qwen_tts 0.1.1

Qwen3-TTS text-to-speech model implementation for Candle
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244

//! Unified attention implementation.
//!
//! This module provides a single attention implementation that handles both
//! standard and multimodal RoPE through the `RopeStrategy` abstraction.

use candle_core::{Result, Tensor};
use candle_nn::{Linear, Module, VarBuilder, linear, linear_no_bias};

use crate::nn::attention::config::AttentionConfig;
use crate::nn::attention::rope_strategy::RopeStrategy;
use crate::nn::attention::{eager_attention_forward_with_sliding_window, flash_attention_forward};
use crate::nn::kv_cache::KVCache;
use crate::nn::norm::RMSNorm;

/// Unified attention with configurable RoPE strategy.
///
/// This struct implements attention with per-head Q/K normalization and
/// supports both standard and multimodal RoPE through the `RopeStrategy` enum.
#[derive(Debug, Clone)]
pub struct UnifiedAttention {
    q_proj: Linear,
    k_proj: Linear,
    v_proj: Linear,
    o_proj: Linear,
    q_norm: RMSNorm,
    k_norm: RMSNorm,
    num_heads: usize,
    num_kv_heads: usize,
    head_dim: usize,
    num_kv_groups: usize,
    scaling: f64,
    rope_strategy: RopeStrategy,
    sliding_window: Option<usize>,
    layer_idx: usize,
    use_flash_attn: bool,
}

impl UnifiedAttention {
    /// Create a new unified attention layer.
    pub fn new<C: AttentionConfig>(
        config: &C,
        rope_strategy: RopeStrategy,
        layer_idx: usize,
        use_flash_attn: bool,
        vb: VarBuilder,
    ) -> Result<Self> {
        let head_dim = config.head_dim();
        let num_heads = config.num_attention_heads();
        let num_kv_heads = config.num_key_value_heads();
        let hidden_size = config.hidden_size();

        let q_proj = if config.attention_bias() {
            linear(hidden_size, num_heads * head_dim, vb.pp("q_proj"))?
        } else {
            linear_no_bias(hidden_size, num_heads * head_dim, vb.pp("q_proj"))?
        };

        let k_proj = if config.attention_bias() {
            linear(hidden_size, num_kv_heads * head_dim, vb.pp("k_proj"))?
        } else {
            linear_no_bias(hidden_size, num_kv_heads * head_dim, vb.pp("k_proj"))?
        };

        let v_proj = if config.attention_bias() {
            linear(hidden_size, num_kv_heads * head_dim, vb.pp("v_proj"))?
        } else {
            linear_no_bias(hidden_size, num_kv_heads * head_dim, vb.pp("v_proj"))?
        };

        let o_proj = if config.attention_bias() {
            linear(num_heads * head_dim, hidden_size, vb.pp("o_proj"))?
        } else {
            linear_no_bias(num_heads * head_dim, hidden_size, vb.pp("o_proj"))?
        };

        let q_norm = RMSNorm::new(head_dim, config.rms_norm_eps(), vb.pp("q_norm"))?;
        let k_norm = RMSNorm::new(head_dim, config.rms_norm_eps(), vb.pp("k_norm"))?;

        let num_kv_groups = num_heads / num_kv_heads;
        let scaling = (head_dim as f64).powf(-0.5);

        Ok(Self {
            q_proj,
            k_proj,
            v_proj,
            o_proj,
            q_norm,
            k_norm,
            num_heads,
            num_kv_heads,
            head_dim,
            num_kv_groups,
            scaling,
            rope_strategy,
            sliding_window: config.sliding_window(),
            layer_idx,
            use_flash_attn,
        })
    }

    /// Get layer index.
    pub fn layer_idx(&self) -> usize {
        self.layer_idx
    }

    /// Get the sliding window size if configured.
    pub fn sliding_window(&self) -> Option<usize> {
        self.sliding_window
    }

    /// Forward pass with RoPE.
    ///
    /// Args:
    ///   hidden_states: Input tensor (batch, seq_len, hidden_size)
    ///   position_embeddings: (cos, sin) from RoPE
    ///   attention_mask: Optional causal mask
    pub fn forward(
        &self,
        hidden_states: &Tensor,
        position_embeddings: (&Tensor, &Tensor),
        attention_mask: Option<&Tensor>,
    ) -> Result<Tensor> {
        let (batch, seq_len, _) = hidden_states.dims3()?;

        // Project Q, K, V
        let query_states = self.q_proj.forward(hidden_states)?;
        let key_states = self.k_proj.forward(hidden_states)?;
        let value_states = self.v_proj.forward(hidden_states)?;

        // Reshape to (batch, seq_len, num_heads, head_dim)
        let query_states = query_states.reshape((batch, seq_len, self.num_heads, self.head_dim))?;
        let key_states = key_states.reshape((batch, seq_len, self.num_kv_heads, self.head_dim))?;
        let value_states =
            value_states.reshape((batch, seq_len, self.num_kv_heads, self.head_dim))?;

        // Apply per-head normalization
        let query_states = self.q_norm.forward(&query_states)?;
        let key_states = self.k_norm.forward(&key_states)?;

        // Transpose to (batch, heads, seq_len, head_dim)
        let query_states = query_states.transpose(1, 2)?;
        let key_states = key_states.transpose(1, 2)?;
        let value_states = value_states.transpose(1, 2)?;

        // Apply RoPE using the configured strategy
        let (cos, sin) = position_embeddings;
        let (query_states, key_states) =
            self.rope_strategy
                .apply(&query_states, &key_states, cos, sin)?;

        // Compute attention - use flash attention if enabled and on CUDA
        let use_flash = self.use_flash_attn && query_states.device().is_cuda();
        let attn_output = if use_flash {
            flash_attention_forward(
                &query_states,
                &key_states,
                &value_states,
                self.scaling,
                self.sliding_window,
            )?
        } else {
            eager_attention_forward_with_sliding_window(
                &query_states,
                &key_states,
                &value_states,
                attention_mask,
                self.num_kv_groups,
                self.scaling,
                self.sliding_window,
            )?
        };

        // Reshape and project output
        let attn_output = attn_output.reshape((batch, seq_len, self.num_heads * self.head_dim))?;
        self.o_proj.forward(&attn_output)
    }

    /// Forward pass with KV-cache for efficient autoregressive generation.
    pub fn forward_with_cache(
        &self,
        hidden_states: &Tensor,
        position_embeddings: (&Tensor, &Tensor),
        attention_mask: Option<&Tensor>,
        cache: &mut KVCache,
    ) -> Result<Tensor> {
        let (batch, seq_len, _) = hidden_states.dims3()?;

        // Project Q, K, V
        let query_states = self.q_proj.forward(hidden_states)?;
        let key_states = self.k_proj.forward(hidden_states)?;
        let value_states = self.v_proj.forward(hidden_states)?;

        // Reshape to (batch, seq_len, num_heads, head_dim)
        let query_states = query_states.reshape((batch, seq_len, self.num_heads, self.head_dim))?;
        let key_states = key_states.reshape((batch, seq_len, self.num_kv_heads, self.head_dim))?;
        let value_states =
            value_states.reshape((batch, seq_len, self.num_kv_heads, self.head_dim))?;

        // Apply per-head normalization
        let query_states = self.q_norm.forward(&query_states)?;
        let key_states = self.k_norm.forward(&key_states)?;

        // Transpose to (batch, heads, seq_len, head_dim)
        let query_states = query_states.transpose(1, 2)?;
        let key_states = key_states.transpose(1, 2)?;
        let value_states = value_states.transpose(1, 2)?;

        // Apply RoPE using the configured strategy
        let (cos, sin) = position_embeddings;
        let (query_states, key_states) =
            self.rope_strategy
                .apply(&query_states, &key_states, cos, sin)?;

        // Update KV-cache and get full key/value tensors
        let (key_states, value_states) =
            cache.update(self.layer_idx, &key_states, &value_states)?;

        // Compute attention - use flash attention if enabled and on CUDA
        let use_flash = self.use_flash_attn && query_states.device().is_cuda();
        let attn_output = if use_flash {
            flash_attention_forward(
                &query_states,
                &key_states,
                &value_states,
                self.scaling,
                self.sliding_window,
            )?
        } else {
            eager_attention_forward_with_sliding_window(
                &query_states,
                &key_states,
                &value_states,
                attention_mask,
                self.num_kv_groups,
                self.scaling,
                self.sliding_window,
            )?
        };

        // Reshape and project output
        let attn_output = attn_output.reshape((batch, seq_len, self.num_heads * self.head_dim))?;
        self.o_proj.forward(&attn_output)
    }
}