agent.voice / models /breeze_config.py
multimodalart's picture
multimodalart HF Staff
Upload folder using huggingface_hub
878a616 verified
Raw History Blame Contribute Delete
2.04 kB
import os
from transformers import AutoConfig, AutoModel
from transformers.models.t5gemma.configuration_t5gemma import T5GemmaModuleConfig
from transformers.models.t5gemma.modeling_t5gemma import T5GemmaEncoder
from .breeze_base_config import BreezeConfig as BreezeConfig_transformers
from .breeze_base_config import BreezeDepthDecoderConfig
from .t5gemma2_compat import T5Gemma2TextConfig, T5Gemma2TextEncoder
class T5GemmaEncoderWrapper(T5GemmaEncoder):
config_class = T5GemmaModuleConfig
# register T5GemmaModuleConfig
AutoConfig.register("t5_gemma_module", T5GemmaModuleConfig)
AutoModel.register(T5GemmaModuleConfig, T5GemmaEncoderWrapper)
AutoConfig.register("t5gemma2_text", T5Gemma2TextConfig)
AutoModel.register(T5Gemma2TextConfig, T5Gemma2TextEncoder)
# Update BreezeConfig to handle text_encoder_config properly
class BreezeConfig(BreezeConfig_transformers):
model_type = "breeze"
sub_configs = {
"codec_config": AutoConfig,
"depth_decoder_config": BreezeDepthDecoderConfig,
}
def __init__(self, **kwargs):
text_encoder_config = kwargs.get("text_encoder_config", None)
super().__init__(**kwargs)
if text_encoder_config is None:
self.text_encoder_config = None
elif isinstance(text_encoder_config, str):
assert os.path.isdir(text_encoder_config), (
f"text_encoder_config as str must be a valid directory path: '{text_encoder_config}'"
)
self.text_encoder_config = AutoConfig.from_pretrained(text_encoder_config)
elif isinstance(text_encoder_config, dict):
self.text_encoder_config = AutoConfig.for_model(**text_encoder_config)
else:
raise ValueError(
f"text_encoder_config must be a str (path), dict, or AutoConfig instance. but got {type(text_encoder_config)}"
)
AutoConfig.register(
"breeze_depth_decoder_model", BreezeDepthDecoderConfig, exist_ok=True
)
AutoConfig.register("breeze", BreezeConfig, exist_ok=True)