Mirror configuration_moss_audio_tokenizer.py
Browse files
configuration_moss_audio_tokenizer.py
ADDED
|
@@ -0,0 +1,475 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2026 OpenMOSS and the HuggingFace Inc. team. All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
+
# See the License for the specific language governing permissions and
|
| 14 |
+
# limitations under the License.
|
| 15 |
+
"""MossAudioTokenizer model configuration."""
|
| 16 |
+
|
| 17 |
+
from typing import Any
|
| 18 |
+
|
| 19 |
+
try:
|
| 20 |
+
from transformers.configuration_utils import PreTrainedConfig
|
| 21 |
+
except ImportError:
|
| 22 |
+
from transformers.configuration_utils import PretrainedConfig as PreTrainedConfig
|
| 23 |
+
from transformers.utils import logging
|
| 24 |
+
|
| 25 |
+
|
| 26 |
+
logger = logging.get_logger(__name__)
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
class MossAudioTokenizerConfig(PreTrainedConfig):
|
| 30 |
+
r"""
|
| 31 |
+
This is the configuration class to store the configuration of a [`MossAudioTokenizerModel`]. It is used to instantiate a
|
| 32 |
+
MossAudioTokenizer model according to the specified arguments, defining the model architecture.
|
| 33 |
+
|
| 34 |
+
Instantiating a configuration with the defaults will yield a similar configuration to that of the
|
| 35 |
+
[OpenMOSS-Team/MOSS-Audio-Tokenizer-v2](https://huggingface.co/OpenMOSS-Team/MOSS-Audio-Tokenizer-v2) architecture.
|
| 36 |
+
|
| 37 |
+
Configuration objects inherit from [`PreTrainedConfig`] and can be used to control the model outputs. Read the
|
| 38 |
+
documentation from [`PreTrainedConfig`] for more information.
|
| 39 |
+
|
| 40 |
+
Args:
|
| 41 |
+
sampling_rate (`int`, *optional*, defaults to 48000):
|
| 42 |
+
The sampling rate at which the audio waveform should be digitalized expressed in hertz (Hz).
|
| 43 |
+
downsample_rate (`int`, *optional*, defaults to 3840):
|
| 44 |
+
Total downsampling rate from waveform to tokens.
|
| 45 |
+
causal_transformer_context_duration (`float`, *optional*, defaults to 10.0):
|
| 46 |
+
Legacy global fallback context duration in seconds for causal transformer. If an individual transformer
|
| 47 |
+
entry in `encoder_kwargs` or `decoder_kwargs` provides `context_duration`, that per-module value takes
|
| 48 |
+
precedence.
|
| 49 |
+
encoder_kwargs (`list[dict]`, *optional*):
|
| 50 |
+
List of encoder module configurations. Each dict specifies a module type and its parameters.
|
| 51 |
+
decoder_kwargs (`list[dict]`, *optional*):
|
| 52 |
+
List of decoder module configurations in execution order.
|
| 53 |
+
number_channels (`int`, *optional*, defaults to 2):
|
| 54 |
+
Number of audio channels exposed by the public waveform interface.
|
| 55 |
+
enable_channel_interleave (`bool`, *optional*, defaults to `True`):
|
| 56 |
+
Whether to flatten multi-channel waveforms into a single internal stream before codec inference.
|
| 57 |
+
attention_implementation (`str`, *optional*, defaults to `"sdpa"`):
|
| 58 |
+
Attention implementation to prefer for transformer layers. Supported values are `"sdpa"` and
|
| 59 |
+
`"flash_attention_2"`.
|
| 60 |
+
compute_dtype (`str`, *optional*, defaults to `"fp32"`):
|
| 61 |
+
Inference compute dtype for non-quantizer modules. Supported values are `"fp32"`, `"bf16"`.
|
| 62 |
+
codec_weight_dtype (`str`, *optional*, defaults to `"fp32"`):
|
| 63 |
+
Parameter dtype for encoder and decoder modules. The quantizer remains fp32 because it explicitly disables
|
| 64 |
+
autocast and performs numerically sensitive codebook operations in fp32.
|
| 65 |
+
quantizer_type (`str`, *optional*, defaults to `"rlfq"`):
|
| 66 |
+
Quantizer type. Options include `"rvq"`, `"spec_rvq"`, `"rlfq"`, `"random_prefix_rlfq"`.
|
| 67 |
+
quantizer_kwargs (`dict`, *optional*):
|
| 68 |
+
Configuration for the quantizer including `input_dim`, `rvq_dim`, `output_dim`, `num_quantizers`,
|
| 69 |
+
`codebook_size`, and `codebook_dim`.
|
| 70 |
+
|
| 71 |
+
Example:
|
| 72 |
+
|
| 73 |
+
```python
|
| 74 |
+
>>> from transformers import MossAudioTokenizerModel, MossAudioTokenizerConfig
|
| 75 |
+
|
| 76 |
+
>>> # Initializing a MossAudioTokenizer style configuration
|
| 77 |
+
>>> configuration = MossAudioTokenizerConfig()
|
| 78 |
+
|
| 79 |
+
>>> # Initializing a model (with random weights) from the configuration
|
| 80 |
+
>>> model = MossAudioTokenizerModel(configuration)
|
| 81 |
+
|
| 82 |
+
>>> # Accessing the model configuration
|
| 83 |
+
>>> configuration = model.config
|
| 84 |
+
```
|
| 85 |
+
"""
|
| 86 |
+
|
| 87 |
+
model_type = "moss-audio-tokenizer"
|
| 88 |
+
|
| 89 |
+
# Backward-compatible alias used by some checkpoints.
|
| 90 |
+
attribute_map = {"sample_rate": "sampling_rate"}
|
| 91 |
+
|
| 92 |
+
sampling_rate: int
|
| 93 |
+
downsample_rate: int
|
| 94 |
+
causal_transformer_context_duration: float
|
| 95 |
+
encoder_kwargs: list[dict[str, Any]]
|
| 96 |
+
decoder_kwargs: list[dict[str, Any]]
|
| 97 |
+
number_channels: int
|
| 98 |
+
enable_channel_interleave: bool
|
| 99 |
+
attention_implementation: str
|
| 100 |
+
compute_dtype: str
|
| 101 |
+
codec_weight_dtype: str
|
| 102 |
+
quantizer_type: str
|
| 103 |
+
quantizer_kwargs: dict[str, Any]
|
| 104 |
+
|
| 105 |
+
def __init__(
|
| 106 |
+
self,
|
| 107 |
+
version: str | None = None,
|
| 108 |
+
sampling_rate: int = 48000,
|
| 109 |
+
downsample_rate: int = 3840,
|
| 110 |
+
causal_transformer_context_duration: float = 10.0,
|
| 111 |
+
encoder_kwargs: list[dict[str, Any]] | None = None,
|
| 112 |
+
decoder_kwargs: list[dict[str, Any]] | None = None,
|
| 113 |
+
number_channels: int = 2,
|
| 114 |
+
enable_channel_interleave: bool = True,
|
| 115 |
+
attention_implementation: str = "sdpa",
|
| 116 |
+
compute_dtype: str = "fp32",
|
| 117 |
+
codec_weight_dtype: str = "fp32",
|
| 118 |
+
quantizer_type: str = "rlfq",
|
| 119 |
+
quantizer_kwargs: dict[str, Any] | None = None,
|
| 120 |
+
**kwargs,
|
| 121 |
+
):
|
| 122 |
+
# Some checkpoints might include an incorrect/legacy `model_type` (e.g. "speech_tokenizer").
|
| 123 |
+
# We drop it to avoid overriding the class-level `model_type`.
|
| 124 |
+
kwargs.pop("model_type", None)
|
| 125 |
+
if "channels_numbers" in kwargs:
|
| 126 |
+
number_channels = kwargs.pop("channels_numbers")
|
| 127 |
+
if "enable_channel_interleave" in kwargs:
|
| 128 |
+
enable_channel_interleave = kwargs.pop("enable_channel_interleave")
|
| 129 |
+
if "attention_backend" in kwargs and attention_implementation == "sdpa":
|
| 130 |
+
attention_implementation = kwargs.pop("attention_backend")
|
| 131 |
+
if "codec_compute_dtype" in kwargs and compute_dtype == "fp32":
|
| 132 |
+
compute_dtype = kwargs.pop("codec_compute_dtype")
|
| 133 |
+
if "codec_load_dtype" in kwargs and codec_weight_dtype == "fp32":
|
| 134 |
+
codec_weight_dtype = kwargs.pop("codec_load_dtype")
|
| 135 |
+
reversed_decoder_kwargs = kwargs.pop("reversed_decoder_kwargs", None)
|
| 136 |
+
|
| 137 |
+
# `version` is accepted for compatibility but not used in modeling.
|
| 138 |
+
self.version = version
|
| 139 |
+
self.sampling_rate = sampling_rate
|
| 140 |
+
self.downsample_rate = downsample_rate
|
| 141 |
+
self.causal_transformer_context_duration = causal_transformer_context_duration
|
| 142 |
+
self.number_channels = number_channels
|
| 143 |
+
self.enable_channel_interleave = enable_channel_interleave
|
| 144 |
+
self.attention_implementation = attention_implementation
|
| 145 |
+
self.compute_dtype = compute_dtype
|
| 146 |
+
self.codec_weight_dtype = codec_weight_dtype
|
| 147 |
+
# Default encoder configuration
|
| 148 |
+
if encoder_kwargs is None:
|
| 149 |
+
encoder_kwargs = [
|
| 150 |
+
{
|
| 151 |
+
"module_type": "PatchedPretransform",
|
| 152 |
+
"patch_size": 240,
|
| 153 |
+
},
|
| 154 |
+
{
|
| 155 |
+
"module_type": "Transformer",
|
| 156 |
+
"input_dimension": 240,
|
| 157 |
+
"output_dimension": 384,
|
| 158 |
+
"d_model": 768,
|
| 159 |
+
"num_heads": 12,
|
| 160 |
+
"num_layers": 12,
|
| 161 |
+
"dim_feedforward": 3072,
|
| 162 |
+
"causal": True,
|
| 163 |
+
"norm": "layer_norm",
|
| 164 |
+
"positional_embedding": "rope",
|
| 165 |
+
"max_period": 10000,
|
| 166 |
+
"gating": "none",
|
| 167 |
+
"layer_scale": 0.01,
|
| 168 |
+
"conv_layout": True,
|
| 169 |
+
"context_duration": 1.0,
|
| 170 |
+
},
|
| 171 |
+
{
|
| 172 |
+
"module_type": "PatchedPretransform",
|
| 173 |
+
"patch_size": 2,
|
| 174 |
+
},
|
| 175 |
+
{
|
| 176 |
+
"module_type": "Transformer",
|
| 177 |
+
"input_dimension": 768,
|
| 178 |
+
"output_dimension": 384,
|
| 179 |
+
"d_model": 768,
|
| 180 |
+
"num_heads": 12,
|
| 181 |
+
"num_layers": 12,
|
| 182 |
+
"dim_feedforward": 3072,
|
| 183 |
+
"causal": True,
|
| 184 |
+
"norm": "layer_norm",
|
| 185 |
+
"positional_embedding": "rope",
|
| 186 |
+
"max_period": 10000,
|
| 187 |
+
"gating": "none",
|
| 188 |
+
"layer_scale": 0.01,
|
| 189 |
+
"conv_layout": True,
|
| 190 |
+
"context_duration": 2.0,
|
| 191 |
+
},
|
| 192 |
+
{
|
| 193 |
+
"module_type": "PatchedPretransform",
|
| 194 |
+
"patch_size": 2,
|
| 195 |
+
},
|
| 196 |
+
{
|
| 197 |
+
"module_type": "Transformer",
|
| 198 |
+
"input_dimension": 768,
|
| 199 |
+
"output_dimension": 384,
|
| 200 |
+
"d_model": 768,
|
| 201 |
+
"num_heads": 12,
|
| 202 |
+
"num_layers": 12,
|
| 203 |
+
"dim_feedforward": 3072,
|
| 204 |
+
"causal": True,
|
| 205 |
+
"norm": "layer_norm",
|
| 206 |
+
"positional_embedding": "rope",
|
| 207 |
+
"max_period": 10000,
|
| 208 |
+
"gating": "none",
|
| 209 |
+
"layer_scale": 0.01,
|
| 210 |
+
"conv_layout": True,
|
| 211 |
+
"context_duration": 4.0,
|
| 212 |
+
},
|
| 213 |
+
{
|
| 214 |
+
"module_type": "PatchedPretransform",
|
| 215 |
+
"patch_size": 2,
|
| 216 |
+
},
|
| 217 |
+
{
|
| 218 |
+
"module_type": "Transformer",
|
| 219 |
+
"input_dimension": 768,
|
| 220 |
+
"output_dimension": 384,
|
| 221 |
+
"d_model": 768,
|
| 222 |
+
"num_heads": 12,
|
| 223 |
+
"num_layers": 12,
|
| 224 |
+
"dim_feedforward": 3072,
|
| 225 |
+
"causal": True,
|
| 226 |
+
"norm": "layer_norm",
|
| 227 |
+
"positional_embedding": "rope",
|
| 228 |
+
"max_period": 10000,
|
| 229 |
+
"gating": "none",
|
| 230 |
+
"layer_scale": 0.01,
|
| 231 |
+
"conv_layout": True,
|
| 232 |
+
"context_duration": 8.0,
|
| 233 |
+
},
|
| 234 |
+
{
|
| 235 |
+
"module_type": "PatchedPretransform",
|
| 236 |
+
"patch_size": 2,
|
| 237 |
+
},
|
| 238 |
+
{
|
| 239 |
+
"module_type": "Transformer",
|
| 240 |
+
"input_dimension": 768,
|
| 241 |
+
"output_dimension": 640,
|
| 242 |
+
"d_model": 768,
|
| 243 |
+
"num_heads": 12,
|
| 244 |
+
"num_layers": 12,
|
| 245 |
+
"dim_feedforward": 3072,
|
| 246 |
+
"causal": True,
|
| 247 |
+
"norm": "layer_norm",
|
| 248 |
+
"positional_embedding": "rope",
|
| 249 |
+
"max_period": 10000,
|
| 250 |
+
"gating": "none",
|
| 251 |
+
"layer_scale": 0.01,
|
| 252 |
+
"conv_layout": True,
|
| 253 |
+
"context_duration": 10.0,
|
| 254 |
+
},
|
| 255 |
+
{
|
| 256 |
+
"module_type": "PatchedPretransform",
|
| 257 |
+
"patch_size": 2,
|
| 258 |
+
},
|
| 259 |
+
{
|
| 260 |
+
"module_type": "Transformer",
|
| 261 |
+
"input_dimension": 1280,
|
| 262 |
+
"output_dimension": 768,
|
| 263 |
+
"d_model": 1280,
|
| 264 |
+
"num_heads": 20,
|
| 265 |
+
"num_layers": 32,
|
| 266 |
+
"dim_feedforward": 5120,
|
| 267 |
+
"causal": True,
|
| 268 |
+
"norm": "layer_norm",
|
| 269 |
+
"positional_embedding": "rope",
|
| 270 |
+
"max_period": 10000,
|
| 271 |
+
"gating": "none",
|
| 272 |
+
"layer_scale": 0.01,
|
| 273 |
+
"conv_layout": True,
|
| 274 |
+
"context_duration": 10.0,
|
| 275 |
+
},
|
| 276 |
+
]
|
| 277 |
+
else:
|
| 278 |
+
encoder_kwargs = [dict(module_kwargs) for module_kwargs in encoder_kwargs]
|
| 279 |
+
for module_kwargs in encoder_kwargs:
|
| 280 |
+
if module_kwargs.get("module_type") == "Transformer":
|
| 281 |
+
module_kwargs.setdefault("context_duration", causal_transformer_context_duration)
|
| 282 |
+
self.encoder_kwargs = encoder_kwargs
|
| 283 |
+
|
| 284 |
+
# Default decoder configuration (execution order)
|
| 285 |
+
if decoder_kwargs is None and reversed_decoder_kwargs is not None:
|
| 286 |
+
reversed_decoder_kwargs = [dict(module_kwargs) for module_kwargs in reversed_decoder_kwargs]
|
| 287 |
+
decoder_kwargs = []
|
| 288 |
+
for module_kwargs in reversed_decoder_kwargs[::-1]:
|
| 289 |
+
if module_kwargs.get("module_type") != "Transformer":
|
| 290 |
+
decoder_kwargs.append(module_kwargs)
|
| 291 |
+
continue
|
| 292 |
+
module_kwargs = dict(module_kwargs)
|
| 293 |
+
module_kwargs["input_dimension"], module_kwargs["output_dimension"] = (
|
| 294 |
+
module_kwargs["output_dimension"],
|
| 295 |
+
module_kwargs["input_dimension"],
|
| 296 |
+
)
|
| 297 |
+
decoder_kwargs.append(module_kwargs)
|
| 298 |
+
|
| 299 |
+
if decoder_kwargs is None:
|
| 300 |
+
decoder_kwargs = [
|
| 301 |
+
{
|
| 302 |
+
"module_type": "Transformer",
|
| 303 |
+
"input_dimension": 768,
|
| 304 |
+
"output_dimension": 1280,
|
| 305 |
+
"d_model": 1280,
|
| 306 |
+
"num_heads": 20,
|
| 307 |
+
"num_layers": 32,
|
| 308 |
+
"dim_feedforward": 5120,
|
| 309 |
+
"causal": True,
|
| 310 |
+
"norm": "layer_norm",
|
| 311 |
+
"positional_embedding": "rope",
|
| 312 |
+
"max_period": 10000,
|
| 313 |
+
"gating": "none",
|
| 314 |
+
"layer_scale": 0.01,
|
| 315 |
+
"conv_layout": True,
|
| 316 |
+
"context_duration": 10.0,
|
| 317 |
+
},
|
| 318 |
+
{
|
| 319 |
+
"module_type": "PatchedPretransform",
|
| 320 |
+
"patch_size": 2,
|
| 321 |
+
},
|
| 322 |
+
{
|
| 323 |
+
"module_type": "Transformer",
|
| 324 |
+
"input_dimension": 640,
|
| 325 |
+
"output_dimension": 768,
|
| 326 |
+
"d_model": 768,
|
| 327 |
+
"num_heads": 12,
|
| 328 |
+
"num_layers": 12,
|
| 329 |
+
"dim_feedforward": 3072,
|
| 330 |
+
"causal": True,
|
| 331 |
+
"norm": "layer_norm",
|
| 332 |
+
"positional_embedding": "rope",
|
| 333 |
+
"max_period": 10000,
|
| 334 |
+
"gating": "none",
|
| 335 |
+
"layer_scale": 0.01,
|
| 336 |
+
"conv_layout": True,
|
| 337 |
+
"context_duration": 10.0,
|
| 338 |
+
},
|
| 339 |
+
{
|
| 340 |
+
"module_type": "PatchedPretransform",
|
| 341 |
+
"patch_size": 2,
|
| 342 |
+
},
|
| 343 |
+
{
|
| 344 |
+
"module_type": "Transformer",
|
| 345 |
+
"input_dimension": 384,
|
| 346 |
+
"output_dimension": 768,
|
| 347 |
+
"d_model": 768,
|
| 348 |
+
"num_heads": 12,
|
| 349 |
+
"num_layers": 12,
|
| 350 |
+
"dim_feedforward": 3072,
|
| 351 |
+
"causal": True,
|
| 352 |
+
"norm": "layer_norm",
|
| 353 |
+
"positional_embedding": "rope",
|
| 354 |
+
"max_period": 10000,
|
| 355 |
+
"gating": "none",
|
| 356 |
+
"layer_scale": 0.01,
|
| 357 |
+
"conv_layout": True,
|
| 358 |
+
"context_duration": 8.0,
|
| 359 |
+
},
|
| 360 |
+
{
|
| 361 |
+
"module_type": "PatchedPretransform",
|
| 362 |
+
"patch_size": 2,
|
| 363 |
+
},
|
| 364 |
+
{
|
| 365 |
+
"module_type": "Transformer",
|
| 366 |
+
"input_dimension": 384,
|
| 367 |
+
"output_dimension": 768,
|
| 368 |
+
"d_model": 768,
|
| 369 |
+
"num_heads": 12,
|
| 370 |
+
"num_layers": 12,
|
| 371 |
+
"dim_feedforward": 3072,
|
| 372 |
+
"causal": True,
|
| 373 |
+
"norm": "layer_norm",
|
| 374 |
+
"positional_embedding": "rope",
|
| 375 |
+
"max_period": 10000,
|
| 376 |
+
"gating": "none",
|
| 377 |
+
"layer_scale": 0.01,
|
| 378 |
+
"conv_layout": True,
|
| 379 |
+
"context_duration": 4.0,
|
| 380 |
+
},
|
| 381 |
+
{
|
| 382 |
+
"module_type": "PatchedPretransform",
|
| 383 |
+
"patch_size": 2,
|
| 384 |
+
},
|
| 385 |
+
{
|
| 386 |
+
"module_type": "Transformer",
|
| 387 |
+
"input_dimension": 384,
|
| 388 |
+
"output_dimension": 768,
|
| 389 |
+
"d_model": 768,
|
| 390 |
+
"num_heads": 12,
|
| 391 |
+
"num_layers": 12,
|
| 392 |
+
"dim_feedforward": 3072,
|
| 393 |
+
"causal": True,
|
| 394 |
+
"norm": "layer_norm",
|
| 395 |
+
"positional_embedding": "rope",
|
| 396 |
+
"max_period": 10000,
|
| 397 |
+
"gating": "none",
|
| 398 |
+
"layer_scale": 0.01,
|
| 399 |
+
"conv_layout": True,
|
| 400 |
+
"context_duration": 2.0,
|
| 401 |
+
},
|
| 402 |
+
{
|
| 403 |
+
"module_type": "PatchedPretransform",
|
| 404 |
+
"patch_size": 2,
|
| 405 |
+
},
|
| 406 |
+
{
|
| 407 |
+
"module_type": "Transformer",
|
| 408 |
+
"input_dimension": 384,
|
| 409 |
+
"output_dimension": 240,
|
| 410 |
+
"d_model": 768,
|
| 411 |
+
"num_heads": 12,
|
| 412 |
+
"num_layers": 12,
|
| 413 |
+
"dim_feedforward": 3072,
|
| 414 |
+
"causal": True,
|
| 415 |
+
"norm": "layer_norm",
|
| 416 |
+
"positional_embedding": "rope",
|
| 417 |
+
"max_period": 10000,
|
| 418 |
+
"gating": "none",
|
| 419 |
+
"layer_scale": 0.01,
|
| 420 |
+
"conv_layout": True,
|
| 421 |
+
"context_duration": 1.0,
|
| 422 |
+
},
|
| 423 |
+
{
|
| 424 |
+
"module_type": "PatchedPretransform",
|
| 425 |
+
"patch_size": 240,
|
| 426 |
+
},
|
| 427 |
+
]
|
| 428 |
+
else:
|
| 429 |
+
decoder_kwargs = [dict(module_kwargs) for module_kwargs in decoder_kwargs]
|
| 430 |
+
for module_kwargs in decoder_kwargs:
|
| 431 |
+
if module_kwargs.get("module_type") == "Transformer":
|
| 432 |
+
module_kwargs.setdefault("context_duration", causal_transformer_context_duration)
|
| 433 |
+
self.decoder_kwargs = decoder_kwargs
|
| 434 |
+
|
| 435 |
+
# Default quantizer configuration
|
| 436 |
+
if quantizer_kwargs is None:
|
| 437 |
+
quantizer_kwargs = {
|
| 438 |
+
"input_dim": 768,
|
| 439 |
+
"rvq_dim": 512,
|
| 440 |
+
"output_dim": 768,
|
| 441 |
+
"num_quantizers": 32,
|
| 442 |
+
"codebook_size": 1024,
|
| 443 |
+
"codebook_dim": 8,
|
| 444 |
+
"quantizer_type": "rlfq",
|
| 445 |
+
}
|
| 446 |
+
|
| 447 |
+
# Handle quantizer_type from kwargs or config
|
| 448 |
+
kw_qtype = quantizer_kwargs.get("quantizer_type", None)
|
| 449 |
+
if kw_qtype is not None:
|
| 450 |
+
self.quantizer_type = kw_qtype
|
| 451 |
+
else:
|
| 452 |
+
self.quantizer_type = quantizer_type
|
| 453 |
+
quantizer_kwargs["quantizer_type"] = quantizer_type
|
| 454 |
+
|
| 455 |
+
self.quantizer_kwargs = quantizer_kwargs
|
| 456 |
+
|
| 457 |
+
super().__init__(**kwargs)
|
| 458 |
+
|
| 459 |
+
@property
|
| 460 |
+
def num_quantizers(self) -> int:
|
| 461 |
+
"""Return the number of quantizers from quantizer_kwargs."""
|
| 462 |
+
return self.quantizer_kwargs.get("num_quantizers", 32)
|
| 463 |
+
|
| 464 |
+
@property
|
| 465 |
+
def codebook_size(self) -> int:
|
| 466 |
+
"""Return the codebook size from quantizer_kwargs."""
|
| 467 |
+
return self.quantizer_kwargs.get("codebook_size", 4096)
|
| 468 |
+
|
| 469 |
+
@property
|
| 470 |
+
def frame_rate(self) -> float:
|
| 471 |
+
"""Return the frame rate (tokens per second)."""
|
| 472 |
+
return self.sampling_rate / self.downsample_rate
|
| 473 |
+
|
| 474 |
+
|
| 475 |
+
__all__ = ["MossAudioTokenizerConfig"]
|