| from dataclasses import dataclass, field |
|
|
| @dataclass |
| class VLMConfig: |
| vit_hidden_dim: int = 768 |
| vit_inter_dim: int = 4 * vit_hidden_dim |
| vit_patch_size: int = 16 |
| vit_img_size: int = 512 |
| vit_n_heads: int = 12 |
| vit_dropout: float = 0.0 |
| vit_n_blocks: int = 12 |
| vit_ln_eps: float = 1e-6 |
| vit_cls_flag: bool = False |
| vit_model_type: str = 'google/siglip2-base-patch16-512' |
|
|
| lm_hidden_dim: int = 960 |
| lm_inter_dim: int = 2560 |
| lm_rms_eps: float = 1e-5 |
| lm_re_base: int = 100000 |
| lm_max_position_embeddings: int = 8192 |
| lm_base_vocab_size: int = 49152 |
| extra_token_amount: int = 66 |
| lm_vocab_size: int = lm_base_vocab_size + extra_token_amount |
| lm_n_heads: int = 15 |
| lm_n_kv_heads: int = 5 |
| lm_dropout: float = 0.0 |
| lm_n_blocks: int = 32 |
| lm_attn_scaling: float = 1.0 |
| lm_max_length: int = 4096 |
| lm_use_tokens: bool = False |
| lm_tie_weights: bool = True |
| lm_model_type: str = 'HuggingFaceTB/SmolLM2-360M-Instruct' |
| lm_tokenizer: str = 'HuggingFaceTB/SmolLM2-360M-Instruct' |
| lm_chat_template: str = "{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}" |
|
|
| mp_pixel_shuffle_factor: int = 4 |
| mp_image_token_length: int = 64 |
|
|
| max_img_size: int = 2048 |
| resize_to_max_side_len: bool = True |
| resize_min_side_len: int | None = None |
|
|
| inference_max_img_size: int | None = None |
|
|
| vlm_extra_tokens: dict[str, str] = field(default_factory=lambda: {"image_token": "<|image|>", "global_image_token": "<|global_image|>", |
| "r1c1": "<row_1_col_1>", "r1c2": "<row_1_col_2>", "r1c3": "<row_1_col_3>", "r1c4": "<row_1_col_4>", "r1c5": "<row_1_col_5>", "r1c6": "<row_1_col_6>", "r1c7": "<row_1_col_7>", "r1c8": "<row_1_col_8>", |
| "r2c1": "<row_2_col_1>", "r2c2": "<row_2_col_2>", "r2c3": "<row_2_col_3>", "r2c4": "<row_2_col_4>", "r2c5": "<row_2_col_5>", "r2c6": "<row_2_col_6>", "r2c7": "<row_2_col_7>", "r2c8": "<row_2_col_8>", |
| "r3c1": "<row_3_col_1>", "r3c2": "<row_3_col_2>", "r3c3": "<row_3_col_3>", "r3c4": "<row_3_col_4>", "r3c5": "<row_3_col_5>", "r3c6": "<row_3_col_6>", "r3c7": "<row_3_col_7>", "r3c8": "<row_3_col_8>", |
| "r4c1": "<row_4_col_1>", "r4c2": "<row_4_col_2>", "r4c3": "<row_4_col_3>", "r4c4": "<row_4_col_4>", "r4c5": "<row_4_col_5>", "r4c6": "<row_4_col_6>", "r4c7": "<row_4_col_7>", "r4c8": "<row_4_col_8>", |
| "r5c1": "<row_5_col_1>", "r5c2": "<row_5_col_2>", "r5c3": "<row_5_col_3>", "r5c4": "<row_5_col_4>", "r5c5": "<row_5_col_5>", "r5c6": "<row_5_col_6>", "r5c7": "<row_5_col_7>", "r5c8": "<row_5_col_8>", |
| "r6c1": "<row_6_col_1>", "r6c2": "<row_6_col_2>", "r6c3": "<row_6_col_3>", "r6c4": "<row_6_col_4>", "r6c5": "<row_6_col_5>", "r6c6": "<row_6_col_6>", "r6c7": "<row_6_col_7>", "r6c8": "<row_6_col_8>", |
| "r7c1": "<row_7_col_1>", "r7c2": "<row_7_col_2>", "r7c3": "<row_7_col_3>", "r7c4": "<row_7_col_4>", "r7c5": "<row_7_col_5>", "r7c6": "<row_7_col_6>", "r7c7": "<row_7_col_7>", "r7c8": "<row_7_col_8>", |
| "r8c1": "<row_8_col_1>", "r8c2": "<row_8_col_2>", "r8c3": "<row_8_col_3>", "r8c4": "<row_8_col_4>", "r8c5": "<row_8_col_5>", "r8c6": "<row_8_col_6>", "r8c7": "<row_8_col_7>", "r8c8": "<row_8_col_8>"}) |
| vlm_load_backbone_weights: bool = True |
| vlm_checkpoint_path: str = 'checkpoints' |
| hf_repo_name: str = 'qvac/VisionPsy-Nano-460M' |
|
|
| compile_inference: bool = True |
| compile_inference_mode: str = 'reduce-overhead' |
| cuda_graphs_cache_quantum: int = 128 |
|
|
| eos_check_interval: int = 16 |
|
|
