Spaces:
Running
on
Zero
Running
on
Zero
update
Browse files- app.py +2 -2
- dia/__init__.py +6 -0
- dia/audio.py +0 -22
- dia/config.py +96 -113
- dia/layers.py +411 -127
- dia/model.py +559 -168
- dia/state.py +82 -69
- requirements.txt +3 -2
app.py
CHANGED
|
@@ -16,7 +16,7 @@ from dia.model import Dia
|
|
| 16 |
print("Loading Nari model...")
|
| 17 |
try:
|
| 18 |
# Use the function from inference.py
|
| 19 |
-
model = Dia.from_pretrained("nari-labs/Dia-1.6B", compute_dtype="
|
| 20 |
except Exception as e:
|
| 21 |
print(f"Error loading Nari model: {e}")
|
| 22 |
raise
|
|
@@ -375,4 +375,4 @@ if __name__ == "__main__":
|
|
| 375 |
|
| 376 |
# set `GRADIO_SERVER_NAME`, `GRADIO_SERVER_PORT` env vars to override default values
|
| 377 |
# use `GRADIO_SERVER_NAME=0.0.0.0` for Docker
|
| 378 |
-
demo.launch()
|
|
|
|
| 16 |
print("Loading Nari model...")
|
| 17 |
try:
|
| 18 |
# Use the function from inference.py
|
| 19 |
+
model = Dia.from_pretrained("nari-labs/Dia-1.6B-0626", compute_dtype="float16")
|
| 20 |
except Exception as e:
|
| 21 |
print(f"Error loading Nari model: {e}")
|
| 22 |
raise
|
|
|
|
| 375 |
|
| 376 |
# set `GRADIO_SERVER_NAME`, `GRADIO_SERVER_PORT` env vars to override default values
|
| 377 |
# use `GRADIO_SERVER_NAME=0.0.0.0` for Docker
|
| 378 |
+
demo.launch()
|
dia/__init__.py
CHANGED
|
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from .model import Dia
|
| 2 |
+
|
| 3 |
+
|
| 4 |
+
__all__ = [
|
| 5 |
+
"Dia",
|
| 6 |
+
]
|
dia/audio.py
CHANGED
|
@@ -179,25 +179,3 @@ def revert_audio_delay(
|
|
| 179 |
) # Changed np.where to torch.where
|
| 180 |
|
| 181 |
return result_BxTxC
|
| 182 |
-
|
| 183 |
-
|
| 184 |
-
@torch.no_grad()
|
| 185 |
-
@torch.inference_mode()
|
| 186 |
-
def decode(
|
| 187 |
-
model,
|
| 188 |
-
audio_codes,
|
| 189 |
-
):
|
| 190 |
-
"""
|
| 191 |
-
Decodes the given frames into an output audio waveform
|
| 192 |
-
"""
|
| 193 |
-
if len(audio_codes) != 1:
|
| 194 |
-
raise ValueError(f"Expected one frame, got {len(audio_codes)}")
|
| 195 |
-
|
| 196 |
-
try:
|
| 197 |
-
audio_values = model.quantizer.from_codes(audio_codes)
|
| 198 |
-
audio_values = model.decode(audio_values[0])
|
| 199 |
-
|
| 200 |
-
return audio_values
|
| 201 |
-
except Exception as e:
|
| 202 |
-
print(f"Error in decode method: {str(e)}")
|
| 203 |
-
raise
|
|
|
|
| 179 |
) # Changed np.where to torch.where
|
| 180 |
|
| 181 |
return result_BxTxC
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
dia/config.py
CHANGED
|
@@ -14,149 +14,132 @@ Key components:
|
|
| 14 |
"""
|
| 15 |
|
| 16 |
import os
|
| 17 |
-
from typing import Annotated
|
| 18 |
|
| 19 |
-
from pydantic import BaseModel,
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
class DataConfig(BaseModel, frozen=True):
|
| 23 |
-
"""Configuration for data loading and preprocessing.
|
| 24 |
-
|
| 25 |
-
Attributes:
|
| 26 |
-
text_length: Maximum length of text sequences (must be multiple of 128).
|
| 27 |
-
audio_length: Maximum length of audio sequences (must be multiple of 128).
|
| 28 |
-
channels: Number of audio channels.
|
| 29 |
-
text_pad_value: Value used for padding text sequences.
|
| 30 |
-
audio_eos_value: Value representing the end of audio sequences.
|
| 31 |
-
audio_bos_value: Value representing the beginning of audio sequences.
|
| 32 |
-
audio_pad_value: Value used for padding audio sequences.
|
| 33 |
-
delay_pattern: List of delay values for each audio channel.
|
| 34 |
-
"""
|
| 35 |
-
|
| 36 |
-
text_length: Annotated[int, BeforeValidator(lambda x: (x + 127) // 128 * 128)] = (
|
| 37 |
-
Field(gt=0, multiple_of=128)
|
| 38 |
-
)
|
| 39 |
-
audio_length: Annotated[int, BeforeValidator(lambda x: (x + 127) // 128 * 128)] = (
|
| 40 |
-
Field(gt=0, multiple_of=128)
|
| 41 |
-
)
|
| 42 |
-
channels: int = Field(default=9, gt=0, multiple_of=1)
|
| 43 |
-
text_pad_value: int = Field(default=0)
|
| 44 |
-
audio_eos_value: int = Field(default=1024)
|
| 45 |
-
audio_pad_value: int = Field(default=1025)
|
| 46 |
-
audio_bos_value: int = Field(default=1026)
|
| 47 |
-
delay_pattern: list[Annotated[int, Field(ge=0)]] = Field(
|
| 48 |
-
default_factory=lambda: [0, 8, 9, 10, 11, 12, 13, 14, 15]
|
| 49 |
-
)
|
| 50 |
-
|
| 51 |
-
def __hash__(self) -> int:
|
| 52 |
-
"""Generate a hash based on all fields of the config."""
|
| 53 |
-
return hash(
|
| 54 |
-
(
|
| 55 |
-
self.text_length,
|
| 56 |
-
self.audio_length,
|
| 57 |
-
self.channels,
|
| 58 |
-
self.text_pad_value,
|
| 59 |
-
self.audio_pad_value,
|
| 60 |
-
self.audio_bos_value,
|
| 61 |
-
self.audio_eos_value,
|
| 62 |
-
tuple(self.delay_pattern),
|
| 63 |
-
)
|
| 64 |
-
)
|
| 65 |
|
| 66 |
|
| 67 |
class EncoderConfig(BaseModel, frozen=True):
|
| 68 |
"""Configuration for the encoder component of the Dia model.
|
| 69 |
|
| 70 |
Attributes:
|
| 71 |
-
|
| 72 |
-
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 76 |
"""
|
| 77 |
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 83 |
|
| 84 |
|
| 85 |
class DecoderConfig(BaseModel, frozen=True):
|
| 86 |
"""Configuration for the decoder component of the Dia model.
|
| 87 |
|
| 88 |
Attributes:
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 97 |
"""
|
| 98 |
|
| 99 |
-
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 107 |
|
| 108 |
|
| 109 |
-
class
|
| 110 |
"""Main configuration container for the Dia model architecture.
|
| 111 |
|
| 112 |
Attributes:
|
|
|
|
|
|
|
| 113 |
encoder: Configuration for the encoder component.
|
| 114 |
decoder: Configuration for the decoder component.
|
| 115 |
src_vocab_size: Size of the source (text) vocabulary.
|
| 116 |
tgt_vocab_size: Size of the target (audio code) vocabulary.
|
| 117 |
-
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 122 |
"""
|
| 123 |
|
| 124 |
-
|
| 125 |
-
|
| 126 |
-
src_vocab_size: int = Field(default=128, gt=0)
|
| 127 |
-
tgt_vocab_size: int = Field(default=1028, gt=0)
|
| 128 |
-
dropout: float = Field(default=0.0, ge=0.0, lt=1.0)
|
| 129 |
-
normalization_layer_epsilon: float = Field(default=1.0e-5, ge=0.0)
|
| 130 |
-
weight_dtype: str = Field(default="float32", description="Weight precision")
|
| 131 |
-
rope_min_timescale: int = Field(
|
| 132 |
-
default=1, description="Timescale For global Attention"
|
| 133 |
)
|
| 134 |
-
|
| 135 |
-
|
|
|
|
|
|
|
| 136 |
)
|
| 137 |
-
|
| 138 |
-
|
| 139 |
-
|
| 140 |
-
|
| 141 |
-
|
| 142 |
-
|
| 143 |
-
|
| 144 |
-
|
| 145 |
-
|
| 146 |
-
Combines all sub-configurations into a single validated object.
|
| 147 |
-
|
| 148 |
-
Attributes:
|
| 149 |
-
version: Configuration version string.
|
| 150 |
-
model: Model architecture configuration.
|
| 151 |
-
training: Training process configuration (precision settings).
|
| 152 |
-
data: Data loading and processing configuration.
|
| 153 |
-
"""
|
| 154 |
-
|
| 155 |
-
version: str = Field(default="1.0")
|
| 156 |
-
model: ModelConfig
|
| 157 |
-
# TODO: remove training. this is just for backwards-compatability
|
| 158 |
-
training: TrainingConfig
|
| 159 |
-
data: DataConfig
|
| 160 |
|
| 161 |
def save(self, path: str) -> None:
|
| 162 |
"""Save the current configuration instance to a JSON file.
|
|
|
|
| 14 |
"""
|
| 15 |
|
| 16 |
import os
|
|
|
|
| 17 |
|
| 18 |
+
from pydantic import BaseModel, Field
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 19 |
|
| 20 |
|
| 21 |
class EncoderConfig(BaseModel, frozen=True):
|
| 22 |
"""Configuration for the encoder component of the Dia model.
|
| 23 |
|
| 24 |
Attributes:
|
| 25 |
+
model_type: Type of the model, defaults to "dia_encoder".
|
| 26 |
+
hidden_size: Size of the encoder layers, defaults to 1024.
|
| 27 |
+
intermediate_size: Size of the "intermediate" (i.e., feed-forward) layer in the encoder, defaults to 4096.
|
| 28 |
+
num_hidden_layers: Number of hidden layers in the encoder, defaults to 12.
|
| 29 |
+
num_attention_heads: Number of attention heads in the encoder, defaults to 16.
|
| 30 |
+
num_key_value_heads: Number of key-value heads in the encoder, defaults to 16.
|
| 31 |
+
head_dim: Dimension of each attention head, defaults to 128.
|
| 32 |
+
hidden_act: Activation function in the encoder, defaults to "silu".
|
| 33 |
+
max_position_embeddings: Maximum number of position embeddings, defaults to 1024.
|
| 34 |
+
initializer_range: Range for initializing weights, defaults to 0.02.
|
| 35 |
+
norm_eps: Epsilon value for normalization layers, defaults to 1e-5.
|
| 36 |
+
rope_theta: Theta value for RoPE, defaults to 10000.0.
|
| 37 |
+
rope_scaling: Optional scaling factor for RoPE.
|
| 38 |
+
vocab_size: Vocabulary size, defaults to 256.
|
| 39 |
"""
|
| 40 |
|
| 41 |
+
head_dim: int = Field(default=128, gt=0)
|
| 42 |
+
hidden_act: str = Field(default="silu")
|
| 43 |
+
hidden_size: int = Field(default=1024, gt=0)
|
| 44 |
+
initializer_range: float = Field(default=0.02)
|
| 45 |
+
intermediate_size: int = Field(default=4096, gt=0)
|
| 46 |
+
max_position_embeddings: int = Field(default=1024, gt=0)
|
| 47 |
+
model_type: str = Field(default="dia_encoder")
|
| 48 |
+
norm_eps: float = Field(default=1e-5)
|
| 49 |
+
num_attention_heads: int = Field(default=16, gt=0)
|
| 50 |
+
num_hidden_layers: int = Field(default=12, gt=0)
|
| 51 |
+
num_key_value_heads: int = Field(default=16, gt=0)
|
| 52 |
+
rope_scaling: float | None = Field(default=None)
|
| 53 |
+
rope_theta: float = Field(default=10000.0)
|
| 54 |
+
vocab_size: int = Field(default=256, gt=0)
|
| 55 |
|
| 56 |
|
| 57 |
class DecoderConfig(BaseModel, frozen=True):
|
| 58 |
"""Configuration for the decoder component of the Dia model.
|
| 59 |
|
| 60 |
Attributes:
|
| 61 |
+
model_type: Type of the model, defaults to "dia_decoder".
|
| 62 |
+
hidden_size: Size of the decoder layers, defaults to 2048.
|
| 63 |
+
intermediate_size: Size of the "intermediate" (i.e., feed-forward) layer in the decoder, defaults to 8192.
|
| 64 |
+
num_hidden_layers: Number of hidden layers in the decoder, defaults to 18.
|
| 65 |
+
num_attention_heads: Number of attention heads in the decoder, defaults to 16.
|
| 66 |
+
num_key_value_heads: Number of key-value heads in the decoder, defaults to 4.
|
| 67 |
+
head_dim: Dimension of each attention head, defaults to 128.
|
| 68 |
+
cross_hidden_size: Size of the cross-attention layers, defaults to 1024.
|
| 69 |
+
cross_num_attention_heads: Number of attention heads in the cross-attention mechanism, defaults to 16.
|
| 70 |
+
cross_num_key_value_heads: Number of key-value heads in the cross-attention mechanism, defaults to 16.
|
| 71 |
+
cross_head_dim: Dimension of each cross-attention head, defaults to 128.
|
| 72 |
+
hidden_act: Activation function in the decoder, defaults to "silu".
|
| 73 |
+
max_position_embeddings: Maximum number of position embeddings in the decoder, defaults to 3072.
|
| 74 |
+
initializer_range: Range for initializing weights in the decoder, defaults to 0.02.
|
| 75 |
+
norm_eps: Epsilon value for normalization layers in the decoder, defaults to 1e-5.
|
| 76 |
+
rope_theta: Theta value for RoPE in the decoder, defaults to 10000.0.
|
| 77 |
+
rope_scaling: Optional scaling factor for RoPE in the decoder.
|
| 78 |
+
vocab_size: Vocabulary size for the decoder, defaults to 1028.
|
| 79 |
+
num_channels: Number of channels in the decoder, defaults to 9.
|
| 80 |
"""
|
| 81 |
|
| 82 |
+
cross_head_dim: int = Field(default=128, gt=0)
|
| 83 |
+
cross_hidden_size: int = Field(default=1024, gt=0)
|
| 84 |
+
cross_num_attention_heads: int = Field(default=16, gt=0)
|
| 85 |
+
cross_num_key_value_heads: int = Field(default=16, gt=0)
|
| 86 |
+
head_dim: int = Field(default=128, gt=0)
|
| 87 |
+
hidden_act: str = Field(default="silu")
|
| 88 |
+
hidden_size: int = Field(default=2048, gt=0)
|
| 89 |
+
initializer_range: float = Field(default=0.02)
|
| 90 |
+
intermediate_size: int = Field(default=8192, gt=0)
|
| 91 |
+
max_position_embeddings: int = Field(default=3072, gt=0)
|
| 92 |
+
model_type: str = Field(default="dia_decoder")
|
| 93 |
+
norm_eps: float = Field(default=1e-5)
|
| 94 |
+
num_attention_heads: int = Field(default=16, gt=0)
|
| 95 |
+
num_channels: int = Field(default=9, gt=0)
|
| 96 |
+
num_hidden_layers: int = Field(default=18, gt=0)
|
| 97 |
+
num_key_value_heads: int = Field(default=4, gt=0)
|
| 98 |
+
rope_scaling: float | None = Field(default=None)
|
| 99 |
+
rope_theta: float = Field(default=10000.0)
|
| 100 |
+
vocab_size: int = Field(default=1028, gt=0)
|
| 101 |
|
| 102 |
|
| 103 |
+
class DiaConfig(BaseModel, frozen=True):
|
| 104 |
"""Main configuration container for the Dia model architecture.
|
| 105 |
|
| 106 |
Attributes:
|
| 107 |
+
model_type: Type of the model, defaults to "dia".
|
| 108 |
+
is_encoder_decoder: Flag indicating if the model is an encoder-decoder type, defaults to True.
|
| 109 |
encoder: Configuration for the encoder component.
|
| 110 |
decoder: Configuration for the decoder component.
|
| 111 |
src_vocab_size: Size of the source (text) vocabulary.
|
| 112 |
tgt_vocab_size: Size of the target (audio code) vocabulary.
|
| 113 |
+
initializer_range: Range for initializing weights, defaults to 0.02.
|
| 114 |
+
norm_eps: Epsilon value for normalization layers, defaults to 1e-5.
|
| 115 |
+
torch_dtype: Data type for model weights in PyTorch, defaults to "float32".
|
| 116 |
+
bos_token_id: Beginning-of-sequence token ID, defaults to 1026.
|
| 117 |
+
eos_token_id: End-of-sequence token ID, defaults to 1024.
|
| 118 |
+
pad_token_id: Padding token ID, defaults to 1025.
|
| 119 |
+
rope_theta: Theta value for RoPE, defaults to 10000.0.
|
| 120 |
+
rope_scaling: Optional scaling factor for RoPE.
|
| 121 |
+
transformers_version: Version of the transformers library, defaults to "4.53.0.dev0".
|
| 122 |
+
architectures: List of model architectures, defaults to ["DiaForConditionalGeneration"].
|
| 123 |
+
delay_pattern: List of delay values for each audio channel, defaults to [0,8,9,10,11,12,13,14,15].
|
| 124 |
"""
|
| 125 |
|
| 126 |
+
architectures: list[str] = Field(
|
| 127 |
+
default_factory=lambda: ["DiaForConditionalGeneration"]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 128 |
)
|
| 129 |
+
bos_token_id: int = Field(default=1026)
|
| 130 |
+
decoder_config: DecoderConfig
|
| 131 |
+
delay_pattern: list[int] = Field(
|
| 132 |
+
default_factory=lambda: [0, 8, 9, 10, 11, 12, 13, 14, 15]
|
| 133 |
)
|
| 134 |
+
encoder_config: EncoderConfig
|
| 135 |
+
eos_token_id: int = Field(default=1024)
|
| 136 |
+
initializer_range: float = Field(default=0.02)
|
| 137 |
+
is_encoder_decoder: bool = Field(default=True)
|
| 138 |
+
model_type: str = Field(default="dia")
|
| 139 |
+
norm_eps: float = Field(default=1e-5)
|
| 140 |
+
pad_token_id: int = Field(default=1025)
|
| 141 |
+
torch_dtype: str = Field(default="float32")
|
| 142 |
+
transformers_version: str = Field(default="4.53.0.dev0")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 143 |
|
| 144 |
def save(self, path: str) -> None:
|
| 145 |
"""Save the current configuration instance to a JSON file.
|
dia/layers.py
CHANGED
|
@@ -1,10 +1,11 @@
|
|
| 1 |
import torch
|
| 2 |
import torch.nn as nn
|
| 3 |
import torch.nn.functional as F
|
|
|
|
| 4 |
from torch import Tensor
|
| 5 |
from torch.nn import RMSNorm
|
| 6 |
|
| 7 |
-
from .config import DiaConfig
|
| 8 |
from .state import DecoderInferenceState, EncoderInferenceState, KVCache
|
| 9 |
|
| 10 |
|
|
@@ -15,12 +16,10 @@ def _normalize_axes(axes: tuple[int, ...], ndim: int) -> tuple[int, ...]:
|
|
| 15 |
class DenseGeneral(nn.Module):
|
| 16 |
"""
|
| 17 |
PyTorch equivalent of flax.linen.DenseGeneral with shapes defined at init.
|
| 18 |
-
|
| 19 |
Stores weights (`kernel`) in the same layout as Jax and uses torch.tensordot
|
| 20 |
for the generalized matrix multiplication. Weight/bias shapes are calculated
|
| 21 |
and parameters created during initialization based on config.
|
| 22 |
`load_weights` validates shapes and copies data.
|
| 23 |
-
|
| 24 |
Attributes:
|
| 25 |
axis (Tuple[int, ...]): Input axis or axes to contract.
|
| 26 |
in_shapes (Tuple[int, ...]): Sizes of the input dimensions specified by `axis`.
|
|
@@ -46,7 +45,6 @@ class DenseGeneral(nn.Module):
|
|
| 46 |
|
| 47 |
factory_kwargs = {"device": device, "dtype": weight_dtype}
|
| 48 |
self.weight = nn.Parameter(torch.empty(self.kernel_shape, **factory_kwargs))
|
| 49 |
-
self.register_parameter("bias", None)
|
| 50 |
|
| 51 |
def forward(self, inputs: Tensor) -> Tensor:
|
| 52 |
norm_axis = _normalize_axes(self.axis, inputs.ndim)
|
|
@@ -112,53 +110,112 @@ class RotaryEmbedding(nn.Module):
|
|
| 112 |
self.embedding_dims = embedding_dims
|
| 113 |
self.min_timescale = min_timescale
|
| 114 |
self.max_timescale = max_timescale
|
| 115 |
-
self.
|
| 116 |
|
| 117 |
half_embedding_dim = embedding_dims // 2
|
| 118 |
fraction = (2.0 * torch.arange(0, half_embedding_dim)) / embedding_dims
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
|
| 122 |
-
|
| 123 |
-
)
|
| 124 |
-
|
| 125 |
-
def extra_repr(self) -> str:
|
| 126 |
-
s = f"{self.timescale.shape}"
|
| 127 |
-
return s
|
| 128 |
|
| 129 |
def forward(self, inputs: torch.Tensor, position: torch.Tensor):
|
| 130 |
"""Applies RoPE."""
|
| 131 |
position = position.unsqueeze(-1).unsqueeze(-1)
|
| 132 |
-
|
| 133 |
-
|
| 134 |
-
|
| 135 |
-
|
| 136 |
-
first_half
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 137 |
first_part = first_half * cos - second_half * sin
|
| 138 |
second_part = second_half * cos + first_half * sin
|
| 139 |
-
return torch.cat(
|
|
|
|
|
|
|
|
|
|
| 140 |
|
| 141 |
|
| 142 |
-
|
| 143 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 144 |
|
| 145 |
def __init__(
|
| 146 |
self,
|
| 147 |
-
config:
|
| 148 |
q_embed_dim: int,
|
| 149 |
kv_embed_dim: int,
|
| 150 |
num_query_heads: int,
|
| 151 |
num_kv_heads: int,
|
| 152 |
head_dim: int,
|
| 153 |
compute_dtype: torch.dtype,
|
| 154 |
-
is_cross_attn: bool = False,
|
| 155 |
out_embed_dim: int | None = None,
|
| 156 |
):
|
| 157 |
super().__init__()
|
| 158 |
self.num_query_heads = num_query_heads
|
| 159 |
self.num_kv_heads = num_kv_heads
|
| 160 |
self.head_dim = head_dim
|
| 161 |
-
self.is_cross_attn = is_cross_attn
|
| 162 |
self.output_dim = out_embed_dim if out_embed_dim is not None else q_embed_dim
|
| 163 |
self.projected_query_dim = num_query_heads * head_dim
|
| 164 |
if num_query_heads % num_kv_heads != 0:
|
|
@@ -196,21 +253,18 @@ class Attention(nn.Module):
|
|
| 196 |
# --- Rotary Embedding ---
|
| 197 |
self.rotary_emb = RotaryEmbedding(
|
| 198 |
embedding_dims=self.head_dim,
|
| 199 |
-
|
| 200 |
-
max_timescale=config.model.rope_max_timescale,
|
| 201 |
dtype=compute_dtype,
|
| 202 |
)
|
| 203 |
|
| 204 |
def forward(
|
| 205 |
self,
|
| 206 |
Xq: torch.Tensor, # (B, T, D) T = 1 in AR generation
|
| 207 |
-
Xkv: torch.Tensor, # (B, S, E) S = 1 in AR generation
|
| 208 |
q_positions: torch.Tensor, # (B, T)
|
| 209 |
kv_positions: torch.Tensor | None = None, # (B, S)
|
| 210 |
attn_mask: torch.Tensor
|
| 211 |
| None = None, # None in Decoder Self Attention, Valid mask in Others
|
| 212 |
cache: KVCache | None = None, # None in Encoder, KVCache in Decoder
|
| 213 |
-
prefill: bool = False,
|
| 214 |
is_causal: bool = False,
|
| 215 |
) -> tuple[torch.Tensor, tuple[torch.Tensor, torch.Tensor] | None]:
|
| 216 |
"""
|
|
@@ -223,7 +277,6 @@ class Attention(nn.Module):
|
|
| 223 |
kv_positions: Positions for keys/values (B, S). If None, uses q_positions.
|
| 224 |
attn_mask: Attention mask.
|
| 225 |
cache: KVCache.
|
| 226 |
-
prefill: If True, use prefill mode.
|
| 227 |
|
| 228 |
Returns:
|
| 229 |
A tuple containing:
|
|
@@ -235,44 +288,266 @@ class Attention(nn.Module):
|
|
| 235 |
original_dtype = Xq.dtype
|
| 236 |
|
| 237 |
Xq_BxTxNxH = self.q_proj(Xq)
|
| 238 |
-
Xq_BxTxNxH = self.rotary_emb(Xq_BxTxNxH, position=q_positions)
|
| 239 |
Xq_BxNxTxH = Xq_BxTxNxH.transpose(1, 2)
|
| 240 |
|
| 241 |
attn_k: torch.Tensor | None = None
|
| 242 |
attn_v: torch.Tensor | None = None
|
| 243 |
|
| 244 |
-
|
| 245 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 246 |
else:
|
| 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 |
attn_output = attn_output.transpose(1, 2).contiguous() # (B, T, N, H)
|
| 277 |
output = self.o_proj(attn_output)
|
| 278 |
|
|
@@ -285,34 +560,33 @@ class EncoderLayer(nn.Module):
|
|
| 285 |
def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
|
| 286 |
super().__init__()
|
| 287 |
self.config = config
|
| 288 |
-
|
| 289 |
-
|
| 290 |
-
|
| 291 |
|
| 292 |
self.pre_sa_norm = RMSNorm(
|
| 293 |
embed_dim,
|
| 294 |
-
eps=
|
| 295 |
dtype=torch.float32,
|
| 296 |
)
|
| 297 |
-
self.self_attention =
|
| 298 |
-
|
| 299 |
q_embed_dim=embed_dim,
|
| 300 |
kv_embed_dim=embed_dim,
|
| 301 |
-
num_query_heads=enc_config.
|
| 302 |
-
num_kv_heads=enc_config.
|
| 303 |
head_dim=enc_config.head_dim,
|
| 304 |
compute_dtype=compute_dtype,
|
| 305 |
-
is_cross_attn=False,
|
| 306 |
out_embed_dim=embed_dim,
|
| 307 |
)
|
| 308 |
self.post_sa_norm = RMSNorm(
|
| 309 |
embed_dim,
|
| 310 |
-
eps=
|
| 311 |
dtype=torch.float32,
|
| 312 |
)
|
| 313 |
self.mlp = MlpBlock(
|
| 314 |
embed_dim=embed_dim,
|
| 315 |
-
intermediate_dim=enc_config.
|
| 316 |
compute_dtype=compute_dtype,
|
| 317 |
)
|
| 318 |
|
|
@@ -322,10 +596,10 @@ class EncoderLayer(nn.Module):
|
|
| 322 |
state: EncoderInferenceState,
|
| 323 |
) -> torch.Tensor:
|
| 324 |
residual = x
|
| 325 |
-
x_norm = self.pre_sa_norm(x)
|
|
|
|
| 326 |
sa_out = self.self_attention(
|
| 327 |
-
|
| 328 |
-
Xkv=x_norm,
|
| 329 |
q_positions=state.positions,
|
| 330 |
kv_positions=state.positions,
|
| 331 |
attn_mask=state.attn_mask,
|
|
@@ -333,7 +607,7 @@ class EncoderLayer(nn.Module):
|
|
| 333 |
x = residual + sa_out
|
| 334 |
|
| 335 |
residual = x
|
| 336 |
-
x_norm = self.post_sa_norm(x)
|
| 337 |
mlp_out = self.mlp(x_norm)
|
| 338 |
x = residual + mlp_out
|
| 339 |
|
|
@@ -346,20 +620,23 @@ class Encoder(nn.Module):
|
|
| 346 |
def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
|
| 347 |
super().__init__()
|
| 348 |
self.config = config
|
| 349 |
-
|
| 350 |
-
|
| 351 |
|
| 352 |
self.embedding = nn.Embedding(
|
| 353 |
-
|
| 354 |
-
enc_config.
|
| 355 |
dtype=compute_dtype,
|
| 356 |
)
|
| 357 |
self.layers = nn.ModuleList(
|
| 358 |
-
[
|
|
|
|
|
|
|
|
|
|
| 359 |
)
|
| 360 |
self.norm = RMSNorm(
|
| 361 |
-
enc_config.
|
| 362 |
-
eps=
|
| 363 |
dtype=torch.float32,
|
| 364 |
)
|
| 365 |
|
|
@@ -373,7 +650,7 @@ class Encoder(nn.Module):
|
|
| 373 |
for layer in self.layers:
|
| 374 |
x = layer(x, state)
|
| 375 |
|
| 376 |
-
x = self.norm(x)
|
| 377 |
return x
|
| 378 |
|
| 379 |
|
|
@@ -383,57 +660,55 @@ class DecoderLayer(nn.Module):
|
|
| 383 |
def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
|
| 384 |
super().__init__()
|
| 385 |
self.config = config
|
| 386 |
-
|
| 387 |
-
|
| 388 |
-
|
| 389 |
-
|
| 390 |
-
|
| 391 |
|
| 392 |
# Norms
|
| 393 |
self.pre_sa_norm = RMSNorm(
|
| 394 |
dec_embed_dim,
|
| 395 |
-
eps=
|
| 396 |
dtype=torch.float32,
|
| 397 |
)
|
| 398 |
self.pre_ca_norm = RMSNorm(
|
| 399 |
dec_embed_dim,
|
| 400 |
-
eps=
|
| 401 |
dtype=torch.float32,
|
| 402 |
)
|
| 403 |
self.pre_mlp_norm = RMSNorm(
|
| 404 |
dec_embed_dim,
|
| 405 |
-
eps=
|
| 406 |
dtype=torch.float32,
|
| 407 |
)
|
| 408 |
|
| 409 |
# Self-Attention (GQA) with Causal Masking
|
| 410 |
-
self.self_attention =
|
| 411 |
-
|
| 412 |
q_embed_dim=dec_embed_dim,
|
| 413 |
kv_embed_dim=dec_embed_dim,
|
| 414 |
-
num_query_heads=dec_config.
|
| 415 |
-
num_kv_heads=dec_config.
|
| 416 |
-
head_dim=dec_config.
|
| 417 |
compute_dtype=compute_dtype,
|
| 418 |
-
is_cross_attn=False,
|
| 419 |
out_embed_dim=dec_embed_dim,
|
| 420 |
)
|
| 421 |
# Cross-Attention (MHA)
|
| 422 |
-
self.cross_attention =
|
| 423 |
-
|
| 424 |
q_embed_dim=dec_embed_dim,
|
| 425 |
kv_embed_dim=enc_embed_dim, # Note kv_embed_dim
|
| 426 |
-
num_query_heads=dec_config.
|
| 427 |
-
num_kv_heads=dec_config.
|
| 428 |
head_dim=dec_config.cross_head_dim,
|
| 429 |
compute_dtype=compute_dtype,
|
| 430 |
-
is_cross_attn=True,
|
| 431 |
out_embed_dim=dec_embed_dim,
|
| 432 |
)
|
| 433 |
# MLP
|
| 434 |
self.mlp = MlpBlock(
|
| 435 |
embed_dim=dec_embed_dim,
|
| 436 |
-
intermediate_dim=dec_config.
|
| 437 |
compute_dtype=compute_dtype,
|
| 438 |
)
|
| 439 |
|
|
@@ -444,37 +719,39 @@ class DecoderLayer(nn.Module):
|
|
| 444 |
self_attn_cache: KVCache | None = None,
|
| 445 |
cross_attn_cache: KVCache | None = None,
|
| 446 |
prefill: bool = False,
|
|
|
|
| 447 |
) -> torch.Tensor:
|
| 448 |
residual = x
|
| 449 |
-
x_norm = self.pre_sa_norm(x)
|
|
|
|
|
|
|
| 450 |
|
| 451 |
sa_out = self.self_attention(
|
| 452 |
-
|
| 453 |
-
Xkv=x_norm, # (2, 1, D)
|
| 454 |
q_positions=state.dec_positions, # (2, 1)
|
| 455 |
kv_positions=state.dec_positions, # (2, 1)
|
| 456 |
-
attn_mask=
|
| 457 |
cache=self_attn_cache,
|
| 458 |
prefill=prefill,
|
| 459 |
is_causal=prefill,
|
|
|
|
| 460 |
)
|
| 461 |
|
| 462 |
x = residual + sa_out
|
| 463 |
|
| 464 |
residual = x
|
| 465 |
-
x_norm = self.pre_ca_norm(x)
|
| 466 |
ca_out = self.cross_attention(
|
| 467 |
Xq=x_norm,
|
| 468 |
-
Xkv=state.enc_out,
|
| 469 |
q_positions=state.dec_positions,
|
| 470 |
kv_positions=state.enc_positions,
|
| 471 |
-
attn_mask=state.
|
| 472 |
cache=cross_attn_cache,
|
| 473 |
)
|
| 474 |
x = residual + ca_out
|
| 475 |
|
| 476 |
residual = x
|
| 477 |
-
x_norm = self.pre_mlp_norm(x)
|
| 478 |
mlp_out = self.mlp(x_norm)
|
| 479 |
x = residual + mlp_out
|
| 480 |
|
|
@@ -487,16 +764,14 @@ class Decoder(nn.Module):
|
|
| 487 |
def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
|
| 488 |
super().__init__()
|
| 489 |
self.config = config
|
| 490 |
-
|
| 491 |
-
|
| 492 |
-
|
| 493 |
-
self.num_channels = data_config.channels
|
| 494 |
-
self.num_layers = dec_config.n_layer
|
| 495 |
|
| 496 |
self.embeddings = nn.ModuleList(
|
| 497 |
[
|
| 498 |
nn.Embedding(
|
| 499 |
-
|
| 500 |
)
|
| 501 |
for _ in range(self.num_channels)
|
| 502 |
]
|
|
@@ -509,14 +784,14 @@ class Decoder(nn.Module):
|
|
| 509 |
)
|
| 510 |
|
| 511 |
self.norm = RMSNorm(
|
| 512 |
-
dec_config.
|
| 513 |
-
eps=
|
| 514 |
dtype=torch.float32,
|
| 515 |
)
|
| 516 |
|
| 517 |
self.logits_dense = DenseGeneral(
|
| 518 |
-
in_shapes=(dec_config.
|
| 519 |
-
out_features=(self.num_channels,
|
| 520 |
axis=(-1,),
|
| 521 |
weight_dtype=compute_dtype,
|
| 522 |
)
|
|
@@ -524,7 +799,6 @@ class Decoder(nn.Module):
|
|
| 524 |
def precompute_cross_attn_cache(
|
| 525 |
self,
|
| 526 |
enc_out: torch.Tensor, # (B, S, E)
|
| 527 |
-
enc_positions: torch.Tensor, # (B, S)
|
| 528 |
) -> list[KVCache]:
|
| 529 |
"""
|
| 530 |
Computes the Key and Value tensors for cross-attention for each layer from the encoder output.
|
|
@@ -536,7 +810,6 @@ class Decoder(nn.Module):
|
|
| 536 |
k_proj = cross_attn_module.k_proj(enc_out)
|
| 537 |
v_proj = cross_attn_module.v_proj(enc_out)
|
| 538 |
|
| 539 |
-
k_proj = cross_attn_module.rotary_emb(k_proj, position=enc_positions)
|
| 540 |
k = k_proj.transpose(1, 2)
|
| 541 |
v = v_proj.transpose(1, 2)
|
| 542 |
|
|
@@ -548,10 +821,10 @@ class Decoder(nn.Module):
|
|
| 548 |
self,
|
| 549 |
tgt_ids_Bx1xC: torch.Tensor, # [B, 1, C]
|
| 550 |
state: DecoderInferenceState,
|
|
|
|
| 551 |
) -> torch.Tensor:
|
| 552 |
"""
|
| 553 |
Performs a single decoding step, managing KV caches layer by layer.
|
| 554 |
-
|
| 555 |
Returns:
|
| 556 |
A tuple containing:
|
| 557 |
- logits_Bx1xCV: The final output logits for the current step (B, 1, C*V), cast to float32.
|
|
@@ -571,6 +844,7 @@ class Decoder(nn.Module):
|
|
| 571 |
state,
|
| 572 |
self_attn_cache=self_cache,
|
| 573 |
cross_attn_cache=cross_cache,
|
|
|
|
| 574 |
)
|
| 575 |
|
| 576 |
x = self.norm(x)
|
|
@@ -583,7 +857,6 @@ class Decoder(nn.Module):
|
|
| 583 |
) -> torch.Tensor:
|
| 584 |
"""
|
| 585 |
Forward pass for the Decoder stack, managing KV caches.
|
| 586 |
-
|
| 587 |
Args:
|
| 588 |
tgt_ids_BxTxC: Target token IDs (B, T, C).
|
| 589 |
encoder_out: Output from the encoder (B, S, E).
|
|
@@ -597,7 +870,6 @@ class Decoder(nn.Module):
|
|
| 597 |
precomputed_cross_attn_kv: A single tuple containing the pre-computed K/V cache
|
| 598 |
derived from `encoder_out`. This is passed identically
|
| 599 |
to all layers.
|
| 600 |
-
|
| 601 |
Returns:
|
| 602 |
A tuple containing:
|
| 603 |
- logits: The final output logits (B, T, C * V), cast to float32.
|
|
@@ -632,7 +904,19 @@ class Decoder(nn.Module):
|
|
| 632 |
return logits_BxTxCxV.to(torch.float32)
|
| 633 |
|
| 634 |
|
| 635 |
-
class DiaModel(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 636 |
"""PyTorch Dia Model using DenseGeneral."""
|
| 637 |
|
| 638 |
def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
|
|
|
|
| 1 |
import torch
|
| 2 |
import torch.nn as nn
|
| 3 |
import torch.nn.functional as F
|
| 4 |
+
from huggingface_hub import PyTorchModelHubMixin
|
| 5 |
from torch import Tensor
|
| 6 |
from torch.nn import RMSNorm
|
| 7 |
|
| 8 |
+
from .config import DecoderConfig, DiaConfig, EncoderConfig
|
| 9 |
from .state import DecoderInferenceState, EncoderInferenceState, KVCache
|
| 10 |
|
| 11 |
|
|
|
|
| 16 |
class DenseGeneral(nn.Module):
|
| 17 |
"""
|
| 18 |
PyTorch equivalent of flax.linen.DenseGeneral with shapes defined at init.
|
|
|
|
| 19 |
Stores weights (`kernel`) in the same layout as Jax and uses torch.tensordot
|
| 20 |
for the generalized matrix multiplication. Weight/bias shapes are calculated
|
| 21 |
and parameters created during initialization based on config.
|
| 22 |
`load_weights` validates shapes and copies data.
|
|
|
|
| 23 |
Attributes:
|
| 24 |
axis (Tuple[int, ...]): Input axis or axes to contract.
|
| 25 |
in_shapes (Tuple[int, ...]): Sizes of the input dimensions specified by `axis`.
|
|
|
|
| 45 |
|
| 46 |
factory_kwargs = {"device": device, "dtype": weight_dtype}
|
| 47 |
self.weight = nn.Parameter(torch.empty(self.kernel_shape, **factory_kwargs))
|
|
|
|
| 48 |
|
| 49 |
def forward(self, inputs: Tensor) -> Tensor:
|
| 50 |
norm_axis = _normalize_axes(self.axis, inputs.ndim)
|
|
|
|
| 110 |
self.embedding_dims = embedding_dims
|
| 111 |
self.min_timescale = min_timescale
|
| 112 |
self.max_timescale = max_timescale
|
| 113 |
+
self.compute_dtype = dtype
|
| 114 |
|
| 115 |
half_embedding_dim = embedding_dims // 2
|
| 116 |
fraction = (2.0 * torch.arange(0, half_embedding_dim)) / embedding_dims
|
| 117 |
+
timescale = (
|
| 118 |
+
self.min_timescale * (self.max_timescale / self.min_timescale) ** fraction
|
| 119 |
+
).to(torch.float32)
|
| 120 |
+
self.register_buffer("timescale", timescale, persistent=False)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 121 |
|
| 122 |
def forward(self, inputs: torch.Tensor, position: torch.Tensor):
|
| 123 |
"""Applies RoPE."""
|
| 124 |
position = position.unsqueeze(-1).unsqueeze(-1)
|
| 125 |
+
sinusoid_inp = position / self.timescale
|
| 126 |
+
sin = torch.sin(sinusoid_inp)
|
| 127 |
+
cos = torch.cos(sinusoid_inp)
|
| 128 |
+
first_half, second_half = torch.chunk(inputs.to(torch.float32), 2, dim=-1)
|
| 129 |
+
first_part = first_half * cos - second_half * sin
|
| 130 |
+
second_part = second_half * cos + first_half * sin
|
| 131 |
+
return torch.cat(
|
| 132 |
+
(first_part.to(self.compute_dtype), second_part.to(self.compute_dtype)),
|
| 133 |
+
dim=-1,
|
| 134 |
+
)
|
| 135 |
+
|
| 136 |
+
def apply_rope(self, inputs: torch.Tensor, sin: torch.Tensor, cos: torch.Tensor):
|
| 137 |
+
first_half, second_half = torch.chunk(inputs.to(torch.float32), 2, dim=-1)
|
| 138 |
first_part = first_half * cos - second_half * sin
|
| 139 |
second_part = second_half * cos + first_half * sin
|
| 140 |
+
return torch.cat(
|
| 141 |
+
(first_part.to(self.compute_dtype), second_part.to(self.compute_dtype)),
|
| 142 |
+
dim=-1,
|
| 143 |
+
)
|
| 144 |
|
| 145 |
|
| 146 |
+
def custom_scaled_dot_product_attention(
|
| 147 |
+
query: torch.Tensor,
|
| 148 |
+
key: torch.Tensor,
|
| 149 |
+
value: torch.Tensor,
|
| 150 |
+
attn_mask: torch.Tensor | None = None,
|
| 151 |
+
scale: float = 1.0,
|
| 152 |
+
is_causal: bool = False,
|
| 153 |
+
num_gqa_groups: int = 1,
|
| 154 |
+
) -> torch.Tensor:
|
| 155 |
+
"""
|
| 156 |
+
Custom scaled dot-product attention with GQA support for MPS compatibility.
|
| 157 |
+
|
| 158 |
+
Args:
|
| 159 |
+
query: (B, N_q, T, H) - Query tensor, N_q = num_query_heads
|
| 160 |
+
key: (B, N_kv, S, H) - Key tensor, N_kv = num_kv_heads
|
| 161 |
+
value: (B, N_kv, S, H) - Value tensor
|
| 162 |
+
attn_mask: (B, 1, T, S) - Attention mask, optional
|
| 163 |
+
scale: Scaling factor for attention scores
|
| 164 |
+
is_causal: If True, apply causal masking
|
| 165 |
+
num_gqa_groups: Number of query groups per KV head (N_q / N_kv)
|
| 166 |
+
|
| 167 |
+
Returns:
|
| 168 |
+
output: (B, N_q, T, H) - Attention output
|
| 169 |
+
"""
|
| 170 |
+
B, N_q, T, H = query.shape
|
| 171 |
+
_, N_kv, S, _ = key.shape
|
| 172 |
+
|
| 173 |
+
# For GQA, repeat key and value tensors to match query heads
|
| 174 |
+
if num_gqa_groups > 1:
|
| 175 |
+
key = key.repeat_interleave(num_gqa_groups, dim=1) # (B, N_q, S, H)
|
| 176 |
+
value = value.repeat_interleave(num_gqa_groups, dim=1) # (B, N_q, S, H)
|
| 177 |
+
|
| 178 |
+
# Compute attention scores: (B, N_q, T, H) @ (B, N_q, H, S) -> (B, N_q, T, S)
|
| 179 |
+
scores = torch.matmul(query, key.transpose(-1, -2)) * scale
|
| 180 |
+
|
| 181 |
+
# Apply causal mask if needed
|
| 182 |
+
if is_causal:
|
| 183 |
+
causal_mask = torch.tril(
|
| 184 |
+
torch.ones(T, S, dtype=torch.bool, device=query.device)
|
| 185 |
+
)
|
| 186 |
+
scores = scores.masked_fill(~causal_mask, float("-inf"))
|
| 187 |
+
|
| 188 |
+
# Apply attention mask if provided
|
| 189 |
+
if attn_mask is not None:
|
| 190 |
+
scores = scores.masked_fill(~attn_mask, float("-inf"))
|
| 191 |
+
|
| 192 |
+
# Softmax over the last dimension (S)
|
| 193 |
+
attn_weights = F.softmax(scores, dim=-1)
|
| 194 |
+
|
| 195 |
+
# Compute output: (B, N_q, T, S) @ (B, N_q, S, H) -> (B, N_q, T, H)
|
| 196 |
+
output = torch.matmul(attn_weights, value)
|
| 197 |
+
|
| 198 |
+
return output
|
| 199 |
+
|
| 200 |
+
|
| 201 |
+
class CrossAttention(nn.Module):
|
| 202 |
+
"""Cross-Attention using DenseGeneral."""
|
| 203 |
|
| 204 |
def __init__(
|
| 205 |
self,
|
| 206 |
+
config: EncoderConfig | DecoderConfig,
|
| 207 |
q_embed_dim: int,
|
| 208 |
kv_embed_dim: int,
|
| 209 |
num_query_heads: int,
|
| 210 |
num_kv_heads: int,
|
| 211 |
head_dim: int,
|
| 212 |
compute_dtype: torch.dtype,
|
|
|
|
| 213 |
out_embed_dim: int | None = None,
|
| 214 |
):
|
| 215 |
super().__init__()
|
| 216 |
self.num_query_heads = num_query_heads
|
| 217 |
self.num_kv_heads = num_kv_heads
|
| 218 |
self.head_dim = head_dim
|
|
|
|
| 219 |
self.output_dim = out_embed_dim if out_embed_dim is not None else q_embed_dim
|
| 220 |
self.projected_query_dim = num_query_heads * head_dim
|
| 221 |
if num_query_heads % num_kv_heads != 0:
|
|
|
|
| 253 |
# --- Rotary Embedding ---
|
| 254 |
self.rotary_emb = RotaryEmbedding(
|
| 255 |
embedding_dims=self.head_dim,
|
| 256 |
+
max_timescale=config.rope_theta,
|
|
|
|
| 257 |
dtype=compute_dtype,
|
| 258 |
)
|
| 259 |
|
| 260 |
def forward(
|
| 261 |
self,
|
| 262 |
Xq: torch.Tensor, # (B, T, D) T = 1 in AR generation
|
|
|
|
| 263 |
q_positions: torch.Tensor, # (B, T)
|
| 264 |
kv_positions: torch.Tensor | None = None, # (B, S)
|
| 265 |
attn_mask: torch.Tensor
|
| 266 |
| None = None, # None in Decoder Self Attention, Valid mask in Others
|
| 267 |
cache: KVCache | None = None, # None in Encoder, KVCache in Decoder
|
|
|
|
| 268 |
is_causal: bool = False,
|
| 269 |
) -> tuple[torch.Tensor, tuple[torch.Tensor, torch.Tensor] | None]:
|
| 270 |
"""
|
|
|
|
| 277 |
kv_positions: Positions for keys/values (B, S). If None, uses q_positions.
|
| 278 |
attn_mask: Attention mask.
|
| 279 |
cache: KVCache.
|
|
|
|
| 280 |
|
| 281 |
Returns:
|
| 282 |
A tuple containing:
|
|
|
|
| 288 |
original_dtype = Xq.dtype
|
| 289 |
|
| 290 |
Xq_BxTxNxH = self.q_proj(Xq)
|
|
|
|
| 291 |
Xq_BxNxTxH = Xq_BxTxNxH.transpose(1, 2)
|
| 292 |
|
| 293 |
attn_k: torch.Tensor | None = None
|
| 294 |
attn_v: torch.Tensor | None = None
|
| 295 |
|
| 296 |
+
attn_k, attn_v = cache.k, cache.v
|
| 297 |
+
|
| 298 |
+
# Use custom attention for MPS backend, otherwise use optimized PyTorch function
|
| 299 |
+
is_mps = Xq.device.type == "mps" and torch.backends.mps.is_available()
|
| 300 |
+
if is_mps:
|
| 301 |
+
attn_output = custom_scaled_dot_product_attention(
|
| 302 |
+
query=Xq_BxNxTxH,
|
| 303 |
+
key=attn_k,
|
| 304 |
+
value=attn_v,
|
| 305 |
+
attn_mask=attn_mask if not is_causal else None,
|
| 306 |
+
scale=1.0,
|
| 307 |
+
is_causal=is_causal,
|
| 308 |
+
num_gqa_groups=self.num_gqa_groups,
|
| 309 |
+
)
|
| 310 |
else:
|
| 311 |
+
attn_output = F.scaled_dot_product_attention(
|
| 312 |
+
Xq_BxNxTxH,
|
| 313 |
+
attn_k,
|
| 314 |
+
attn_v,
|
| 315 |
+
attn_mask=attn_mask if not is_causal else None,
|
| 316 |
+
scale=1.0,
|
| 317 |
+
enable_gqa=self.num_gqa_groups > 1,
|
| 318 |
+
is_causal=is_causal,
|
| 319 |
+
)
|
| 320 |
+
|
| 321 |
+
attn_output = attn_output.transpose(1, 2).contiguous() # (B, T, N, H)
|
| 322 |
+
output = self.o_proj(attn_output)
|
| 323 |
+
|
| 324 |
+
return output.to(original_dtype)
|
| 325 |
+
|
| 326 |
+
|
| 327 |
+
class FusedQKV(nn.Module):
|
| 328 |
+
def __init__(
|
| 329 |
+
self,
|
| 330 |
+
in_features: int,
|
| 331 |
+
out_features: int,
|
| 332 |
+
bias: bool = False,
|
| 333 |
+
num_q_heads: int = 1,
|
| 334 |
+
q_head_dim: int = 1,
|
| 335 |
+
num_kv_heads: int = 1,
|
| 336 |
+
kv_head_dim: int = 1,
|
| 337 |
+
):
|
| 338 |
+
super().__init__()
|
| 339 |
+
self.num_q_heads = num_q_heads
|
| 340 |
+
self.q_head_dim = q_head_dim
|
| 341 |
+
self.num_kv_heads = num_kv_heads
|
| 342 |
+
self.kv_head_dim = kv_head_dim
|
| 343 |
+
self.q_output_dim = num_q_heads * q_head_dim
|
| 344 |
+
self.kv_output_dim = num_kv_heads * kv_head_dim
|
| 345 |
+
self.linear = nn.Linear(in_features, out_features, bias=bias)
|
| 346 |
+
|
| 347 |
+
def forward(
|
| 348 |
+
self, inputs: torch.Tensor
|
| 349 |
+
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
|
| 350 |
+
x = self.linear(inputs)
|
| 351 |
+
|
| 352 |
+
q, k, v = x.split(
|
| 353 |
+
[self.q_output_dim, self.kv_output_dim, self.kv_output_dim], dim=-1
|
| 354 |
)
|
| 355 |
|
| 356 |
+
q = q.reshape(q.shape[:-1] + (self.num_q_heads, self.q_head_dim))
|
| 357 |
+
k = k.reshape(k.shape[:-1] + (self.num_kv_heads, self.kv_head_dim))
|
| 358 |
+
v = v.reshape(v.shape[:-1] + (self.num_kv_heads, self.kv_head_dim))
|
| 359 |
+
|
| 360 |
+
return q, k, v
|
| 361 |
+
|
| 362 |
+
|
| 363 |
+
class SelfAttention(nn.Module):
|
| 364 |
+
"""Attention using DenseGeneral."""
|
| 365 |
+
|
| 366 |
+
def __init__(
|
| 367 |
+
self,
|
| 368 |
+
config: DiaConfig,
|
| 369 |
+
q_embed_dim: int,
|
| 370 |
+
kv_embed_dim: int,
|
| 371 |
+
num_query_heads: int,
|
| 372 |
+
num_kv_heads: int,
|
| 373 |
+
head_dim: int,
|
| 374 |
+
compute_dtype: torch.dtype,
|
| 375 |
+
out_embed_dim: int | None = None,
|
| 376 |
+
):
|
| 377 |
+
super().__init__()
|
| 378 |
+
self.num_query_heads = num_query_heads
|
| 379 |
+
self.num_kv_heads = num_kv_heads
|
| 380 |
+
self.head_dim = head_dim
|
| 381 |
+
self.output_dim = out_embed_dim if out_embed_dim is not None else q_embed_dim
|
| 382 |
+
self.projected_query_dim = num_query_heads * head_dim
|
| 383 |
+
if num_query_heads % num_kv_heads != 0:
|
| 384 |
+
raise ValueError(
|
| 385 |
+
f"num_query_heads ({num_query_heads}) must be divisible by num_kv_heads ({num_kv_heads})"
|
| 386 |
+
)
|
| 387 |
+
self.num_gqa_groups = num_query_heads // num_kv_heads
|
| 388 |
+
self.kv_embed_dim = kv_embed_dim
|
| 389 |
+
self.q_embed_dim = q_embed_dim
|
| 390 |
+
|
| 391 |
+
# --- Projection Layers using DenseGeneral ---
|
| 392 |
+
self.q_proj = DenseGeneral(
|
| 393 |
+
in_shapes=(q_embed_dim,),
|
| 394 |
+
out_features=(num_query_heads, head_dim),
|
| 395 |
+
axis=(-1,),
|
| 396 |
+
weight_dtype=compute_dtype,
|
| 397 |
+
)
|
| 398 |
+
self.k_proj = DenseGeneral(
|
| 399 |
+
in_shapes=(kv_embed_dim,),
|
| 400 |
+
out_features=(num_kv_heads, head_dim),
|
| 401 |
+
axis=(-1,),
|
| 402 |
+
weight_dtype=compute_dtype,
|
| 403 |
+
)
|
| 404 |
+
self.v_proj = DenseGeneral(
|
| 405 |
+
in_shapes=(kv_embed_dim,),
|
| 406 |
+
out_features=(num_kv_heads, head_dim),
|
| 407 |
+
axis=(-1,),
|
| 408 |
+
weight_dtype=compute_dtype,
|
| 409 |
+
)
|
| 410 |
+
self.o_proj = DenseGeneral(
|
| 411 |
+
in_shapes=(num_query_heads, head_dim),
|
| 412 |
+
out_features=(self.output_dim,),
|
| 413 |
+
axis=(-2, -1),
|
| 414 |
+
weight_dtype=compute_dtype,
|
| 415 |
+
)
|
| 416 |
+
|
| 417 |
+
# --- Rotary Embedding ---
|
| 418 |
+
self.rotary_emb = RotaryEmbedding(
|
| 419 |
+
embedding_dims=self.head_dim,
|
| 420 |
+
max_timescale=config.rope_theta,
|
| 421 |
+
dtype=compute_dtype,
|
| 422 |
+
)
|
| 423 |
+
|
| 424 |
+
self.is_fused_qkv = False
|
| 425 |
+
|
| 426 |
+
def get_linear_weight(self, dense: DenseGeneral):
|
| 427 |
+
W_dg = dense.weight.data
|
| 428 |
+
|
| 429 |
+
out_features = 1
|
| 430 |
+
input_features = 1
|
| 431 |
+
for dim in dense.out_features:
|
| 432 |
+
out_features *= dim
|
| 433 |
+
for dim in dense.in_shapes:
|
| 434 |
+
input_features *= dim
|
| 435 |
+
|
| 436 |
+
W_dg_reshaped_for_linear_T = W_dg.reshape(input_features, out_features)
|
| 437 |
+
linear_weight = W_dg_reshaped_for_linear_T.transpose(0, 1).contiguous()
|
| 438 |
+
return linear_weight
|
| 439 |
+
|
| 440 |
+
def patch_fused_qkv(self):
|
| 441 |
+
q_proj_weight = self.get_linear_weight(self.q_proj)
|
| 442 |
+
k_proj_weight = self.get_linear_weight(self.k_proj)
|
| 443 |
+
v_proj_weight = self.get_linear_weight(self.v_proj)
|
| 444 |
+
|
| 445 |
+
self.qkv = FusedQKV(
|
| 446 |
+
self.kv_embed_dim,
|
| 447 |
+
(
|
| 448 |
+
self.num_query_heads * self.head_dim
|
| 449 |
+
+ 2 * (self.num_kv_heads * self.head_dim)
|
| 450 |
+
),
|
| 451 |
+
bias=False,
|
| 452 |
+
num_q_heads=self.num_query_heads,
|
| 453 |
+
q_head_dim=self.head_dim,
|
| 454 |
+
num_kv_heads=self.num_kv_heads,
|
| 455 |
+
kv_head_dim=self.head_dim,
|
| 456 |
+
)
|
| 457 |
+
self.qkv.linear.weight.data = torch.cat(
|
| 458 |
+
[q_proj_weight, k_proj_weight, v_proj_weight], dim=0
|
| 459 |
+
)
|
| 460 |
+
|
| 461 |
+
# print(f"qkv.weight.shape: {self.qkv.linear.weight.shape}")
|
| 462 |
+
self.is_fused_qkv = True
|
| 463 |
+
|
| 464 |
+
def forward(
|
| 465 |
+
self,
|
| 466 |
+
X: torch.Tensor, # (B, T, D) T = 1 in AR generation
|
| 467 |
+
q_positions: torch.Tensor, # (B, T)
|
| 468 |
+
kv_positions: torch.Tensor | None = None, # (B, S)
|
| 469 |
+
attn_mask: torch.Tensor
|
| 470 |
+
| None = None, # None in Decoder Self Attention, Valid mask in Others
|
| 471 |
+
cache: KVCache | None = None, # None in Encoder, KVCache in Decoder
|
| 472 |
+
prefill: bool = False,
|
| 473 |
+
is_causal: bool = False,
|
| 474 |
+
current_idx: torch.Tensor | None = None,
|
| 475 |
+
) -> tuple[torch.Tensor, tuple[torch.Tensor, torch.Tensor] | None]:
|
| 476 |
+
"""
|
| 477 |
+
Performs attention calculation with optional KV caching.
|
| 478 |
+
Args:
|
| 479 |
+
Xq: Query tensor (B, T, D). T=1 during single-step decoding.
|
| 480 |
+
Xkv: Key/Value source tensor (B, S, E). S=1 during single-step decoding for self-attn.
|
| 481 |
+
q_positions: Positions for queries (B, T).
|
| 482 |
+
kv_positions: Positions for keys/values (B, S). If None, uses q_positions.
|
| 483 |
+
attn_mask: Attention mask.
|
| 484 |
+
cache: KVCache.
|
| 485 |
+
prefill: If True, use prefill mode.
|
| 486 |
+
Returns:
|
| 487 |
+
A tuple containing:
|
| 488 |
+
- output: The attention output tensor (B, T, output_dim).
|
| 489 |
+
- present_kv: The K/V state to be cached for the next step ((B, N, S_new, H), (B, N, S_new, H)). For self-attn, S_new = S_past + S. For cross-attn, S_new = S_kv.
|
| 490 |
+
"""
|
| 491 |
+
if kv_positions is None:
|
| 492 |
+
kv_positions = q_positions
|
| 493 |
+
|
| 494 |
+
original_dtype = X.dtype
|
| 495 |
+
|
| 496 |
+
if self.is_fused_qkv:
|
| 497 |
+
Xq_BxTxNxH, Xk_BxSxKxH, Xv_BxSxKxH = self.qkv(X)
|
| 498 |
+
else:
|
| 499 |
+
Xq_BxTxNxH = self.q_proj(X)
|
| 500 |
+
Xk_BxSxKxH = self.k_proj(X)
|
| 501 |
+
Xv_BxSxKxH = self.v_proj(X)
|
| 502 |
+
|
| 503 |
+
position = q_positions.unsqueeze(-1).unsqueeze(-1)
|
| 504 |
+
sinusoid_inp = position / self.rotary_emb.timescale
|
| 505 |
+
sin = torch.sin(sinusoid_inp)
|
| 506 |
+
cos = torch.cos(sinusoid_inp)
|
| 507 |
+
|
| 508 |
+
Xq_BxTxNxH = self.rotary_emb.apply_rope(Xq_BxTxNxH, sin, cos)
|
| 509 |
+
Xk_BxSxKxH = self.rotary_emb.apply_rope(Xk_BxSxKxH, sin, cos)
|
| 510 |
+
|
| 511 |
+
Xq_BxNxTxH = Xq_BxTxNxH.transpose(1, 2)
|
| 512 |
+
|
| 513 |
+
attn_k: torch.Tensor | None = None
|
| 514 |
+
attn_v: torch.Tensor | None = None
|
| 515 |
+
|
| 516 |
+
Xk_BxKxSxH = Xk_BxSxKxH.transpose(1, 2) # (B, K, S, H)
|
| 517 |
+
Xv_BxKxSxH = Xv_BxSxKxH.transpose(1, 2) # (B, K, S, H)
|
| 518 |
+
|
| 519 |
+
if cache is None:
|
| 520 |
+
attn_k = Xk_BxKxSxH
|
| 521 |
+
attn_v = Xv_BxKxSxH
|
| 522 |
+
elif prefill:
|
| 523 |
+
attn_k, attn_v = Xk_BxKxSxH, Xv_BxKxSxH
|
| 524 |
+
cache.prefill(attn_k, attn_v)
|
| 525 |
+
else:
|
| 526 |
+
attn_k, attn_v = cache.update(Xk_BxKxSxH, Xv_BxKxSxH, current_idx)
|
| 527 |
+
|
| 528 |
+
# Use custom attention for MPS backend, otherwise use optimized PyTorch function
|
| 529 |
+
is_mps = Xv_BxSxKxH.device.type == "mps" and torch.backends.mps.is_available()
|
| 530 |
+
if is_mps:
|
| 531 |
+
attn_output = custom_scaled_dot_product_attention(
|
| 532 |
+
query=Xq_BxNxTxH,
|
| 533 |
+
key=attn_k,
|
| 534 |
+
value=attn_v,
|
| 535 |
+
attn_mask=attn_mask if not is_causal else None,
|
| 536 |
+
scale=1.0,
|
| 537 |
+
is_causal=is_causal,
|
| 538 |
+
num_gqa_groups=self.num_gqa_groups,
|
| 539 |
+
)
|
| 540 |
+
else:
|
| 541 |
+
attn_output = F.scaled_dot_product_attention(
|
| 542 |
+
Xq_BxNxTxH,
|
| 543 |
+
attn_k,
|
| 544 |
+
attn_v,
|
| 545 |
+
attn_mask=attn_mask if not is_causal else None,
|
| 546 |
+
scale=1.0,
|
| 547 |
+
enable_gqa=self.num_gqa_groups > 1,
|
| 548 |
+
is_causal=is_causal,
|
| 549 |
+
)
|
| 550 |
+
|
| 551 |
attn_output = attn_output.transpose(1, 2).contiguous() # (B, T, N, H)
|
| 552 |
output = self.o_proj(attn_output)
|
| 553 |
|
|
|
|
| 560 |
def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
|
| 561 |
super().__init__()
|
| 562 |
self.config = config
|
| 563 |
+
enc_config = config.encoder_config
|
| 564 |
+
embed_dim = enc_config.hidden_size
|
| 565 |
+
self.compute_dtype = compute_dtype
|
| 566 |
|
| 567 |
self.pre_sa_norm = RMSNorm(
|
| 568 |
embed_dim,
|
| 569 |
+
eps=enc_config.norm_eps,
|
| 570 |
dtype=torch.float32,
|
| 571 |
)
|
| 572 |
+
self.self_attention = SelfAttention(
|
| 573 |
+
enc_config,
|
| 574 |
q_embed_dim=embed_dim,
|
| 575 |
kv_embed_dim=embed_dim,
|
| 576 |
+
num_query_heads=enc_config.num_attention_heads,
|
| 577 |
+
num_kv_heads=enc_config.num_key_value_heads,
|
| 578 |
head_dim=enc_config.head_dim,
|
| 579 |
compute_dtype=compute_dtype,
|
|
|
|
| 580 |
out_embed_dim=embed_dim,
|
| 581 |
)
|
| 582 |
self.post_sa_norm = RMSNorm(
|
| 583 |
embed_dim,
|
| 584 |
+
eps=enc_config.norm_eps,
|
| 585 |
dtype=torch.float32,
|
| 586 |
)
|
| 587 |
self.mlp = MlpBlock(
|
| 588 |
embed_dim=embed_dim,
|
| 589 |
+
intermediate_dim=enc_config.intermediate_size,
|
| 590 |
compute_dtype=compute_dtype,
|
| 591 |
)
|
| 592 |
|
|
|
|
| 596 |
state: EncoderInferenceState,
|
| 597 |
) -> torch.Tensor:
|
| 598 |
residual = x
|
| 599 |
+
x_norm = self.pre_sa_norm(x).to(self.compute_dtype)
|
| 600 |
+
|
| 601 |
sa_out = self.self_attention(
|
| 602 |
+
X=x_norm,
|
|
|
|
| 603 |
q_positions=state.positions,
|
| 604 |
kv_positions=state.positions,
|
| 605 |
attn_mask=state.attn_mask,
|
|
|
|
| 607 |
x = residual + sa_out
|
| 608 |
|
| 609 |
residual = x
|
| 610 |
+
x_norm = self.post_sa_norm(x).to(self.compute_dtype)
|
| 611 |
mlp_out = self.mlp(x_norm)
|
| 612 |
x = residual + mlp_out
|
| 613 |
|
|
|
|
| 620 |
def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
|
| 621 |
super().__init__()
|
| 622 |
self.config = config
|
| 623 |
+
enc_config = config.encoder_config
|
| 624 |
+
self.compute_dtype = compute_dtype
|
| 625 |
|
| 626 |
self.embedding = nn.Embedding(
|
| 627 |
+
enc_config.vocab_size,
|
| 628 |
+
enc_config.hidden_size,
|
| 629 |
dtype=compute_dtype,
|
| 630 |
)
|
| 631 |
self.layers = nn.ModuleList(
|
| 632 |
+
[
|
| 633 |
+
EncoderLayer(config, compute_dtype)
|
| 634 |
+
for _ in range(enc_config.num_hidden_layers)
|
| 635 |
+
]
|
| 636 |
)
|
| 637 |
self.norm = RMSNorm(
|
| 638 |
+
enc_config.hidden_size,
|
| 639 |
+
eps=enc_config.norm_eps,
|
| 640 |
dtype=torch.float32,
|
| 641 |
)
|
| 642 |
|
|
|
|
| 650 |
for layer in self.layers:
|
| 651 |
x = layer(x, state)
|
| 652 |
|
| 653 |
+
x = self.norm(x).to(self.compute_dtype)
|
| 654 |
return x
|
| 655 |
|
| 656 |
|
|
|
|
| 660 |
def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
|
| 661 |
super().__init__()
|
| 662 |
self.config = config
|
| 663 |
+
dec_config = config.decoder_config
|
| 664 |
+
enc_config = config.encoder_config
|
| 665 |
+
dec_embed_dim = dec_config.hidden_size
|
| 666 |
+
enc_embed_dim = enc_config.hidden_size
|
| 667 |
+
self.compute_dtype = compute_dtype
|
| 668 |
|
| 669 |
# Norms
|
| 670 |
self.pre_sa_norm = RMSNorm(
|
| 671 |
dec_embed_dim,
|
| 672 |
+
eps=dec_config.norm_eps,
|
| 673 |
dtype=torch.float32,
|
| 674 |
)
|
| 675 |
self.pre_ca_norm = RMSNorm(
|
| 676 |
dec_embed_dim,
|
| 677 |
+
eps=dec_config.norm_eps,
|
| 678 |
dtype=torch.float32,
|
| 679 |
)
|
| 680 |
self.pre_mlp_norm = RMSNorm(
|
| 681 |
dec_embed_dim,
|
| 682 |
+
eps=dec_config.norm_eps,
|
| 683 |
dtype=torch.float32,
|
| 684 |
)
|
| 685 |
|
| 686 |
# Self-Attention (GQA) with Causal Masking
|
| 687 |
+
self.self_attention = SelfAttention(
|
| 688 |
+
dec_config,
|
| 689 |
q_embed_dim=dec_embed_dim,
|
| 690 |
kv_embed_dim=dec_embed_dim,
|
| 691 |
+
num_query_heads=dec_config.num_attention_heads,
|
| 692 |
+
num_kv_heads=dec_config.num_key_value_heads,
|
| 693 |
+
head_dim=dec_config.head_dim,
|
| 694 |
compute_dtype=compute_dtype,
|
|
|
|
| 695 |
out_embed_dim=dec_embed_dim,
|
| 696 |
)
|
| 697 |
# Cross-Attention (MHA)
|
| 698 |
+
self.cross_attention = CrossAttention(
|
| 699 |
+
dec_config,
|
| 700 |
q_embed_dim=dec_embed_dim,
|
| 701 |
kv_embed_dim=enc_embed_dim, # Note kv_embed_dim
|
| 702 |
+
num_query_heads=dec_config.cross_num_attention_heads,
|
| 703 |
+
num_kv_heads=dec_config.cross_num_key_value_heads,
|
| 704 |
head_dim=dec_config.cross_head_dim,
|
| 705 |
compute_dtype=compute_dtype,
|
|
|
|
| 706 |
out_embed_dim=dec_embed_dim,
|
| 707 |
)
|
| 708 |
# MLP
|
| 709 |
self.mlp = MlpBlock(
|
| 710 |
embed_dim=dec_embed_dim,
|
| 711 |
+
intermediate_dim=dec_config.intermediate_size,
|
| 712 |
compute_dtype=compute_dtype,
|
| 713 |
)
|
| 714 |
|
|
|
|
| 719 |
self_attn_cache: KVCache | None = None,
|
| 720 |
cross_attn_cache: KVCache | None = None,
|
| 721 |
prefill: bool = False,
|
| 722 |
+
current_idx: int = 0,
|
| 723 |
) -> torch.Tensor:
|
| 724 |
residual = x
|
| 725 |
+
x_norm = self.pre_sa_norm(x).to(self.compute_dtype)
|
| 726 |
+
|
| 727 |
+
self_attn_mask = state.casual_attn_mask[None, None, current_idx]
|
| 728 |
|
| 729 |
sa_out = self.self_attention(
|
| 730 |
+
X=x_norm, # (2, 1, D)
|
|
|
|
| 731 |
q_positions=state.dec_positions, # (2, 1)
|
| 732 |
kv_positions=state.dec_positions, # (2, 1)
|
| 733 |
+
attn_mask=self_attn_mask,
|
| 734 |
cache=self_attn_cache,
|
| 735 |
prefill=prefill,
|
| 736 |
is_causal=prefill,
|
| 737 |
+
current_idx=current_idx,
|
| 738 |
)
|
| 739 |
|
| 740 |
x = residual + sa_out
|
| 741 |
|
| 742 |
residual = x
|
| 743 |
+
x_norm = self.pre_ca_norm(x).to(self.compute_dtype)
|
| 744 |
ca_out = self.cross_attention(
|
| 745 |
Xq=x_norm,
|
|
|
|
| 746 |
q_positions=state.dec_positions,
|
| 747 |
kv_positions=state.enc_positions,
|
| 748 |
+
attn_mask=state.cross_attn_mask,
|
| 749 |
cache=cross_attn_cache,
|
| 750 |
)
|
| 751 |
x = residual + ca_out
|
| 752 |
|
| 753 |
residual = x
|
| 754 |
+
x_norm = self.pre_mlp_norm(x).to(self.compute_dtype)
|
| 755 |
mlp_out = self.mlp(x_norm)
|
| 756 |
x = residual + mlp_out
|
| 757 |
|
|
|
|
| 764 |
def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
|
| 765 |
super().__init__()
|
| 766 |
self.config = config
|
| 767 |
+
dec_config = config.decoder_config
|
| 768 |
+
self.num_channels = dec_config.num_channels
|
| 769 |
+
self.num_layers = dec_config.num_hidden_layers
|
|
|
|
|
|
|
| 770 |
|
| 771 |
self.embeddings = nn.ModuleList(
|
| 772 |
[
|
| 773 |
nn.Embedding(
|
| 774 |
+
dec_config.vocab_size, dec_config.hidden_size, dtype=compute_dtype
|
| 775 |
)
|
| 776 |
for _ in range(self.num_channels)
|
| 777 |
]
|
|
|
|
| 784 |
)
|
| 785 |
|
| 786 |
self.norm = RMSNorm(
|
| 787 |
+
dec_config.hidden_size,
|
| 788 |
+
eps=dec_config.norm_eps,
|
| 789 |
dtype=torch.float32,
|
| 790 |
)
|
| 791 |
|
| 792 |
self.logits_dense = DenseGeneral(
|
| 793 |
+
in_shapes=(dec_config.hidden_size,),
|
| 794 |
+
out_features=(self.num_channels, dec_config.vocab_size),
|
| 795 |
axis=(-1,),
|
| 796 |
weight_dtype=compute_dtype,
|
| 797 |
)
|
|
|
|
| 799 |
def precompute_cross_attn_cache(
|
| 800 |
self,
|
| 801 |
enc_out: torch.Tensor, # (B, S, E)
|
|
|
|
| 802 |
) -> list[KVCache]:
|
| 803 |
"""
|
| 804 |
Computes the Key and Value tensors for cross-attention for each layer from the encoder output.
|
|
|
|
| 810 |
k_proj = cross_attn_module.k_proj(enc_out)
|
| 811 |
v_proj = cross_attn_module.v_proj(enc_out)
|
| 812 |
|
|
|
|
| 813 |
k = k_proj.transpose(1, 2)
|
| 814 |
v = v_proj.transpose(1, 2)
|
| 815 |
|
|
|
|
| 821 |
self,
|
| 822 |
tgt_ids_Bx1xC: torch.Tensor, # [B, 1, C]
|
| 823 |
state: DecoderInferenceState,
|
| 824 |
+
current_idx: int,
|
| 825 |
) -> torch.Tensor:
|
| 826 |
"""
|
| 827 |
Performs a single decoding step, managing KV caches layer by layer.
|
|
|
|
| 828 |
Returns:
|
| 829 |
A tuple containing:
|
| 830 |
- logits_Bx1xCV: The final output logits for the current step (B, 1, C*V), cast to float32.
|
|
|
|
| 844 |
state,
|
| 845 |
self_attn_cache=self_cache,
|
| 846 |
cross_attn_cache=cross_cache,
|
| 847 |
+
current_idx=current_idx,
|
| 848 |
)
|
| 849 |
|
| 850 |
x = self.norm(x)
|
|
|
|
| 857 |
) -> torch.Tensor:
|
| 858 |
"""
|
| 859 |
Forward pass for the Decoder stack, managing KV caches.
|
|
|
|
| 860 |
Args:
|
| 861 |
tgt_ids_BxTxC: Target token IDs (B, T, C).
|
| 862 |
encoder_out: Output from the encoder (B, S, E).
|
|
|
|
| 870 |
precomputed_cross_attn_kv: A single tuple containing the pre-computed K/V cache
|
| 871 |
derived from `encoder_out`. This is passed identically
|
| 872 |
to all layers.
|
|
|
|
| 873 |
Returns:
|
| 874 |
A tuple containing:
|
| 875 |
- logits: The final output logits (B, T, C * V), cast to float32.
|
|
|
|
| 904 |
return logits_BxTxCxV.to(torch.float32)
|
| 905 |
|
| 906 |
|
| 907 |
+
class DiaModel(
|
| 908 |
+
nn.Module,
|
| 909 |
+
PyTorchModelHubMixin,
|
| 910 |
+
repo_url="https://github.com/nari-labs/dia",
|
| 911 |
+
pipeline_tag="text-to-speech",
|
| 912 |
+
license="apache-2.0",
|
| 913 |
+
coders={
|
| 914 |
+
DiaConfig: (
|
| 915 |
+
lambda x: x.model_dump(),
|
| 916 |
+
lambda data: DiaConfig.model_validate(data),
|
| 917 |
+
),
|
| 918 |
+
},
|
| 919 |
+
):
|
| 920 |
"""PyTorch Dia Model using DenseGeneral."""
|
| 921 |
|
| 922 |
def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
|
dia/model.py
CHANGED
|
@@ -1,17 +1,16 @@
|
|
| 1 |
import time
|
| 2 |
from enum import Enum
|
|
|
|
| 3 |
|
| 4 |
-
import dac
|
| 5 |
import numpy as np
|
| 6 |
import torch
|
|
|
|
| 7 |
import torchaudio
|
| 8 |
-
from huggingface_hub import hf_hub_download
|
| 9 |
|
| 10 |
from .audio import (
|
| 11 |
apply_audio_delay,
|
| 12 |
build_delay_indices,
|
| 13 |
build_revert_indices,
|
| 14 |
-
decode,
|
| 15 |
revert_audio_delay,
|
| 16 |
)
|
| 17 |
from .config import DiaConfig
|
|
@@ -20,6 +19,7 @@ from .state import DecoderInferenceState, DecoderOutput, EncoderInferenceState
|
|
| 20 |
|
| 21 |
|
| 22 |
DEFAULT_SAMPLE_RATE = 44100
|
|
|
|
| 23 |
|
| 24 |
|
| 25 |
def _get_default_device():
|
|
@@ -34,16 +34,29 @@ def _sample_next_token(
|
|
| 34 |
logits_BCxV: torch.Tensor,
|
| 35 |
temperature: float,
|
| 36 |
top_p: float,
|
| 37 |
-
|
|
|
|
| 38 |
) -> torch.Tensor:
|
| 39 |
if temperature == 0.0:
|
| 40 |
return torch.argmax(logits_BCxV, dim=-1)
|
| 41 |
|
| 42 |
logits_BCxV = logits_BCxV / temperature
|
| 43 |
-
|
| 44 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 45 |
mask = torch.ones_like(logits_BCxV, dtype=torch.bool)
|
| 46 |
-
mask.
|
| 47 |
logits_BCxV = logits_BCxV.masked_fill(mask, -torch.inf)
|
| 48 |
|
| 49 |
if top_p < 1.0:
|
|
@@ -54,13 +67,15 @@ def _sample_next_token(
|
|
| 54 |
cumulative_probs_BCxV = torch.cumsum(sorted_probs_BCxV, dim=-1)
|
| 55 |
|
| 56 |
sorted_indices_to_remove_BCxV = cumulative_probs_BCxV > top_p
|
| 57 |
-
sorted_indices_to_remove_BCxV
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
sorted_indices_to_remove_BCxV[..., 0] =
|
|
|
|
|
|
|
| 61 |
|
| 62 |
indices_to_remove_BCxV = torch.zeros_like(sorted_indices_to_remove_BCxV)
|
| 63 |
-
indices_to_remove_BCxV.
|
| 64 |
dim=-1, index=sorted_indices_BCxV, src=sorted_indices_to_remove_BCxV
|
| 65 |
)
|
| 66 |
logits_BCxV = logits_BCxV.masked_fill(indices_to_remove_BCxV, -torch.inf)
|
|
@@ -94,12 +109,15 @@ class Dia:
|
|
| 94 |
config: DiaConfig,
|
| 95 |
compute_dtype: str | ComputeDtype = ComputeDtype.FLOAT32,
|
| 96 |
device: torch.device | None = None,
|
|
|
|
| 97 |
):
|
| 98 |
"""Initializes the Dia model.
|
| 99 |
|
| 100 |
Args:
|
| 101 |
config: The configuration object for the model.
|
|
|
|
| 102 |
device: The device to load the model onto. If None, will automatically select the best available device.
|
|
|
|
| 103 |
|
| 104 |
Raises:
|
| 105 |
RuntimeError: If there is an error loading the DAC model.
|
|
@@ -110,8 +128,16 @@ class Dia:
|
|
| 110 |
if isinstance(compute_dtype, str):
|
| 111 |
compute_dtype = ComputeDtype(compute_dtype)
|
| 112 |
self.compute_dtype = compute_dtype.to_dtype()
|
| 113 |
-
self.model = DiaModel(config, self.compute_dtype)
|
| 114 |
self.dac_model = None
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 115 |
|
| 116 |
@classmethod
|
| 117 |
def from_local(
|
|
@@ -120,13 +146,16 @@ class Dia:
|
|
| 120 |
checkpoint_path: str,
|
| 121 |
compute_dtype: str | ComputeDtype = ComputeDtype.FLOAT32,
|
| 122 |
device: torch.device | None = None,
|
|
|
|
| 123 |
) -> "Dia":
|
| 124 |
"""Loads the Dia model from local configuration and checkpoint files.
|
| 125 |
|
| 126 |
Args:
|
| 127 |
config_path: Path to the configuration JSON file.
|
| 128 |
checkpoint_path: Path to the model checkpoint (.pth) file.
|
|
|
|
| 129 |
device: The device to load the model onto. If None, will automatically select the best available device.
|
|
|
|
| 130 |
|
| 131 |
Returns:
|
| 132 |
An instance of the Dia model loaded with weights and set to eval mode.
|
|
@@ -139,7 +168,7 @@ class Dia:
|
|
| 139 |
if config is None:
|
| 140 |
raise FileNotFoundError(f"Config file not found at {config_path}")
|
| 141 |
|
| 142 |
-
dia = cls(config, compute_dtype, device)
|
| 143 |
|
| 144 |
try:
|
| 145 |
state_dict = torch.load(checkpoint_path, map_location=dia.device)
|
|
@@ -153,15 +182,17 @@ class Dia:
|
|
| 153 |
|
| 154 |
dia.model.to(dia.device)
|
| 155 |
dia.model.eval()
|
| 156 |
-
|
|
|
|
| 157 |
return dia
|
| 158 |
|
| 159 |
@classmethod
|
| 160 |
def from_pretrained(
|
| 161 |
cls,
|
| 162 |
-
model_name: str = "nari-labs/Dia-1.6B",
|
| 163 |
compute_dtype: str | ComputeDtype = ComputeDtype.FLOAT32,
|
| 164 |
device: torch.device | None = None,
|
|
|
|
| 165 |
) -> "Dia":
|
| 166 |
"""Loads the Dia model from a Hugging Face Hub repository.
|
| 167 |
|
|
@@ -169,8 +200,10 @@ class Dia:
|
|
| 169 |
repository ID and then loads the model.
|
| 170 |
|
| 171 |
Args:
|
| 172 |
-
model_name: The Hugging Face Hub repository ID (e.g., "
|
|
|
|
| 173 |
device: The device to load the model onto. If None, will automatically select the best available device.
|
|
|
|
| 174 |
|
| 175 |
Returns:
|
| 176 |
An instance of the Dia model loaded with weights and set to eval mode.
|
|
@@ -179,110 +212,192 @@ class Dia:
|
|
| 179 |
FileNotFoundError: If config or checkpoint download/loading fails.
|
| 180 |
RuntimeError: If there is an error loading the checkpoint.
|
| 181 |
"""
|
| 182 |
-
|
| 183 |
-
|
| 184 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 185 |
|
| 186 |
def _load_dac_model(self):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 187 |
try:
|
| 188 |
dac_model_path = dac.utils.download()
|
| 189 |
dac_model = dac.DAC.load(dac_model_path).to(self.device)
|
|
|
|
| 190 |
except Exception as e:
|
| 191 |
raise RuntimeError("Failed to load DAC model") from e
|
| 192 |
self.dac_model = dac_model
|
| 193 |
|
| 194 |
-
def
|
| 195 |
-
"""Encodes text
|
| 196 |
-
|
| 197 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 198 |
|
| 199 |
byte_text = text.encode("utf-8")
|
|
|
|
|
|
|
| 200 |
replaced_bytes = byte_text.replace(b"[S1]", b"\x01").replace(b"[S2]", b"\x02")
|
| 201 |
text_tokens = list(replaced_bytes)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 202 |
|
| 203 |
-
|
| 204 |
-
|
| 205 |
-
|
| 206 |
-
|
| 207 |
-
|
| 208 |
-
|
| 209 |
-
|
| 210 |
-
|
| 211 |
-
|
| 212 |
-
|
| 213 |
-
|
| 214 |
-
|
| 215 |
-
|
| 216 |
-
|
| 217 |
-
|
| 218 |
-
) # [1, S]
|
| 219 |
return src_tokens
|
| 220 |
|
| 221 |
def _prepare_audio_prompt(
|
| 222 |
-
self,
|
| 223 |
-
) -> tuple[torch.Tensor, int]:
|
| 224 |
-
|
| 225 |
-
audio_bos_value = self.config.data.audio_bos_value
|
| 226 |
-
audio_pad_value = self.config.data.audio_pad_value
|
| 227 |
-
delay_pattern = self.config.data.delay_pattern
|
| 228 |
-
max_delay_pattern = max(delay_pattern)
|
| 229 |
|
| 230 |
-
|
| 231 |
-
|
| 232 |
-
|
| 233 |
-
|
| 234 |
-
|
| 235 |
-
|
|
|
|
| 236 |
|
| 237 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 238 |
|
| 239 |
-
|
| 240 |
-
|
| 241 |
-
|
|
|
|
|
|
|
| 242 |
|
| 243 |
-
|
| 244 |
-
(
|
| 245 |
fill_value=-1,
|
| 246 |
dtype=torch.int,
|
| 247 |
device=self.device,
|
| 248 |
)
|
| 249 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 250 |
|
| 251 |
delay_precomp = build_delay_indices(
|
| 252 |
-
B=
|
| 253 |
-
T=
|
| 254 |
C=num_channels,
|
| 255 |
delay_pattern=delay_pattern,
|
| 256 |
)
|
| 257 |
|
| 258 |
-
|
| 259 |
-
audio_BxTxC=prefill
|
| 260 |
-
pad_value
|
| 261 |
bos_value=audio_bos_value,
|
| 262 |
precomp=delay_precomp,
|
| 263 |
-
)
|
| 264 |
|
| 265 |
-
return
|
| 266 |
|
| 267 |
def _prepare_generation(
|
| 268 |
-
self,
|
|
|
|
|
|
|
|
|
|
|
|
|
| 269 |
):
|
| 270 |
-
|
| 271 |
-
enc_input_uncond = torch.zeros_like(enc_input_cond)
|
| 272 |
-
enc_input = torch.cat([enc_input_uncond, enc_input_cond], dim=0)
|
| 273 |
|
| 274 |
-
|
| 275 |
-
|
| 276 |
-
|
|
|
|
| 277 |
|
| 278 |
-
|
| 279 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 280 |
|
| 281 |
enc_state = EncoderInferenceState.new(self.config, enc_input_cond)
|
| 282 |
encoder_out = self.model.encoder(enc_input, enc_state)
|
| 283 |
|
| 284 |
dec_cross_attn_cache = self.model.decoder.precompute_cross_attn_cache(
|
| 285 |
-
encoder_out
|
| 286 |
)
|
| 287 |
dec_state = DecoderInferenceState.new(
|
| 288 |
self.config,
|
|
@@ -290,15 +405,18 @@ class Dia:
|
|
| 290 |
encoder_out,
|
| 291 |
dec_cross_attn_cache,
|
| 292 |
self.compute_dtype,
|
|
|
|
| 293 |
)
|
| 294 |
-
|
| 295 |
-
dec_output.prefill(prefill, prefill_step)
|
| 296 |
|
| 297 |
-
|
|
|
|
|
|
|
|
|
|
| 298 |
if dec_step > 0:
|
| 299 |
dec_state.prepare_step(0, dec_step)
|
| 300 |
-
tokens_BxTxC = (
|
| 301 |
-
|
| 302 |
)
|
| 303 |
self.model.decoder.forward(tokens_BxTxC, dec_state)
|
| 304 |
|
|
@@ -311,43 +429,114 @@ class Dia:
|
|
| 311 |
cfg_scale: float,
|
| 312 |
temperature: float,
|
| 313 |
top_p: float,
|
| 314 |
-
|
|
|
|
| 315 |
) -> torch.Tensor:
|
| 316 |
-
|
| 317 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 318 |
|
| 319 |
-
|
| 320 |
-
|
| 321 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 322 |
|
| 323 |
-
|
| 324 |
-
|
| 325 |
-
|
| 326 |
|
| 327 |
-
|
| 328 |
-
|
| 329 |
temperature=temperature,
|
| 330 |
top_p=top_p,
|
| 331 |
-
|
|
|
|
| 332 |
)
|
| 333 |
-
return pred_C
|
| 334 |
|
| 335 |
-
|
| 336 |
-
|
| 337 |
-
|
| 338 |
-
|
| 339 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 340 |
max_delay_pattern = max(delay_pattern)
|
| 341 |
|
| 342 |
revert_precomp = build_revert_indices(
|
| 343 |
-
B=
|
| 344 |
T=seq_length,
|
| 345 |
C=num_channels,
|
| 346 |
delay_pattern=delay_pattern,
|
| 347 |
)
|
| 348 |
|
| 349 |
codebook = revert_audio_delay(
|
| 350 |
-
audio_BxTxC=generated_codes
|
| 351 |
pad_value=audio_pad_value,
|
| 352 |
precomp=revert_precomp,
|
| 353 |
T=seq_length,
|
|
@@ -358,20 +547,85 @@ class Dia:
|
|
| 358 |
invalid_mask = (codebook < min_valid_index) | (codebook > max_valid_index)
|
| 359 |
codebook[invalid_mask] = 0
|
| 360 |
|
| 361 |
-
|
| 362 |
|
| 363 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 364 |
|
| 365 |
def load_audio(self, audio_path: str) -> torch.Tensor:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 366 |
audio, sr = torchaudio.load(audio_path, channels_first=True) # C, T
|
| 367 |
if sr != DEFAULT_SAMPLE_RATE:
|
| 368 |
audio = torchaudio.functional.resample(audio, sr, DEFAULT_SAMPLE_RATE)
|
| 369 |
-
|
| 370 |
-
|
| 371 |
-
|
| 372 |
-
|
|
|
|
|
|
|
| 373 |
|
| 374 |
def save_audio(self, path: str, audio: np.ndarray):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 375 |
import soundfile as sf
|
| 376 |
|
| 377 |
sf.write(path, audio, DEFAULT_SAMPLE_RATE)
|
|
@@ -379,23 +633,63 @@ class Dia:
|
|
| 379 |
@torch.inference_mode()
|
| 380 |
def generate(
|
| 381 |
self,
|
| 382 |
-
text: str,
|
| 383 |
-
max_tokens: int
|
| 384 |
cfg_scale: float = 3.0,
|
| 385 |
-
temperature: float = 1.
|
| 386 |
top_p: float = 0.95,
|
| 387 |
use_torch_compile: bool = False,
|
| 388 |
-
cfg_filter_top_k: int =
|
| 389 |
-
audio_prompt: str | torch.Tensor | None
|
| 390 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 391 |
use_cfg_filter: bool | None = None,
|
| 392 |
verbose: bool = False,
|
| 393 |
-
) -> np.ndarray:
|
| 394 |
-
|
| 395 |
-
|
| 396 |
-
|
| 397 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 398 |
max_delay_pattern = max(delay_pattern)
|
|
|
|
|
|
|
|
|
|
| 399 |
self.model.eval()
|
| 400 |
|
| 401 |
if audio_prompt_path:
|
|
@@ -407,82 +701,179 @@ class Dia:
|
|
| 407 |
if verbose:
|
| 408 |
total_start_time = time.time()
|
| 409 |
|
| 410 |
-
|
| 411 |
-
|
| 412 |
-
|
| 413 |
-
|
| 414 |
-
|
| 415 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 416 |
|
| 417 |
-
if
|
| 418 |
-
|
| 419 |
else:
|
| 420 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 421 |
|
| 422 |
if verbose:
|
| 423 |
print("generate: starting generation loop")
|
| 424 |
if use_torch_compile:
|
| 425 |
print(
|
| 426 |
-
"generate:
|
| 427 |
)
|
| 428 |
start_time = time.time()
|
| 429 |
|
|
|
|
| 430 |
while dec_step < max_tokens:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 431 |
dec_state.prepare_step(dec_step)
|
| 432 |
-
tokens_Bx1xC = (
|
| 433 |
-
|
| 434 |
-
)
|
| 435 |
-
|
|
|
|
| 436 |
tokens_Bx1xC,
|
| 437 |
dec_state,
|
| 438 |
cfg_scale,
|
| 439 |
temperature,
|
| 440 |
top_p,
|
| 441 |
cfg_filter_top_k,
|
|
|
|
| 442 |
)
|
| 443 |
|
| 444 |
-
|
| 445 |
-
|
| 446 |
-
|
| 447 |
-
|
| 448 |
-
|
| 449 |
-
|
| 450 |
-
|
| 451 |
-
|
| 452 |
-
|
| 453 |
-
|
| 454 |
-
|
| 455 |
-
|
| 456 |
-
|
| 457 |
-
|
| 458 |
-
|
| 459 |
-
|
| 460 |
-
|
| 461 |
-
|
| 462 |
-
|
| 463 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 464 |
|
| 465 |
dec_step += 1
|
|
|
|
| 466 |
if verbose and dec_step % 86 == 0:
|
| 467 |
duration = time.time() - start_time
|
| 468 |
-
|
| 469 |
-
|
| 470 |
-
|
|
|
|
| 471 |
start_time = time.time()
|
| 472 |
|
| 473 |
-
|
| 474 |
-
|
| 475 |
-
return None
|
| 476 |
|
| 477 |
-
|
| 478 |
-
dec_output.prefill_step : dec_step + 1, :
|
| 479 |
-
]
|
| 480 |
|
| 481 |
-
|
| 482 |
-
|
| 483 |
-
|
| 484 |
-
|
| 485 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 486 |
)
|
| 487 |
|
| 488 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
import time
|
| 2 |
from enum import Enum
|
| 3 |
+
from typing import Callable
|
| 4 |
|
|
|
|
| 5 |
import numpy as np
|
| 6 |
import torch
|
| 7 |
+
import torch.nn.functional as F
|
| 8 |
import torchaudio
|
|
|
|
| 9 |
|
| 10 |
from .audio import (
|
| 11 |
apply_audio_delay,
|
| 12 |
build_delay_indices,
|
| 13 |
build_revert_indices,
|
|
|
|
| 14 |
revert_audio_delay,
|
| 15 |
)
|
| 16 |
from .config import DiaConfig
|
|
|
|
| 19 |
|
| 20 |
|
| 21 |
DEFAULT_SAMPLE_RATE = 44100
|
| 22 |
+
SAMPLE_RATE_RATIO = 512
|
| 23 |
|
| 24 |
|
| 25 |
def _get_default_device():
|
|
|
|
| 34 |
logits_BCxV: torch.Tensor,
|
| 35 |
temperature: float,
|
| 36 |
top_p: float,
|
| 37 |
+
top_k: int | None,
|
| 38 |
+
audio_eos_value: int,
|
| 39 |
) -> torch.Tensor:
|
| 40 |
if temperature == 0.0:
|
| 41 |
return torch.argmax(logits_BCxV, dim=-1)
|
| 42 |
|
| 43 |
logits_BCxV = logits_BCxV / temperature
|
| 44 |
+
|
| 45 |
+
if audio_eos_value is not None and audio_eos_value >= 0:
|
| 46 |
+
top_logit_indices_BC = torch.argmax(logits_BCxV, dim=-1)
|
| 47 |
+
eos_not_highest_mask_BC = top_logit_indices_BC != audio_eos_value
|
| 48 |
+
mask_eos_unless_highest_BCxV = torch.zeros_like(logits_BCxV, dtype=torch.bool)
|
| 49 |
+
mask_eos_unless_highest_BCxV[eos_not_highest_mask_BC, audio_eos_value] = True
|
| 50 |
+
logits_BCxV = logits_BCxV.masked_fill(mask_eos_unless_highest_BCxV, -torch.inf)
|
| 51 |
+
eos_highest_mask_BC = top_logit_indices_BC == audio_eos_value
|
| 52 |
+
mask_eos_highest_BCxV = torch.zeros_like(logits_BCxV, dtype=torch.bool)
|
| 53 |
+
mask_eos_highest_BCxV[eos_highest_mask_BC, :audio_eos_value] = True
|
| 54 |
+
logits_BCxV = logits_BCxV.masked_fill(mask_eos_highest_BCxV, -torch.inf)
|
| 55 |
+
|
| 56 |
+
if top_k is not None:
|
| 57 |
+
_, top_k_indices_BCxV = torch.topk(logits_BCxV, k=top_k, dim=-1)
|
| 58 |
mask = torch.ones_like(logits_BCxV, dtype=torch.bool)
|
| 59 |
+
mask = mask.scatter(dim=-1, index=top_k_indices_BCxV, value=False)
|
| 60 |
logits_BCxV = logits_BCxV.masked_fill(mask, -torch.inf)
|
| 61 |
|
| 62 |
if top_p < 1.0:
|
|
|
|
| 67 |
cumulative_probs_BCxV = torch.cumsum(sorted_probs_BCxV, dim=-1)
|
| 68 |
|
| 69 |
sorted_indices_to_remove_BCxV = cumulative_probs_BCxV > top_p
|
| 70 |
+
sorted_indices_to_remove_BCxV = torch.roll(
|
| 71 |
+
sorted_indices_to_remove_BCxV, shifts=1, dims=-1
|
| 72 |
+
)
|
| 73 |
+
sorted_indices_to_remove_BCxV[..., 0] = torch.zeros_like(
|
| 74 |
+
sorted_indices_to_remove_BCxV[..., 0]
|
| 75 |
+
)
|
| 76 |
|
| 77 |
indices_to_remove_BCxV = torch.zeros_like(sorted_indices_to_remove_BCxV)
|
| 78 |
+
indices_to_remove_BCxV = indices_to_remove_BCxV.scatter(
|
| 79 |
dim=-1, index=sorted_indices_BCxV, src=sorted_indices_to_remove_BCxV
|
| 80 |
)
|
| 81 |
logits_BCxV = logits_BCxV.masked_fill(indices_to_remove_BCxV, -torch.inf)
|
|
|
|
| 109 |
config: DiaConfig,
|
| 110 |
compute_dtype: str | ComputeDtype = ComputeDtype.FLOAT32,
|
| 111 |
device: torch.device | None = None,
|
| 112 |
+
load_dac: bool = True,
|
| 113 |
):
|
| 114 |
"""Initializes the Dia model.
|
| 115 |
|
| 116 |
Args:
|
| 117 |
config: The configuration object for the model.
|
| 118 |
+
compute_dtype: The computation dtype to use.
|
| 119 |
device: The device to load the model onto. If None, will automatically select the best available device.
|
| 120 |
+
load_dac: Whether to load the DAC model.
|
| 121 |
|
| 122 |
Raises:
|
| 123 |
RuntimeError: If there is an error loading the DAC model.
|
|
|
|
| 128 |
if isinstance(compute_dtype, str):
|
| 129 |
compute_dtype = ComputeDtype(compute_dtype)
|
| 130 |
self.compute_dtype = compute_dtype.to_dtype()
|
| 131 |
+
self.model: DiaModel = DiaModel(config, self.compute_dtype)
|
| 132 |
self.dac_model = None
|
| 133 |
+
self._compiled_step = None
|
| 134 |
+
self.load_dac = load_dac
|
| 135 |
+
|
| 136 |
+
if not self.load_dac:
|
| 137 |
+
print("Warning: DAC model will not be loaded. This is not recommended.")
|
| 138 |
+
|
| 139 |
+
if torch.cuda.is_available():
|
| 140 |
+
torch.backends.cuda.matmul.allow_tf32 = True
|
| 141 |
|
| 142 |
@classmethod
|
| 143 |
def from_local(
|
|
|
|
| 146 |
checkpoint_path: str,
|
| 147 |
compute_dtype: str | ComputeDtype = ComputeDtype.FLOAT32,
|
| 148 |
device: torch.device | None = None,
|
| 149 |
+
load_dac: bool = True,
|
| 150 |
) -> "Dia":
|
| 151 |
"""Loads the Dia model from local configuration and checkpoint files.
|
| 152 |
|
| 153 |
Args:
|
| 154 |
config_path: Path to the configuration JSON file.
|
| 155 |
checkpoint_path: Path to the model checkpoint (.pth) file.
|
| 156 |
+
compute_dtype: The computation dtype to use.
|
| 157 |
device: The device to load the model onto. If None, will automatically select the best available device.
|
| 158 |
+
load_dac: Whether to load the DAC model.
|
| 159 |
|
| 160 |
Returns:
|
| 161 |
An instance of the Dia model loaded with weights and set to eval mode.
|
|
|
|
| 168 |
if config is None:
|
| 169 |
raise FileNotFoundError(f"Config file not found at {config_path}")
|
| 170 |
|
| 171 |
+
dia = cls(config, compute_dtype, device, load_dac)
|
| 172 |
|
| 173 |
try:
|
| 174 |
state_dict = torch.load(checkpoint_path, map_location=dia.device)
|
|
|
|
| 182 |
|
| 183 |
dia.model.to(dia.device)
|
| 184 |
dia.model.eval()
|
| 185 |
+
if load_dac:
|
| 186 |
+
dia._load_dac_model()
|
| 187 |
return dia
|
| 188 |
|
| 189 |
@classmethod
|
| 190 |
def from_pretrained(
|
| 191 |
cls,
|
| 192 |
+
model_name: str = "nari-labs/Dia-1.6B-0626",
|
| 193 |
compute_dtype: str | ComputeDtype = ComputeDtype.FLOAT32,
|
| 194 |
device: torch.device | None = None,
|
| 195 |
+
load_dac: bool = True,
|
| 196 |
) -> "Dia":
|
| 197 |
"""Loads the Dia model from a Hugging Face Hub repository.
|
| 198 |
|
|
|
|
| 200 |
repository ID and then loads the model.
|
| 201 |
|
| 202 |
Args:
|
| 203 |
+
model_name: The Hugging Face Hub repository ID (e.g., "nari-labs/Dia-1.6B-0626").
|
| 204 |
+
compute_dtype: The computation dtype to use.
|
| 205 |
device: The device to load the model onto. If None, will automatically select the best available device.
|
| 206 |
+
load_dac: Whether to load the DAC model.
|
| 207 |
|
| 208 |
Returns:
|
| 209 |
An instance of the Dia model loaded with weights and set to eval mode.
|
|
|
|
| 212 |
FileNotFoundError: If config or checkpoint download/loading fails.
|
| 213 |
RuntimeError: If there is an error loading the checkpoint.
|
| 214 |
"""
|
| 215 |
+
if isinstance(compute_dtype, str):
|
| 216 |
+
compute_dtype = ComputeDtype(compute_dtype)
|
| 217 |
+
|
| 218 |
+
# Load model directly using DiaModel's from_pretrained which handles HF download
|
| 219 |
+
try:
|
| 220 |
+
loaded_model = DiaModel.from_pretrained(
|
| 221 |
+
model_name, compute_dtype=compute_dtype.to_dtype()
|
| 222 |
+
)
|
| 223 |
+
except Exception as e:
|
| 224 |
+
raise RuntimeError(
|
| 225 |
+
f"Error loading model from Hugging Face Hub ({model_name})"
|
| 226 |
+
) from e
|
| 227 |
+
|
| 228 |
+
config = loaded_model.config # Get config from the loaded model
|
| 229 |
+
dia = cls(config, compute_dtype, device, load_dac)
|
| 230 |
+
|
| 231 |
+
dia.model = loaded_model # Assign the already loaded model
|
| 232 |
+
dia.model.to(dia.device)
|
| 233 |
+
dia.model.eval()
|
| 234 |
+
if load_dac:
|
| 235 |
+
dia._load_dac_model()
|
| 236 |
+
return dia
|
| 237 |
|
| 238 |
def _load_dac_model(self):
|
| 239 |
+
"""Loads the Descript Audio Codec (DAC) model.
|
| 240 |
+
|
| 241 |
+
Downloads the DAC model if necessary and loads it onto the specified device.
|
| 242 |
+
Sets the DAC model to evaluation mode.
|
| 243 |
+
|
| 244 |
+
Raises:
|
| 245 |
+
RuntimeError: If downloading or loading the DAC model fails.
|
| 246 |
+
"""
|
| 247 |
+
import dac
|
| 248 |
+
|
| 249 |
try:
|
| 250 |
dac_model_path = dac.utils.download()
|
| 251 |
dac_model = dac.DAC.load(dac_model_path).to(self.device)
|
| 252 |
+
dac_model.eval() # Ensure DAC is in eval mode
|
| 253 |
except Exception as e:
|
| 254 |
raise RuntimeError("Failed to load DAC model") from e
|
| 255 |
self.dac_model = dac_model
|
| 256 |
|
| 257 |
+
def _encode_text(self, text: str) -> torch.Tensor:
|
| 258 |
+
"""Encodes the input text string into a tensor of token IDs using byte-level encoding.
|
| 259 |
+
|
| 260 |
+
Special tokens [S1] and [S2] are replaced by their byte values. The resulting
|
| 261 |
+
sequence is truncated to the maximum configured text length.
|
| 262 |
+
|
| 263 |
+
Args:
|
| 264 |
+
text: The input text string.
|
| 265 |
+
|
| 266 |
+
Returns:
|
| 267 |
+
A tensor containing the encoded byte token IDs.
|
| 268 |
+
"""
|
| 269 |
+
max_len = self.config.encoder_config.max_position_embeddings
|
| 270 |
|
| 271 |
byte_text = text.encode("utf-8")
|
| 272 |
+
# Replace special tokens with their byte values if needed by the specific tokenizer/config
|
| 273 |
+
# Assuming byte values 1 and 2 are correct placeholders based on original code
|
| 274 |
replaced_bytes = byte_text.replace(b"[S1]", b"\x01").replace(b"[S2]", b"\x02")
|
| 275 |
text_tokens = list(replaced_bytes)
|
| 276 |
+
return torch.tensor(
|
| 277 |
+
text_tokens[:max_len],
|
| 278 |
+
dtype=torch.long,
|
| 279 |
+
device=self.device,
|
| 280 |
+
)
|
| 281 |
|
| 282 |
+
def _pad_text_input(self, text_tokens: list[torch.Tensor]) -> torch.Tensor:
|
| 283 |
+
"""Pads the text input to the maximum length."""
|
| 284 |
+
text_pad_value = 0
|
| 285 |
+
max_len = self.config.encoder_config.max_position_embeddings
|
| 286 |
+
batch_size = len(text_tokens)
|
| 287 |
+
|
| 288 |
+
src_tokens = torch.full(
|
| 289 |
+
(batch_size, 1, max_len),
|
| 290 |
+
fill_value=text_pad_value,
|
| 291 |
+
dtype=torch.long,
|
| 292 |
+
device=self.device,
|
| 293 |
+
)
|
| 294 |
+
for i in range(batch_size):
|
| 295 |
+
current_len = len(text_tokens[i])
|
| 296 |
+
src_tokens[i, 0, :current_len] = text_tokens[i]
|
|
|
|
| 297 |
return src_tokens
|
| 298 |
|
| 299 |
def _prepare_audio_prompt(
|
| 300 |
+
self, audio_prompts: list[torch.Tensor | None]
|
| 301 |
+
) -> tuple[torch.Tensor, list[int]]:
|
| 302 |
+
"""Prepares the audio prompt tensor for the decoder.
|
|
|
|
|
|
|
|
|
|
|
|
|
| 303 |
|
| 304 |
+
Handles padding, adds the beginning-of-sequence (BOS) token, applies the
|
| 305 |
+
delay pattern, and determines the number of prefill steps for each item
|
| 306 |
+
in the batch.
|
| 307 |
+
|
| 308 |
+
Args:
|
| 309 |
+
audio_prompts: A list of audio prompt tensors (encoded DAC frames) or None.
|
| 310 |
+
Each tensor should have shape [T, C].
|
| 311 |
|
| 312 |
+
Returns:
|
| 313 |
+
A tuple containing:
|
| 314 |
+
- delayed_batch (torch.Tensor): The prepared audio prompt tensor with
|
| 315 |
+
delays applied, shape [B, T_max_padded, C].
|
| 316 |
+
- prefill_steps (list[int]): A list containing the number of valid
|
| 317 |
+
tokens (including BOS) for each prompt in the batch.
|
| 318 |
+
"""
|
| 319 |
+
num_channels = self.config.decoder_config.num_channels
|
| 320 |
+
audio_bos_value = self.config.bos_token_id
|
| 321 |
+
delay_pattern = self.config.delay_pattern
|
| 322 |
+
max_delay_pattern = max(delay_pattern)
|
| 323 |
+
batch_size = len(audio_prompts)
|
| 324 |
|
| 325 |
+
max_len = (
|
| 326 |
+
max(p.shape[0] if p is not None else 0 for p in audio_prompts)
|
| 327 |
+
+ max_delay_pattern
|
| 328 |
+
)
|
| 329 |
+
prefill_steps = []
|
| 330 |
|
| 331 |
+
prefill = torch.full(
|
| 332 |
+
(batch_size, max_len, num_channels),
|
| 333 |
fill_value=-1,
|
| 334 |
dtype=torch.int,
|
| 335 |
device=self.device,
|
| 336 |
)
|
| 337 |
+
|
| 338 |
+
prefill[:, 0, :] = audio_bos_value
|
| 339 |
+
|
| 340 |
+
for i in range(batch_size):
|
| 341 |
+
prompt = audio_prompts[i]
|
| 342 |
+
if prompt is not None:
|
| 343 |
+
prompt = prompt.to(device=self.device, dtype=torch.int)
|
| 344 |
+
prefill[i, 1 : prompt.shape[0] + 1, :] = prompt
|
| 345 |
+
prefill_steps.append(prompt.shape[0] + 1)
|
| 346 |
+
else:
|
| 347 |
+
prefill_steps.append(1)
|
| 348 |
|
| 349 |
delay_precomp = build_delay_indices(
|
| 350 |
+
B=batch_size,
|
| 351 |
+
T=max_len,
|
| 352 |
C=num_channels,
|
| 353 |
delay_pattern=delay_pattern,
|
| 354 |
)
|
| 355 |
|
| 356 |
+
delayed_batch = apply_audio_delay(
|
| 357 |
+
audio_BxTxC=prefill,
|
| 358 |
+
pad_value=-1,
|
| 359 |
bos_value=audio_bos_value,
|
| 360 |
precomp=delay_precomp,
|
| 361 |
+
)
|
| 362 |
|
| 363 |
+
return delayed_batch, prefill_steps
|
| 364 |
|
| 365 |
def _prepare_generation(
|
| 366 |
+
self,
|
| 367 |
+
text: torch.Tensor,
|
| 368 |
+
audio_prompts: list[torch.Tensor | None],
|
| 369 |
+
max_tokens: int | None = None,
|
| 370 |
+
attn_fn: Callable = F.scaled_dot_product_attention,
|
| 371 |
):
|
| 372 |
+
"""Initializes the model state for generation.
|
|
|
|
|
|
|
| 373 |
|
| 374 |
+
Encodes the text input (conditional and unconditional), prepares the
|
| 375 |
+
encoder and decoder states (including KV caches and cross-attention),
|
| 376 |
+
prepares the audio prompt, and performs the initial decoder prefill steps
|
| 377 |
+
based on the audio prompts.
|
| 378 |
|
| 379 |
+
Args:
|
| 380 |
+
text: The padded text input tensor, shape [B, 1, T_text].
|
| 381 |
+
audio_prompts: A list of prepared audio prompt tensors or None.
|
| 382 |
+
|
| 383 |
+
Returns:
|
| 384 |
+
A tuple containing:
|
| 385 |
+
- dec_state (DecoderInferenceState): The initialized decoder state.
|
| 386 |
+
- dec_output (DecoderOutput): The initialized decoder output manager,
|
| 387 |
+
containing the prefilled audio tokens.
|
| 388 |
+
"""
|
| 389 |
+
batch_size = text.shape[0]
|
| 390 |
+
|
| 391 |
+
enc_input_uncond = torch.zeros_like(text)
|
| 392 |
+
enc_input_cond = text
|
| 393 |
+
stacked_inputs = torch.stack([enc_input_uncond, enc_input_cond], dim=1)
|
| 394 |
+
enc_input = stacked_inputs.view(2 * batch_size, -1)
|
| 395 |
|
| 396 |
enc_state = EncoderInferenceState.new(self.config, enc_input_cond)
|
| 397 |
encoder_out = self.model.encoder(enc_input, enc_state)
|
| 398 |
|
| 399 |
dec_cross_attn_cache = self.model.decoder.precompute_cross_attn_cache(
|
| 400 |
+
encoder_out
|
| 401 |
)
|
| 402 |
dec_state = DecoderInferenceState.new(
|
| 403 |
self.config,
|
|
|
|
| 405 |
encoder_out,
|
| 406 |
dec_cross_attn_cache,
|
| 407 |
self.compute_dtype,
|
| 408 |
+
max_generation_length=max_tokens,
|
| 409 |
)
|
| 410 |
+
prefill, prefill_steps = self._prepare_audio_prompt(audio_prompts)
|
|
|
|
| 411 |
|
| 412 |
+
dec_output = DecoderOutput.new(batch_size, self.config, self.device)
|
| 413 |
+
dec_output.prefill(prefill, prefill_steps)
|
| 414 |
+
|
| 415 |
+
dec_step = min(prefill_steps) - 1
|
| 416 |
if dec_step > 0:
|
| 417 |
dec_state.prepare_step(0, dec_step)
|
| 418 |
+
tokens_BxTxC = dec_output.get_tokens_at(0, dec_step).repeat_interleave(
|
| 419 |
+
2, dim=0
|
| 420 |
)
|
| 421 |
self.model.decoder.forward(tokens_BxTxC, dec_state)
|
| 422 |
|
|
|
|
| 429 |
cfg_scale: float,
|
| 430 |
temperature: float,
|
| 431 |
top_p: float,
|
| 432 |
+
top_k: int,
|
| 433 |
+
current_idx: int,
|
| 434 |
) -> torch.Tensor:
|
| 435 |
+
"""Performs a single step of the decoder inference.
|
| 436 |
+
|
| 437 |
+
Takes the tokens from the previous step, runs them through the decoder
|
| 438 |
+
(for both conditional and unconditional paths), applies classifier-free
|
| 439 |
+
guidance (CFG), samples the next token using temperature, top-p, and top-k
|
| 440 |
+
sampling, and applies constraints (e.g., preventing EOS in certain channels).
|
| 441 |
+
|
| 442 |
+
Args:
|
| 443 |
+
tokens_Bx1xC: The input tokens for the current step, shape [2*B, 1, C].
|
| 444 |
+
Repeated for CFG (unconditional and conditional).
|
| 445 |
+
dec_state: The current state of the decoder (KV caches, etc.).
|
| 446 |
+
cfg_scale: The scale factor for classifier-free guidance.
|
| 447 |
+
temperature: The temperature for sampling.
|
| 448 |
+
top_p: The cumulative probability threshold for top-p sampling.
|
| 449 |
+
top_k: The number of top logits to consider for top-k sampling.
|
| 450 |
+
current_idx: The current generation step index.
|
| 451 |
+
|
| 452 |
+
Returns:
|
| 453 |
+
torch.Tensor: The sampled next tokens for each item in the batch,
|
| 454 |
+
shape [B, C].
|
| 455 |
+
"""
|
| 456 |
+
B = tokens_Bx1xC.shape[0] // 2
|
| 457 |
+
|
| 458 |
+
audio_eos_value = self.config.eos_token_id
|
| 459 |
+
logits_Bx1xCxV = self.model.decoder.decode_step(
|
| 460 |
+
tokens_Bx1xC, dec_state, current_idx
|
| 461 |
+
)
|
| 462 |
+
|
| 463 |
+
logits_last_2BxCxV = logits_Bx1xCxV[:, -1]
|
| 464 |
+
logits_last_Bx2xCxV = logits_last_2BxCxV.view(
|
| 465 |
+
B, 2, *logits_last_2BxCxV.shape[1:]
|
| 466 |
+
)
|
| 467 |
|
| 468 |
+
uncond_logits_BxCxV = logits_last_Bx2xCxV[:, 0, :, :] # Shape [B, C, V]
|
| 469 |
+
cond_logits_BxCxV = logits_last_Bx2xCxV[:, 1, :, :] # Shape [B, C, V]
|
| 470 |
+
logits_BxCxV = cond_logits_BxCxV + cfg_scale * (
|
| 471 |
+
cond_logits_BxCxV - uncond_logits_BxCxV
|
| 472 |
+
)
|
| 473 |
+
|
| 474 |
+
_, top_k_indices_BxCxk = torch.topk(logits_BxCxV, k=top_k, dim=-1)
|
| 475 |
+
mask_BxCxV = torch.ones_like(logits_BxCxV, dtype=torch.bool)
|
| 476 |
+
mask_BxCxV = mask_BxCxV.scatter(dim=-1, index=top_k_indices_BxCxk, value=False)
|
| 477 |
+
logits_BxCxV = cond_logits_BxCxV.masked_fill(mask_BxCxV, -torch.inf)
|
| 478 |
+
|
| 479 |
+
logits_BxCxV[:, :, audio_eos_value + 1 :] = torch.full_like(
|
| 480 |
+
logits_BxCxV[:, :, audio_eos_value + 1 :],
|
| 481 |
+
fill_value=-torch.inf,
|
| 482 |
+
)
|
| 483 |
+
logits_BxCxV[:, 1:, audio_eos_value:] = torch.full_like(
|
| 484 |
+
logits_BxCxV[:, 1:, audio_eos_value:],
|
| 485 |
+
fill_value=-torch.inf,
|
| 486 |
+
)
|
| 487 |
|
| 488 |
+
flat_logits_BCxV = logits_BxCxV.view(
|
| 489 |
+
B * self.config.decoder_config.num_channels, -1
|
| 490 |
+
)
|
| 491 |
|
| 492 |
+
pred_BC = _sample_next_token(
|
| 493 |
+
flat_logits_BCxV.float(),
|
| 494 |
temperature=temperature,
|
| 495 |
top_p=top_p,
|
| 496 |
+
top_k=top_k,
|
| 497 |
+
audio_eos_value=audio_eos_value,
|
| 498 |
)
|
|
|
|
| 499 |
|
| 500 |
+
pred_BxC = pred_BC.view(B, self.config.decoder_config.num_channels)
|
| 501 |
+
return pred_BxC
|
| 502 |
+
|
| 503 |
+
def _generate_output(
|
| 504 |
+
self, generated_codes: torch.Tensor, lengths_Bx: torch.Tensor
|
| 505 |
+
) -> list[np.ndarray]:
|
| 506 |
+
"""Converts generated delayed codes into audio waveforms.
|
| 507 |
+
|
| 508 |
+
Reverts the delay pattern applied during generation, decodes the resulting
|
| 509 |
+
codebook using the DAC model (if loaded), and returns a list of audio
|
| 510 |
+
waveforms as NumPy arrays. If DAC is not loaded, returns the raw codebook indices.
|
| 511 |
+
|
| 512 |
+
Args:
|
| 513 |
+
generated_codes: The tensor of generated audio codes with delays,
|
| 514 |
+
shape [B, T_gen, C].
|
| 515 |
+
lengths_Bx: A tensor containing the valid length of generated codes
|
| 516 |
+
(excluding padding and BOS/EOS markers) for each item
|
| 517 |
+
in the batch, shape [B].
|
| 518 |
+
|
| 519 |
+
Returns:
|
| 520 |
+
A list of NumPy arrays, where each array represents the generated audio
|
| 521 |
+
waveform for one item in the batch. If DAC is not loaded, returns the
|
| 522 |
+
raw, reverted codebook indices as NumPy arrays.
|
| 523 |
+
"""
|
| 524 |
+
num_channels = self.config.decoder_config.num_channels
|
| 525 |
+
batch_size = generated_codes.shape[0]
|
| 526 |
+
seq_length = generated_codes.shape[1]
|
| 527 |
+
delay_pattern = self.config.delay_pattern
|
| 528 |
+
audio_pad_value = self.config.pad_token_id
|
| 529 |
max_delay_pattern = max(delay_pattern)
|
| 530 |
|
| 531 |
revert_precomp = build_revert_indices(
|
| 532 |
+
B=batch_size,
|
| 533 |
T=seq_length,
|
| 534 |
C=num_channels,
|
| 535 |
delay_pattern=delay_pattern,
|
| 536 |
)
|
| 537 |
|
| 538 |
codebook = revert_audio_delay(
|
| 539 |
+
audio_BxTxC=generated_codes,
|
| 540 |
pad_value=audio_pad_value,
|
| 541 |
precomp=revert_precomp,
|
| 542 |
T=seq_length,
|
|
|
|
| 547 |
invalid_mask = (codebook < min_valid_index) | (codebook > max_valid_index)
|
| 548 |
codebook[invalid_mask] = 0
|
| 549 |
|
| 550 |
+
audios = []
|
| 551 |
|
| 552 |
+
if self.load_dac:
|
| 553 |
+
for i in range(batch_size):
|
| 554 |
+
audio = self._decode(codebook[i, : lengths_Bx[i], :])
|
| 555 |
+
audio_np = audio.cpu().numpy()
|
| 556 |
+
audios.append(audio_np)
|
| 557 |
+
else:
|
| 558 |
+
for i in range(batch_size):
|
| 559 |
+
audios.append(codebook[i, : lengths_Bx[i], :].cpu().numpy())
|
| 560 |
+
return audios
|
| 561 |
+
|
| 562 |
+
@torch.no_grad()
|
| 563 |
+
@torch.inference_mode()
|
| 564 |
+
def _encode(self, audio: torch.Tensor) -> torch.Tensor:
|
| 565 |
+
"""
|
| 566 |
+
Encodes the given audio waveform into a tensor of DAC codebook indices
|
| 567 |
+
"""
|
| 568 |
+
audio = audio.unsqueeze(0)
|
| 569 |
+
audio_data = self.dac_model.preprocess(audio, DEFAULT_SAMPLE_RATE)
|
| 570 |
+
_, encoded_frame, _, _, _ = self.dac_model.encode(audio_data)
|
| 571 |
+
encoded_frame: torch.Tensor
|
| 572 |
+
return encoded_frame.squeeze(0).transpose(0, 1)
|
| 573 |
+
|
| 574 |
+
@torch.no_grad()
|
| 575 |
+
@torch.inference_mode()
|
| 576 |
+
def _decode(self, audio_codes: torch.Tensor) -> torch.Tensor:
|
| 577 |
+
"""
|
| 578 |
+
Decodes the given frames into an output audio waveform
|
| 579 |
+
"""
|
| 580 |
+
audio_codes = audio_codes.unsqueeze(0).transpose(1, 2)
|
| 581 |
+
audio_values, _, _ = self.dac_model.quantizer.from_codes(audio_codes)
|
| 582 |
+
audio_values = self.dac_model.decode(audio_values)
|
| 583 |
+
audio_values: torch.Tensor
|
| 584 |
+
return audio_values.squeeze()
|
| 585 |
|
| 586 |
def load_audio(self, audio_path: str) -> torch.Tensor:
|
| 587 |
+
"""Loads and preprocesses an audio file for use as a prompt.
|
| 588 |
+
|
| 589 |
+
Loads the audio file, resamples it to the target sample rate if necessary,
|
| 590 |
+
preprocesses it using the DAC model's preprocessing, and encodes it into
|
| 591 |
+
DAC codebook indices.
|
| 592 |
+
|
| 593 |
+
Args:
|
| 594 |
+
audio_path: Path to the audio file.
|
| 595 |
+
|
| 596 |
+
Returns:
|
| 597 |
+
torch.Tensor: The encoded audio prompt as DAC codebook indices,
|
| 598 |
+
shape [T, C].
|
| 599 |
+
|
| 600 |
+
Raises:
|
| 601 |
+
RuntimeError: If the DAC model is not loaded (`load_dac=False` during init).
|
| 602 |
+
FileNotFoundError: If the audio file cannot be found.
|
| 603 |
+
Exception: If there's an error during loading or processing.
|
| 604 |
+
"""
|
| 605 |
+
if self.dac_model is None:
|
| 606 |
+
raise RuntimeError(
|
| 607 |
+
"DAC model is required for loading audio prompts but was not loaded."
|
| 608 |
+
)
|
| 609 |
audio, sr = torchaudio.load(audio_path, channels_first=True) # C, T
|
| 610 |
if sr != DEFAULT_SAMPLE_RATE:
|
| 611 |
audio = torchaudio.functional.resample(audio, sr, DEFAULT_SAMPLE_RATE)
|
| 612 |
+
# Convert to mono if stereo
|
| 613 |
+
if audio.shape[0] > 1:
|
| 614 |
+
audio = torch.mean(
|
| 615 |
+
audio, dim=0, keepdim=True
|
| 616 |
+
) # Average channels to get mono
|
| 617 |
+
return self._encode(audio.to(self.device))
|
| 618 |
|
| 619 |
def save_audio(self, path: str, audio: np.ndarray):
|
| 620 |
+
"""Saves the generated audio waveform to a file.
|
| 621 |
+
|
| 622 |
+
Uses the soundfile library to write the NumPy audio array to the specified
|
| 623 |
+
path with the default sample rate.
|
| 624 |
+
|
| 625 |
+
Args:
|
| 626 |
+
path: The path where the audio file will be saved.
|
| 627 |
+
audio: The audio waveform as a NumPy array.
|
| 628 |
+
"""
|
| 629 |
import soundfile as sf
|
| 630 |
|
| 631 |
sf.write(path, audio, DEFAULT_SAMPLE_RATE)
|
|
|
|
| 633 |
@torch.inference_mode()
|
| 634 |
def generate(
|
| 635 |
self,
|
| 636 |
+
text: str | list[str],
|
| 637 |
+
max_tokens: int = 3072,
|
| 638 |
cfg_scale: float = 3.0,
|
| 639 |
+
temperature: float = 1.2,
|
| 640 |
top_p: float = 0.95,
|
| 641 |
use_torch_compile: bool = False,
|
| 642 |
+
cfg_filter_top_k: int = 45,
|
| 643 |
+
audio_prompt: list[str | torch.Tensor | None]
|
| 644 |
+
| str
|
| 645 |
+
| torch.Tensor
|
| 646 |
+
| None = None,
|
| 647 |
+
audio_prompt_path: list[str | torch.Tensor | None]
|
| 648 |
+
| str
|
| 649 |
+
| torch.Tensor
|
| 650 |
+
| None = None,
|
| 651 |
use_cfg_filter: bool | None = None,
|
| 652 |
verbose: bool = False,
|
| 653 |
+
) -> np.ndarray | list[np.ndarray]:
|
| 654 |
+
"""Generates audio corresponding to the input text.
|
| 655 |
+
|
| 656 |
+
Args:
|
| 657 |
+
text: The input text prompt, or a list of text prompts for batch generation.
|
| 658 |
+
max_tokens: The maximum number of audio tokens to generate per prompt.
|
| 659 |
+
Defaults to the model's configured audio length if None.
|
| 660 |
+
cfg_scale: The scale factor for classifier-free guidance (CFG). Higher values
|
| 661 |
+
lead to stronger guidance towards the text prompt.
|
| 662 |
+
temperature: The temperature for sampling. Higher values increase randomness.
|
| 663 |
+
top_p: The cumulative probability threshold for nucleus (top-p) sampling.
|
| 664 |
+
use_torch_compile: Whether to compile the generation steps using torch.compile.
|
| 665 |
+
Can significantly speed up generation after the initial
|
| 666 |
+
compilation overhead. Defaults to False.
|
| 667 |
+
cfg_filter_top_k: The number of top logits to consider during CFG filtering.
|
| 668 |
+
(Note: This parameter name might be slightly misleading based
|
| 669 |
+
on the code; it's used in the `_sample_next_token` function.)
|
| 670 |
+
audio_prompt: An audio prompt or list of prompts to condition the generation.
|
| 671 |
+
Can be a file path (str), a pre-loaded tensor (DAC codes), or None.
|
| 672 |
+
If a list, its length must match the batch size of the text input.
|
| 673 |
+
audio_prompt_path: (Deprecated) Use `audio_prompt` instead.
|
| 674 |
+
use_cfg_filter: (Deprecated) This parameter is no longer used.
|
| 675 |
+
verbose: If True, prints progress information during generation, including
|
| 676 |
+
speed metrics.
|
| 677 |
+
|
| 678 |
+
Returns:
|
| 679 |
+
If a single text prompt was provided, returns a NumPy array containing the
|
| 680 |
+
generated audio waveform.
|
| 681 |
+
If a list of text prompts was provided, returns a list of NumPy arrays,
|
| 682 |
+
each corresponding to a prompt in the input list. Returns None for a
|
| 683 |
+
sequence if no audio was generated for it.
|
| 684 |
+
"""
|
| 685 |
+
batch_size = len(text) if isinstance(text, list) else 1
|
| 686 |
+
audio_eos_value = self.config.eos_token_id
|
| 687 |
+
audio_pad_value = self.config.pad_token_id
|
| 688 |
+
delay_pattern = self.config.delay_pattern
|
| 689 |
max_delay_pattern = max(delay_pattern)
|
| 690 |
+
delay_pattern_Cx = torch.tensor(
|
| 691 |
+
delay_pattern, device=self.device, dtype=torch.long
|
| 692 |
+
)
|
| 693 |
self.model.eval()
|
| 694 |
|
| 695 |
if audio_prompt_path:
|
|
|
|
| 701 |
if verbose:
|
| 702 |
total_start_time = time.time()
|
| 703 |
|
| 704 |
+
if use_torch_compile and not hasattr(self, "_compiled"):
|
| 705 |
+
# Compilation can take about a minute.
|
| 706 |
+
self._prepare_generation = torch.compile(
|
| 707 |
+
self._prepare_generation, dynamic=True, fullgraph=True
|
| 708 |
+
)
|
| 709 |
+
self._decoder_step = torch.compile(
|
| 710 |
+
self._decoder_step, fullgraph=True, mode="max-autotune"
|
| 711 |
+
)
|
| 712 |
+
self._compiled = True
|
| 713 |
+
|
| 714 |
+
if isinstance(audio_prompt, list):
|
| 715 |
+
audio_prompt = [
|
| 716 |
+
self.load_audio(p) if isinstance(p, str) else p for p in audio_prompt
|
| 717 |
+
]
|
| 718 |
+
elif isinstance(audio_prompt, str):
|
| 719 |
+
audio_prompt = [self.load_audio(audio_prompt)]
|
| 720 |
+
elif isinstance(audio_prompt, torch.Tensor):
|
| 721 |
+
audio_prompt = [audio_prompt]
|
| 722 |
+
elif audio_prompt is None:
|
| 723 |
+
audio_prompt = [None] * batch_size
|
| 724 |
+
|
| 725 |
+
assert len(audio_prompt) == batch_size, (
|
| 726 |
+
"Number of audio prompts must match batch size"
|
| 727 |
+
)
|
| 728 |
|
| 729 |
+
if isinstance(text, list):
|
| 730 |
+
text = [self._encode_text(t) for t in text]
|
| 731 |
else:
|
| 732 |
+
text = [self._encode_text(text)]
|
| 733 |
+
text = self._pad_text_input(text)
|
| 734 |
+
|
| 735 |
+
dec_state, dec_output = self._prepare_generation(
|
| 736 |
+
text, audio_prompt, max_tokens=max_tokens
|
| 737 |
+
)
|
| 738 |
+
dec_step = min(dec_output.prefill_steps) - 1
|
| 739 |
+
current_idx = torch.tensor([dec_step], device=self.device)
|
| 740 |
+
|
| 741 |
+
eos_detected_Bx = torch.zeros(
|
| 742 |
+
(batch_size,), dtype=torch.bool, device=self.device
|
| 743 |
+
)
|
| 744 |
+
eos_countdown_Bx = torch.full(
|
| 745 |
+
(batch_size,), -1, dtype=torch.long, device=self.device
|
| 746 |
+
)
|
| 747 |
+
finished_step_Bx = torch.full(
|
| 748 |
+
(batch_size,), -1, dtype=torch.long, device=self.device
|
| 749 |
+
)
|
| 750 |
+
|
| 751 |
+
bos_over = False
|
| 752 |
|
| 753 |
if verbose:
|
| 754 |
print("generate: starting generation loop")
|
| 755 |
if use_torch_compile:
|
| 756 |
print(
|
| 757 |
+
"generate: using use_torch_compile=True, the first step may be slow"
|
| 758 |
)
|
| 759 |
start_time = time.time()
|
| 760 |
|
| 761 |
+
# --- Generation Loop ---
|
| 762 |
while dec_step < max_tokens:
|
| 763 |
+
if (eos_countdown_Bx == 0).all():
|
| 764 |
+
break
|
| 765 |
+
|
| 766 |
+
current_step_idx = dec_step + 1
|
| 767 |
+
torch.compiler.cudagraph_mark_step_begin()
|
| 768 |
dec_state.prepare_step(dec_step)
|
| 769 |
+
tokens_Bx1xC = dec_output.get_tokens_at(dec_step).repeat_interleave(
|
| 770 |
+
2, dim=0
|
| 771 |
+
) # Repeat for CFG
|
| 772 |
+
|
| 773 |
+
pred_BxC = self._decoder_step(
|
| 774 |
tokens_Bx1xC,
|
| 775 |
dec_state,
|
| 776 |
cfg_scale,
|
| 777 |
temperature,
|
| 778 |
top_p,
|
| 779 |
cfg_filter_top_k,
|
| 780 |
+
current_idx,
|
| 781 |
)
|
| 782 |
|
| 783 |
+
current_idx += 1
|
| 784 |
+
|
| 785 |
+
active_mask_Bx = eos_countdown_Bx != 0
|
| 786 |
+
eos_trigger_Bx = torch.zeros_like(active_mask_Bx)
|
| 787 |
+
if active_mask_Bx.any():
|
| 788 |
+
is_eos_token = (~eos_detected_Bx[active_mask_Bx]) & (
|
| 789 |
+
pred_BxC[active_mask_Bx, 0] == audio_eos_value
|
| 790 |
+
)
|
| 791 |
+
is_max_len = current_step_idx >= max_tokens - max_delay_pattern
|
| 792 |
+
eos_trigger_Bx[active_mask_Bx] = is_eos_token | is_max_len
|
| 793 |
+
eos_detected_Bx |= eos_trigger_Bx
|
| 794 |
+
start_countdown_mask_Bx = eos_trigger_Bx & (eos_countdown_Bx < 0)
|
| 795 |
+
if start_countdown_mask_Bx.any():
|
| 796 |
+
eos_countdown_Bx[start_countdown_mask_Bx] = max_delay_pattern
|
| 797 |
+
finished_step_Bx[start_countdown_mask_Bx] = current_step_idx
|
| 798 |
+
|
| 799 |
+
padding_mask_Bx = eos_countdown_Bx > 0
|
| 800 |
+
if padding_mask_Bx.any():
|
| 801 |
+
pred_active_BxC = pred_BxC[padding_mask_Bx].clone()
|
| 802 |
+
countdown_active_Bx = eos_countdown_Bx[padding_mask_Bx]
|
| 803 |
+
step_after_eos_Bx = max_delay_pattern - countdown_active_Bx
|
| 804 |
+
step_after_eos_Bx_ = step_after_eos_Bx.unsqueeze(1)
|
| 805 |
+
delay_pattern_Cx_ = delay_pattern_Cx.unsqueeze(0)
|
| 806 |
+
eos_mask_NxC = step_after_eos_Bx_ == delay_pattern_Cx_
|
| 807 |
+
pad_mask_NxC = step_after_eos_Bx_ > delay_pattern_Cx_
|
| 808 |
+
pred_active_BxC[eos_mask_NxC] = audio_eos_value
|
| 809 |
+
pred_active_BxC[pad_mask_NxC] = audio_pad_value
|
| 810 |
+
pred_BxC[padding_mask_Bx] = pred_active_BxC
|
| 811 |
+
eos_countdown_Bx[padding_mask_Bx] -= 1
|
| 812 |
+
|
| 813 |
+
# --- Update BOS flag (Original) ---
|
| 814 |
+
if not bos_over:
|
| 815 |
+
bos_over = all(
|
| 816 |
+
dec_step - prefill_step > max_delay_pattern
|
| 817 |
+
for prefill_step in dec_output.prefill_steps
|
| 818 |
+
)
|
| 819 |
+
|
| 820 |
+
dec_output.update_one(pred_BxC, current_step_idx, not bos_over)
|
| 821 |
|
| 822 |
dec_step += 1
|
| 823 |
+
|
| 824 |
if verbose and dec_step % 86 == 0:
|
| 825 |
duration = time.time() - start_time
|
| 826 |
+
if duration > 0:
|
| 827 |
+
print(
|
| 828 |
+
f"generate step {dec_step}: speed={86 * batch_size / duration:.3f} tokens/s, realtime factor={batch_size / duration:.3f}x"
|
| 829 |
+
)
|
| 830 |
start_time = time.time()
|
| 831 |
|
| 832 |
+
# --- Finalize and Extract Output ---
|
| 833 |
+
final_step = dec_step + 1
|
|
|
|
| 834 |
|
| 835 |
+
finished_step_Bx[finished_step_Bx == -1] = final_step - max_delay_pattern
|
|
|
|
|
|
|
| 836 |
|
| 837 |
+
prefill_steps_tensor = torch.tensor(
|
| 838 |
+
dec_output.prefill_steps, device=self.device
|
| 839 |
+
)
|
| 840 |
+
lengths_Bx = finished_step_Bx - prefill_steps_tensor
|
| 841 |
+
lengths_Bx = torch.clamp(lengths_Bx, min=0)
|
| 842 |
+
|
| 843 |
+
max_len = lengths_Bx.max().item() + max_delay_pattern
|
| 844 |
+
outputs = []
|
| 845 |
+
|
| 846 |
+
if max_len > 0:
|
| 847 |
+
num_channels = self.config.decoder_config.num_channels
|
| 848 |
+
audio_pad_value = self.config.pad_token_id
|
| 849 |
+
generated_codes = torch.full(
|
| 850 |
+
(batch_size, max_len, num_channels),
|
| 851 |
+
fill_value=audio_pad_value,
|
| 852 |
+
dtype=torch.long,
|
| 853 |
+
device=self.device,
|
| 854 |
)
|
| 855 |
|
| 856 |
+
for i in range(batch_size):
|
| 857 |
+
start_step = dec_output.prefill_steps[i]
|
| 858 |
+
actual_len = lengths_Bx[i].item() + max_delay_pattern
|
| 859 |
+
if actual_len > 0:
|
| 860 |
+
tokens_to_copy = dec_output.generated_tokens[
|
| 861 |
+
i, start_step : start_step + actual_len, :
|
| 862 |
+
]
|
| 863 |
+
generated_codes[i, :actual_len, :] = tokens_to_copy
|
| 864 |
+
|
| 865 |
+
if verbose:
|
| 866 |
+
avg_steps = lengths_Bx.float().mean().item()
|
| 867 |
+
total_duration = time.time() - total_start_time
|
| 868 |
+
print(
|
| 869 |
+
f"generate: avg steps={avg_steps:.1f}, total duration={total_duration:.3f}s"
|
| 870 |
+
)
|
| 871 |
+
|
| 872 |
+
del dec_state
|
| 873 |
+
|
| 874 |
+
outputs = self._generate_output(generated_codes, lengths_Bx)
|
| 875 |
+
else:
|
| 876 |
+
print("Warning: Nothing generated for any sequence in the batch.")
|
| 877 |
+
outputs = [None] * batch_size
|
| 878 |
+
|
| 879 |
+
return outputs if batch_size > 1 else outputs[0]
|
dia/state.py
CHANGED
|
@@ -1,4 +1,5 @@
|
|
| 1 |
from dataclasses import dataclass
|
|
|
|
| 2 |
|
| 3 |
import torch
|
| 4 |
|
|
@@ -14,29 +15,18 @@ def create_attn_mask(
|
|
| 14 |
"""
|
| 15 |
Creates the attention mask (self or cross) mimicking JAX segment ID logic.
|
| 16 |
"""
|
| 17 |
-
B1, Tq = q_padding_mask_1d.shape
|
| 18 |
-
B2, Tk = k_padding_mask_1d.shape
|
| 19 |
-
assert B1 == B2, "Query and key batch dimensions must match"
|
| 20 |
|
| 21 |
p_mask_q = q_padding_mask_1d.unsqueeze(2) # Shape [B, Tq, 1]
|
| 22 |
p_mask_k = k_padding_mask_1d.unsqueeze(1) # Shape [B, 1, Tk]
|
| 23 |
|
| 24 |
-
|
| 25 |
-
non_pad_attends_non_pad = p_mask_q & p_mask_k # Shape [B, Tq, Tk]
|
| 26 |
-
|
| 27 |
-
# Condition B: Padding query attends to padding key
|
| 28 |
-
pad_attends_pad = (~p_mask_q) & (~p_mask_k) # Shape [B, Tq, Tk]
|
| 29 |
-
|
| 30 |
-
# Combine: True if padding status is compatible (both non-pad OR both pad)
|
| 31 |
-
mask = non_pad_attends_non_pad | pad_attends_pad # Shape [B, Tq, Tk]
|
| 32 |
-
|
| 33 |
if is_causal:
|
| 34 |
-
assert Tq == Tk,
|
| 35 |
-
"Causal mask requires query and key sequence lengths to be equal"
|
| 36 |
-
)
|
| 37 |
causal_mask_2d = torch.tril(
|
| 38 |
-
torch.
|
| 39 |
-
) # Shape [Tq, Tk]
|
| 40 |
causal_mask = mask & causal_mask_2d # Shape [B, Tq, Tk]
|
| 41 |
return causal_mask.unsqueeze(1) # Shape [B, 1, Tq, Tk]
|
| 42 |
else:
|
|
@@ -58,19 +48,18 @@ class EncoderInferenceState:
|
|
| 58 |
"""Creates EtorchrInferenceParams from DiaConfig and a device."""
|
| 59 |
device = cond_src.device
|
| 60 |
|
| 61 |
-
positions = (
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
)
|
| 67 |
-
padding_mask = (cond_src != config.data.text_pad_value).to(device).expand(2, -1)
|
| 68 |
attn_mask = create_attn_mask(
|
| 69 |
padding_mask, padding_mask, device, is_causal=False
|
| 70 |
)
|
| 71 |
|
| 72 |
return cls(
|
| 73 |
-
max_seq_len=config.
|
| 74 |
device=device,
|
| 75 |
positions=positions,
|
| 76 |
padding_mask=padding_mask,
|
|
@@ -78,9 +67,13 @@ class EncoderInferenceState:
|
|
| 78 |
)
|
| 79 |
|
| 80 |
|
| 81 |
-
class KVCache:
|
|
|
|
|
|
|
|
|
|
| 82 |
def __init__(
|
| 83 |
self,
|
|
|
|
| 84 |
num_heads: int,
|
| 85 |
max_len: int,
|
| 86 |
head_dim: int,
|
|
@@ -89,21 +82,33 @@ class KVCache:
|
|
| 89 |
k: torch.Tensor | None = None,
|
| 90 |
v: torch.Tensor | None = None,
|
| 91 |
):
|
| 92 |
-
|
| 93 |
-
torch.zeros(
|
|
|
|
|
|
|
|
|
|
|
|
|
| 94 |
if k is None
|
| 95 |
else k
|
| 96 |
)
|
| 97 |
-
|
| 98 |
-
torch.zeros(
|
|
|
|
|
|
|
|
|
|
|
|
|
| 99 |
if v is None
|
| 100 |
else v
|
| 101 |
)
|
| 102 |
-
|
|
|
|
|
|
|
|
|
|
| 103 |
|
| 104 |
@classmethod
|
| 105 |
def from_kv(cls, k: torch.Tensor, v: torch.Tensor) -> "KVCache":
|
| 106 |
return cls(
|
|
|
|
| 107 |
num_heads=k.shape[1],
|
| 108 |
max_len=k.shape[2],
|
| 109 |
head_dim=k.shape[3],
|
|
@@ -114,20 +119,17 @@ class KVCache:
|
|
| 114 |
)
|
| 115 |
|
| 116 |
def update(
|
| 117 |
-
self, k: torch.Tensor, v: torch.Tensor
|
| 118 |
) -> tuple[torch.Tensor, torch.Tensor]:
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
|
| 122 |
-
return self.k
|
| 123 |
|
| 124 |
-
def prefill(
|
| 125 |
-
self, k: torch.Tensor, v: torch.Tensor
|
| 126 |
-
) -> tuple[torch.Tensor, torch.Tensor]:
|
| 127 |
prefill_len = k.shape[2]
|
| 128 |
self.k[:, :, :prefill_len, :] = k
|
| 129 |
self.v[:, :, :prefill_len, :] = v
|
| 130 |
-
self.current_idx = prefill_len - 1
|
| 131 |
|
| 132 |
|
| 133 |
@dataclass
|
|
@@ -139,9 +141,10 @@ class DecoderInferenceState:
|
|
| 139 |
enc_out: torch.Tensor
|
| 140 |
enc_positions: torch.Tensor
|
| 141 |
dec_positions: torch.Tensor
|
| 142 |
-
dec_cross_attn_mask: torch.Tensor
|
| 143 |
self_attn_cache: list[KVCache]
|
| 144 |
cross_attn_cache: list[KVCache]
|
|
|
|
|
|
|
| 145 |
|
| 146 |
@classmethod
|
| 147 |
def new(
|
|
@@ -151,28 +154,36 @@ class DecoderInferenceState:
|
|
| 151 |
enc_out: torch.Tensor,
|
| 152 |
dec_cross_attn_cache: list[KVCache],
|
| 153 |
compute_dtype: torch.dtype,
|
|
|
|
| 154 |
) -> "DecoderInferenceState":
|
| 155 |
"""Creates DecoderInferenceParams from DiaConfig and a device."""
|
| 156 |
device = enc_out.device
|
| 157 |
-
max_audio_len =
|
|
|
|
|
|
|
|
|
|
| 158 |
|
| 159 |
dec_positions = torch.full(
|
| 160 |
-
(2, 1), fill_value=0, dtype=torch.
|
|
|
|
|
|
|
|
|
|
| 161 |
)
|
| 162 |
-
|
| 163 |
-
|
| 164 |
-
|
| 165 |
)
|
| 166 |
|
| 167 |
self_attn_cache = [
|
| 168 |
KVCache(
|
| 169 |
-
|
|
|
|
| 170 |
max_audio_len,
|
| 171 |
-
config.
|
| 172 |
compute_dtype,
|
| 173 |
device,
|
| 174 |
)
|
| 175 |
-
for _ in range(config.
|
| 176 |
]
|
| 177 |
|
| 178 |
return cls(
|
|
@@ -181,54 +192,56 @@ class DecoderInferenceState:
|
|
| 181 |
enc_out=enc_out,
|
| 182 |
enc_positions=enc_state.positions,
|
| 183 |
dec_positions=dec_positions,
|
| 184 |
-
dec_cross_attn_mask=dec_cross_attn_mask,
|
| 185 |
self_attn_cache=self_attn_cache,
|
| 186 |
cross_attn_cache=dec_cross_attn_cache,
|
|
|
|
|
|
|
| 187 |
)
|
| 188 |
|
| 189 |
def prepare_step(self, step_from: int, step_to: int | None = None) -> None:
|
| 190 |
if step_to is None:
|
| 191 |
step_to = step_from + 1
|
| 192 |
-
self.dec_positions = (
|
| 193 |
-
|
| 194 |
-
|
| 195 |
-
.expand(2, -1)
|
| 196 |
-
)
|
| 197 |
|
| 198 |
|
| 199 |
@dataclass
|
| 200 |
class DecoderOutput:
|
| 201 |
generated_tokens: torch.Tensor
|
| 202 |
-
|
| 203 |
|
| 204 |
@classmethod
|
| 205 |
-
def new(
|
| 206 |
-
|
|
|
|
|
|
|
| 207 |
return cls(
|
| 208 |
generated_tokens=torch.full(
|
| 209 |
-
(max_audio_len, config.
|
| 210 |
fill_value=-1,
|
| 211 |
dtype=torch.int,
|
| 212 |
device=device,
|
| 213 |
),
|
| 214 |
-
|
| 215 |
)
|
| 216 |
|
| 217 |
def get_tokens_at(self, step_from: int, step_to: int | None = None) -> torch.Tensor:
|
| 218 |
if step_to is None:
|
| 219 |
step_to = step_from + 1
|
| 220 |
-
return self.generated_tokens[step_from:step_to, :]
|
| 221 |
|
| 222 |
def update_one(self, dec_out: torch.Tensor, step: int, apply_mask: bool = False):
|
|
|
|
| 223 |
if apply_mask:
|
| 224 |
-
mask = self.generated_tokens[
|
| 225 |
-
self.generated_tokens[
|
| 226 |
-
mask, dec_out, self.generated_tokens[
|
| 227 |
)
|
| 228 |
else:
|
| 229 |
-
self.generated_tokens[
|
| 230 |
|
| 231 |
-
def prefill(self, dec_out: torch.Tensor,
|
| 232 |
-
length = dec_out.shape[
|
| 233 |
-
self.generated_tokens[
|
| 234 |
-
self.
|
|
|
|
| 1 |
from dataclasses import dataclass
|
| 2 |
+
from typing import Optional
|
| 3 |
|
| 4 |
import torch
|
| 5 |
|
|
|
|
| 15 |
"""
|
| 16 |
Creates the attention mask (self or cross) mimicking JAX segment ID logic.
|
| 17 |
"""
|
| 18 |
+
# B1, Tq = q_padding_mask_1d.shape
|
| 19 |
+
# B2, Tk = k_padding_mask_1d.shape
|
|
|
|
| 20 |
|
| 21 |
p_mask_q = q_padding_mask_1d.unsqueeze(2) # Shape [B, Tq, 1]
|
| 22 |
p_mask_k = k_padding_mask_1d.unsqueeze(1) # Shape [B, 1, Tk]
|
| 23 |
|
| 24 |
+
mask = p_mask_q & p_mask_k
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 25 |
if is_causal:
|
| 26 |
+
# assert Tq == Tk, "Causal mask requires query and key sequence lengths to be equal"
|
|
|
|
|
|
|
| 27 |
causal_mask_2d = torch.tril(
|
| 28 |
+
torch.ones_like(mask[0], dtype=torch.bool, device=device)
|
| 29 |
+
) # Shape [B, Tq, Tk]
|
| 30 |
causal_mask = mask & causal_mask_2d # Shape [B, Tq, Tk]
|
| 31 |
return causal_mask.unsqueeze(1) # Shape [B, 1, Tq, Tk]
|
| 32 |
else:
|
|
|
|
| 48 |
"""Creates EtorchrInferenceParams from DiaConfig and a device."""
|
| 49 |
device = cond_src.device
|
| 50 |
|
| 51 |
+
positions = torch.arange(
|
| 52 |
+
config.encoder_config.max_position_embeddings,
|
| 53 |
+
dtype=torch.float32,
|
| 54 |
+
device=device,
|
| 55 |
+
).unsqueeze(0)
|
| 56 |
+
padding_mask = (cond_src.squeeze(1) != 0).to(device).repeat_interleave(2, dim=0)
|
|
|
|
| 57 |
attn_mask = create_attn_mask(
|
| 58 |
padding_mask, padding_mask, device, is_causal=False
|
| 59 |
)
|
| 60 |
|
| 61 |
return cls(
|
| 62 |
+
max_seq_len=config.encoder_config.max_position_embeddings,
|
| 63 |
device=device,
|
| 64 |
positions=positions,
|
| 65 |
padding_mask=padding_mask,
|
|
|
|
| 67 |
)
|
| 68 |
|
| 69 |
|
| 70 |
+
class KVCache(torch.nn.Module):
|
| 71 |
+
k: torch.Tensor
|
| 72 |
+
v: torch.Tensor
|
| 73 |
+
|
| 74 |
def __init__(
|
| 75 |
self,
|
| 76 |
+
batch_size: int,
|
| 77 |
num_heads: int,
|
| 78 |
max_len: int,
|
| 79 |
head_dim: int,
|
|
|
|
| 82 |
k: torch.Tensor | None = None,
|
| 83 |
v: torch.Tensor | None = None,
|
| 84 |
):
|
| 85 |
+
k = (
|
| 86 |
+
torch.zeros(
|
| 87 |
+
(2 * batch_size, num_heads, max_len, head_dim),
|
| 88 |
+
dtype=dtype,
|
| 89 |
+
device=device,
|
| 90 |
+
)
|
| 91 |
if k is None
|
| 92 |
else k
|
| 93 |
)
|
| 94 |
+
v = (
|
| 95 |
+
torch.zeros(
|
| 96 |
+
(2 * batch_size, num_heads, max_len, head_dim),
|
| 97 |
+
dtype=dtype,
|
| 98 |
+
device=device,
|
| 99 |
+
)
|
| 100 |
if v is None
|
| 101 |
else v
|
| 102 |
)
|
| 103 |
+
super().__init__()
|
| 104 |
+
|
| 105 |
+
self.register_buffer("k", k)
|
| 106 |
+
self.register_buffer("v", v)
|
| 107 |
|
| 108 |
@classmethod
|
| 109 |
def from_kv(cls, k: torch.Tensor, v: torch.Tensor) -> "KVCache":
|
| 110 |
return cls(
|
| 111 |
+
batch_size=k.shape[0] // 2,
|
| 112 |
num_heads=k.shape[1],
|
| 113 |
max_len=k.shape[2],
|
| 114 |
head_dim=k.shape[3],
|
|
|
|
| 119 |
)
|
| 120 |
|
| 121 |
def update(
|
| 122 |
+
self, k: torch.Tensor, v: torch.Tensor, current_idx: torch.Tensor
|
| 123 |
) -> tuple[torch.Tensor, torch.Tensor]:
|
| 124 |
+
k_out, v_out = self.k, self.v
|
| 125 |
+
k_out[:, :, current_idx, :] = k
|
| 126 |
+
v_out[:, :, current_idx, :] = v
|
| 127 |
+
return self.k, self.v
|
| 128 |
|
| 129 |
+
def prefill(self, k: torch.Tensor, v: torch.Tensor):
|
|
|
|
|
|
|
| 130 |
prefill_len = k.shape[2]
|
| 131 |
self.k[:, :, :prefill_len, :] = k
|
| 132 |
self.v[:, :, :prefill_len, :] = v
|
|
|
|
| 133 |
|
| 134 |
|
| 135 |
@dataclass
|
|
|
|
| 141 |
enc_out: torch.Tensor
|
| 142 |
enc_positions: torch.Tensor
|
| 143 |
dec_positions: torch.Tensor
|
|
|
|
| 144 |
self_attn_cache: list[KVCache]
|
| 145 |
cross_attn_cache: list[KVCache]
|
| 146 |
+
casual_attn_mask: torch.Tensor
|
| 147 |
+
cross_attn_mask: torch.Tensor
|
| 148 |
|
| 149 |
@classmethod
|
| 150 |
def new(
|
|
|
|
| 154 |
enc_out: torch.Tensor,
|
| 155 |
dec_cross_attn_cache: list[KVCache],
|
| 156 |
compute_dtype: torch.dtype,
|
| 157 |
+
max_generation_length: Optional[int] = None,
|
| 158 |
) -> "DecoderInferenceState":
|
| 159 |
"""Creates DecoderInferenceParams from DiaConfig and a device."""
|
| 160 |
device = enc_out.device
|
| 161 |
+
max_audio_len = (
|
| 162 |
+
max_generation_length or config.decoder_config.max_position_embeddings
|
| 163 |
+
)
|
| 164 |
+
batch_size = enc_out.shape[0] // 2
|
| 165 |
|
| 166 |
dec_positions = torch.full(
|
| 167 |
+
(2 * batch_size, 1), fill_value=0, dtype=torch.int32, device=device
|
| 168 |
+
)
|
| 169 |
+
causal_mask = torch.tril(
|
| 170 |
+
torch.ones(max_audio_len, max_audio_len, dtype=torch.bool, device=device)
|
| 171 |
)
|
| 172 |
+
dec_mask = torch.ones((2 * batch_size, 1), dtype=torch.bool, device=device)
|
| 173 |
+
cross_attn_mask = create_attn_mask(
|
| 174 |
+
dec_mask, enc_state.padding_mask, device, is_causal=False
|
| 175 |
)
|
| 176 |
|
| 177 |
self_attn_cache = [
|
| 178 |
KVCache(
|
| 179 |
+
batch_size,
|
| 180 |
+
config.decoder_config.num_key_value_heads,
|
| 181 |
max_audio_len,
|
| 182 |
+
config.decoder_config.head_dim,
|
| 183 |
compute_dtype,
|
| 184 |
device,
|
| 185 |
)
|
| 186 |
+
for _ in range(config.decoder_config.num_hidden_layers)
|
| 187 |
]
|
| 188 |
|
| 189 |
return cls(
|
|
|
|
| 192 |
enc_out=enc_out,
|
| 193 |
enc_positions=enc_state.positions,
|
| 194 |
dec_positions=dec_positions,
|
|
|
|
| 195 |
self_attn_cache=self_attn_cache,
|
| 196 |
cross_attn_cache=dec_cross_attn_cache,
|
| 197 |
+
casual_attn_mask=causal_mask,
|
| 198 |
+
cross_attn_mask=cross_attn_mask,
|
| 199 |
)
|
| 200 |
|
| 201 |
def prepare_step(self, step_from: int, step_to: int | None = None) -> None:
|
| 202 |
if step_to is None:
|
| 203 |
step_to = step_from + 1
|
| 204 |
+
self.dec_positions = torch.arange(
|
| 205 |
+
step_from, step_to, dtype=torch.int32, device=self.device
|
| 206 |
+
).unsqueeze(0)
|
|
|
|
|
|
|
| 207 |
|
| 208 |
|
| 209 |
@dataclass
|
| 210 |
class DecoderOutput:
|
| 211 |
generated_tokens: torch.Tensor
|
| 212 |
+
prefill_steps: list[int]
|
| 213 |
|
| 214 |
@classmethod
|
| 215 |
+
def new(
|
| 216 |
+
cls, batch_size: int, config: DiaConfig, device: torch.device
|
| 217 |
+
) -> "DecoderOutput":
|
| 218 |
+
max_audio_len = config.decoder_config.max_position_embeddings
|
| 219 |
return cls(
|
| 220 |
generated_tokens=torch.full(
|
| 221 |
+
(batch_size, max_audio_len, config.decoder_config.num_channels),
|
| 222 |
fill_value=-1,
|
| 223 |
dtype=torch.int,
|
| 224 |
device=device,
|
| 225 |
),
|
| 226 |
+
prefill_steps=[],
|
| 227 |
)
|
| 228 |
|
| 229 |
def get_tokens_at(self, step_from: int, step_to: int | None = None) -> torch.Tensor:
|
| 230 |
if step_to is None:
|
| 231 |
step_to = step_from + 1
|
| 232 |
+
return self.generated_tokens[:, step_from:step_to, :]
|
| 233 |
|
| 234 |
def update_one(self, dec_out: torch.Tensor, step: int, apply_mask: bool = False):
|
| 235 |
+
dec_out = dec_out.to(self.generated_tokens.dtype)
|
| 236 |
if apply_mask:
|
| 237 |
+
mask = self.generated_tokens[:, step, :] == -1
|
| 238 |
+
self.generated_tokens[:, step, :] = torch.where(
|
| 239 |
+
mask, dec_out, self.generated_tokens[:, step, :]
|
| 240 |
)
|
| 241 |
else:
|
| 242 |
+
self.generated_tokens[:, step, :] = dec_out
|
| 243 |
|
| 244 |
+
def prefill(self, dec_out: torch.Tensor, prefill_steps: list[int]):
|
| 245 |
+
length = dec_out.shape[1]
|
| 246 |
+
self.generated_tokens[:, :length, :] = dec_out
|
| 247 |
+
self.prefill_steps = prefill_steps
|
requirements.txt
CHANGED
|
@@ -4,6 +4,7 @@ huggingface-hub>=0.30.2
|
|
| 4 |
numpy>=2.2.4
|
| 5 |
pydantic>=2.11.3
|
| 6 |
soundfile>=0.13.1
|
| 7 |
-
torchaudio
|
| 8 |
-
torch
|
|
|
|
| 9 |
gradio-dialogue>=0.0.4
|
|
|
|
| 4 |
numpy>=2.2.4
|
| 5 |
pydantic>=2.11.3
|
| 6 |
soundfile>=0.13.1
|
| 7 |
+
torchaudio==2.6.0
|
| 8 |
+
torch==2.6.0
|
| 9 |
+
triton==3.2.0
|
| 10 |
gradio-dialogue>=0.0.4
|