File size: 8,811 Bytes
2bbfbb7 0393dfa 2bbfbb7 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 |
//! GPT-based sequence generation model
use crate::{Error, Result};
use ndarray::{Array, Array1, Array2, IxDyn};
use std::collections::HashMap;
use std::path::Path;
use super::{OnnxSession, SamplingStrategy, sample_from_logits, apply_repetition_penalty};
/// GPT model configuration
#[derive(Debug, Clone)]
pub struct GptConfig {
/// Number of transformer layers
pub num_layers: usize,
/// Model dimension
pub hidden_size: usize,
/// Number of attention heads
pub num_heads: usize,
/// Maximum sequence length
pub max_seq_len: usize,
/// Vocabulary size
pub vocab_size: usize,
/// Stop token ID
pub stop_token: usize,
/// Start token ID
pub start_token: usize,
}
impl Default for GptConfig {
fn default() -> Self {
Self {
num_layers: 8,
hidden_size: 512,
num_heads: 8,
max_seq_len: 250,
vocab_size: 8194,
stop_token: 8193,
start_token: 8192,
}
}
}
/// GPT model for autoregressive generation
pub struct GptModel {
session: OnnxSession,
config: GptConfig,
}
impl GptModel {
/// Load GPT model from ONNX file
pub fn load<P: AsRef<Path>>(path: P, config: GptConfig) -> Result<Self> {
let session = OnnxSession::load(path)?;
Ok(Self { session, config })
}
/// Generate mel tokens from semantic tokens
pub fn generate(
&self,
semantic_tokens: &[i64],
speaker_embedding: &Array1<f32>,
max_length: usize,
strategy: &SamplingStrategy,
repetition_penalty: f32,
) -> Result<Vec<i64>> {
let mut generated_tokens = vec![self.config.start_token as i64];
let mut past_tokens = Vec::new();
for _ in 0..max_length {
// Prepare input
let input_tokens = Array::from_shape_vec(
IxDyn(&[1, generated_tokens.len()]),
generated_tokens.clone(),
)?;
let speaker_emb = speaker_embedding
.clone()
.into_shape(IxDyn(&[1, speaker_embedding.len()]))?;
let semantic_input = Array::from_shape_vec(
IxDyn(&[1, semantic_tokens.len()]),
semantic_tokens.to_vec(),
)?;
// Create input map
let mut inputs = HashMap::new();
inputs.insert("input_ids".to_string(), input_tokens.mapv(|x| x as f32));
inputs.insert("speaker_embedding".to_string(), speaker_emb);
inputs.insert("semantic_tokens".to_string(), semantic_input.mapv(|x| x as f32));
// Run inference
let outputs = self.session.run(inputs)?;
// Get logits for next token
let logits = outputs
.get("logits")
.ok_or_else(|| Error::Model("Missing logits output".into()))?;
// Get last token logits
let seq_len = logits.shape()[1];
let vocab_size = logits.shape()[2];
let last_logits: Vec<f32> = (0..vocab_size)
.map(|i| logits[[0, seq_len - 1, i]])
.collect();
// Apply repetition penalty
let mut logits_vec = last_logits;
let past_usize: Vec<usize> = past_tokens.iter().map(|&x| x as usize).collect();
apply_repetition_penalty(&mut logits_vec, &past_usize, repetition_penalty);
// Sample next token
let next_token = sample_from_logits(&logits_vec, strategy) as i64;
// Check for stop token
if next_token == self.config.stop_token as i64 {
break;
}
generated_tokens.push(next_token);
past_tokens.push(next_token);
}
Ok(generated_tokens)
}
/// Generate with KV cache for efficiency
pub fn generate_with_cache(
&self,
semantic_tokens: &[i64],
speaker_embedding: &Array1<f32>,
max_length: usize,
strategy: &SamplingStrategy,
repetition_penalty: f32,
) -> Result<Vec<i64>> {
// For models with KV cache support
// This is a simplified version - full implementation would maintain cache state
self.generate(
semantic_tokens,
speaker_embedding,
max_length,
strategy,
repetition_penalty,
)
}
/// Get model config
pub fn config(&self) -> &GptConfig {
&self.config
}
/// Estimate memory usage
pub fn estimate_memory_mb(&self) -> f32 {
let params = self.config.num_layers
* self.config.hidden_size
* self.config.hidden_size
* 4; // Approximate
(params * 4) as f32 / 1_000_000.0 // 4 bytes per param
}
}
/// Simplified GPT model using pure Rust (fallback when ONNX not available)
pub struct SimpleGptModel {
config: GptConfig,
/// Token embeddings
token_embeddings: Array2<f32>,
/// Position embeddings
position_embeddings: Array2<f32>,
/// Output projection
output_projection: Array2<f32>,
}
impl SimpleGptModel {
/// Create random initialized model (for testing)
pub fn new_random(config: GptConfig) -> Self {
use rand::Rng;
let mut rng = rand::thread_rng();
let token_embeddings = Array2::from_shape_fn(
(config.vocab_size, config.hidden_size),
|_| rng.gen_range(-0.1..0.1),
);
let position_embeddings = Array2::from_shape_fn(
(config.max_seq_len, config.hidden_size),
|_| rng.gen_range(-0.1..0.1),
);
let output_projection = Array2::from_shape_fn(
(config.hidden_size, config.vocab_size),
|_| rng.gen_range(-0.1..0.1),
);
Self {
config,
token_embeddings,
position_embeddings,
output_projection,
}
}
/// Simple forward pass (for demonstration)
pub fn forward(&self, tokens: &[i64]) -> Vec<f32> {
// Get embeddings
let mut hidden = vec![0.0f32; self.config.hidden_size];
for (pos, &token) in tokens.iter().enumerate().take(self.config.max_seq_len) {
let token_idx = (token as usize).min(self.config.vocab_size - 1);
for i in 0..self.config.hidden_size {
hidden[i] += self.token_embeddings[[token_idx, i]]
+ self.position_embeddings[[pos, i]];
}
}
// Normalize
let norm: f32 = hidden.iter().map(|x| x * x).sum::<f32>().sqrt();
if norm > 1e-8 {
for h in hidden.iter_mut() {
*h /= norm;
}
}
// Project to vocab
let mut logits = vec![0.0f32; self.config.vocab_size];
for (i, logit) in logits.iter_mut().enumerate() {
for j in 0..self.config.hidden_size {
*logit += hidden[j] * self.output_projection[[j, i]];
}
}
logits
}
/// Generate tokens
pub fn generate(
&self,
prompt: &[i64],
max_length: usize,
strategy: &SamplingStrategy,
) -> Vec<i64> {
let mut tokens = prompt.to_vec();
for _ in 0..max_length {
let logits = self.forward(&tokens);
let next_token = sample_from_logits(&logits, strategy) as i64;
if next_token == self.config.stop_token as i64 {
break;
}
tokens.push(next_token);
if tokens.len() >= self.config.max_seq_len {
break;
}
}
tokens
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn test_gpt_config_default() {
let config = GptConfig::default();
assert_eq!(config.num_layers, 8);
assert_eq!(config.hidden_size, 512);
}
#[test]
fn test_simple_gpt_forward() {
let config = GptConfig {
vocab_size: 100,
hidden_size: 32,
max_seq_len: 10,
..Default::default()
};
let model = SimpleGptModel::new_random(config);
let tokens = vec![1i64, 2, 3];
let logits = model.forward(&tokens);
assert_eq!(logits.len(), 100);
}
#[test]
fn test_simple_gpt_generate() {
let config = GptConfig {
vocab_size: 100,
hidden_size: 32,
max_seq_len: 20,
stop_token: 99,
..Default::default()
};
let model = SimpleGptModel::new_random(config);
let prompt = vec![1i64, 2, 3];
let generated = model.generate(&prompt, 10, &SamplingStrategy::Greedy);
assert!(generated.len() >= 3);
assert!(generated.len() <= 20);
}
}
|