| from dataclasses import dataclass, field |
| from typing import Optional |
|
|
| from transformers import TrainingArguments |
|
|
|
|
| @dataclass |
| class ModelArguments: |
| model_id: Optional[str] = field(default="Qwen/Qwen2.5-VL-7B-Instruct") |
| model_path: Optional[str] = field(default="Qwen/Qwen2.5-VL-7B-Instruct") |
| anchor_model_id: str = field(default=None, metadata={"help": "List of anchor model ids"}) |
| anchor_loss_weight: str = field(default='[1.0, 1.0, 1.0, 1.0, 1.0]', metadata={"help": "List of anchor loss weights"}) |
| anchor_tokens: str = field(default='[64, 64, 64, 64, 64]', metadata={"help": "List of anchor tokens"}) |
| |
| |
| vision_tower: str = field(default=None, metadata={"help": "Vision tower"}) |
| pretrain_mm_mlp_adapter: str = field(default=None, metadata={"help": "Pretrain mm mlp adapter"}) |
| mm_projector_type: str = field(default=None, metadata={"help": "Mm projector type"}) |
| mm_vision_select_layer: int = field(default=None, metadata={"help": "Mm vision select layer"}) |
| mm_vision_select_feature: str = field(default=None, metadata={"help": "Mm vision select feature"}) |
| mm_patch_merge_type: str = field(default=None, metadata={"help": "Mm patch merge type"}) |
| mm_use_im_patch_token: bool = field(default=False, metadata={"help": "Mm use im patch token"}) |
| mm_use_im_start_end: bool = field(default=False, metadata={"help": "Mm use im start end"}) |
| |
|
|
| @dataclass |
| class TrainingArguments(TrainingArguments): |
| cache_dir: Optional[str] = field(default=None) |
| optim: str = field(default="adamw_torch") |
| adam_beta1: float = field(default=0.9) |
| adam_beta2: float = field(default=0.999) |
| adam_epsilon: float = field(default=1e-8) |
|
|
| freeze_vision_tower: bool = field(default=False) |
| freeze_llm: bool = field(default=False) |
| tune_merger: bool = field(default=False) |
| disable_flash_attn2: bool = field(default=False) |
|
|
| max_seq_length: int = field( |
| default=32768, |
| metadata={ |
| "help": |
| "Maximum sequence length. Sequences will be right padded (and possibly truncated)." |
| }, |
| ) |
|
|
| double_quant: bool = field( |
| default=True, |
| metadata={"help": "Compress the quantization statistics through double quantization."} |
| ) |
| quant_type: str = field( |
| default="nf4", |
| metadata={"help": "Quantization data type to use. Should be one of `fp4` or `nf4`."} |
| ) |
| bits: int = field( |
| default=16, |
| metadata={"help": "How many bits to use."} |
| ) |
| lora_enable: bool = False |
| vision_lora: bool = False |
| use_dora: bool = False |
| lora_rank: int = 64 |
| lora_alpha: int = 16 |
| lora_dropout: float = 0.05 |
| lora_weight_path: str = "" |
| lora_bias: str = "none" |
| vision_lr: Optional[float] = None |
| merger_lr: Optional[float] = None |
| lora_namespan_exclude: str = field(default=None, metadata={"help": "List of namespan to exclude for LoRA"}) |
| num_lora_modules: int = -1 |
| use_liger: bool = True |
| |
| |
| training_stage: str = field(default="full", metadata={"help": "Training stage, should be one of `start` or `full`."}) |
| projection_layer_lr: Optional[float] = None |
| vqa_only_stage: int = field(default=4000, metadata={"help": "VQA only stage."}) |
|
|
|
|
| @dataclass |
| class DataArguments: |
| data_path: str = field( |
| default=None, metadata={"help": "Path to the training data."} |
| ) |
| lazy_preprocess: bool = False |
| image_folder: Optional[str] = field(default=None) |
| image_min_pixels: Optional[int] = field(default=3136) |
| image_max_pixels: Optional[int] = field(default=12845056) |
| video_min_pixels: Optional[int] = field(default=100352) |
| video_max_pixels: Optional[int] = field(default=602112) |
| image_resized_width: int = field(default=None) |
| image_resized_height: int = field(default=None) |
| fps: float = 1.0 |
| |
| |
| stage_0_step: int = field(default=0, metadata={"help": "Stage 0 step."}) |
| stage_1_step: int = field(default=2000, metadata={"help": "Stage 1 step."}) |
| stage_2_step: int = field(default=4000, metadata={"help": "Stage 2 step."}) |
| |
