File size: 5,401 Bytes
9d49701 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 | # Copyright 2026 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import math
from huggingface_hub.dataclasses import strict
from transformers.configuration_utils import PreTrainedConfig
from transformers.models.auto import CONFIG_MAPPING
from transformers.utils import auto_docstring
from .configuration_lfm3 import Lfm3Config
def _config_class(model_type):
# the LFM3 configs ship with the checkpoint instead of the transformers registry
remote = {"lfm3": Lfm3Config}
return remote.get(model_type) or CONFIG_MAPPING[model_type]
@auto_docstring(checkpoint="LiquidAI/LFM3-Audio-2B-Preview")
@strict
class LiquidCodecConfig(PreTrainedConfig):
r"""
Configuration of the [`LiquidCodecModel`] bundled with LFM3-Audio-2B and LFM3-Omni-2B in their
`audio_tokenizer/` directory.
encoder_dim (`int`, *optional*, defaults to 512):
The width of the encoder and decoder LFM3 models.
encoder_conv_dim (`int`, *optional*, defaults to 256):
The number of channels of the STFT front end.
stft_n_fft (`int`, *optional*, defaults to 1280):
The FFT size of the analysis and synthesis transforms.
stft_hop_length (`int`, *optional*, defaults to 320):
The hop of the analysis and synthesis transforms.
stft_win_length (`int`, *optional*, defaults to 1280):
The window length of the analysis and synthesis transforms.
stft_frame_stack (`int`, *optional*, defaults to 6):
The number of STFT frames per codec frame.
quantizer_dim (`int`, *optional*, defaults to 256):
The width of the latent handed to the quantizer.
fsq_levels (`list[int]`, *optional*, defaults to `[8, 7, 6, 6]`):
The number of levels of each finite-scalar-quantization dimension; their product is the codebook size.
n_q (`int`, *optional*, defaults to 12):
The number of codebooks.
decoder_window (`int`, *optional*, defaults to 30):
The sliding attention window of the decoder LFM3 model, in 75 Hz frames.
sampling_rate (`int`, *optional*, defaults to 24000):
The sampling rate of the audio.
initializer_range (`float`, *optional*, defaults to 0.02):
The standard deviation of the truncated normal initializer of the weights.
"""
model_type = "liquid_codec"
encoder_dim: int = 512
encoder_conv_dim: int = 256
stft_n_fft: int = 1280
stft_hop_length: int = 320
stft_win_length: int = 1280
stft_frame_stack: int = 6
quantizer_dim: int = 256
fsq_levels: list[int] | None = None
n_q: int = 12
decoder_window: int = 30
sampling_rate: int = 24_000
initializer_range: float = 0.02
def __post_init__(self, **kwargs):
# the released configs nest the fields under `model` and dump the two LFM3 sub-configs
dropped = ("encoder_lfm_config", "decoder_lfm_config", "n_output_samples")
for key, value in kwargs.pop("model", {}).items():
if key in dropped:
continue
if key not in self.__dataclass_fields__:
raise ValueError(f"Unknown Liquid codec setting `{key}`")
setattr(self, key, value)
for key in dropped:
kwargs.pop(key, None)
if self.fsq_levels is None:
self.fsq_levels = [8, 7, 6, 6]
self.fsq_levels = list(self.fsq_levels)
super().__post_init__(**kwargs)
@property
def codebook_size(self) -> int:
return math.prod(self.fsq_levels)
@property
def frame_size(self) -> int:
return self.stft_hop_length * self.stft_frame_stack
def lfm_config(self, sliding_window: int | None = None) -> PreTrainedConfig:
attention_type = "full_attention" if sliding_window is None else "sliding_attention"
return _config_class("lfm3")(
# the backbones are not sub-configs, so the requested attention implementation is handed over here
attn_implementation=self._attn_implementation,
vocab_size=1,
pad_token_id=None,
bos_token_id=None,
eos_token_id=None,
hidden_size=self.encoder_dim,
intermediate_size=2304,
num_hidden_layers=8,
num_attention_heads=8,
num_key_value_heads=2,
head_dim=64,
max_position_embeddings=131_072,
norm_eps=1e-5,
conv_L_cache=7,
conv_bias=False,
use_sandwich_norm=True,
layer_types=[attention_type if i in (2, 4, 6) else "conv" for i in range(8)],
sliding_window=sliding_window,
rope_parameters={
attention_type: {"rope_type": "default", "rope_theta": 1_000_000.0, "partial_rotary_factor": 1.0}
},
use_cache=False,
tie_word_embeddings=False,
)
__all__ = ["LiquidCodecConfig"]
|