File size: 12,933 Bytes
8b576d1 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 |
# coding=utf-8
# Copyright 2026 OpenMOSS and the HuggingFace Inc. 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.
"""MossAudioTokenizer model configuration"""
from typing import Any
from transformers.configuration_utils import PreTrainedConfig
from transformers.utils import logging
logger = logging.get_logger(__name__)
class MossAudioTokenizerConfig(PreTrainedConfig):
r"""
This is the configuration class to store the configuration of a [`MossAudioTokenizerModel`]. It is used to instantiate a
MossAudioTokenizer model according to the specified arguments, defining the model architecture.
Instantiating a configuration with the defaults will yield a similar configuration to that of the
[VoiceAgentGroup/moss_audio_tokenizer](https://huggingface.co/VoiceAgentGroup/moss_audio_tokenizer) architecture.
Configuration objects inherit from [`PreTrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PreTrainedConfig`] for more information.
Args:
sampling_rate (`int`, *optional*, defaults to 24000):
The sampling rate at which the audio waveform should be digitalized expressed in hertz (Hz).
downsample_rate (`int`, *optional*, defaults to 1920):
Total downsampling rate from waveform to tokens.
causal_transformer_context_duration (`float`, *optional*, defaults to 10.0):
Context duration in seconds for causal transformer.
encoder_kwargs (`list[dict]`, *optional*):
List of encoder module configurations. Each dict specifies a module type and its parameters.
decoder_kwargs (`list[dict]`, *optional*):
List of decoder module configurations in execution order.
quantizer_type (`str`, *optional*, defaults to `"rvq"`):
Quantizer type. Options include `"rvq"`, `"spec_rvq"`, `"rlfq"`, `"random_prefix_rlfq"`.
quantizer_kwargs (`dict`, *optional*):
Configuration for the quantizer including `input_dim`, `rvq_dim`, `output_dim`, `num_quantizers`,
`codebook_size`, and `codebook_dim`.
Example:
```python
>>> from transformers import MossAudioTokenizerModel, MossAudioTokenizerConfig
>>> # Initializing a MossAudioTokenizer style configuration
>>> configuration = MossAudioTokenizerConfig()
>>> # Initializing a model (with random weights) from the configuration
>>> model = MossAudioTokenizerModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
```
"""
model_type = "moss-audio-tokenizer"
# Backward-compatible alias used by some checkpoints.
attribute_map = {"sample_rate": "sampling_rate"}
sampling_rate: int
downsample_rate: int
causal_transformer_context_duration: float
encoder_kwargs: list[dict[str, Any]]
decoder_kwargs: list[dict[str, Any]]
quantizer_type: str
quantizer_kwargs: dict[str, Any]
def __init__(
self,
version: str | None = None,
sampling_rate: int = 24000,
downsample_rate: int = 1920,
causal_transformer_context_duration: float = 10.0,
encoder_kwargs: list[dict[str, Any]] | None = None,
decoder_kwargs: list[dict[str, Any]] | None = None,
quantizer_type: str = "rlfq",
quantizer_kwargs: dict[str, Any] | None = None,
**kwargs,
):
# Some checkpoints might include an incorrect/legacy `model_type` (e.g. "speech_tokenizer").
# We drop it to avoid overriding the class-level `model_type`.
kwargs.pop("model_type", None)
# `version` is accepted for compatibility but not used in modeling.
self.version = version
self.sampling_rate = sampling_rate
self.downsample_rate = downsample_rate
self.causal_transformer_context_duration = causal_transformer_context_duration
# Default encoder configuration
if encoder_kwargs is None:
encoder_kwargs = [
{
"module_type": "PatchedPretransform",
"patch_size": 240,
},
{
"module_type": "Transformer",
"input_dimension": 240,
"output_dimension": 384,
"d_model": 768,
"num_heads": 12,
"num_layers": 12,
"dim_feedforward": 3072,
"causal": True,
"norm": "layer_norm",
"positional_embedding": "rope",
"max_period": 10000,
"gating": "none",
"layer_scale": 0.01,
"conv_layout": True,
},
{
"module_type": "PatchedPretransform",
"patch_size": 2,
},
{
"module_type": "Transformer",
"input_dimension": 768,
"output_dimension": 384,
"d_model": 768,
"num_heads": 12,
"num_layers": 12,
"dim_feedforward": 3072,
"causal": True,
"norm": "layer_norm",
"positional_embedding": "rope",
"max_period": 10000,
"gating": "none",
"layer_scale": 0.01,
"conv_layout": True,
},
{
"module_type": "PatchedPretransform",
"patch_size": 2,
},
{
"module_type": "Transformer",
"input_dimension": 768,
"output_dimension": 640,
"d_model": 768,
"num_heads": 12,
"num_layers": 12,
"dim_feedforward": 3072,
"causal": True,
"norm": "layer_norm",
"positional_embedding": "rope",
"max_period": 10000,
"gating": "none",
"layer_scale": 0.01,
"conv_layout": True,
},
{
"module_type": "PatchedPretransform",
"patch_size": 2,
},
{
"module_type": "Transformer",
"input_dimension": 1280,
"output_dimension": 768,
"d_model": 1280,
"num_heads": 20,
"num_layers": 32,
"dim_feedforward": 5120,
"causal": True,
"norm": "layer_norm",
"positional_embedding": "rope",
"max_period": 10000,
"gating": "none",
"layer_scale": 0.01,
"conv_layout": True,
},
]
self.encoder_kwargs = encoder_kwargs
# Default decoder configuration (execution order)
if decoder_kwargs is None:
decoder_kwargs = [
{
"module_type": "Transformer",
"input_dimension": 768,
"output_dimension": 1280,
"d_model": 1280,
"num_heads": 20,
"num_layers": 32,
"dim_feedforward": 5120,
"causal": True,
"norm": "layer_norm",
"positional_embedding": "rope",
"max_period": 10000,
"gating": "none",
"layer_scale": 0.01,
"conv_layout": True,
},
{
"module_type": "PatchedPretransform",
"patch_size": 2,
},
{
"module_type": "Transformer",
"input_dimension": 640,
"output_dimension": 768,
"d_model": 768,
"num_heads": 12,
"num_layers": 12,
"dim_feedforward": 3072,
"causal": True,
"norm": "layer_norm",
"positional_embedding": "rope",
"max_period": 10000,
"gating": "none",
"layer_scale": 0.01,
"conv_layout": True,
},
{
"module_type": "PatchedPretransform",
"patch_size": 2,
},
{
"module_type": "Transformer",
"input_dimension": 384,
"output_dimension": 768,
"d_model": 768,
"num_heads": 12,
"num_layers": 12,
"dim_feedforward": 3072,
"causal": True,
"norm": "layer_norm",
"positional_embedding": "rope",
"max_period": 10000,
"gating": "none",
"layer_scale": 0.01,
"conv_layout": True,
},
{
"module_type": "PatchedPretransform",
"patch_size": 2,
},
{
"module_type": "Transformer",
"input_dimension": 384,
"output_dimension": 768,
"d_model": 768,
"num_heads": 12,
"num_layers": 12,
"dim_feedforward": 3072,
"causal": True,
"norm": "layer_norm",
"positional_embedding": "rope",
"max_period": 10000,
"gating": "none",
"layer_scale": 0.01,
"conv_layout": True,
},
{
"module_type": "PatchedPretransform",
"patch_size": 2,
},
{
"module_type": "Transformer",
"input_dimension": 384,
"output_dimension": 240,
"d_model": 768,
"num_heads": 12,
"num_layers": 12,
"dim_feedforward": 3072,
"causal": True,
"norm": "layer_norm",
"positional_embedding": "rope",
"max_period": 10000,
"gating": "none",
"layer_scale": 0.01,
"conv_layout": True,
},
{
"module_type": "PatchedPretransform",
"patch_size": 240,
},
]
self.decoder_kwargs = decoder_kwargs
# Default quantizer configuration
if quantizer_kwargs is None:
quantizer_kwargs = {
"input_dim": 768,
"rvq_dim": 512,
"output_dim": 768,
"num_quantizers": 32,
"codebook_size": 1024,
"codebook_dim": 8,
"quantizer_type": "rlfq",
}
# Handle quantizer_type from kwargs or config
kw_qtype = quantizer_kwargs.get("quantizer_type", None)
if kw_qtype is not None:
self.quantizer_type = kw_qtype
else:
self.quantizer_type = quantizer_type
quantizer_kwargs["quantizer_type"] = quantizer_type
self.quantizer_kwargs = quantizer_kwargs
super().__init__(**kwargs)
@property
def num_quantizers(self) -> int:
"""Return the number of quantizers from quantizer_kwargs."""
return self.quantizer_kwargs.get("num_quantizers", 32)
@property
def codebook_size(self) -> int:
"""Return the codebook size from quantizer_kwargs."""
return self.quantizer_kwargs.get("codebook_size", 4096)
@property
def frame_rate(self) -> float:
"""Return the frame rate (tokens per second)."""
return self.sampling_rate / self.downsample_rate
__all__ = ["MossAudioTokenizerConfig"]
|