musicgpt 0.3.22

Generate music based on natural language prompts using LLMs running locally
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256

use std::fmt::Debug;
use std::marker::PhantomData;
use std::sync::mpsc::Receiver;
use std::sync::Arc;

use crate::delay_pattern_mask_ids::DelayedPatternMaskIds;
use crate::music_gen_config::MusicGenConfig;
use crate::music_gen_inputs::MusicGenInputs;
use crate::music_gen_outputs::MusicGenOutputs;
use crate::tensor_ops::{dupe_zeros_along_first_dim, zeros_tensor};
use num_traits::Zero;
use ort::session::Session;
use ort::tensor::PrimitiveTensorElementType;
use ort::value::{DynValue, Tensor};

pub trait MusicGenType: PrimitiveTensorElementType + Debug + Clone + Zero {}

impl MusicGenType for u8 {}
impl MusicGenType for i8 {}
impl MusicGenType for f32 {}
impl MusicGenType for half::f16 {}

// TODO: is this configurable?
const GUIDANCE_SCALE: usize = 3;

pub trait MusicGenDecoder: Send + Sync {
    fn generate_tokens(
        &self,
        last_hidden_state: DynValue,
        encoder_attention_mask: DynValue,
        max_len: usize,
    ) -> ort::Result<Receiver<ort::Result<[i64; 4]>>>;
}

pub struct MusicGenMergedDecoder<T: MusicGenType> {
    pub decoder_model_merged: Arc<Session>,
    pub config: MusicGenConfig,
    pub _phantom_data: PhantomData<T>,
}

unsafe impl<T: MusicGenType> Send for MusicGenMergedDecoder<T> {}
unsafe impl<T: MusicGenType> Sync for MusicGenMergedDecoder<T> {}

impl<T: MusicGenType + 'static> MusicGenDecoder for MusicGenMergedDecoder<T> {
    fn generate_tokens(
        &self,
        last_hidden_state: DynValue,
        encoder_attention_mask: DynValue,
        max_len: usize,
    ) -> ort::Result<Receiver<ort::Result<[i64; 4]>>> {
        // Apparently, there's a setting in huggingface's transformers that says that
        // if `guidance_scale` > 1 then you should concatenate 0 along the first axis.
        let encoder_hidden_states = dupe_zeros_along_first_dim::<T>(last_hidden_state.downcast()?)?;
        let encoder_attention_mask =
            dupe_zeros_along_first_dim::<i64>(encoder_attention_mask.downcast()?)?;

        let mut delay_pattern_mask_ids = DelayedPatternMaskIds::<4>::new();

        let decoder_model_merged = self.decoder_model_merged.clone();

        let mut inputs = MusicGenInputs::new();
        inputs.encoder_attention_mask(encoder_attention_mask)?;
        inputs.encoder_hidden_states(encoder_hidden_states)?;

        let num_hidden_layers = self.config.decoder.num_hidden_layers;
        let num_attention_heads = self.config.decoder.num_attention_heads;
        let pad_token_id = self.config.decoder.pad_token_id;
        let d_kv = self.config.text_encoder.d_kv;
        let top_k = self.config.decoder.top_k;
        let decoder_dims = [1, num_attention_heads, 0, d_kv];
        let encoder_dims = [1, num_attention_heads, 0, d_kv];

        // TODO: 100?
        let (tx, rx) = std::sync::mpsc::channel::<ort::Result<[i64; 4]>>();
        let tx2 = tx.clone();

        std::thread::spawn(move || {
            let result = {
                inputs.input_ids(Tensor::from_array(([8, 1], vec![pad_token_id; 8]))?)?;

                for i in 0..num_hidden_layers {
                    inputs.past_key_value_decoder_key(i, zeros_tensor::<T>(&decoder_dims))?;
                    inputs.past_key_value_decoder_value(i, zeros_tensor::<T>(&decoder_dims))?;
                    inputs.past_key_value_encoder_key(i, zeros_tensor::<T>(&encoder_dims))?;
                    inputs.past_key_value_encoder_value(i, zeros_tensor::<T>(&encoder_dims))?;
                }
                inputs.use_cache_branch(false);
                for _ in 0..max_len {
                    let outputs = decoder_model_merged.run(inputs.ort())?;
                    let mut outputs = MusicGenOutputs::new(outputs);

                    delay_pattern_mask_ids.push(
                        outputs
                            .take_logits()?
                            .apply_free_guidance(GUIDANCE_SCALE)
                            .sample(top_k)
                            .iter()
                            .map(|e| e.0),
                    );

                    let [a, b, c, d] = delay_pattern_mask_ids.last_delayed_masked(pad_token_id);
                    inputs.input_ids(ort::value::Value::from_array((
                        [8, 1],
                        vec![a, b, c, d, a, b, c, d],
                    ))?)?;

                    if let Some(last_de_delayed) = delay_pattern_mask_ids.last_de_delayed() {
                        let sent = tx.send(Ok(last_de_delayed));
                        if sent.is_err() {
                            break;
                        }
                    }

                    for j in 0..num_hidden_layers {
                        let v = outputs.take_present_decoder_key(j);
                        inputs.past_key_value_decoder_key(j, v)?;
                        let v = outputs.take_present_decoder_value(j);
                        inputs.past_key_value_decoder_value(j, v)?;
                        if !inputs.use_cache_branch {
                            // Optimization introduced by optimum to reuse past key values. So, we just replace the constant
                            // outputs with the previous past key values.
                            // https://github.com/huggingface/optimum/blob/0bf2c05fb7e1182b52d21b703cfc95fd9e4ea3dc/optimum/onnxruntime/base.py#L677-L704
                            let v = outputs.take_present_encoder_key(j);
                            inputs.past_key_value_encoder_key(j, v)?;
                            let v = outputs.take_present_encoder_value(j);
                            inputs.past_key_value_encoder_value(j, v)?;
                        }
                    }

                    inputs.use_cache_branch(true);
                }
                Ok::<(), ort::Error>(())
            };
            if let Err(err) = result {
                let _ = tx2.send(Err(err));
            }
            Ok::<(), ort::Error>(())
        });

        Ok(rx)
    }
}

pub struct MusicGenSplitDecoder<T: MusicGenType> {
    pub decoder_model: Session,
    pub decoder_with_past_model: Arc<Session>,
    pub config: MusicGenConfig,
    pub _phantom_data: PhantomData<T>,
}

unsafe impl<T: MusicGenType> Send for MusicGenSplitDecoder<T> {}
unsafe impl<T: MusicGenType> Sync for MusicGenSplitDecoder<T> {}

impl<T: MusicGenType + 'static> MusicGenDecoder for MusicGenSplitDecoder<T> {
    fn generate_tokens(
        &self,
        last_hidden_state: DynValue,
        encoder_attention_mask: DynValue,
        max_len: usize,
    ) -> ort::Result<Receiver<ort::Result<[i64; 4]>>> {
        // Apparently, there's a setting in huggingface's transformers that says that
        // if `guidance_scale` > 1 then you should concatenate 0 along the first axis.
        let encoder_hidden_states = dupe_zeros_along_first_dim::<T>(last_hidden_state.downcast()?)?;
        let encoder_attention_mask =
            dupe_zeros_along_first_dim::<i64>(encoder_attention_mask.downcast()?)?;

        let mut delay_pattern_mask_ids = DelayedPatternMaskIds::<4>::new();

        let num_hidden_layers = self.config.decoder.num_hidden_layers;
        let pad_token_id = self.config.decoder.pad_token_id;
        let top_k = self.config.decoder.top_k;

        let mut inputs = MusicGenInputs::new();
        inputs.encoder_attention_mask(encoder_attention_mask)?;
        inputs.input_ids(Tensor::from_array(([8, 1], vec![pad_token_id; 8]))?)?;
        inputs.encoder_hidden_states(encoder_hidden_states)?;

        let outputs = self.decoder_model.run(inputs.ort())?;
        let mut outputs = MusicGenOutputs::new(outputs);

        delay_pattern_mask_ids.push(
            outputs
                .take_logits()?
                .apply_free_guidance(GUIDANCE_SCALE)
                .sample(top_k)
                .iter()
                .map(|e| e.0),
        );

        for j in 0..num_hidden_layers {
            let v = outputs.take_present_decoder_key(j);
            inputs.past_key_value_decoder_key(j, v)?;
            let v = outputs.take_present_decoder_value(j);
            inputs.past_key_value_decoder_value(j, v)?;
            let v = outputs.take_present_encoder_key(j);
            inputs.past_key_value_encoder_key(j, v)?;
            let v = outputs.take_present_encoder_value(j);
            inputs.past_key_value_encoder_value(j, v)?;
        }

        inputs.remove_encoder_hidden_states();

        let decoder_with_past = self.decoder_with_past_model.clone();

        // TODO: 100?
        let (tx, rx) = std::sync::mpsc::channel::<ort::Result<[i64; 4]>>();
        let tx2 = tx.clone();
        std::thread::spawn(move || {
            let result = {
                for _ in 0..max_len {
                    let [a, b, c, d] = delay_pattern_mask_ids.last_delayed_masked(pad_token_id);
                    inputs
                        .input_ids(Tensor::from_array(([8, 1], vec![a, b, c, d, a, b, c, d]))?)?;
                    let outputs = decoder_with_past.run(inputs.ort())?;
                    let mut outputs = MusicGenOutputs::new(outputs);

                    delay_pattern_mask_ids.push(
                        outputs
                            .take_logits()?
                            .apply_free_guidance(GUIDANCE_SCALE)
                            .sample(top_k)
                            .iter()
                            .map(|e| e.0),
                    );

                    if let Some(last_de_delayed) = delay_pattern_mask_ids.last_de_delayed() {
                        let sent = tx.send(Ok(last_de_delayed));
                        if sent.is_err() {
                            break;
                        }
                    }

                    for j in 0..num_hidden_layers {
                        let v = outputs.take_present_decoder_key(j);
                        inputs.past_key_value_decoder_key(j, v)?;
                        let v = outputs.take_present_decoder_value(j);
                        inputs.past_key_value_decoder_value(j, v)?;
                        // NOTE: No need to propagate encoder values.
                        //
                        // let v = outputs.take_present_encoder_key(j);
                        // inputs.past_key_value_encoder_key(j, v)?;
                        // let v = outputs.take_present_encoder_value(j);
                        // inputs.past_key_value_encoder_value(j, v)?;
                    }
                }
                Ok::<(), ort::Error>(())
            };
            if let Err(err) = result {
                let _ = tx2.send(Err(err));
            }
            Ok::<(), ort::Error>(())
        });

        Ok(rx)
    }
}