Upload configuration_dots.py
Browse files- configuration_dots.py +76 -0
configuration_dots.py
ADDED
|
@@ -0,0 +1,76 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from typing import Any, Optional
|
| 2 |
+
from transformers.configuration_utils import PretrainedConfig
|
| 3 |
+
from transformers.models.qwen2 import Qwen2Config
|
| 4 |
+
from transformers import Qwen2_5_VLProcessor, AutoProcessor
|
| 5 |
+
from transformers.models.auto.configuration_auto import CONFIG_MAPPING
|
| 6 |
+
|
| 7 |
+
|
| 8 |
+
class DotsVisionConfig(PretrainedConfig):
|
| 9 |
+
model_type: str = "dots_vit"
|
| 10 |
+
|
| 11 |
+
def __init__(
|
| 12 |
+
self,
|
| 13 |
+
embed_dim: int = 1536, # vision encoder embed size
|
| 14 |
+
hidden_size: int = 1536, # after merger hidden size
|
| 15 |
+
intermediate_size: int = 4224,
|
| 16 |
+
num_hidden_layers: int = 42,
|
| 17 |
+
num_attention_heads: int = 12,
|
| 18 |
+
num_channels: int = 3,
|
| 19 |
+
patch_size: int = 14,
|
| 20 |
+
spatial_merge_size: int = 2,
|
| 21 |
+
temporal_patch_size: int = 1,
|
| 22 |
+
rms_norm_eps: float = 1e-5,
|
| 23 |
+
use_bias: bool = False,
|
| 24 |
+
attn_implementation="flash_attention_2", # "eager","sdpa","flash_attention_2"
|
| 25 |
+
initializer_range=0.02,
|
| 26 |
+
init_merger_std=0.02,
|
| 27 |
+
is_causal=False, # ve causal forward
|
| 28 |
+
post_norm=True,
|
| 29 |
+
gradient_checkpointing=False,
|
| 30 |
+
**kwargs: Any,
|
| 31 |
+
):
|
| 32 |
+
super().__init__(**kwargs)
|
| 33 |
+
self.embed_dim = embed_dim
|
| 34 |
+
self.hidden_size = hidden_size
|
| 35 |
+
self.intermediate_size = intermediate_size
|
| 36 |
+
self.num_hidden_layers = num_hidden_layers
|
| 37 |
+
self.num_attention_heads = num_attention_heads
|
| 38 |
+
self.num_channels = num_channels
|
| 39 |
+
self.patch_size = patch_size
|
| 40 |
+
self.spatial_merge_size = spatial_merge_size
|
| 41 |
+
self.temporal_patch_size = temporal_patch_size
|
| 42 |
+
self.rms_norm_eps = rms_norm_eps
|
| 43 |
+
self.use_bias = use_bias
|
| 44 |
+
self.attn_implementation = attn_implementation
|
| 45 |
+
self.initializer_range = initializer_range
|
| 46 |
+
self.init_merger_std = init_merger_std
|
| 47 |
+
self.is_causal = is_causal
|
| 48 |
+
self.post_norm = post_norm
|
| 49 |
+
self.gradient_checkpointing = gradient_checkpointing
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
|
| 53 |
+
class DotsOCRConfig(Qwen2Config):
|
| 54 |
+
model_type = "dots_ocr"
|
| 55 |
+
def __init__(self,
|
| 56 |
+
image_token_id = 151665,
|
| 57 |
+
video_token_id = 151656,
|
| 58 |
+
vision_config: Optional[dict] = None, *args, **kwargs):
|
| 59 |
+
super().__init__(*args, **kwargs)
|
| 60 |
+
self.image_token_id = image_token_id
|
| 61 |
+
self.video_token_id = video_token_id
|
| 62 |
+
self.vision_config = DotsVisionConfig(**(vision_config or {}))
|
| 63 |
+
|
| 64 |
+
def save_pretrained(self, save_directory, **kwargs):
|
| 65 |
+
self._auto_class = None
|
| 66 |
+
super().save_pretrained(save_directory, **kwargs)
|
| 67 |
+
|
| 68 |
+
|
| 69 |
+
class DotsVLProcessor(Qwen2_5_VLProcessor):
|
| 70 |
+
def __init__(self, image_processor=None, tokenizer=None, chat_template=None, **kwargs):
|
| 71 |
+
super().__init__(image_processor, tokenizer, chat_template=chat_template)
|
| 72 |
+
self.image_token = "<|imgpad|>" if not hasattr(tokenizer, "image_token") else tokenizer.image_token
|
| 73 |
+
|
| 74 |
+
|
| 75 |
+
AutoProcessor.register("dots_ocr", DotsVLProcessor)
|
| 76 |
+
CONFIG_MAPPING.register("dots_ocr", DotsOCRConfig)
|