias)

mgie_train.py:129

↓ 1 callersFunctionget_peft_state_non_lora_maybe_zero_3

(named_params, require_grad_only=True)

mgie_train.py:154

↓ 1 callersMethodget_vision_tower

(self)

mgie_llava.py:233

↓ 1 callersMethodinitialize_vision_modules

(self, vision_tower, mm_vision_select_layer, pretrain_mm_mlp_adapter=None, f

mgie_llava.py:52

↓ 1 callersMethodinitialize_vision_tokenizer

(self, mm_use_im_start_end, tokenizer, device, tune_mm_mlp_adapter=False,

mgie_llava.py:362

↓ 1 callersFunctionmake_supervised_data_module

Make dataset and collator for supervised fine-tuning.

mgie_train.py:581

↓ 1 callersFunctionpatch_FSDP_use_orig_params

(func)

mgie_train.py:761

↓ 1 callersFunctionpreprocess_mpt

( sources, tokenizer: transformers.PreTrainedTokenizer, )

mgie_train.py:373

↓ 1 callersFunctionpreprocess_multimodal

( sources: Sequence[str], multimodal_cfg: dict, cur_token_len: int, )

mgie_train.py:278

↓ 1 callersFunctionpreprocess_v1

( sources, tokenizer: transformers.PreTrainedTokenizer, )

mgie_train.py:303

↓ 1 callersFunctionsafe_save_model_for_hf_trainer

Collects the state dict and dump to disk.

mgie_train.py:176

↓ 1 callersFunctionsmart_tokenizer_and_embedding_resize

Resize tokenizer and embedding. Note: This is the unoptimized version that may make your embedding size not be divisible by 64.

mgie_train.py:194

↓ 1 callersFunctiontrain

()

mgie_train.py:602

Method__call__

(self, instances: Sequence[Dict])

(self, data_path: str, tokenizer: transformers.PreTrainedTokenizer, multimod

mgie_train.py:504

Method__init__

(self, config: LlamaConfig)

( self, input_ids: torch.LongTensor = None, attention_mask: Optional[torch.Tensor] = N

mgie_llava.py:89

Methodforward

( self, input_ids: torch.LongTensor = None, attention_mask: Optional[torch.Tensor] = N

mgie_llava.py:243

Functionmake_inputs_require_grad

(module, input, output)

mgie_train.py:658

Methodprepare_inputs_for_generation

( self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs )