1use super::with_tracing::{linear_b as linear, Linear, RmsNorm};
6use candle::{DType, Device, Result, Tensor, D};
7use candle_nn::{Module, VarBuilder};
8use std::sync::Arc;
9
10fn default_use_flash_attn() -> bool {
11 false
12}
13
14#[derive(Debug, Clone, serde::Deserialize)]
15pub struct Config {
16 pub attention_bias: bool,
17 pub bos_token_id: u32,
18 pub eos_token_id: u32,
19 pub head_dim: usize,
20 pub hidden_act: candle_nn::Activation,
21 pub hidden_size: usize,
22 pub intermediate_size: usize,
23 pub max_position_embeddings: usize,
24 pub mlp_bias: bool,
25 pub num_attention_heads: usize,
26 pub num_hidden_layers: usize,
27 pub num_key_value_heads: usize,
28 pub rms_norm_eps: f64,
29 pub rope_theta: f64,
30 pub tie_word_embeddings: bool,
31 pub vocab_size: usize,
32 #[serde(default = "default_use_flash_attn")]
33 pub use_flash_attn: bool,
34}
35
36impl Config {
37 pub fn config_2b(use_flash_attn: bool) -> Self {
38 Self {
39 attention_bias: false,
40 bos_token_id: 1,
41 eos_token_id: 2,
42 head_dim: 128,
43 hidden_act: candle_nn::Activation::Silu,
44 hidden_size: 2560,
45 intermediate_size: 7040,
46 max_position_embeddings: 4096,
47 mlp_bias: false,
48 num_attention_heads: 20,
49 num_hidden_layers: 24,
50 num_key_value_heads: 20,
51 rms_norm_eps: 1e-08,
52 rope_theta: 100000.0,
53 tie_word_embeddings: false,
54 vocab_size: 48000,
55 use_flash_attn,
56 }
57 }
58}
59
60#[derive(Debug, Clone)]
61struct RotaryEmbedding {
62 sin: Tensor,
63 cos: Tensor,
64}
65
66impl RotaryEmbedding {
67 fn new(dtype: DType, cfg: &Config, dev: &Device) -> Result<Self> {
68 let rope_theta = cfg.rope_theta as f32;
69 let dim = cfg.head_dim;
70 let max_seq_len = cfg.max_position_embeddings;
71 let inv_freq: Vec<_> = (0..dim)
72 .step_by(2)
73 .map(|i| 1f32 / rope_theta.powf(i as f32 / dim as f32))
74 .collect();
75 let inv_freq_len = inv_freq.len();
76 let inv_freq = Tensor::from_vec(inv_freq, (1, inv_freq_len), dev)?.to_dtype(DType::F32)?;
77 let t = Tensor::arange(0u32, max_seq_len as u32, dev)?
78 .to_dtype(DType::F32)?
79 .reshape((max_seq_len, 1))?;
80 let freqs = t.matmul(&inv_freq)?;
81 Ok(Self {
82 sin: freqs.sin()?.to_dtype(dtype)?,
83 cos: freqs.cos()?.to_dtype(dtype)?,
84 })
85 }
86
87 fn apply_rotary_emb_qkv(
88 &self,
89 q: &Tensor,
90 k: &Tensor,
91 seqlen_offset: usize,
92 ) -> Result<(Tensor, Tensor)> {
93 let (_b_sz, _h, seq_len, _n_embd) = q.dims4()?;
94 let cos = self.cos.narrow(0, seqlen_offset, seq_len)?;
95 let sin = self.sin.narrow(0, seqlen_offset, seq_len)?;
96 let q_embed = candle_nn::rotary_emb::rope_i(q, &cos, &sin)?;
97 let k_embed = candle_nn::rotary_emb::rope_i(k, &cos, &sin)?;
98 Ok((q_embed, k_embed))
99 }
100}
101
102#[derive(Debug, Clone)]
103#[allow(clippy::upper_case_acronyms)]
104struct MLP {
105 gate_proj: Linear,
106 up_proj: Linear,
107 down_proj: Linear,
108 act_fn: candle_nn::Activation,
109}
110
111impl MLP {
112 fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
113 let hidden_sz = cfg.hidden_size;
114 let intermediate_sz = cfg.intermediate_size;
115 let bias = cfg.mlp_bias;
116 let gate_proj = linear(hidden_sz, intermediate_sz, bias, vb.pp("gate_proj"))?;
117 let up_proj = linear(hidden_sz, intermediate_sz, bias, vb.pp("up_proj"))?;
118 let down_proj = linear(intermediate_sz, hidden_sz, bias, vb.pp("down_proj"))?;
119 Ok(Self {
120 gate_proj,
121 up_proj,
122 down_proj,
123 act_fn: cfg.hidden_act,
124 })
125 }
126}
127
128impl Module for MLP {
129 fn forward(&self, xs: &Tensor) -> Result<Tensor> {
130 let lhs = xs.apply(&self.gate_proj)?.apply(&self.act_fn)?;
131 let rhs = xs.apply(&self.up_proj)?;
132 (lhs * rhs)?.apply(&self.down_proj)
133 }
134}
135
136#[cfg(feature = "flash-attn")]
137fn flash_attn(
138 q: &Tensor,
139 k: &Tensor,
140 v: &Tensor,
141 softmax_scale: f32,
142 causal: bool,
143) -> Result<Tensor> {
144 candle_flash_attn::flash_attn(q, k, v, softmax_scale, causal)
145}
146
147#[cfg(not(feature = "flash-attn"))]
148fn flash_attn(_: &Tensor, _: &Tensor, _: &Tensor, _: f32, _: bool) -> Result<Tensor> {
149 unimplemented!("compile with '--features flash-attn'")
150}
151
152#[derive(Debug, Clone)]
153struct Attention {
154 q_proj: Linear,
155 k_proj: Linear,
156 v_proj: Linear,
157 o_proj: Linear,
158 num_heads: usize,
159 num_kv_heads: usize,
160 num_kv_groups: usize,
161 head_dim: usize,
162 rotary_emb: Arc<RotaryEmbedding>,
163 kv_cache: Option<(Tensor, Tensor)>,
164 use_flash_attn: bool,
165}
166
167impl Attention {
168 fn new(rotary_emb: Arc<RotaryEmbedding>, cfg: &Config, vb: VarBuilder) -> Result<Self> {
169 let hidden_sz = cfg.hidden_size;
170 let num_heads = cfg.num_attention_heads;
171 let num_kv_heads = cfg.num_key_value_heads;
172 let num_kv_groups = num_heads / num_kv_heads;
173 let head_dim = cfg.head_dim;
174 let bias = cfg.attention_bias;
175 let q_proj = linear(hidden_sz, num_heads * head_dim, bias, vb.pp("q_proj"))?;
176 let k_proj = linear(hidden_sz, num_kv_heads * head_dim, bias, vb.pp("k_proj"))?;
177 let v_proj = linear(hidden_sz, num_kv_heads * head_dim, bias, vb.pp("v_proj"))?;
178 let o_proj = linear(num_heads * head_dim, hidden_sz, bias, vb.pp("o_proj"))?;
179 Ok(Self {
180 q_proj,
181 k_proj,
182 v_proj,
183 o_proj,
184 num_heads,
185 num_kv_heads,
186 num_kv_groups,
187 head_dim,
188 rotary_emb,
189 kv_cache: None,
190 use_flash_attn: cfg.use_flash_attn,
191 })
192 }
193
194 fn forward(
195 &mut self,
196 xs: &Tensor,
197 attention_mask: Option<&Tensor>,
198 seqlen_offset: usize,
199 ) -> Result<Tensor> {
200 let (b_sz, q_len, _) = xs.dims3()?;
201
202 let query_states = self.q_proj.forward(xs)?;
203 let key_states = self.k_proj.forward(xs)?;
204 let value_states = self.v_proj.forward(xs)?;
205
206 let query_states = query_states
207 .reshape((b_sz, q_len, self.num_heads, self.head_dim))?
208 .transpose(1, 2)?
209 .contiguous()?;
210 let key_states = key_states
211 .reshape((b_sz, q_len, self.num_kv_heads, self.head_dim))?
212 .transpose(1, 2)?
213 .contiguous()?;
214 let value_states = value_states
215 .reshape((b_sz, q_len, self.num_kv_heads, self.head_dim))?
216 .transpose(1, 2)?
217 .contiguous()?;
218
219 let (query_states, key_states) =
220 self.rotary_emb
221 .apply_rotary_emb_qkv(&query_states, &key_states, seqlen_offset)?;
222
223 let (key_states, value_states) = match &self.kv_cache {
224 None => (key_states, value_states),
225 Some((prev_k, prev_v)) => {
226 let key_states = Tensor::cat(&[prev_k, &key_states], 2)?;
227 let value_states = Tensor::cat(&[prev_v, &value_states], 2)?;
228 (key_states, value_states)
229 }
230 };
231 self.kv_cache = Some((key_states.clone(), value_states.clone()));
232
233 let key_states = crate::utils::repeat_kv(key_states, self.num_kv_groups)?;
234 let value_states = crate::utils::repeat_kv(value_states, self.num_kv_groups)?;
235
236 let attn_output = if self.use_flash_attn {
237 let q = query_states.transpose(1, 2)?;
239 let k = key_states.transpose(1, 2)?;
240 let v = value_states.transpose(1, 2)?;
241 let softmax_scale = 1f32 / (self.head_dim as f32).sqrt();
242 flash_attn(&q, &k, &v, softmax_scale, q_len > 1)?.transpose(1, 2)?
243 } else {
244 let scale = 1f64 / f64::sqrt(self.head_dim as f64);
245 let attn_weights = (query_states.matmul(&key_states.transpose(2, 3)?)? * scale)?;
246
247 let attn_weights = match attention_mask {
248 None => attn_weights,
249 Some(mask) => attn_weights.broadcast_add(mask)?,
250 };
251 let attn_weights = candle_nn::ops::softmax_last_dim(&attn_weights)?;
252 attn_weights.matmul(&value_states)?
253 };
254 attn_output
255 .transpose(1, 2)?
256 .reshape((b_sz, q_len, self.num_heads * self.head_dim))?
257 .apply(&self.o_proj)
258 }
259
260 fn clear_kv_cache(&mut self) {
261 self.kv_cache = None
262 }
263}
264
265#[derive(Debug, Clone)]
266struct DecoderLayer {
267 self_attn: Attention,
268 mlp: MLP,
269 input_layernorm: RmsNorm,
270 post_attention_layernorm: RmsNorm,
271}
272
273impl DecoderLayer {
274 fn new(rotary_emb: Arc<RotaryEmbedding>, cfg: &Config, vb: VarBuilder) -> Result<Self> {
275 let self_attn = Attention::new(rotary_emb, cfg, vb.pp("self_attn"))?;
276 let mlp = MLP::new(cfg, vb.pp("mlp"))?;
277 let input_layernorm =
278 RmsNorm::new(cfg.hidden_size, cfg.rms_norm_eps, vb.pp("input_layernorm"))?;
279 let post_attention_layernorm = RmsNorm::new(
280 cfg.hidden_size,
281 cfg.rms_norm_eps,
282 vb.pp("post_attention_layernorm"),
283 )?;
284 Ok(Self {
285 self_attn,
286 mlp,
287 input_layernorm,
288 post_attention_layernorm,
289 })
290 }
291
292 fn forward(
293 &mut self,
294 xs: &Tensor,
295 attention_mask: Option<&Tensor>,
296 seqlen_offset: usize,
297 ) -> Result<Tensor> {
298 let residual = xs;
299 let xs = self.input_layernorm.forward(xs)?;
300 let xs = self.self_attn.forward(&xs, attention_mask, seqlen_offset)?;
301 let xs = (xs + residual)?;
302 let residual = &xs;
303 let xs = xs.apply(&self.post_attention_layernorm)?.apply(&self.mlp)?;
304 residual + xs
305 }
306
307 fn clear_kv_cache(&mut self) {
308 self.self_attn.clear_kv_cache()
309 }
310}
311
312#[derive(Debug, Clone)]
313pub struct Model {
314 embed_tokens: candle_nn::Embedding,
315 layers: Vec<DecoderLayer>,
316 norm: RmsNorm,
317 lm_head: Linear,
318 device: Device,
319 dtype: DType,
320}
321
322impl Model {
323 pub fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
324 let vb_m = vb.pp("model");
325 let embed_tokens =
326 candle_nn::embedding(cfg.vocab_size, cfg.hidden_size, vb_m.pp("embed_tokens"))?;
327 let rotary_emb = Arc::new(RotaryEmbedding::new(vb.dtype(), cfg, vb_m.device())?);
328 let mut layers = Vec::with_capacity(cfg.num_hidden_layers);
329 let vb_l = vb_m.pp("layers");
330 for layer_idx in 0..cfg.num_hidden_layers {
331 let layer = DecoderLayer::new(rotary_emb.clone(), cfg, vb_l.pp(layer_idx))?;
332 layers.push(layer)
333 }
334 let norm = RmsNorm::new(cfg.hidden_size, cfg.rms_norm_eps, vb_m.pp("norm"))?;
335 let lm_head = if cfg.tie_word_embeddings {
336 Linear::from_weights(embed_tokens.embeddings().clone(), None)
337 } else {
338 linear(cfg.hidden_size, cfg.vocab_size, false, vb.pp("lm_head"))?
339 };
340 Ok(Self {
341 embed_tokens,
342 layers,
343 norm,
344 lm_head,
345 device: vb.device().clone(),
346 dtype: vb.dtype(),
347 })
348 }
349
350 fn prepare_decoder_attention_mask(
351 &self,
352 tgt_len: usize,
353 seqlen_offset: usize,
354 ) -> Result<Tensor> {
355 let mask: Vec<_> = (0..tgt_len)
356 .flat_map(|i| (0..tgt_len).map(move |j| if i < j { f32::NEG_INFINITY } else { 0. }))
357 .collect();
358 let mask = Tensor::from_slice(&mask, (tgt_len, tgt_len), &self.device)?;
359 let mask = if seqlen_offset > 0 {
360 let mask0 = Tensor::zeros((tgt_len, seqlen_offset), DType::F32, &self.device)?;
361 Tensor::cat(&[&mask0, &mask], D::Minus1)?
362 } else {
363 mask
364 };
365 mask.expand((1, 1, tgt_len, tgt_len + seqlen_offset))?
366 .to_dtype(self.dtype)
367 }
368
369 pub fn embed_tokens(&self) -> &candle_nn::Embedding {
370 &self.embed_tokens
371 }
372
373 pub fn forward(&mut self, input_ids: &Tensor, seqlen_offset: usize) -> Result<Tensor> {
374 let (_b_size, seq_len) = input_ids.dims2()?;
375 let attention_mask = if seq_len <= 1 {
376 None
377 } else {
378 let mask = self.prepare_decoder_attention_mask(seq_len, seqlen_offset)?;
379 Some(mask)
380 };
381 let mut xs = self.embed_tokens.forward(input_ids)?;
382 for layer in self.layers.iter_mut() {
383 xs = layer.forward(&xs, attention_mask.as_ref(), seqlen_offset)?
384 }
385 xs.narrow(1, seq_len - 1, 1)?
386 .apply(&self.norm)?
387 .apply(&self.lm_head)
388 }
389
390 pub fn clear_kv_cache(&mut self) {
391 for layer in self.layers.iter_mut() {
392 layer.clear_kv_cache()
393 }
394 }
395}