Spaces:
Running
Running
Michael Hu
commited on
Commit
·
9c4b958
1
Parent(s):
1a3633a
add dia tts model. Since dia is not yet released to pypi, we pull in the source directly
Browse files- dia/__init__.py +6 -0
- dia/audio.py +185 -0
- dia/config.py +187 -0
- dia/layers.py +624 -0
- dia/model.py +455 -0
- dia/state.py +207 -0
dia/__init__.py
ADDED
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from .model import Dia
|
2 |
+
|
3 |
+
|
4 |
+
__all__ = [
|
5 |
+
"Dia",
|
6 |
+
]
|
dia/audio.py
ADDED
@@ -0,0 +1,185 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import typing as tp
|
2 |
+
|
3 |
+
import torch
|
4 |
+
|
5 |
+
|
6 |
+
def build_delay_indices(B: int, T: int, C: int, delay_pattern: tp.List[int]) -> tp.Tuple[torch.Tensor, torch.Tensor]:
|
7 |
+
"""
|
8 |
+
Precompute (t_idx_BxTxC, indices_BTCx3) so that out[t, c] = in[t - delay[c], c].
|
9 |
+
Negative t_idx => BOS; t_idx >= T => PAD.
|
10 |
+
"""
|
11 |
+
delay_arr = torch.tensor(delay_pattern, dtype=torch.int32)
|
12 |
+
|
13 |
+
t_idx_BxT = torch.broadcast_to(
|
14 |
+
torch.arange(T, dtype=torch.int32)[None, :],
|
15 |
+
[B, T],
|
16 |
+
)
|
17 |
+
t_idx_BxTx1 = t_idx_BxT[..., None]
|
18 |
+
t_idx_BxTxC = t_idx_BxTx1 - delay_arr.view(1, 1, C)
|
19 |
+
|
20 |
+
b_idx_BxTxC = torch.broadcast_to(
|
21 |
+
torch.arange(B, dtype=torch.int32).view(B, 1, 1),
|
22 |
+
[B, T, C],
|
23 |
+
)
|
24 |
+
c_idx_BxTxC = torch.broadcast_to(
|
25 |
+
torch.arange(C, dtype=torch.int32).view(1, 1, C),
|
26 |
+
[B, T, C],
|
27 |
+
)
|
28 |
+
|
29 |
+
# We must clamp time indices to [0..T-1] so gather_nd equivalent won't fail
|
30 |
+
t_clamped_BxTxC = torch.clamp(t_idx_BxTxC, 0, T - 1)
|
31 |
+
|
32 |
+
indices_BTCx3 = torch.stack(
|
33 |
+
[
|
34 |
+
b_idx_BxTxC.reshape(-1),
|
35 |
+
t_clamped_BxTxC.reshape(-1),
|
36 |
+
c_idx_BxTxC.reshape(-1),
|
37 |
+
],
|
38 |
+
dim=1,
|
39 |
+
).long() # Ensure indices are long type for indexing
|
40 |
+
|
41 |
+
return t_idx_BxTxC, indices_BTCx3
|
42 |
+
|
43 |
+
|
44 |
+
def apply_audio_delay(
|
45 |
+
audio_BxTxC: torch.Tensor,
|
46 |
+
pad_value: int,
|
47 |
+
bos_value: int,
|
48 |
+
precomp: tp.Tuple[torch.Tensor, torch.Tensor],
|
49 |
+
) -> torch.Tensor:
|
50 |
+
"""
|
51 |
+
Applies the delay pattern to batched audio tokens using precomputed indices,
|
52 |
+
inserting BOS where t_idx < 0 and PAD where t_idx >= T.
|
53 |
+
|
54 |
+
Args:
|
55 |
+
audio_BxTxC: [B, T, C] int16 audio tokens (or int32/float)
|
56 |
+
pad_value: the padding token
|
57 |
+
bos_value: the BOS token
|
58 |
+
precomp: (t_idx_BxTxC, indices_BTCx3) from build_delay_indices
|
59 |
+
|
60 |
+
Returns:
|
61 |
+
result_BxTxC: [B, T, C] delayed audio tokens
|
62 |
+
"""
|
63 |
+
device = audio_BxTxC.device # Get device from input tensor
|
64 |
+
t_idx_BxTxC, indices_BTCx3 = precomp
|
65 |
+
t_idx_BxTxC = t_idx_BxTxC.to(device) # Move precomputed indices to device
|
66 |
+
indices_BTCx3 = indices_BTCx3.to(device)
|
67 |
+
|
68 |
+
# Equivalent of tf.gather_nd using advanced indexing
|
69 |
+
# Ensure indices are long type if not already (build_delay_indices should handle this)
|
70 |
+
gathered_flat = audio_BxTxC[indices_BTCx3[:, 0], indices_BTCx3[:, 1], indices_BTCx3[:, 2]]
|
71 |
+
gathered_BxTxC = gathered_flat.view(audio_BxTxC.shape)
|
72 |
+
|
73 |
+
# Create masks on the correct device
|
74 |
+
mask_bos = t_idx_BxTxC < 0 # => place bos_value
|
75 |
+
mask_pad = t_idx_BxTxC >= audio_BxTxC.shape[1] # => place pad_value
|
76 |
+
|
77 |
+
# Create scalar tensors on the correct device
|
78 |
+
bos_tensor = torch.tensor(bos_value, dtype=audio_BxTxC.dtype, device=device)
|
79 |
+
pad_tensor = torch.tensor(pad_value, dtype=audio_BxTxC.dtype, device=device)
|
80 |
+
|
81 |
+
# If mask_bos, BOS; else if mask_pad, PAD; else original gather
|
82 |
+
# All tensors should now be on the same device
|
83 |
+
result_BxTxC = torch.where(mask_bos, bos_tensor, torch.where(mask_pad, pad_tensor, gathered_BxTxC))
|
84 |
+
|
85 |
+
return result_BxTxC
|
86 |
+
|
87 |
+
|
88 |
+
def build_revert_indices(B: int, T: int, C: int, delay_pattern: tp.List[int]) -> tp.Tuple[torch.Tensor, torch.Tensor]:
|
89 |
+
"""
|
90 |
+
Precompute indices for the revert operation using PyTorch.
|
91 |
+
|
92 |
+
Returns:
|
93 |
+
A tuple (t_idx_BxTxC, indices_BTCx3) where:
|
94 |
+
- t_idx_BxTxC is a tensor of shape [B, T, C] computed as time indices plus the delay.
|
95 |
+
- indices_BTCx3 is a tensor of shape [B*T*C, 3] used for gathering, computed from:
|
96 |
+
batch indices, clamped time indices, and channel indices.
|
97 |
+
"""
|
98 |
+
# Use default device unless specified otherwise; assumes inputs might define device later
|
99 |
+
device = None # Or determine dynamically if needed, e.g., from a model parameter
|
100 |
+
|
101 |
+
delay_arr = torch.tensor(delay_pattern, dtype=torch.int32, device=device)
|
102 |
+
|
103 |
+
t_idx_BT1 = torch.broadcast_to(torch.arange(T, device=device).unsqueeze(0), [B, T])
|
104 |
+
t_idx_BT1 = t_idx_BT1.unsqueeze(-1)
|
105 |
+
|
106 |
+
t_idx_BxTxC = torch.minimum(
|
107 |
+
t_idx_BT1 + delay_arr.view(1, 1, C),
|
108 |
+
torch.tensor(T - 1, device=device),
|
109 |
+
)
|
110 |
+
b_idx_BxTxC = torch.broadcast_to(torch.arange(B, device=device).view(B, 1, 1), [B, T, C])
|
111 |
+
c_idx_BxTxC = torch.broadcast_to(torch.arange(C, device=device).view(1, 1, C), [B, T, C])
|
112 |
+
|
113 |
+
indices_BTCx3 = torch.stack(
|
114 |
+
[
|
115 |
+
b_idx_BxTxC.reshape(-1),
|
116 |
+
t_idx_BxTxC.reshape(-1),
|
117 |
+
c_idx_BxTxC.reshape(-1),
|
118 |
+
],
|
119 |
+
axis=1,
|
120 |
+
).long() # Ensure indices are long type
|
121 |
+
|
122 |
+
return t_idx_BxTxC, indices_BTCx3
|
123 |
+
|
124 |
+
|
125 |
+
def revert_audio_delay(
|
126 |
+
audio_BxTxC: torch.Tensor,
|
127 |
+
pad_value: int,
|
128 |
+
precomp: tp.Tuple[torch.Tensor, torch.Tensor],
|
129 |
+
T: int,
|
130 |
+
) -> torch.Tensor:
|
131 |
+
"""
|
132 |
+
Reverts a delay pattern from batched audio tokens using precomputed indices (PyTorch version).
|
133 |
+
|
134 |
+
Args:
|
135 |
+
audio_BxTxC: Input delayed audio tensor
|
136 |
+
pad_value: Padding value for out-of-bounds indices
|
137 |
+
precomp: Precomputed revert indices tuple containing:
|
138 |
+
- t_idx_BxTxC: Time offset indices tensor
|
139 |
+
- indices_BTCx3: Gather indices tensor for original audio
|
140 |
+
T: Original sequence length before padding
|
141 |
+
|
142 |
+
Returns:
|
143 |
+
Reverted audio tensor with same shape as input
|
144 |
+
"""
|
145 |
+
t_idx_BxTxC, indices_BTCx3 = precomp
|
146 |
+
device = audio_BxTxC.device # Get device from input tensor
|
147 |
+
|
148 |
+
# Move precomputed indices to the same device as audio_BxTxC if they aren't already
|
149 |
+
t_idx_BxTxC = t_idx_BxTxC.to(device)
|
150 |
+
indices_BTCx3 = indices_BTCx3.to(device)
|
151 |
+
|
152 |
+
# Using PyTorch advanced indexing (equivalent to tf.gather_nd or np equivalent)
|
153 |
+
gathered_flat = audio_BxTxC[indices_BTCx3[:, 0], indices_BTCx3[:, 1], indices_BTCx3[:, 2]]
|
154 |
+
gathered_BxTxC = gathered_flat.view(audio_BxTxC.size()) # Use .size() for robust reshaping
|
155 |
+
|
156 |
+
# Create pad_tensor on the correct device
|
157 |
+
pad_tensor = torch.tensor(pad_value, dtype=audio_BxTxC.dtype, device=device)
|
158 |
+
# Create T tensor on the correct device for comparison
|
159 |
+
T_tensor = torch.tensor(T, device=device)
|
160 |
+
|
161 |
+
result_BxTxC = torch.where(t_idx_BxTxC >= T_tensor, pad_tensor, gathered_BxTxC) # Changed np.where to torch.where
|
162 |
+
|
163 |
+
return result_BxTxC
|
164 |
+
|
165 |
+
|
166 |
+
@torch.no_grad()
|
167 |
+
@torch.inference_mode()
|
168 |
+
def decode(
|
169 |
+
model,
|
170 |
+
audio_codes,
|
171 |
+
):
|
172 |
+
"""
|
173 |
+
Decodes the given frames into an output audio waveform
|
174 |
+
"""
|
175 |
+
if len(audio_codes) != 1:
|
176 |
+
raise ValueError(f"Expected one frame, got {len(audio_codes)}")
|
177 |
+
|
178 |
+
try:
|
179 |
+
audio_values = model.quantizer.from_codes(audio_codes)
|
180 |
+
audio_values = model.decode(audio_values[0])
|
181 |
+
|
182 |
+
return audio_values
|
183 |
+
except Exception as e:
|
184 |
+
print(f"Error in decode method: {str(e)}")
|
185 |
+
raise
|
dia/config.py
ADDED
@@ -0,0 +1,187 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""Configuration management module for the Dia model.
|
2 |
+
|
3 |
+
This module provides comprehensive configuration management for the Dia model,
|
4 |
+
utilizing Pydantic for validation. It defines configurations for data processing,
|
5 |
+
model architecture (encoder and decoder), and training settings.
|
6 |
+
|
7 |
+
Key components:
|
8 |
+
- DataConfig: Parameters for data loading and preprocessing.
|
9 |
+
- EncoderConfig: Architecture details for the encoder module.
|
10 |
+
- DecoderConfig: Architecture details for the decoder module.
|
11 |
+
- ModelConfig: Combined model architecture settings.
|
12 |
+
- TrainingConfig: Training hyperparameters and settings.
|
13 |
+
- DiaConfig: Master configuration combining all components.
|
14 |
+
"""
|
15 |
+
|
16 |
+
import os
|
17 |
+
from typing import Annotated
|
18 |
+
|
19 |
+
from pydantic import BaseModel, BeforeValidator, Field
|
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)] = Field(gt=0, multiple_of=128)
|
37 |
+
audio_length: Annotated[int, BeforeValidator(lambda x: (x + 127) // 128 * 128)] = Field(gt=0, multiple_of=128)
|
38 |
+
channels: int = Field(default=9, gt=0, multiple_of=1)
|
39 |
+
text_pad_value: int = Field(default=0)
|
40 |
+
audio_eos_value: int = Field(default=1024)
|
41 |
+
audio_pad_value: int = Field(default=1025)
|
42 |
+
audio_bos_value: int = Field(default=1026)
|
43 |
+
delay_pattern: list[Annotated[int, Field(ge=0)]] = Field(default_factory=lambda: [0, 8, 9, 10, 11, 12, 13, 14, 15])
|
44 |
+
|
45 |
+
def __hash__(self) -> int:
|
46 |
+
"""Generate a hash based on all fields of the config."""
|
47 |
+
return hash(
|
48 |
+
(
|
49 |
+
self.text_length,
|
50 |
+
self.audio_length,
|
51 |
+
self.channels,
|
52 |
+
self.text_pad_value,
|
53 |
+
self.audio_pad_value,
|
54 |
+
self.audio_bos_value,
|
55 |
+
self.audio_eos_value,
|
56 |
+
tuple(self.delay_pattern),
|
57 |
+
)
|
58 |
+
)
|
59 |
+
|
60 |
+
|
61 |
+
class EncoderConfig(BaseModel, frozen=True):
|
62 |
+
"""Configuration for the encoder component of the Dia model.
|
63 |
+
|
64 |
+
Attributes:
|
65 |
+
n_layer: Number of transformer layers.
|
66 |
+
n_embd: Embedding dimension.
|
67 |
+
n_hidden: Hidden dimension size in the MLP layers.
|
68 |
+
n_head: Number of attention heads.
|
69 |
+
head_dim: Dimension per attention head.
|
70 |
+
"""
|
71 |
+
|
72 |
+
n_layer: int = Field(gt=0)
|
73 |
+
n_embd: int = Field(gt=0)
|
74 |
+
n_hidden: int = Field(gt=0)
|
75 |
+
n_head: int = Field(gt=0)
|
76 |
+
head_dim: int = Field(gt=0)
|
77 |
+
|
78 |
+
|
79 |
+
class DecoderConfig(BaseModel, frozen=True):
|
80 |
+
"""Configuration for the decoder component of the Dia model.
|
81 |
+
|
82 |
+
Attributes:
|
83 |
+
n_layer: Number of transformer layers.
|
84 |
+
n_embd: Embedding dimension.
|
85 |
+
n_hidden: Hidden dimension size in the MLP layers.
|
86 |
+
gqa_query_heads: Number of query heads for grouped-query self-attention.
|
87 |
+
kv_heads: Number of key/value heads for grouped-query self-attention.
|
88 |
+
gqa_head_dim: Dimension per query head for grouped-query self-attention.
|
89 |
+
cross_query_heads: Number of query heads for cross-attention.
|
90 |
+
cross_head_dim: Dimension per cross-attention head.
|
91 |
+
"""
|
92 |
+
|
93 |
+
n_layer: int = Field(gt=0)
|
94 |
+
n_embd: int = Field(gt=0)
|
95 |
+
n_hidden: int = Field(gt=0)
|
96 |
+
gqa_query_heads: int = Field(gt=0)
|
97 |
+
kv_heads: int = Field(gt=0)
|
98 |
+
gqa_head_dim: int = Field(gt=0)
|
99 |
+
cross_query_heads: int = Field(gt=0)
|
100 |
+
cross_head_dim: int = Field(gt=0)
|
101 |
+
|
102 |
+
|
103 |
+
class ModelConfig(BaseModel, frozen=True):
|
104 |
+
"""Main configuration container for the Dia model architecture.
|
105 |
+
|
106 |
+
Attributes:
|
107 |
+
encoder: Configuration for the encoder component.
|
108 |
+
decoder: Configuration for the decoder component.
|
109 |
+
src_vocab_size: Size of the source (text) vocabulary.
|
110 |
+
tgt_vocab_size: Size of the target (audio code) vocabulary.
|
111 |
+
dropout: Dropout probability applied within the model.
|
112 |
+
normalization_layer_epsilon: Epsilon value for normalization layers (e.g., LayerNorm).
|
113 |
+
weight_dtype: Data type for model weights (e.g., "float32", "bfloat16").
|
114 |
+
rope_min_timescale: Minimum timescale for Rotary Positional Embeddings (RoPE).
|
115 |
+
rope_max_timescale: Maximum timescale for Rotary Positional Embeddings (RoPE).
|
116 |
+
"""
|
117 |
+
|
118 |
+
encoder: EncoderConfig
|
119 |
+
decoder: DecoderConfig
|
120 |
+
src_vocab_size: int = Field(default=128, gt=0)
|
121 |
+
tgt_vocab_size: int = Field(default=1028, gt=0)
|
122 |
+
dropout: float = Field(default=0.0, ge=0.0, lt=1.0)
|
123 |
+
normalization_layer_epsilon: float = Field(default=1.0e-5, ge=0.0)
|
124 |
+
weight_dtype: str = Field(default="float32", description="Weight precision")
|
125 |
+
rope_min_timescale: int = Field(default=1, description="Timescale For global Attention")
|
126 |
+
rope_max_timescale: int = Field(default=10_000, description="Timescale For global Attention")
|
127 |
+
|
128 |
+
|
129 |
+
class TrainingConfig(BaseModel, frozen=True):
|
130 |
+
pass
|
131 |
+
|
132 |
+
|
133 |
+
class DiaConfig(BaseModel, frozen=True):
|
134 |
+
"""Master configuration for the Dia model.
|
135 |
+
|
136 |
+
Combines all sub-configurations into a single validated object.
|
137 |
+
|
138 |
+
Attributes:
|
139 |
+
version: Configuration version string.
|
140 |
+
model: Model architecture configuration.
|
141 |
+
training: Training process configuration (precision settings).
|
142 |
+
data: Data loading and processing configuration.
|
143 |
+
"""
|
144 |
+
|
145 |
+
version: str = Field(default="1.0")
|
146 |
+
model: ModelConfig
|
147 |
+
# TODO: remove training. this is just for backward compatibility
|
148 |
+
training: TrainingConfig | None = Field(default=None)
|
149 |
+
data: DataConfig
|
150 |
+
|
151 |
+
def save(self, path: str) -> None:
|
152 |
+
"""Save the current configuration instance to a JSON file.
|
153 |
+
|
154 |
+
Ensures the parent directory exists and the file has a .json extension.
|
155 |
+
|
156 |
+
Args:
|
157 |
+
path: The target file path to save the configuration.
|
158 |
+
|
159 |
+
Raises:
|
160 |
+
ValueError: If the path is not a file with a .json extension.
|
161 |
+
"""
|
162 |
+
os.makedirs(os.path.dirname(path), exist_ok=True)
|
163 |
+
config_json = self.model_dump_json(indent=2)
|
164 |
+
with open(path, "w") as f:
|
165 |
+
f.write(config_json)
|
166 |
+
|
167 |
+
@classmethod
|
168 |
+
def load(cls, path: str) -> "DiaConfig | None":
|
169 |
+
"""Load and validate a Dia configuration from a JSON file.
|
170 |
+
|
171 |
+
Args:
|
172 |
+
path: The path to the configuration file.
|
173 |
+
|
174 |
+
Returns:
|
175 |
+
A validated DiaConfig instance if the file exists and is valid,
|
176 |
+
otherwise None if the file is not found.
|
177 |
+
|
178 |
+
Raises:
|
179 |
+
ValueError: If the path does not point to an existing .json file.
|
180 |
+
pydantic.ValidationError: If the JSON content fails validation against the DiaConfig schema.
|
181 |
+
"""
|
182 |
+
try:
|
183 |
+
with open(path, "r") as f:
|
184 |
+
content = f.read()
|
185 |
+
return cls.model_validate_json(content)
|
186 |
+
except FileNotFoundError:
|
187 |
+
return None
|
dia/layers.py
ADDED
@@ -0,0 +1,624 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
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 DiaConfig
|
9 |
+
from .state import DecoderInferenceState, EncoderInferenceState, KVCache
|
10 |
+
|
11 |
+
|
12 |
+
def _normalize_axes(axes: tuple[int, ...], ndim: int) -> tuple[int, ...]:
|
13 |
+
return tuple(ax if ax >= 0 else ndim + ax for ax in axes)
|
14 |
+
|
15 |
+
|
16 |
+
class DenseGeneral(nn.Module):
|
17 |
+
"""
|
18 |
+
PyTorch equivalent of flax.linen.DenseGeneral with shapes defined at init.
|
19 |
+
|
20 |
+
Stores weights (`kernel`) in the same layout as Jax and uses torch.tensordot
|
21 |
+
for the generalized matrix multiplication. Weight/bias shapes are calculated
|
22 |
+
and parameters created during initialization based on config.
|
23 |
+
`load_weights` validates shapes and copies data.
|
24 |
+
|
25 |
+
Attributes:
|
26 |
+
axis (Tuple[int, ...]): Input axis or axes to contract.
|
27 |
+
in_shapes (Tuple[int, ...]): Sizes of the input dimensions specified by `axis`.
|
28 |
+
out_features (Tuple[int, ...]): Shape of the output features (non-contracted dims).
|
29 |
+
use_bias (bool): Whether to add a bias term.
|
30 |
+
weight (nn.Parameter): The kernel parameter.
|
31 |
+
bias (Optional[nn.Parameter]): The bias parameter (if use_bias=True).
|
32 |
+
"""
|
33 |
+
|
34 |
+
def __init__(
|
35 |
+
self,
|
36 |
+
in_shapes: tuple[int, ...],
|
37 |
+
out_features: tuple[int, ...],
|
38 |
+
axis: tuple[int, ...] = (-1,),
|
39 |
+
weight_dtype: torch.dtype | None = None,
|
40 |
+
device: torch.device | None = None,
|
41 |
+
):
|
42 |
+
super().__init__()
|
43 |
+
self.in_shapes = in_shapes
|
44 |
+
self.out_features = out_features
|
45 |
+
self.axis = axis
|
46 |
+
self.kernel_shape = self.in_shapes + self.out_features
|
47 |
+
|
48 |
+
factory_kwargs = {"device": device, "dtype": weight_dtype}
|
49 |
+
self.weight = nn.Parameter(torch.empty(self.kernel_shape, **factory_kwargs))
|
50 |
+
|
51 |
+
def forward(self, inputs: Tensor) -> Tensor:
|
52 |
+
norm_axis = _normalize_axes(self.axis, inputs.ndim)
|
53 |
+
kernel_contract_axes = tuple(range(len(norm_axis)))
|
54 |
+
|
55 |
+
output = torch.tensordot(
|
56 |
+
inputs.to(self.weight.dtype),
|
57 |
+
self.weight,
|
58 |
+
dims=(norm_axis, kernel_contract_axes),
|
59 |
+
).to(inputs.dtype)
|
60 |
+
return output
|
61 |
+
|
62 |
+
|
63 |
+
class MlpBlock(nn.Module):
|
64 |
+
"""MLP block using DenseGeneral."""
|
65 |
+
|
66 |
+
def __init__(self, embed_dim: int, intermediate_dim: int, compute_dtype: torch.dtype):
|
67 |
+
super().__init__()
|
68 |
+
self.dtype = compute_dtype
|
69 |
+
|
70 |
+
self.wi_fused = DenseGeneral(
|
71 |
+
in_shapes=(embed_dim,),
|
72 |
+
out_features=(2, intermediate_dim),
|
73 |
+
axis=(-1,),
|
74 |
+
weight_dtype=compute_dtype,
|
75 |
+
)
|
76 |
+
|
77 |
+
self.wo = DenseGeneral(
|
78 |
+
in_shapes=(intermediate_dim,),
|
79 |
+
out_features=(embed_dim,),
|
80 |
+
axis=(-1,),
|
81 |
+
weight_dtype=compute_dtype,
|
82 |
+
)
|
83 |
+
|
84 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
85 |
+
"""Forward pass."""
|
86 |
+
fused_x = self.wi_fused(x)
|
87 |
+
|
88 |
+
gate = fused_x[..., 0, :]
|
89 |
+
up = fused_x[..., 1, :]
|
90 |
+
|
91 |
+
hidden = torch.mul(F.silu(gate), up).to(self.dtype)
|
92 |
+
|
93 |
+
output = self.wo(hidden)
|
94 |
+
return output
|
95 |
+
|
96 |
+
|
97 |
+
class RotaryEmbedding(nn.Module):
|
98 |
+
"""Rotary Position Embedding (RoPE) implementation in PyTorch."""
|
99 |
+
|
100 |
+
def __init__(
|
101 |
+
self,
|
102 |
+
embedding_dims: int,
|
103 |
+
min_timescale: int = 1,
|
104 |
+
max_timescale: int = 10000,
|
105 |
+
dtype: torch.dtype = torch.float32,
|
106 |
+
):
|
107 |
+
super().__init__()
|
108 |
+
if embedding_dims % 2 != 0:
|
109 |
+
raise ValueError("Embedding dim must be even for RoPE.")
|
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 = (self.min_timescale * (self.max_timescale / self.min_timescale) ** fraction).to(torch.float32)
|
118 |
+
self.register_buffer("timescale", timescale, persistent=False)
|
119 |
+
|
120 |
+
def forward(self, inputs: torch.Tensor, position: torch.Tensor):
|
121 |
+
"""Applies RoPE."""
|
122 |
+
position = position.unsqueeze(-1).unsqueeze(-1)
|
123 |
+
sinusoid_inp = position / self.timescale
|
124 |
+
sin = torch.sin(sinusoid_inp)
|
125 |
+
cos = torch.cos(sinusoid_inp)
|
126 |
+
first_half, second_half = torch.chunk(inputs.to(torch.float32), 2, dim=-1)
|
127 |
+
first_part = first_half * cos - second_half * sin
|
128 |
+
second_part = second_half * cos + first_half * sin
|
129 |
+
return torch.cat((first_part.to(self.compute_dtype), second_part.to(self.compute_dtype)), dim=-1)
|
130 |
+
|
131 |
+
|
132 |
+
class Attention(nn.Module):
|
133 |
+
"""Attention using DenseGeneral."""
|
134 |
+
|
135 |
+
def __init__(
|
136 |
+
self,
|
137 |
+
config: DiaConfig,
|
138 |
+
q_embed_dim: int,
|
139 |
+
kv_embed_dim: int,
|
140 |
+
num_query_heads: int,
|
141 |
+
num_kv_heads: int,
|
142 |
+
head_dim: int,
|
143 |
+
compute_dtype: torch.dtype,
|
144 |
+
is_cross_attn: bool = False,
|
145 |
+
out_embed_dim: int | None = None,
|
146 |
+
):
|
147 |
+
super().__init__()
|
148 |
+
self.num_query_heads = num_query_heads
|
149 |
+
self.num_kv_heads = num_kv_heads
|
150 |
+
self.head_dim = head_dim
|
151 |
+
self.is_cross_attn = is_cross_attn
|
152 |
+
self.output_dim = out_embed_dim if out_embed_dim is not None else q_embed_dim
|
153 |
+
self.projected_query_dim = num_query_heads * head_dim
|
154 |
+
if num_query_heads % num_kv_heads != 0:
|
155 |
+
raise ValueError(f"num_query_heads ({num_query_heads}) must be divisible by num_kv_heads ({num_kv_heads})")
|
156 |
+
self.num_gqa_groups = num_query_heads // num_kv_heads
|
157 |
+
|
158 |
+
# --- Projection Layers using DenseGeneral ---
|
159 |
+
self.q_proj = DenseGeneral(
|
160 |
+
in_shapes=(q_embed_dim,),
|
161 |
+
out_features=(num_query_heads, head_dim),
|
162 |
+
axis=(-1,),
|
163 |
+
weight_dtype=compute_dtype,
|
164 |
+
)
|
165 |
+
self.k_proj = DenseGeneral(
|
166 |
+
in_shapes=(kv_embed_dim,),
|
167 |
+
out_features=(num_kv_heads, head_dim),
|
168 |
+
axis=(-1,),
|
169 |
+
weight_dtype=compute_dtype,
|
170 |
+
)
|
171 |
+
self.v_proj = DenseGeneral(
|
172 |
+
in_shapes=(kv_embed_dim,),
|
173 |
+
out_features=(num_kv_heads, head_dim),
|
174 |
+
axis=(-1,),
|
175 |
+
weight_dtype=compute_dtype,
|
176 |
+
)
|
177 |
+
self.o_proj = DenseGeneral(
|
178 |
+
in_shapes=(num_query_heads, head_dim),
|
179 |
+
out_features=(self.output_dim,),
|
180 |
+
axis=(-2, -1),
|
181 |
+
weight_dtype=compute_dtype,
|
182 |
+
)
|
183 |
+
|
184 |
+
# --- Rotary Embedding ---
|
185 |
+
self.rotary_emb = RotaryEmbedding(
|
186 |
+
embedding_dims=self.head_dim,
|
187 |
+
min_timescale=config.model.rope_min_timescale,
|
188 |
+
max_timescale=config.model.rope_max_timescale,
|
189 |
+
dtype=compute_dtype,
|
190 |
+
)
|
191 |
+
|
192 |
+
def forward(
|
193 |
+
self,
|
194 |
+
Xq: torch.Tensor, # (B, T, D) T = 1 in AR generation
|
195 |
+
Xkv: torch.Tensor, # (B, S, E) S = 1 in AR generation
|
196 |
+
q_positions: torch.Tensor, # (B, T)
|
197 |
+
kv_positions: torch.Tensor | None = None, # (B, S)
|
198 |
+
attn_mask: torch.Tensor | None = None, # None in Decoder Self Attention, Valid mask in Others
|
199 |
+
cache: KVCache | None = None, # None in Encoder, KVCache in Decoder
|
200 |
+
prefill: bool = False,
|
201 |
+
is_causal: bool = False,
|
202 |
+
) -> tuple[torch.Tensor, tuple[torch.Tensor, torch.Tensor] | None]:
|
203 |
+
"""
|
204 |
+
Performs attention calculation with optional KV caching.
|
205 |
+
|
206 |
+
Args:
|
207 |
+
Xq: Query tensor (B, T, D). T=1 during single-step decoding.
|
208 |
+
Xkv: Key/Value source tensor (B, S, E). S=1 during single-step decoding for self-attn.
|
209 |
+
q_positions: Positions for queries (B, T).
|
210 |
+
kv_positions: Positions for keys/values (B, S). If None, uses q_positions.
|
211 |
+
attn_mask: Attention mask.
|
212 |
+
cache: KVCache.
|
213 |
+
prefill: If True, use prefill mode.
|
214 |
+
|
215 |
+
Returns:
|
216 |
+
A tuple containing:
|
217 |
+
- output: The attention output tensor (B, T, output_dim).
|
218 |
+
- 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.
|
219 |
+
"""
|
220 |
+
if kv_positions is None:
|
221 |
+
kv_positions = q_positions
|
222 |
+
original_dtype = Xq.dtype
|
223 |
+
|
224 |
+
Xq_BxTxNxH = self.q_proj(Xq)
|
225 |
+
Xq_BxTxNxH = self.rotary_emb(Xq_BxTxNxH, position=q_positions)
|
226 |
+
Xq_BxNxTxH = Xq_BxTxNxH.transpose(1, 2)
|
227 |
+
|
228 |
+
attn_k: torch.Tensor | None = None
|
229 |
+
attn_v: torch.Tensor | None = None
|
230 |
+
|
231 |
+
if self.is_cross_attn:
|
232 |
+
attn_k, attn_v = cache.k, cache.v
|
233 |
+
else:
|
234 |
+
Xk_BxSxKxH = self.k_proj(Xkv) # (B, S, K, H)
|
235 |
+
Xv_BxSxKxH = self.v_proj(Xkv) # (B, S, K, H)
|
236 |
+
Xk_BxSxKxH = self.rotary_emb(Xk_BxSxKxH, position=kv_positions) # (B, S, K, H)
|
237 |
+
|
238 |
+
Xk_BxKxSxH = Xk_BxSxKxH.transpose(1, 2) # (B, K, S, H)
|
239 |
+
Xv_BxKxSxH = Xv_BxSxKxH.transpose(1, 2) # (B, K, S, H)
|
240 |
+
|
241 |
+
if cache is None:
|
242 |
+
attn_k = Xk_BxKxSxH
|
243 |
+
attn_v = Xv_BxKxSxH
|
244 |
+
else:
|
245 |
+
if prefill:
|
246 |
+
attn_k, attn_v = Xk_BxKxSxH, Xv_BxKxSxH
|
247 |
+
cache.prefill(attn_k, attn_v)
|
248 |
+
else:
|
249 |
+
attn_k, attn_v = cache.update(Xk_BxKxSxH, Xv_BxKxSxH)
|
250 |
+
|
251 |
+
attn_output = F.scaled_dot_product_attention(
|
252 |
+
Xq_BxNxTxH,
|
253 |
+
attn_k,
|
254 |
+
attn_v,
|
255 |
+
attn_mask=attn_mask,
|
256 |
+
scale=1.0,
|
257 |
+
enable_gqa=self.num_gqa_groups > 1,
|
258 |
+
is_causal=is_causal,
|
259 |
+
)
|
260 |
+
|
261 |
+
attn_output = attn_output.transpose(1, 2).contiguous() # (B, T, N, H)
|
262 |
+
output = self.o_proj(attn_output)
|
263 |
+
|
264 |
+
return output.to(original_dtype)
|
265 |
+
|
266 |
+
|
267 |
+
class EncoderLayer(nn.Module):
|
268 |
+
"""Transformer Encoder Layer using DenseGeneral."""
|
269 |
+
|
270 |
+
def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
|
271 |
+
super().__init__()
|
272 |
+
self.config = config
|
273 |
+
model_config = config.model
|
274 |
+
enc_config = config.model.encoder
|
275 |
+
embed_dim = enc_config.n_embd
|
276 |
+
self.compute_dtype = compute_dtype
|
277 |
+
|
278 |
+
self.pre_sa_norm = RMSNorm(
|
279 |
+
embed_dim,
|
280 |
+
eps=model_config.normalization_layer_epsilon,
|
281 |
+
dtype=torch.float32,
|
282 |
+
)
|
283 |
+
self.self_attention = Attention(
|
284 |
+
config,
|
285 |
+
q_embed_dim=embed_dim,
|
286 |
+
kv_embed_dim=embed_dim,
|
287 |
+
num_query_heads=enc_config.n_head,
|
288 |
+
num_kv_heads=enc_config.n_head,
|
289 |
+
head_dim=enc_config.head_dim,
|
290 |
+
compute_dtype=compute_dtype,
|
291 |
+
is_cross_attn=False,
|
292 |
+
out_embed_dim=embed_dim,
|
293 |
+
)
|
294 |
+
self.post_sa_norm = RMSNorm(
|
295 |
+
embed_dim,
|
296 |
+
eps=model_config.normalization_layer_epsilon,
|
297 |
+
dtype=torch.float32,
|
298 |
+
)
|
299 |
+
self.mlp = MlpBlock(embed_dim=embed_dim, intermediate_dim=enc_config.n_hidden, compute_dtype=compute_dtype)
|
300 |
+
|
301 |
+
def forward(
|
302 |
+
self,
|
303 |
+
x: torch.Tensor,
|
304 |
+
state: EncoderInferenceState,
|
305 |
+
) -> torch.Tensor:
|
306 |
+
residual = x
|
307 |
+
x_norm = self.pre_sa_norm(x).to(self.compute_dtype)
|
308 |
+
|
309 |
+
sa_out = self.self_attention(
|
310 |
+
Xq=x_norm,
|
311 |
+
Xkv=x_norm,
|
312 |
+
q_positions=state.positions,
|
313 |
+
kv_positions=state.positions,
|
314 |
+
attn_mask=state.attn_mask,
|
315 |
+
)
|
316 |
+
x = residual + sa_out
|
317 |
+
|
318 |
+
residual = x
|
319 |
+
x_norm = self.post_sa_norm(x).to(self.compute_dtype)
|
320 |
+
mlp_out = self.mlp(x_norm)
|
321 |
+
x = residual + mlp_out
|
322 |
+
|
323 |
+
return x
|
324 |
+
|
325 |
+
|
326 |
+
class Encoder(nn.Module):
|
327 |
+
"""Transformer Encoder Stack using DenseGeneral."""
|
328 |
+
|
329 |
+
def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
|
330 |
+
super().__init__()
|
331 |
+
self.config = config
|
332 |
+
model_config = config.model
|
333 |
+
enc_config = config.model.encoder
|
334 |
+
self.compute_dtype = compute_dtype
|
335 |
+
|
336 |
+
self.embedding = nn.Embedding(
|
337 |
+
model_config.src_vocab_size,
|
338 |
+
enc_config.n_embd,
|
339 |
+
dtype=compute_dtype,
|
340 |
+
)
|
341 |
+
self.layers = nn.ModuleList([EncoderLayer(config, compute_dtype) for _ in range(enc_config.n_layer)])
|
342 |
+
self.norm = RMSNorm(
|
343 |
+
enc_config.n_embd,
|
344 |
+
eps=model_config.normalization_layer_epsilon,
|
345 |
+
dtype=torch.float32,
|
346 |
+
)
|
347 |
+
|
348 |
+
def forward(
|
349 |
+
self,
|
350 |
+
x_ids: torch.Tensor,
|
351 |
+
state: EncoderInferenceState,
|
352 |
+
) -> torch.Tensor:
|
353 |
+
x = self.embedding(x_ids)
|
354 |
+
|
355 |
+
for layer in self.layers:
|
356 |
+
x = layer(x, state)
|
357 |
+
|
358 |
+
x = self.norm(x).to(self.compute_dtype)
|
359 |
+
return x
|
360 |
+
|
361 |
+
|
362 |
+
class DecoderLayer(nn.Module):
|
363 |
+
"""Transformer Decoder Layer using DenseGeneral."""
|
364 |
+
|
365 |
+
def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
|
366 |
+
super().__init__()
|
367 |
+
self.config = config
|
368 |
+
model_config = config.model
|
369 |
+
dec_config = config.model.decoder
|
370 |
+
enc_config = config.model.encoder
|
371 |
+
dec_embed_dim = dec_config.n_embd
|
372 |
+
enc_embed_dim = enc_config.n_embd
|
373 |
+
self.compute_dtype = compute_dtype
|
374 |
+
|
375 |
+
# Norms
|
376 |
+
self.pre_sa_norm = RMSNorm(
|
377 |
+
dec_embed_dim,
|
378 |
+
eps=model_config.normalization_layer_epsilon,
|
379 |
+
dtype=torch.float32,
|
380 |
+
)
|
381 |
+
self.pre_ca_norm = RMSNorm(
|
382 |
+
dec_embed_dim,
|
383 |
+
eps=model_config.normalization_layer_epsilon,
|
384 |
+
dtype=torch.float32,
|
385 |
+
)
|
386 |
+
self.pre_mlp_norm = RMSNorm(
|
387 |
+
dec_embed_dim,
|
388 |
+
eps=model_config.normalization_layer_epsilon,
|
389 |
+
dtype=torch.float32,
|
390 |
+
)
|
391 |
+
|
392 |
+
# Self-Attention (GQA) with Causal Masking
|
393 |
+
self.self_attention = Attention(
|
394 |
+
config,
|
395 |
+
q_embed_dim=dec_embed_dim,
|
396 |
+
kv_embed_dim=dec_embed_dim,
|
397 |
+
num_query_heads=dec_config.gqa_query_heads,
|
398 |
+
num_kv_heads=dec_config.kv_heads,
|
399 |
+
head_dim=dec_config.gqa_head_dim,
|
400 |
+
compute_dtype=compute_dtype,
|
401 |
+
is_cross_attn=False,
|
402 |
+
out_embed_dim=dec_embed_dim,
|
403 |
+
)
|
404 |
+
# Cross-Attention (MHA)
|
405 |
+
self.cross_attention = Attention(
|
406 |
+
config=config,
|
407 |
+
q_embed_dim=dec_embed_dim,
|
408 |
+
kv_embed_dim=enc_embed_dim, # Note kv_embed_dim
|
409 |
+
num_query_heads=dec_config.cross_query_heads,
|
410 |
+
num_kv_heads=dec_config.cross_query_heads,
|
411 |
+
head_dim=dec_config.cross_head_dim,
|
412 |
+
compute_dtype=compute_dtype,
|
413 |
+
is_cross_attn=True,
|
414 |
+
out_embed_dim=dec_embed_dim,
|
415 |
+
)
|
416 |
+
# MLP
|
417 |
+
self.mlp = MlpBlock(
|
418 |
+
embed_dim=dec_embed_dim,
|
419 |
+
intermediate_dim=dec_config.n_hidden,
|
420 |
+
compute_dtype=compute_dtype,
|
421 |
+
)
|
422 |
+
|
423 |
+
def forward(
|
424 |
+
self,
|
425 |
+
x: torch.Tensor,
|
426 |
+
state: DecoderInferenceState,
|
427 |
+
self_attn_cache: KVCache | None = None,
|
428 |
+
cross_attn_cache: KVCache | None = None,
|
429 |
+
prefill: bool = False,
|
430 |
+
) -> torch.Tensor:
|
431 |
+
residual = x
|
432 |
+
x_norm = self.pre_sa_norm(x).to(self.compute_dtype)
|
433 |
+
|
434 |
+
sa_out = self.self_attention(
|
435 |
+
Xq=x_norm, # (2, 1, D)
|
436 |
+
Xkv=x_norm, # (2, 1, D)
|
437 |
+
q_positions=state.dec_positions, # (2, 1)
|
438 |
+
kv_positions=state.dec_positions, # (2, 1)
|
439 |
+
attn_mask=None,
|
440 |
+
cache=self_attn_cache,
|
441 |
+
prefill=prefill,
|
442 |
+
is_causal=prefill,
|
443 |
+
)
|
444 |
+
|
445 |
+
x = residual + sa_out
|
446 |
+
|
447 |
+
residual = x
|
448 |
+
x_norm = self.pre_ca_norm(x).to(self.compute_dtype)
|
449 |
+
ca_out = self.cross_attention(
|
450 |
+
Xq=x_norm,
|
451 |
+
Xkv=state.enc_out,
|
452 |
+
q_positions=state.dec_positions,
|
453 |
+
kv_positions=state.enc_positions,
|
454 |
+
attn_mask=state.dec_cross_attn_mask,
|
455 |
+
cache=cross_attn_cache,
|
456 |
+
)
|
457 |
+
x = residual + ca_out
|
458 |
+
|
459 |
+
residual = x
|
460 |
+
x_norm = self.pre_mlp_norm(x).to(self.compute_dtype)
|
461 |
+
mlp_out = self.mlp(x_norm)
|
462 |
+
x = residual + mlp_out
|
463 |
+
|
464 |
+
return x
|
465 |
+
|
466 |
+
|
467 |
+
class Decoder(nn.Module):
|
468 |
+
"""Transformer Decoder Stack using DenseGeneral."""
|
469 |
+
|
470 |
+
def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
|
471 |
+
super().__init__()
|
472 |
+
self.config = config
|
473 |
+
model_config = config.model
|
474 |
+
dec_config = config.model.decoder
|
475 |
+
data_config = config.data
|
476 |
+
self.num_channels = data_config.channels
|
477 |
+
self.num_layers = dec_config.n_layer
|
478 |
+
|
479 |
+
self.embeddings = nn.ModuleList(
|
480 |
+
[
|
481 |
+
nn.Embedding(model_config.tgt_vocab_size, dec_config.n_embd, dtype=compute_dtype)
|
482 |
+
for _ in range(self.num_channels)
|
483 |
+
]
|
484 |
+
)
|
485 |
+
self.layers = nn.ModuleList(
|
486 |
+
[DecoderLayer(config=config, compute_dtype=compute_dtype) for _ in range(self.num_layers)]
|
487 |
+
)
|
488 |
+
|
489 |
+
self.norm = RMSNorm(
|
490 |
+
dec_config.n_embd,
|
491 |
+
eps=model_config.normalization_layer_epsilon,
|
492 |
+
dtype=torch.float32,
|
493 |
+
)
|
494 |
+
|
495 |
+
self.logits_dense = DenseGeneral(
|
496 |
+
in_shapes=(dec_config.n_embd,),
|
497 |
+
out_features=(self.num_channels, model_config.tgt_vocab_size),
|
498 |
+
axis=(-1,),
|
499 |
+
weight_dtype=compute_dtype,
|
500 |
+
)
|
501 |
+
|
502 |
+
def precompute_cross_attn_cache(
|
503 |
+
self,
|
504 |
+
enc_out: torch.Tensor, # (B, S, E)
|
505 |
+
enc_positions: torch.Tensor, # (B, S)
|
506 |
+
) -> list[KVCache]:
|
507 |
+
"""
|
508 |
+
Computes the Key and Value tensors for cross-attention for each layer from the encoder output.
|
509 |
+
"""
|
510 |
+
per_layer_kv_cache: list[KVCache] = []
|
511 |
+
|
512 |
+
for layer in self.layers:
|
513 |
+
cross_attn_module = layer.cross_attention
|
514 |
+
k_proj = cross_attn_module.k_proj(enc_out)
|
515 |
+
v_proj = cross_attn_module.v_proj(enc_out)
|
516 |
+
|
517 |
+
k_proj = cross_attn_module.rotary_emb(k_proj, position=enc_positions)
|
518 |
+
k = k_proj.transpose(1, 2)
|
519 |
+
v = v_proj.transpose(1, 2)
|
520 |
+
|
521 |
+
per_layer_kv_cache.append(KVCache.from_kv(k, v))
|
522 |
+
|
523 |
+
return per_layer_kv_cache
|
524 |
+
|
525 |
+
def decode_step(
|
526 |
+
self,
|
527 |
+
tgt_ids_Bx1xC: torch.Tensor, # [B, 1, C]
|
528 |
+
state: DecoderInferenceState,
|
529 |
+
) -> torch.Tensor:
|
530 |
+
"""
|
531 |
+
Performs a single decoding step, managing KV caches layer by layer.
|
532 |
+
|
533 |
+
Returns:
|
534 |
+
A tuple containing:
|
535 |
+
- logits_Bx1xCV: The final output logits for the current step (B, 1, C*V), cast to float32.
|
536 |
+
"""
|
537 |
+
|
538 |
+
x = None
|
539 |
+
for i in range(self.num_channels):
|
540 |
+
channel_tokens = tgt_ids_Bx1xC[..., i]
|
541 |
+
channel_embed = self.embeddings[i](channel_tokens)
|
542 |
+
x = channel_embed if x is None else x + channel_embed
|
543 |
+
|
544 |
+
for i, layer in enumerate(self.layers):
|
545 |
+
self_cache = state.self_attn_cache[i]
|
546 |
+
cross_cache = state.cross_attn_cache[i]
|
547 |
+
x = layer(
|
548 |
+
x, # (2, 1, D)
|
549 |
+
state,
|
550 |
+
self_attn_cache=self_cache,
|
551 |
+
cross_attn_cache=cross_cache,
|
552 |
+
)
|
553 |
+
|
554 |
+
x = self.norm(x)
|
555 |
+
logits_Bx1xCxV = self.logits_dense(x)
|
556 |
+
|
557 |
+
return logits_Bx1xCxV.to(torch.float32)
|
558 |
+
|
559 |
+
def forward(self, tgt_ids_BxTxC: torch.Tensor, state: DecoderInferenceState) -> torch.Tensor:
|
560 |
+
"""
|
561 |
+
Forward pass for the Decoder stack, managing KV caches.
|
562 |
+
|
563 |
+
Args:
|
564 |
+
tgt_ids_BxTxC: Target token IDs (B, T, C).
|
565 |
+
encoder_out: Output from the encoder (B, S, E).
|
566 |
+
tgt_positions: Positions for target sequence (B, T).
|
567 |
+
src_positions: Positions for source sequence (B, S).
|
568 |
+
self_attn_mask: Mask for self-attention.
|
569 |
+
cross_attn_mask: Mask for cross-attention.
|
570 |
+
past_key_values: List containing the self-attention KV cache for each layer
|
571 |
+
from the previous decoding step. `len(past_key_values)` should
|
572 |
+
equal `num_layers`.
|
573 |
+
precomputed_cross_attn_kv: A single tuple containing the pre-computed K/V cache
|
574 |
+
derived from `encoder_out`. This is passed identically
|
575 |
+
to all layers.
|
576 |
+
|
577 |
+
Returns:
|
578 |
+
A tuple containing:
|
579 |
+
- logits: The final output logits (B, T, C * V), cast to float32.
|
580 |
+
- present_key_values: A list containing the updated self-attention KV cache
|
581 |
+
for each layer for the *current* decoding step.
|
582 |
+
"""
|
583 |
+
_, _, num_channels_in = tgt_ids_BxTxC.shape
|
584 |
+
assert num_channels_in == self.num_channels, "Input channels mismatch"
|
585 |
+
|
586 |
+
# Embeddings
|
587 |
+
x = None
|
588 |
+
for i in range(self.num_channels):
|
589 |
+
channel_tokens = tgt_ids_BxTxC[..., i]
|
590 |
+
channel_embed = self.embeddings[i](channel_tokens)
|
591 |
+
x = channel_embed if x is None else x + channel_embed
|
592 |
+
|
593 |
+
for i, layer in enumerate(self.layers):
|
594 |
+
self_cache = state.self_attn_cache[i]
|
595 |
+
cross_cache = state.cross_attn_cache[i]
|
596 |
+
x = layer(x, state, self_attn_cache=self_cache, cross_attn_cache=cross_cache, prefill=True)
|
597 |
+
|
598 |
+
# Final Norm
|
599 |
+
x = self.norm(x)
|
600 |
+
logits_BxTxCxV = self.logits_dense(x)
|
601 |
+
|
602 |
+
return logits_BxTxCxV.to(torch.float32)
|
603 |
+
|
604 |
+
|
605 |
+
class DiaModel(
|
606 |
+
nn.Module,
|
607 |
+
PyTorchModelHubMixin,
|
608 |
+
repo_url="https://github.com/nari-labs/dia",
|
609 |
+
pipeline_tag="text-to-speech",
|
610 |
+
license="apache-2.0",
|
611 |
+
coders={
|
612 |
+
DiaConfig: (
|
613 |
+
lambda x: x.model_dump(),
|
614 |
+
lambda data: DiaConfig.model_validate(data),
|
615 |
+
),
|
616 |
+
},
|
617 |
+
):
|
618 |
+
"""PyTorch Dia Model using DenseGeneral."""
|
619 |
+
|
620 |
+
def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
|
621 |
+
super().__init__()
|
622 |
+
self.config = config
|
623 |
+
self.encoder = Encoder(config, compute_dtype)
|
624 |
+
self.decoder = Decoder(config, compute_dtype)
|
dia/model.py
ADDED
@@ -0,0 +1,455 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import time
|
2 |
+
from enum import Enum
|
3 |
+
|
4 |
+
import dac
|
5 |
+
import numpy as np
|
6 |
+
import torch
|
7 |
+
import torchaudio
|
8 |
+
|
9 |
+
from .audio import apply_audio_delay, build_delay_indices, build_revert_indices, decode, revert_audio_delay
|
10 |
+
from .config import DiaConfig
|
11 |
+
from .layers import DiaModel
|
12 |
+
from .state import DecoderInferenceState, DecoderOutput, EncoderInferenceState
|
13 |
+
|
14 |
+
|
15 |
+
DEFAULT_SAMPLE_RATE = 44100
|
16 |
+
|
17 |
+
|
18 |
+
def _get_default_device():
|
19 |
+
if torch.cuda.is_available():
|
20 |
+
return torch.device("cuda")
|
21 |
+
elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
|
22 |
+
return torch.device("mps")
|
23 |
+
return torch.device("cpu")
|
24 |
+
|
25 |
+
|
26 |
+
def _sample_next_token(
|
27 |
+
logits_BCxV: torch.Tensor,
|
28 |
+
temperature: float,
|
29 |
+
top_p: float,
|
30 |
+
cfg_filter_top_k: int | None = None,
|
31 |
+
) -> torch.Tensor:
|
32 |
+
if temperature == 0.0:
|
33 |
+
return torch.argmax(logits_BCxV, dim=-1)
|
34 |
+
|
35 |
+
logits_BCxV = logits_BCxV / temperature
|
36 |
+
if cfg_filter_top_k is not None:
|
37 |
+
_, top_k_indices_BCxV = torch.topk(logits_BCxV, k=cfg_filter_top_k, dim=-1)
|
38 |
+
mask = torch.ones_like(logits_BCxV, dtype=torch.bool)
|
39 |
+
mask.scatter_(dim=-1, index=top_k_indices_BCxV, value=False)
|
40 |
+
logits_BCxV = logits_BCxV.masked_fill(mask, -torch.inf)
|
41 |
+
|
42 |
+
if top_p < 1.0:
|
43 |
+
probs_BCxV = torch.softmax(logits_BCxV, dim=-1)
|
44 |
+
sorted_probs_BCxV, sorted_indices_BCxV = torch.sort(probs_BCxV, dim=-1, descending=True)
|
45 |
+
cumulative_probs_BCxV = torch.cumsum(sorted_probs_BCxV, dim=-1)
|
46 |
+
|
47 |
+
sorted_indices_to_remove_BCxV = cumulative_probs_BCxV > top_p
|
48 |
+
sorted_indices_to_remove_BCxV[..., 1:] = sorted_indices_to_remove_BCxV[..., :-1].clone()
|
49 |
+
sorted_indices_to_remove_BCxV[..., 0] = 0
|
50 |
+
|
51 |
+
indices_to_remove_BCxV = torch.zeros_like(sorted_indices_to_remove_BCxV)
|
52 |
+
indices_to_remove_BCxV.scatter_(dim=-1, index=sorted_indices_BCxV, src=sorted_indices_to_remove_BCxV)
|
53 |
+
logits_BCxV = logits_BCxV.masked_fill(indices_to_remove_BCxV, -torch.inf)
|
54 |
+
|
55 |
+
final_probs_BCxV = torch.softmax(logits_BCxV, dim=-1)
|
56 |
+
|
57 |
+
sampled_indices_BC = torch.multinomial(final_probs_BCxV, num_samples=1)
|
58 |
+
sampled_indices_C = sampled_indices_BC.squeeze(-1)
|
59 |
+
return sampled_indices_C
|
60 |
+
|
61 |
+
|
62 |
+
class ComputeDtype(str, Enum):
|
63 |
+
FLOAT32 = "float32"
|
64 |
+
FLOAT16 = "float16"
|
65 |
+
BFLOAT16 = "bfloat16"
|
66 |
+
|
67 |
+
def to_dtype(self) -> torch.dtype:
|
68 |
+
if self == ComputeDtype.FLOAT32:
|
69 |
+
return torch.float32
|
70 |
+
elif self == ComputeDtype.FLOAT16:
|
71 |
+
return torch.float16
|
72 |
+
elif self == ComputeDtype.BFLOAT16:
|
73 |
+
return torch.bfloat16
|
74 |
+
else:
|
75 |
+
raise ValueError(f"Unsupported compute dtype: {self}")
|
76 |
+
|
77 |
+
|
78 |
+
class Dia:
|
79 |
+
def __init__(
|
80 |
+
self,
|
81 |
+
config: DiaConfig,
|
82 |
+
compute_dtype: str | ComputeDtype = ComputeDtype.FLOAT32,
|
83 |
+
device: torch.device | None = None,
|
84 |
+
):
|
85 |
+
"""Initializes the Dia model.
|
86 |
+
|
87 |
+
Args:
|
88 |
+
config: The configuration object for the model.
|
89 |
+
device: The device to load the model onto. If None, will automatically select the best available device.
|
90 |
+
|
91 |
+
Raises:
|
92 |
+
RuntimeError: If there is an error loading the DAC model.
|
93 |
+
"""
|
94 |
+
super().__init__()
|
95 |
+
self.config = config
|
96 |
+
self.device = device if device is not None else _get_default_device()
|
97 |
+
if isinstance(compute_dtype, str):
|
98 |
+
compute_dtype = ComputeDtype(compute_dtype)
|
99 |
+
self.compute_dtype = compute_dtype.to_dtype()
|
100 |
+
self.model = DiaModel(config, self.compute_dtype)
|
101 |
+
self.dac_model = None
|
102 |
+
|
103 |
+
@classmethod
|
104 |
+
def from_local(
|
105 |
+
cls,
|
106 |
+
config_path: str,
|
107 |
+
checkpoint_path: str,
|
108 |
+
compute_dtype: str | ComputeDtype = ComputeDtype.FLOAT32,
|
109 |
+
device: torch.device | None = None,
|
110 |
+
) -> "Dia":
|
111 |
+
"""Loads the Dia model from local configuration and checkpoint files.
|
112 |
+
|
113 |
+
Args:
|
114 |
+
config_path: Path to the configuration JSON file.
|
115 |
+
checkpoint_path: Path to the model checkpoint (.pth) file.
|
116 |
+
device: The device to load the model onto. If None, will automatically select the best available device.
|
117 |
+
|
118 |
+
Returns:
|
119 |
+
An instance of the Dia model loaded with weights and set to eval mode.
|
120 |
+
|
121 |
+
Raises:
|
122 |
+
FileNotFoundError: If the config or checkpoint file is not found.
|
123 |
+
RuntimeError: If there is an error loading the checkpoint.
|
124 |
+
"""
|
125 |
+
config = DiaConfig.load(config_path)
|
126 |
+
if config is None:
|
127 |
+
raise FileNotFoundError(f"Config file not found at {config_path}")
|
128 |
+
|
129 |
+
dia = cls(config, compute_dtype, device)
|
130 |
+
|
131 |
+
try:
|
132 |
+
state_dict = torch.load(checkpoint_path, map_location=dia.device)
|
133 |
+
dia.model.load_state_dict(state_dict)
|
134 |
+
except FileNotFoundError:
|
135 |
+
raise FileNotFoundError(f"Checkpoint file not found at {checkpoint_path}")
|
136 |
+
except Exception as e:
|
137 |
+
raise RuntimeError(f"Error loading checkpoint from {checkpoint_path}") from e
|
138 |
+
|
139 |
+
dia.model.to(dia.device)
|
140 |
+
dia.model.eval()
|
141 |
+
dia._load_dac_model()
|
142 |
+
return dia
|
143 |
+
|
144 |
+
@classmethod
|
145 |
+
def from_pretrained(
|
146 |
+
cls,
|
147 |
+
model_name: str = "nari-labs/Dia-1.6B",
|
148 |
+
compute_dtype: str | ComputeDtype = ComputeDtype.FLOAT32,
|
149 |
+
device: torch.device | None = None,
|
150 |
+
) -> "Dia":
|
151 |
+
"""Loads the Dia model from a Hugging Face Hub repository.
|
152 |
+
|
153 |
+
Downloads the configuration and checkpoint files from the specified
|
154 |
+
repository ID and then loads the model.
|
155 |
+
|
156 |
+
Args:
|
157 |
+
model_name: The Hugging Face Hub repository ID (e.g., "nari-labs/Dia-1.6B").
|
158 |
+
compute_dtype: The computation dtype to use.
|
159 |
+
device: The device to load the model onto. If None, will automatically select the best available device.
|
160 |
+
|
161 |
+
Returns:
|
162 |
+
An instance of the Dia model loaded with weights and set to eval mode.
|
163 |
+
|
164 |
+
Raises:
|
165 |
+
FileNotFoundError: If config or checkpoint download/loading fails.
|
166 |
+
RuntimeError: If there is an error loading the checkpoint.
|
167 |
+
"""
|
168 |
+
if isinstance(compute_dtype, str):
|
169 |
+
compute_dtype = ComputeDtype(compute_dtype)
|
170 |
+
loaded_model = DiaModel.from_pretrained(model_name, compute_dtype=compute_dtype.to_dtype())
|
171 |
+
config = loaded_model.config
|
172 |
+
dia = cls(config, compute_dtype, device)
|
173 |
+
|
174 |
+
dia.model = loaded_model
|
175 |
+
dia.model.to(dia.device)
|
176 |
+
dia.model.eval()
|
177 |
+
dia._load_dac_model()
|
178 |
+
return dia
|
179 |
+
|
180 |
+
def _load_dac_model(self):
|
181 |
+
try:
|
182 |
+
dac_model_path = dac.utils.download()
|
183 |
+
dac_model = dac.DAC.load(dac_model_path).to(self.device)
|
184 |
+
except Exception as e:
|
185 |
+
raise RuntimeError("Failed to load DAC model") from e
|
186 |
+
self.dac_model = dac_model
|
187 |
+
|
188 |
+
def _prepare_text_input(self, text: str) -> torch.Tensor:
|
189 |
+
"""Encodes text prompt, pads, and creates attention mask and positions."""
|
190 |
+
text_pad_value = self.config.data.text_pad_value
|
191 |
+
max_len = self.config.data.text_length
|
192 |
+
|
193 |
+
byte_text = text.encode("utf-8")
|
194 |
+
replaced_bytes = byte_text.replace(b"[S1]", b"\x01").replace(b"[S2]", b"\x02")
|
195 |
+
text_tokens = list(replaced_bytes)
|
196 |
+
|
197 |
+
current_len = len(text_tokens)
|
198 |
+
padding_needed = max_len - current_len
|
199 |
+
if padding_needed <= 0:
|
200 |
+
text_tokens = text_tokens[:max_len]
|
201 |
+
padded_text_np = np.array(text_tokens, dtype=np.uint8)
|
202 |
+
else:
|
203 |
+
padded_text_np = np.pad(
|
204 |
+
text_tokens,
|
205 |
+
(0, padding_needed),
|
206 |
+
mode="constant",
|
207 |
+
constant_values=text_pad_value,
|
208 |
+
).astype(np.uint8)
|
209 |
+
|
210 |
+
src_tokens = torch.from_numpy(padded_text_np).to(torch.long).to(self.device).unsqueeze(0) # [1, S]
|
211 |
+
return src_tokens
|
212 |
+
|
213 |
+
def _prepare_audio_prompt(self, audio_prompt: torch.Tensor | None) -> tuple[torch.Tensor, int]:
|
214 |
+
num_channels = self.config.data.channels
|
215 |
+
audio_bos_value = self.config.data.audio_bos_value
|
216 |
+
audio_pad_value = self.config.data.audio_pad_value
|
217 |
+
delay_pattern = self.config.data.delay_pattern
|
218 |
+
max_delay_pattern = max(delay_pattern)
|
219 |
+
|
220 |
+
prefill = torch.full(
|
221 |
+
(1, num_channels),
|
222 |
+
fill_value=audio_bos_value,
|
223 |
+
dtype=torch.int,
|
224 |
+
device=self.device,
|
225 |
+
)
|
226 |
+
|
227 |
+
prefill_step = 1
|
228 |
+
|
229 |
+
if audio_prompt is not None:
|
230 |
+
prefill_step += audio_prompt.shape[0]
|
231 |
+
prefill = torch.cat([prefill, audio_prompt], dim=0)
|
232 |
+
|
233 |
+
delay_pad_tensor = torch.full(
|
234 |
+
(max_delay_pattern, num_channels), fill_value=-1, dtype=torch.int, device=self.device
|
235 |
+
)
|
236 |
+
prefill = torch.cat([prefill, delay_pad_tensor], dim=0)
|
237 |
+
|
238 |
+
delay_precomp = build_delay_indices(
|
239 |
+
B=1,
|
240 |
+
T=prefill.shape[0],
|
241 |
+
C=num_channels,
|
242 |
+
delay_pattern=delay_pattern,
|
243 |
+
)
|
244 |
+
|
245 |
+
prefill = apply_audio_delay(
|
246 |
+
audio_BxTxC=prefill.unsqueeze(0),
|
247 |
+
pad_value=audio_pad_value,
|
248 |
+
bos_value=audio_bos_value,
|
249 |
+
precomp=delay_precomp,
|
250 |
+
).squeeze(0)
|
251 |
+
|
252 |
+
return prefill, prefill_step
|
253 |
+
|
254 |
+
def _prepare_generation(self, text: str, audio_prompt: str | torch.Tensor | None, verbose: bool):
|
255 |
+
enc_input_cond = self._prepare_text_input(text)
|
256 |
+
enc_input_uncond = torch.zeros_like(enc_input_cond)
|
257 |
+
enc_input = torch.cat([enc_input_uncond, enc_input_cond], dim=0)
|
258 |
+
|
259 |
+
if isinstance(audio_prompt, str):
|
260 |
+
audio_prompt = self.load_audio(audio_prompt)
|
261 |
+
prefill, prefill_step = self._prepare_audio_prompt(audio_prompt)
|
262 |
+
|
263 |
+
if verbose:
|
264 |
+
print("generate: data loaded")
|
265 |
+
|
266 |
+
enc_state = EncoderInferenceState.new(self.config, enc_input_cond)
|
267 |
+
encoder_out = self.model.encoder(enc_input, enc_state)
|
268 |
+
|
269 |
+
dec_cross_attn_cache = self.model.decoder.precompute_cross_attn_cache(encoder_out, enc_state.positions)
|
270 |
+
dec_state = DecoderInferenceState.new(
|
271 |
+
self.config, enc_state, encoder_out, dec_cross_attn_cache, self.compute_dtype
|
272 |
+
)
|
273 |
+
dec_output = DecoderOutput.new(self.config, self.device)
|
274 |
+
dec_output.prefill(prefill, prefill_step)
|
275 |
+
|
276 |
+
dec_step = prefill_step - 1
|
277 |
+
if dec_step > 0:
|
278 |
+
dec_state.prepare_step(0, dec_step)
|
279 |
+
tokens_BxTxC = dec_output.get_tokens_at(0, dec_step).unsqueeze(0).expand(2, -1, -1)
|
280 |
+
self.model.decoder.forward(tokens_BxTxC, dec_state)
|
281 |
+
|
282 |
+
return dec_state, dec_output
|
283 |
+
|
284 |
+
def _decoder_step(
|
285 |
+
self,
|
286 |
+
tokens_Bx1xC: torch.Tensor,
|
287 |
+
dec_state: DecoderInferenceState,
|
288 |
+
cfg_scale: float,
|
289 |
+
temperature: float,
|
290 |
+
top_p: float,
|
291 |
+
cfg_filter_top_k: int,
|
292 |
+
) -> torch.Tensor:
|
293 |
+
audio_eos_value = self.config.data.audio_eos_value
|
294 |
+
logits_Bx1xCxV = self.model.decoder.decode_step(tokens_Bx1xC, dec_state)
|
295 |
+
|
296 |
+
logits_last_BxCxV = logits_Bx1xCxV[:, -1, :, :]
|
297 |
+
uncond_logits_CxV = logits_last_BxCxV[0, :, :]
|
298 |
+
cond_logits_CxV = logits_last_BxCxV[1, :, :]
|
299 |
+
|
300 |
+
logits_CxV = cond_logits_CxV + cfg_scale * (cond_logits_CxV - uncond_logits_CxV)
|
301 |
+
logits_CxV[:, audio_eos_value + 1 :] = -torch.inf
|
302 |
+
logits_CxV[1:, audio_eos_value:] = -torch.inf
|
303 |
+
|
304 |
+
pred_C = _sample_next_token(
|
305 |
+
logits_CxV.float(),
|
306 |
+
temperature=temperature,
|
307 |
+
top_p=top_p,
|
308 |
+
cfg_filter_top_k=cfg_filter_top_k,
|
309 |
+
)
|
310 |
+
return pred_C
|
311 |
+
|
312 |
+
def _generate_output(self, generated_codes: torch.Tensor) -> np.ndarray:
|
313 |
+
num_channels = self.config.data.channels
|
314 |
+
seq_length = generated_codes.shape[0]
|
315 |
+
delay_pattern = self.config.data.delay_pattern
|
316 |
+
audio_pad_value = self.config.data.audio_pad_value
|
317 |
+
max_delay_pattern = max(delay_pattern)
|
318 |
+
|
319 |
+
revert_precomp = build_revert_indices(
|
320 |
+
B=1,
|
321 |
+
T=seq_length,
|
322 |
+
C=num_channels,
|
323 |
+
delay_pattern=delay_pattern,
|
324 |
+
)
|
325 |
+
|
326 |
+
codebook = revert_audio_delay(
|
327 |
+
audio_BxTxC=generated_codes.unsqueeze(0),
|
328 |
+
pad_value=audio_pad_value,
|
329 |
+
precomp=revert_precomp,
|
330 |
+
T=seq_length,
|
331 |
+
)[:, :-max_delay_pattern, :]
|
332 |
+
|
333 |
+
min_valid_index = 0
|
334 |
+
max_valid_index = 1023
|
335 |
+
invalid_mask = (codebook < min_valid_index) | (codebook > max_valid_index)
|
336 |
+
codebook[invalid_mask] = 0
|
337 |
+
|
338 |
+
audio = decode(self.dac_model, codebook.transpose(1, 2))
|
339 |
+
|
340 |
+
return audio.squeeze().cpu().numpy()
|
341 |
+
|
342 |
+
def load_audio(self, audio_path: str) -> torch.Tensor:
|
343 |
+
audio, sr = torchaudio.load(audio_path, channels_first=True) # C, T
|
344 |
+
if sr != DEFAULT_SAMPLE_RATE:
|
345 |
+
audio = torchaudio.functional.resample(audio, sr, DEFAULT_SAMPLE_RATE)
|
346 |
+
audio = audio.to(self.device).unsqueeze(0) # 1, C, T
|
347 |
+
audio_data = self.dac_model.preprocess(audio, DEFAULT_SAMPLE_RATE)
|
348 |
+
_, encoded_frame, _, _, _ = self.dac_model.encode(audio_data) # 1, C, T
|
349 |
+
return encoded_frame.squeeze(0).transpose(0, 1)
|
350 |
+
|
351 |
+
def save_audio(self, path: str, audio: np.ndarray):
|
352 |
+
import soundfile as sf
|
353 |
+
|
354 |
+
sf.write(path, audio, DEFAULT_SAMPLE_RATE)
|
355 |
+
|
356 |
+
@torch.inference_mode()
|
357 |
+
def generate(
|
358 |
+
self,
|
359 |
+
text: str,
|
360 |
+
max_tokens: int | None = None,
|
361 |
+
cfg_scale: float = 3.0,
|
362 |
+
temperature: float = 1.3,
|
363 |
+
top_p: float = 0.95,
|
364 |
+
use_torch_compile: bool = False,
|
365 |
+
cfg_filter_top_k: int = 35,
|
366 |
+
audio_prompt: str | torch.Tensor | None = None,
|
367 |
+
audio_prompt_path: str | None = None,
|
368 |
+
use_cfg_filter: bool | None = None,
|
369 |
+
verbose: bool = False,
|
370 |
+
) -> np.ndarray:
|
371 |
+
audio_eos_value = self.config.data.audio_eos_value
|
372 |
+
audio_pad_value = self.config.data.audio_pad_value
|
373 |
+
delay_pattern = self.config.data.delay_pattern
|
374 |
+
max_tokens = self.config.data.audio_length if max_tokens is None else max_tokens
|
375 |
+
max_delay_pattern = max(delay_pattern)
|
376 |
+
self.model.eval()
|
377 |
+
|
378 |
+
if audio_prompt_path:
|
379 |
+
print("Warning: audio_prompt_path is deprecated. Use audio_prompt instead.")
|
380 |
+
audio_prompt = audio_prompt_path
|
381 |
+
if use_cfg_filter is not None:
|
382 |
+
print("Warning: use_cfg_filter is deprecated.")
|
383 |
+
|
384 |
+
if verbose:
|
385 |
+
total_start_time = time.time()
|
386 |
+
|
387 |
+
dec_state, dec_output = self._prepare_generation(text, audio_prompt, verbose)
|
388 |
+
dec_step = dec_output.prefill_step - 1
|
389 |
+
|
390 |
+
bos_countdown = max_delay_pattern
|
391 |
+
eos_detected = False
|
392 |
+
eos_countdown = -1
|
393 |
+
|
394 |
+
if use_torch_compile:
|
395 |
+
step_fn = torch.compile(self._decoder_step, mode="default")
|
396 |
+
else:
|
397 |
+
step_fn = self._decoder_step
|
398 |
+
|
399 |
+
if verbose:
|
400 |
+
print("generate: starting generation loop")
|
401 |
+
if use_torch_compile:
|
402 |
+
print("generate: by using use_torch_compile=True, the first step would take long")
|
403 |
+
start_time = time.time()
|
404 |
+
|
405 |
+
while dec_step < max_tokens:
|
406 |
+
dec_state.prepare_step(dec_step)
|
407 |
+
tokens_Bx1xC = dec_output.get_tokens_at(dec_step).unsqueeze(0).expand(2, -1, -1)
|
408 |
+
pred_C = step_fn(
|
409 |
+
tokens_Bx1xC,
|
410 |
+
dec_state,
|
411 |
+
cfg_scale,
|
412 |
+
temperature,
|
413 |
+
top_p,
|
414 |
+
cfg_filter_top_k,
|
415 |
+
)
|
416 |
+
|
417 |
+
if (not eos_detected and pred_C[0] == audio_eos_value) or dec_step == max_tokens - max_delay_pattern - 1:
|
418 |
+
eos_detected = True
|
419 |
+
eos_countdown = max_delay_pattern
|
420 |
+
|
421 |
+
if eos_countdown > 0:
|
422 |
+
step_after_eos = max_delay_pattern - eos_countdown
|
423 |
+
for i, d in enumerate(delay_pattern):
|
424 |
+
if step_after_eos == d:
|
425 |
+
pred_C[i] = audio_eos_value
|
426 |
+
elif step_after_eos > d:
|
427 |
+
pred_C[i] = audio_pad_value
|
428 |
+
eos_countdown -= 1
|
429 |
+
|
430 |
+
bos_countdown = max(0, bos_countdown - 1)
|
431 |
+
dec_output.update_one(pred_C, dec_step + 1, bos_countdown > 0)
|
432 |
+
|
433 |
+
if eos_countdown == 0:
|
434 |
+
break
|
435 |
+
|
436 |
+
dec_step += 1
|
437 |
+
if verbose and dec_step % 86 == 0:
|
438 |
+
duration = time.time() - start_time
|
439 |
+
print(
|
440 |
+
f"generate step {dec_step}: speed={86 / duration:.3f} tokens/s, realtime factor={1 / duration:.3f}x"
|
441 |
+
)
|
442 |
+
start_time = time.time()
|
443 |
+
|
444 |
+
if dec_output.prefill_step >= dec_step + 1:
|
445 |
+
print("Warning: Nothing generated")
|
446 |
+
return None
|
447 |
+
|
448 |
+
generated_codes = dec_output.generated_tokens[dec_output.prefill_step : dec_step + 1, :]
|
449 |
+
|
450 |
+
if verbose:
|
451 |
+
total_step = dec_step + 1 - dec_output.prefill_step
|
452 |
+
total_duration = time.time() - total_start_time
|
453 |
+
print(f"generate: total step={total_step}, total duration={total_duration:.3f}s")
|
454 |
+
|
455 |
+
return self._generate_output(generated_codes)
|
dia/state.py
ADDED
@@ -0,0 +1,207 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from dataclasses import dataclass
|
2 |
+
|
3 |
+
import torch
|
4 |
+
|
5 |
+
from .config import DiaConfig
|
6 |
+
|
7 |
+
|
8 |
+
def create_attn_mask(
|
9 |
+
q_padding_mask_1d: torch.Tensor,
|
10 |
+
k_padding_mask_1d: torch.Tensor,
|
11 |
+
device: torch.device,
|
12 |
+
is_causal: bool = False,
|
13 |
+
) -> torch.Tensor:
|
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 |
+
# Condition A: Non-padding query attends to non-padding key
|
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, "Causal mask requires query and key sequence lengths to be equal"
|
35 |
+
causal_mask_2d = torch.tril(torch.ones((Tq, Tk), dtype=torch.bool, device=device)) # Shape [Tq, Tk]
|
36 |
+
causal_mask = mask & causal_mask_2d # Shape [B, Tq, Tk]
|
37 |
+
return causal_mask.unsqueeze(1) # Shape [B, 1, Tq, Tk]
|
38 |
+
else:
|
39 |
+
return mask.unsqueeze(1) # Shape [B, 1, Tq, Tk]
|
40 |
+
|
41 |
+
|
42 |
+
@dataclass
|
43 |
+
class EncoderInferenceState:
|
44 |
+
"""Parameters specifically for encoder inference."""
|
45 |
+
|
46 |
+
max_seq_len: int
|
47 |
+
device: torch.device
|
48 |
+
positions: torch.Tensor
|
49 |
+
padding_mask: torch.Tensor
|
50 |
+
attn_mask: torch.Tensor
|
51 |
+
|
52 |
+
@classmethod
|
53 |
+
def new(cls, config: DiaConfig, cond_src: torch.Tensor) -> "EncoderInferenceState":
|
54 |
+
"""Creates EtorchrInferenceParams from DiaConfig and a device."""
|
55 |
+
device = cond_src.device
|
56 |
+
|
57 |
+
positions = (
|
58 |
+
torch.arange(config.data.text_length, dtype=torch.float32, device=device).unsqueeze(0).expand(2, -1)
|
59 |
+
)
|
60 |
+
padding_mask = (cond_src != config.data.text_pad_value).to(device).expand(2, -1)
|
61 |
+
attn_mask = create_attn_mask(padding_mask, padding_mask, device, is_causal=False)
|
62 |
+
|
63 |
+
return cls(
|
64 |
+
max_seq_len=config.data.text_length,
|
65 |
+
device=device,
|
66 |
+
positions=positions,
|
67 |
+
padding_mask=padding_mask,
|
68 |
+
attn_mask=attn_mask,
|
69 |
+
)
|
70 |
+
|
71 |
+
|
72 |
+
class KVCache:
|
73 |
+
def __init__(
|
74 |
+
self,
|
75 |
+
num_heads: int,
|
76 |
+
max_len: int,
|
77 |
+
head_dim: int,
|
78 |
+
dtype: torch.dtype,
|
79 |
+
device: torch.device,
|
80 |
+
k: torch.Tensor | None = None,
|
81 |
+
v: torch.Tensor | None = None,
|
82 |
+
):
|
83 |
+
self.k = torch.zeros((2, num_heads, max_len, head_dim), dtype=dtype, device=device) if k is None else k
|
84 |
+
self.v = torch.zeros((2, num_heads, max_len, head_dim), dtype=dtype, device=device) if v is None else v
|
85 |
+
self.current_idx = torch.tensor(0)
|
86 |
+
|
87 |
+
@classmethod
|
88 |
+
def from_kv(cls, k: torch.Tensor, v: torch.Tensor) -> "KVCache":
|
89 |
+
return cls(
|
90 |
+
num_heads=k.shape[1],
|
91 |
+
max_len=k.shape[2],
|
92 |
+
head_dim=k.shape[3],
|
93 |
+
dtype=k.dtype,
|
94 |
+
device=k.device,
|
95 |
+
k=k,
|
96 |
+
v=v,
|
97 |
+
)
|
98 |
+
|
99 |
+
def update(self, k: torch.Tensor, v: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
|
100 |
+
self.k[:, :, self.current_idx : self.current_idx + 1, :] = k
|
101 |
+
self.v[:, :, self.current_idx : self.current_idx + 1, :] = v
|
102 |
+
self.current_idx += 1
|
103 |
+
return self.k[:, :, : self.current_idx, :], self.v[:, :, : self.current_idx, :]
|
104 |
+
|
105 |
+
def prefill(self, k: torch.Tensor, v: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
|
106 |
+
prefill_len = k.shape[2]
|
107 |
+
self.k[:, :, :prefill_len, :] = k
|
108 |
+
self.v[:, :, :prefill_len, :] = v
|
109 |
+
self.current_idx = prefill_len - 1
|
110 |
+
|
111 |
+
|
112 |
+
@dataclass
|
113 |
+
class DecoderInferenceState:
|
114 |
+
"""Parameters specifically for decoder inference."""
|
115 |
+
|
116 |
+
device: torch.device
|
117 |
+
dtype: torch.dtype
|
118 |
+
enc_out: torch.Tensor
|
119 |
+
enc_positions: torch.Tensor
|
120 |
+
dec_positions: torch.Tensor
|
121 |
+
dec_cross_attn_mask: torch.Tensor
|
122 |
+
self_attn_cache: list[KVCache]
|
123 |
+
cross_attn_cache: list[KVCache]
|
124 |
+
|
125 |
+
@classmethod
|
126 |
+
def new(
|
127 |
+
cls,
|
128 |
+
config: DiaConfig,
|
129 |
+
enc_state: EncoderInferenceState,
|
130 |
+
enc_out: torch.Tensor,
|
131 |
+
dec_cross_attn_cache: list[KVCache],
|
132 |
+
compute_dtype: torch.dtype,
|
133 |
+
) -> "DecoderInferenceState":
|
134 |
+
"""Creates DecoderInferenceParams from DiaConfig and a device."""
|
135 |
+
device = enc_out.device
|
136 |
+
max_audio_len = config.data.audio_length
|
137 |
+
|
138 |
+
dec_positions = torch.full((2, 1), fill_value=0, dtype=torch.long, device=device)
|
139 |
+
tgt_padding_mask = torch.ones((2, 1), dtype=torch.bool, device=device)
|
140 |
+
dec_cross_attn_mask = create_attn_mask(tgt_padding_mask, enc_state.padding_mask, device, is_causal=False)
|
141 |
+
|
142 |
+
self_attn_cache = [
|
143 |
+
KVCache(
|
144 |
+
config.model.decoder.kv_heads,
|
145 |
+
max_audio_len,
|
146 |
+
config.model.decoder.gqa_head_dim,
|
147 |
+
compute_dtype,
|
148 |
+
device,
|
149 |
+
)
|
150 |
+
for _ in range(config.model.decoder.n_layer)
|
151 |
+
]
|
152 |
+
|
153 |
+
return cls(
|
154 |
+
device=device,
|
155 |
+
dtype=compute_dtype,
|
156 |
+
enc_out=enc_out,
|
157 |
+
enc_positions=enc_state.positions,
|
158 |
+
dec_positions=dec_positions,
|
159 |
+
dec_cross_attn_mask=dec_cross_attn_mask,
|
160 |
+
self_attn_cache=self_attn_cache,
|
161 |
+
cross_attn_cache=dec_cross_attn_cache,
|
162 |
+
)
|
163 |
+
|
164 |
+
def prepare_step(self, step_from: int, step_to: int | None = None) -> None:
|
165 |
+
if step_to is None:
|
166 |
+
step_to = step_from + 1
|
167 |
+
self.dec_positions = (
|
168 |
+
torch.arange(step_from, step_to, dtype=torch.float32, device=self.device).unsqueeze(0).expand(2, -1)
|
169 |
+
)
|
170 |
+
|
171 |
+
|
172 |
+
@dataclass
|
173 |
+
class DecoderOutput:
|
174 |
+
generated_tokens: torch.Tensor
|
175 |
+
prefill_step: int
|
176 |
+
|
177 |
+
@classmethod
|
178 |
+
def new(cls, config: DiaConfig, device: torch.device) -> "DecoderOutput":
|
179 |
+
max_audio_len = config.data.audio_length
|
180 |
+
return cls(
|
181 |
+
generated_tokens=torch.full(
|
182 |
+
(max_audio_len, config.data.channels),
|
183 |
+
fill_value=-1,
|
184 |
+
dtype=torch.int,
|
185 |
+
device=device,
|
186 |
+
),
|
187 |
+
prefill_step=0,
|
188 |
+
)
|
189 |
+
|
190 |
+
def get_tokens_at(self, step_from: int, step_to: int | None = None) -> torch.Tensor:
|
191 |
+
if step_to is None:
|
192 |
+
step_to = step_from + 1
|
193 |
+
return self.generated_tokens[step_from:step_to, :]
|
194 |
+
|
195 |
+
def update_one(self, dec_out: torch.Tensor, step: int, apply_mask: bool = False):
|
196 |
+
if apply_mask:
|
197 |
+
mask = self.generated_tokens[step : step + 1, :] == -1
|
198 |
+
self.generated_tokens[step : step + 1, :] = torch.where(
|
199 |
+
mask, dec_out, self.generated_tokens[step : step + 1, :]
|
200 |
+
)
|
201 |
+
else:
|
202 |
+
self.generated_tokens[step : step + 1, :] = dec_out
|
203 |
+
|
204 |
+
def prefill(self, dec_out: torch.Tensor, prefill_step: int):
|
205 |
+
length = dec_out.shape[0]
|
206 |
+
self.generated_tokens[0:length, :] = dec_out
|
207 |
+
self.prefill_step = prefill_step
|