1use crate::model_config;
10use super::traits::*;
11use anyhow::Result;
12use serde::{Serialize, Deserialize};
13
14model_config!(CodeLlamaConfig {
16 vocab_size: usize = 32016,
17 hidden_size: usize = 4096,
18 intermediate_size: usize = 11008,
19 num_hidden_layers: usize = 32,
20 num_attention_heads: usize = 32,
21 num_key_value_heads: usize = 32,
22 hidden_act: String = "silu".to_string(),
23 max_position_embeddings: usize = 16384,
24 initializer_range: f32 = 0.02,
25 rms_norm_eps: f32 = 1e-5,
26 use_cache: bool = true,
27 pad_token_id: i64 = 0,
28 bos_token_id: i64 = 1,
29 eos_token_id: i64 = 2,
30 tie_word_embeddings: bool = false,
31 rope_theta: f32 = 1000000.0, rope_scaling: f32 = 1.0,
33});
34
35impl CodeLlamaConfig {
36 pub fn from_gguf_config(gguf: &crate::weight_loader_core::GGUFModelConfig) -> Self {
37 Self {
38 vocab_size: gguf.vocab_size,
39 hidden_size: gguf.hidden_size,
40 intermediate_size: gguf.intermediate_size,
41 num_hidden_layers: gguf.num_hidden_layers,
42 num_attention_heads: gguf.num_attention_heads,
43 num_key_value_heads: gguf.num_key_value_heads,
44 max_position_embeddings: gguf.max_position_embeddings,
45 rope_theta: gguf.rope_theta,
46 ..Default::default()
47 }
48 }
49}
50
51pub struct CodeLlamaModelV2 {
52 config: CodeLlamaConfig,
53 device: Device,
54 embed_tokens: Tensor,
55 layers: Vec<CodeLlamaLayer>,
56 norm: Tensor,
57 lm_head: Tensor,
58}
59
60pub struct CodeLlamaLayer {
61 self_attn: CodeLlamaAttention,
62 mlp: CodeLlamaMLP,
63 input_layernorm: Tensor,
64 post_attention_layernorm: Tensor,
65}
66
67pub struct CodeLlamaAttention {
68 q_proj: Tensor,
69 k_proj: Tensor,
70 v_proj: Tensor,
71 o_proj: Tensor,
72 num_heads: usize,
73 num_kv_heads: usize,
74 head_dim: usize,
75 scale: f32,
76}
77
78pub struct CodeLlamaMLP {
79 gate_proj: Tensor,
80 up_proj: Tensor,
81 down_proj: Tensor,
82}
83
84fn apply_rope_codellama(
85 q: &candle_core::Tensor,
86 k: &candle_core::Tensor,
87 seq_len: usize,
88 head_dim: usize,
89 rope_theta: f32,
90) -> Result<(candle_core::Tensor, candle_core::Tensor)> {
91 let device = q.device();
92 let half_dim = head_dim / 2;
93 let inv_freq: Vec<f32> = (0..half_dim)
94 .map(|i| 1.0 / rope_theta.powf((2 * i) as f32 / head_dim as f32))
95 .collect();
96
97 let positions: Vec<f32> = (0..seq_len).map(|p| p as f32).collect();
98 let mut angles = Vec::with_capacity(seq_len * half_dim);
99 for pos in &positions {
100 for freq in &inv_freq {
101 angles.push(pos * freq);
102 }
103 }
104
105 let angles_tensor = candle_core::Tensor::from_vec(angles, &[seq_len, half_dim], device)?;
106 let cos = angles_tensor.cos()?.unsqueeze(0)?.unsqueeze(0)?;
107 let sin = angles_tensor.sin()?.unsqueeze(0)?.unsqueeze(0)?;
108
109 let q_half1 = q.narrow(3, 0, half_dim)?;
110 let q_half2 = q.narrow(3, half_dim, half_dim)?;
111 let k_half1 = k.narrow(3, 0, half_dim)?;
112 let k_half2 = k.narrow(3, half_dim, half_dim)?;
113
114 let q_rot1 = (q_half1.broadcast_mul(&cos)? - q_half2.broadcast_mul(&sin)?)?;
115 let q_rot2 = (q_half1.broadcast_mul(&sin)? + q_half2.broadcast_mul(&cos)?)?;
116 let k_rot1 = (k_half1.broadcast_mul(&cos)? - k_half2.broadcast_mul(&sin)?)?;
117 let k_rot2 = (k_half1.broadcast_mul(&sin)? + k_half2.broadcast_mul(&cos)?)?;
118
119 Ok((
120 candle_core::Tensor::cat(&[&q_rot1, &q_rot2], 3)?,
121 candle_core::Tensor::cat(&[&k_rot1, &k_rot2], 3)?
122 ))
123}
124
125impl Model for CodeLlamaModelV2 {
126 type Config = CodeLlamaConfig;
127
128 fn new(config: CodeLlamaConfig) -> Result<Self> {
129 let device = Device::CPU;
130 let embed_tokens = ops_fn::zeros(&[config.vocab_size, config.hidden_size], DataType::Float32, &device)?;
131 let norm = ops_fn::zeros(&[config.hidden_size], DataType::Float32, &device)?;
132 let lm_head = ops_fn::zeros(&[config.vocab_size, config.hidden_size], DataType::Float32, &device)?;
133
134 let mut layers = Vec::with_capacity(config.num_hidden_layers);
135 for _ in 0..config.num_hidden_layers {
136 layers.push(CodeLlamaLayer::new(&config, &device)?);
137 }
138
139 Ok(Self { config, device, embed_tokens, layers, norm, lm_head })
140 }
141
142 fn from_weights(config: CodeLlamaConfig, weights: ModelWeights) -> Result<Self> {
143 let mut model = Self::new(config)?;
144 if let Some(w) = weights.get("model.embed_tokens.weight") { model.embed_tokens = w.clone(); }
145 if let Some(w) = weights.get("model.norm.weight") { model.norm = w.clone(); }
146 if let Some(w) = weights.get("lm_head.weight") { model.lm_head = w.clone(); }
147 for (i, layer) in model.layers.iter_mut().enumerate() { layer.load_weights(&weights, i)?; }
148 Ok(model)
149 }
150
151 fn forward(&self, inputs: &ModelInputs) -> Result<ModelOutputs> {
152 match inputs {
153 ModelInputs::Text { input_ids, .. } => {
154 let seq_len = input_ids.shape()[1];
155 let mut hidden = ops_fn::embedding(input_ids, &self.embed_tokens)?;
156
157 for layer in &self.layers {
158 hidden = layer.forward(&hidden, seq_len, self.config.rope_theta)?;
159 }
160
161 hidden = ops_fn::rms_norm(&hidden, &self.norm, self.config.rms_norm_eps)?;
162 let logits = ops_fn::matmul(&hidden, &ops_fn::transpose(&self.lm_head)?)?;
163
164 Ok(ModelOutputs::Logits { logits, hidden_states: None })
165 }
166 _ => Err(anyhow::anyhow!("CodeLlama only supports text inputs")),
167 }
168 }
169
170 fn generate(&self, prompt: &str, config: &GenerationConfig) -> Result<String> {
171 use crate::tokenizer::Tokenizer;
172 use rand::Rng;
173 let tokenizer = Tokenizer::new();
174 let mut tokens: Vec<u32> = tokenizer.encode(prompt);
175 for _ in 0..config.max_new_tokens {
176 let tokens_i64: Vec<i64> = tokens.iter().map(|&t| t as i64).collect();
177 let input = Tensor::from_i64_slice(&tokens_i64, &[1, tokens.len()], &self.device)?;
178 let outputs = self.forward(&ModelInputs::text(input))?;
179 let logits = match outputs { ModelOutputs::Logits { logits, .. } => logits, _ => return Err(anyhow::anyhow!("Expected logits")) };
180 let logits_candle = logits.to_candle()?;
181 let last = logits_candle.narrow(1, logits_candle.dims()[1] - 1, 1)?.squeeze(1)?.squeeze(0)?;
182 let logits_vec: Vec<f32> = last.to_vec1()?;
183 let next = if config.do_sample && config.temperature > 0.0 {
184 let scaled: Vec<f32> = logits_vec.iter().map(|&x| x / config.temperature).collect();
185 let max_v = scaled.iter().cloned().fold(f32::NEG_INFINITY, f32::max);
186 let exp_sum: f32 = scaled.iter().map(|&x| (x - max_v).exp()).sum();
187 let probs: Vec<f32> = scaled.iter().map(|&x| (x - max_v).exp() / exp_sum).collect();
188 let mut rng = rand::thread_rng();
189 let r: f32 = rng.gen();
190 let mut cum = 0.0;
191 let mut s = 0u32;
192 for (i, &p) in probs.iter().enumerate() { cum += p; if r <= cum { s = i as u32; break; } }
193 s
194 } else {
195 logits_vec.iter().enumerate().max_by(|a, b| a.1.partial_cmp(b.1).unwrap()).map(|(i, _)| i as u32).unwrap_or(0)
196 };
197 if next == config.eos_token_id { break; }
198 tokens.push(next);
199 }
200 Ok(tokenizer.decode(&tokens))
201 }
202
203 fn config(&self) -> &Self::Config { &self.config }
204 fn memory_requirements(&self) -> MemoryRequirements {
205 let p = self.config.vocab_size * self.config.hidden_size + self.config.num_hidden_layers * 8 * self.config.hidden_size.pow(2);
206 MemoryRequirements { gpu_memory: p * 4, cpu_memory: p, kv_cache_memory: 2 * self.config.num_hidden_layers * self.config.max_position_embeddings * self.config.hidden_size * 4, peak_memory: p * 5 }
207 }
208 fn to_device(&mut self, device: &Device) -> Result<()> {
209 self.embed_tokens = self.embed_tokens.to_device(device)?;
210 self.norm = self.norm.to_device(device)?;
211 self.lm_head = self.lm_head.to_device(device)?;
212 for layer in &mut self.layers { layer.to_device(device)?; }
213 self.device = device.clone();
214 Ok(())
215 }
216}
217
218impl CodeLlamaLayer {
219 fn new(config: &CodeLlamaConfig, device: &Device) -> Result<Self> {
220 Ok(Self {
221 self_attn: CodeLlamaAttention::new(config, device)?,
222 mlp: CodeLlamaMLP::new(config, device)?,
223 input_layernorm: ops_fn::zeros(&[config.hidden_size], DataType::Float32, device)?,
224 post_attention_layernorm: ops_fn::zeros(&[config.hidden_size], DataType::Float32, device)?,
225 })
226 }
227
228 fn forward(&self, hidden_states: &Tensor, seq_len: usize, rope_theta: f32) -> Result<Tensor> {
229 let residual = hidden_states.clone();
230 let h = ops_fn::rms_norm(hidden_states, &self.input_layernorm, 1e-5)?;
231 let attn_out = self.self_attn.forward(&h, seq_len, rope_theta)?;
232 let h = ops_fn::add(&residual, &attn_out)?;
233
234 let residual = h.clone();
235 let h = ops_fn::rms_norm(&h, &self.post_attention_layernorm, 1e-5)?;
236 let mlp_out = self.mlp.forward(&h)?;
237 ops_fn::add(&residual, &mlp_out)
238 }
239
240 fn load_weights(&mut self, weights: &ModelWeights, idx: usize) -> Result<()> {
241 let p = format!("model.layers.{}", idx);
242 if let Some(w) = weights.get(&format!("{}.self_attn.q_proj.weight", p)) { self.self_attn.q_proj = ops_fn::transpose(w)?; }
243 if let Some(w) = weights.get(&format!("{}.self_attn.k_proj.weight", p)) { self.self_attn.k_proj = ops_fn::transpose(w)?; }
244 if let Some(w) = weights.get(&format!("{}.self_attn.v_proj.weight", p)) { self.self_attn.v_proj = ops_fn::transpose(w)?; }
245 if let Some(w) = weights.get(&format!("{}.self_attn.o_proj.weight", p)) { self.self_attn.o_proj = ops_fn::transpose(w)?; }
246 if let Some(w) = weights.get(&format!("{}.mlp.gate_proj.weight", p)) { self.mlp.gate_proj = ops_fn::transpose(w)?; }
247 if let Some(w) = weights.get(&format!("{}.mlp.up_proj.weight", p)) { self.mlp.up_proj = ops_fn::transpose(w)?; }
248 if let Some(w) = weights.get(&format!("{}.mlp.down_proj.weight", p)) { self.mlp.down_proj = ops_fn::transpose(w)?; }
249 if let Some(w) = weights.get(&format!("{}.input_layernorm.weight", p)) { self.input_layernorm = w.clone(); }
250 if let Some(w) = weights.get(&format!("{}.post_attention_layernorm.weight", p)) { self.post_attention_layernorm = w.clone(); }
251 Ok(())
252 }
253
254 fn to_device(&mut self, device: &Device) -> Result<()> {
255 self.self_attn.to_device(device)?;
256 self.mlp.to_device(device)?;
257 self.input_layernorm = self.input_layernorm.to_device(device)?;
258 self.post_attention_layernorm = self.post_attention_layernorm.to_device(device)?;
259 Ok(())
260 }
261}
262
263impl CodeLlamaAttention {
264 fn new(config: &CodeLlamaConfig, device: &Device) -> Result<Self> {
265 let head_dim = config.hidden_size / config.num_attention_heads;
266 Ok(Self {
267 q_proj: ops_fn::zeros(&[config.hidden_size, config.num_attention_heads * head_dim], DataType::Float32, device)?,
268 k_proj: ops_fn::zeros(&[config.hidden_size, config.num_key_value_heads * head_dim], DataType::Float32, device)?,
269 v_proj: ops_fn::zeros(&[config.hidden_size, config.num_key_value_heads * head_dim], DataType::Float32, device)?,
270 o_proj: ops_fn::zeros(&[config.num_attention_heads * head_dim, config.hidden_size], DataType::Float32, device)?,
271 num_heads: config.num_attention_heads,
272 num_kv_heads: config.num_key_value_heads,
273 head_dim,
274 scale: 1.0 / (head_dim as f32).sqrt(),
275 })
276 }
277
278 fn forward(&self, hidden_states: &Tensor, seq_len: usize, rope_theta: f32) -> Result<Tensor> {
279 let shape = hidden_states.shape();
280 let (batch, _seq, _hidden) = (shape[0], shape[1], shape[2]);
281
282 let q = ops_fn::matmul(hidden_states, &self.q_proj)?.to_candle()?;
283 let k = ops_fn::matmul(hidden_states, &self.k_proj)?.to_candle()?;
284 let v = ops_fn::matmul(hidden_states, &self.v_proj)?.to_candle()?;
285
286 let q = q.reshape(&[batch, seq_len, self.num_heads, self.head_dim])?.transpose(1, 2)?;
287 let k = k.reshape(&[batch, seq_len, self.num_kv_heads, self.head_dim])?.transpose(1, 2)?;
288 let v = v.reshape(&[batch, seq_len, self.num_kv_heads, self.head_dim])?.transpose(1, 2)?;
289
290 let (q, k) = apply_rope_codellama(&q, &k, seq_len, self.head_dim, rope_theta)?;
291
292 let num_groups = self.num_heads / self.num_kv_heads;
294 let (k, v) = if num_groups > 1 {
295 let k_exp = k.unsqueeze(2)?.broadcast_as(&[batch, self.num_kv_heads, num_groups, seq_len, self.head_dim])?.reshape(&[batch, self.num_heads, seq_len, self.head_dim])?;
296 let v_exp = v.unsqueeze(2)?.broadcast_as(&[batch, self.num_kv_heads, num_groups, seq_len, self.head_dim])?.reshape(&[batch, self.num_heads, seq_len, self.head_dim])?;
297 (k_exp, v_exp)
298 } else {
299 (k, v)
300 };
301
302 let q = q.contiguous()?;
303 let k_t = k.transpose(2, 3)?.contiguous()?;
304 let scores = (q.matmul(&k_t)? * (self.scale as f64))?;
305
306 let device = scores.device();
307 let mask = {
308 let mut m = vec![0.0f32; seq_len * seq_len];
309 for i in 0..seq_len { for j in (i+1)..seq_len { m[i*seq_len+j] = f32::NEG_INFINITY; } }
310 candle_core::Tensor::from_vec(m, &[1, 1, seq_len, seq_len], device)?
311 };
312 let scores = scores.broadcast_add(&mask)?;
313
314 let v = v.contiguous()?;
315 let attn = candle_nn::ops::softmax_last_dim(&scores)?.matmul(&v)?;
316 let out = attn.transpose(1, 2)?.reshape(&[batch, seq_len, self.num_heads * self.head_dim])?;
317 ops_fn::matmul(&Tensor::from_candle(out), &self.o_proj)
318 }
319
320 fn to_device(&mut self, device: &Device) -> Result<()> {
321 self.q_proj = self.q_proj.to_device(device)?;
322 self.k_proj = self.k_proj.to_device(device)?;
323 self.v_proj = self.v_proj.to_device(device)?;
324 self.o_proj = self.o_proj.to_device(device)?;
325 Ok(())
326 }
327}
328
329impl CodeLlamaMLP {
330 fn new(config: &CodeLlamaConfig, device: &Device) -> Result<Self> {
331 Ok(Self {
332 gate_proj: ops_fn::zeros(&[config.hidden_size, config.intermediate_size], DataType::Float32, device)?,
333 up_proj: ops_fn::zeros(&[config.hidden_size, config.intermediate_size], DataType::Float32, device)?,
334 down_proj: ops_fn::zeros(&[config.intermediate_size, config.hidden_size], DataType::Float32, device)?,
335 })
336 }
337 fn forward(&self, x: &Tensor) -> Result<Tensor> {
338 let gate = ops_fn::matmul(x, &self.gate_proj)?;
339 let up = ops_fn::matmul(x, &self.up_proj)?;
340 let h = ops_fn::mul(&ops_fn::silu(&gate)?, &up)?;
341 ops_fn::matmul(&h, &self.down_proj)
342 }
343 fn to_device(&mut self, device: &Device) -> Result<()> {
344 self.gate_proj = self.gate_proj.to_device(device)?;
345 self.up_proj = self.up_proj.to_device(device)?;
346 self.down_proj = self.down_proj.to_device(device)?;
347 Ok(())
348 }
349}
350
351#[cfg(test)]
352mod tests {
353 use super::*;
354 #[test]
355 fn test_codellama_creation() {
356 let config = CodeLlamaConfig { vocab_size: 1000, hidden_size: 128, intermediate_size: 512, num_hidden_layers: 2, num_attention_heads: 4, num_key_value_heads: 4, ..Default::default() };
357 let model = CodeLlamaModelV2::new(config).unwrap();
358 assert_eq!(model.config().vocab_size(), 1000);
359 }
360 #[test]
361 fn test_codellama_forward() {
362 let config = CodeLlamaConfig { vocab_size: 100, hidden_size: 64, intermediate_size: 256, num_hidden_layers: 1, num_attention_heads: 4, num_key_value_heads: 4, ..Default::default() };
363 let model = CodeLlamaModelV2::new(config).unwrap();
364 let inputs = ModelInputs::text(ops_fn::zeros(&[2, 8], DataType::Int64, &Device::CPU).unwrap());
365 match model.forward(&inputs).unwrap() { ModelOutputs::Logits { logits, .. } => assert_eq!(logits.shape(), &[2, 8, 100]), _ => panic!() }
366 }
367}