candle_transformers/models/
yi.rs

1//! Yi model implementation.
2//!
3//! This candle implementation uses a pre-trained Yi decoder-only large language model for inference.
4//! The model was trained by 01.AI and follows a standard transformer architecture similar to LLaMA.
5//!
6//! Original code:
7//! - 💻 [Yi Model](https://huggingface.co/01-ai/Yi-6B)
8//! - 💻 [Yi Modeling Code](https://huggingface.co/01-ai/Yi-6B/blob/main/modeling_yi.py)
9//! - 📝 [Technical Report](https://arxiv.org/abs/2403.04652) Yi: Open Foundation Models by 01.AI
10//!
11//! Key characteristics:
12//! - Multi-head attention with rotary positional embeddings
13//! - RMS normalization
14//! - SwiGLU activation in feed-forward layers
15//! - Grouped-query attention for efficient inference
16//!
17
18use crate::models::with_tracing::{linear_no_bias, Linear, RmsNorm};
19use candle::{DType, Device, Module, Result, Tensor, D};
20use candle_nn::{Activation, VarBuilder};
21use std::sync::Arc;
22
23#[derive(Debug, Clone, PartialEq)]
24pub struct Config {
25    pub(crate) vocab_size: usize,
26    pub(crate) hidden_size: usize,
27    pub(crate) intermediate_size: usize,
28    pub(crate) num_hidden_layers: usize,
29    pub(crate) num_attention_heads: usize,
30    pub(crate) num_key_value_heads: usize,
31    pub(crate) hidden_act: Activation,
32    pub(crate) max_position_embeddings: usize,
33    pub(crate) rms_norm_eps: f64,
34    pub(crate) rope_theta: f64,
35}
36
37impl Config {
38    pub fn config_6b() -> Self {
39        Self {
40            vocab_size: 64000,
41            hidden_size: 4096,
42            intermediate_size: 11008,
43            num_hidden_layers: 32,
44            num_attention_heads: 32,
45            num_key_value_heads: 4,
46            hidden_act: Activation::Silu,
47            max_position_embeddings: 4096,
48            rms_norm_eps: 1e-5,
49            rope_theta: 5_000_000.,
50        }
51    }
52
53    pub fn config_34b() -> Self {
54        Self {
55            vocab_size: 64000,
56            hidden_size: 7168,
57            intermediate_size: 20480,
58            num_hidden_layers: 60,
59            num_attention_heads: 56,
60            num_key_value_heads: 8,
61            hidden_act: Activation::Silu,
62            max_position_embeddings: 4096,
63            rms_norm_eps: 1e-5,
64            rope_theta: 5_000_000.,
65        }
66    }
67}
68
69#[derive(Debug, Clone)]
70struct RotaryEmbedding {
71    sin: Tensor,
72    cos: Tensor,
73}
74
75fn rotate_half(xs: &Tensor) -> Result<Tensor> {
76    let last_dim = xs.dim(D::Minus1)?;
77    let xs1 = xs.narrow(D::Minus1, 0, last_dim / 2)?;
78    let xs2 = xs.narrow(D::Minus1, last_dim / 2, last_dim - last_dim / 2)?;
79    Tensor::cat(&[&xs2.neg()?, &xs1], D::Minus1)
80}
81
82impl RotaryEmbedding {
83    fn new(dtype: DType, cfg: &Config, dev: &Device) -> Result<Self> {
84        let dim = cfg.hidden_size / cfg.num_attention_heads;
85        let max_seq_len = cfg.max_position_embeddings;
86        let inv_freq: Vec<_> = (0..dim)
87            .step_by(2)
88            .map(|i| 1f32 / 10000f32.powf(i as f32 / dim as f32))
89            .collect();
90        let inv_freq_len = inv_freq.len();
91        let inv_freq = Tensor::from_vec(inv_freq, (1, inv_freq_len), dev)?.to_dtype(dtype)?;
92        let t = Tensor::arange(0u32, max_seq_len as u32, dev)?
93            .to_dtype(dtype)?
94            .reshape((max_seq_len, 1))?;
95        let freqs = t.matmul(&inv_freq)?;
96        let freqs = Tensor::cat(&[&freqs, &freqs], D::Minus1)?;
97        Ok(Self {
98            sin: freqs.sin()?,
99            cos: freqs.cos()?,
100        })
101    }
102
103    fn apply_rotary_emb_qkv(
104        &self,
105        q: &Tensor,
106        k: &Tensor,
107        seqlen_offset: usize,
108    ) -> Result<(Tensor, Tensor)> {
109        let (_b_sz, _h, seq_len, _n_embd) = q.dims4()?;
110        let cos = self.cos.narrow(0, seqlen_offset, seq_len)?;
111        let sin = self.sin.narrow(0, seqlen_offset, seq_len)?;
112        let cos = cos.unsqueeze(0)?.unsqueeze(0)?; // (1, 1, seq_len, dim)
113        let sin = sin.unsqueeze(0)?.unsqueeze(0)?; // (1, 1, seq_len, dim)
114        let q_embed = (q.broadcast_mul(&cos)? + rotate_half(q)?.broadcast_mul(&sin))?;
115        let k_embed = (k.broadcast_mul(&cos)? + rotate_half(k)?.broadcast_mul(&sin))?;
116        Ok((q_embed, k_embed))
117    }
118}
119
120#[derive(Debug, Clone)]
121#[allow(clippy::upper_case_acronyms)]
122struct MLP {
123    gate_proj: Linear,
124    up_proj: Linear,
125    down_proj: Linear,
126    act_fn: Activation,
127}
128
129impl MLP {
130    fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
131        let hidden_sz = cfg.hidden_size;
132        let intermediate_sz = cfg.intermediate_size;
133        let gate_proj = linear_no_bias(hidden_sz, intermediate_sz, vb.pp("gate_proj"))?;
134        let up_proj = linear_no_bias(hidden_sz, intermediate_sz, vb.pp("up_proj"))?;
135        let down_proj = linear_no_bias(intermediate_sz, hidden_sz, vb.pp("down_proj"))?;
136        Ok(Self {
137            gate_proj,
138            up_proj,
139            down_proj,
140            act_fn: cfg.hidden_act,
141        })
142    }
143}
144
145impl Module for MLP {
146    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
147        let lhs = xs.apply(&self.gate_proj)?.apply(&self.act_fn)?;
148        let rhs = xs.apply(&self.up_proj)?;
149        (lhs * rhs)?.apply(&self.down_proj)
150    }
151}
152
153#[derive(Debug, Clone)]
154struct Attention {
155    q_proj: Linear,
156    k_proj: Linear,
157    v_proj: Linear,
158    o_proj: Linear,
159    num_heads: usize,
160    num_kv_heads: usize,
161    num_kv_groups: usize,
162    head_dim: usize,
163    hidden_size: usize,
164    rotary_emb: Arc<RotaryEmbedding>,
165    kv_cache: Option<(Tensor, Tensor)>,
166}
167
168impl Attention {
169    fn new(rotary_emb: Arc<RotaryEmbedding>, cfg: &Config, vb: VarBuilder) -> Result<Self> {
170        let hidden_sz = cfg.hidden_size;
171        let num_heads = cfg.num_attention_heads;
172        let num_kv_heads = cfg.num_key_value_heads;
173        let num_kv_groups = num_heads / num_kv_heads;
174        let head_dim = hidden_sz / num_heads;
175        let q_proj = linear_no_bias(hidden_sz, num_heads * head_dim, vb.pp("q_proj"))?;
176        let k_proj = linear_no_bias(hidden_sz, num_kv_heads * head_dim, vb.pp("k_proj"))?;
177        let v_proj = linear_no_bias(hidden_sz, num_kv_heads * head_dim, vb.pp("v_proj"))?;
178        let o_proj = linear_no_bias(num_heads * head_dim, hidden_sz, 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            hidden_size: hidden_sz,
189            rotary_emb,
190            kv_cache: None,
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        let key_states = key_states
210            .reshape((b_sz, q_len, self.num_kv_heads, self.head_dim))?
211            .transpose(1, 2)?;
212        let value_states = value_states
213            .reshape((b_sz, q_len, self.num_kv_heads, self.head_dim))?
214            .transpose(1, 2)?;
215
216        let (query_states, key_states) =
217            self.rotary_emb
218                .apply_rotary_emb_qkv(&query_states, &key_states, seqlen_offset)?;
219
220        let (key_states, value_states) = match &self.kv_cache {
221            None => (key_states, value_states),
222            Some((prev_k, prev_v)) => {
223                let key_states = Tensor::cat(&[prev_k, &key_states], 2)?;
224                let value_states = Tensor::cat(&[prev_v, &value_states], 2)?;
225                (key_states, value_states)
226            }
227        };
228        self.kv_cache = Some((key_states.clone(), value_states.clone()));
229
230        let key_states = crate::utils::repeat_kv(key_states, self.num_kv_groups)?;
231        let value_states = crate::utils::repeat_kv(value_states, self.num_kv_groups)?;
232
233        let attn_output = {
234            let scale = 1f64 / f64::sqrt(self.head_dim as f64);
235            let attn_weights = (query_states.matmul(&key_states.transpose(2, 3)?)? * scale)?;
236
237            let attn_weights = match attention_mask {
238                None => attn_weights,
239                Some(mask) => attn_weights.broadcast_add(mask)?,
240            };
241            let attn_weights = candle_nn::ops::softmax_last_dim(&attn_weights)?;
242            attn_weights.matmul(&value_states)?
243        };
244        attn_output
245            .transpose(1, 2)?
246            .reshape((b_sz, q_len, self.hidden_size))?
247            .apply(&self.o_proj)
248    }
249}
250
251#[derive(Debug, Clone)]
252struct DecoderLayer {
253    self_attn: Attention,
254    mlp: MLP,
255    ln1: RmsNorm,
256    ln2: RmsNorm,
257}
258
259impl DecoderLayer {
260    fn new(rotary_emb: Arc<RotaryEmbedding>, cfg: &Config, vb: VarBuilder) -> Result<Self> {
261        let self_attn = Attention::new(rotary_emb, cfg, vb.pp("self_attn"))?;
262        let mlp = MLP::new(cfg, vb.pp("mlp"))?;
263        let ln1 = RmsNorm::new(cfg.hidden_size, cfg.rms_norm_eps, vb.pp("input_layernorm"))?;
264        let ln2 = RmsNorm::new(
265            cfg.hidden_size,
266            cfg.rms_norm_eps,
267            vb.pp("post_attention_layernorm"),
268        )?;
269        Ok(Self {
270            self_attn,
271            mlp,
272            ln1,
273            ln2,
274        })
275    }
276
277    fn forward(
278        &mut self,
279        xs: &Tensor,
280        attention_mask: Option<&Tensor>,
281        seqlen_offset: usize,
282    ) -> Result<Tensor> {
283        let residual = xs;
284        let xs = self.ln1.forward(xs)?;
285        let xs = self.self_attn.forward(&xs, attention_mask, seqlen_offset)?;
286        let xs = (xs + residual)?;
287        let residual = &xs;
288        let xs = xs.apply(&self.ln2)?.apply(&self.mlp)?;
289        residual + xs
290    }
291}
292
293#[derive(Debug, Clone)]
294pub struct Model {
295    embed_tokens: candle_nn::Embedding,
296    layers: Vec<DecoderLayer>,
297    norm: RmsNorm,
298    lm_head: Linear,
299    device: Device,
300    dtype: DType,
301}
302
303impl Model {
304    pub fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
305        let vb_m = vb.pp("model");
306        let embed_tokens =
307            candle_nn::embedding(cfg.vocab_size, cfg.hidden_size, vb_m.pp("embed_tokens"))?;
308        let rotary_emb = Arc::new(RotaryEmbedding::new(vb.dtype(), cfg, vb_m.device())?);
309        let mut layers = Vec::with_capacity(cfg.num_hidden_layers);
310        let vb_l = vb_m.pp("layers");
311        for layer_idx in 0..cfg.num_hidden_layers {
312            let layer = DecoderLayer::new(rotary_emb.clone(), cfg, vb_l.pp(layer_idx))?;
313            layers.push(layer)
314        }
315        let norm = RmsNorm::new(cfg.hidden_size, cfg.rms_norm_eps, vb_m.pp("norm"))?;
316        let lm_head = linear_no_bias(cfg.hidden_size, cfg.vocab_size, vb.pp("lm_head"))?;
317        Ok(Self {
318            embed_tokens,
319            layers,
320            norm,
321            lm_head,
322            device: vb.device().clone(),
323            dtype: vb.dtype(),
324        })
325    }
326
327    fn prepare_decoder_attention_mask(
328        &self,
329        b_size: usize,
330        tgt_len: usize,
331        seqlen_offset: usize,
332    ) -> Result<Tensor> {
333        // Sliding window mask?
334        let mask: Vec<_> = (0..tgt_len)
335            .flat_map(|i| (0..tgt_len).map(move |j| if i < j { f32::NEG_INFINITY } else { 0. }))
336            .collect();
337        let mask = Tensor::from_slice(&mask, (tgt_len, tgt_len), &self.device)?;
338        let mask = if seqlen_offset > 0 {
339            let mask0 = Tensor::zeros((tgt_len, seqlen_offset), DType::F32, &self.device)?;
340            Tensor::cat(&[&mask0, &mask], D::Minus1)?
341        } else {
342            mask
343        };
344        mask.expand((b_size, 1, tgt_len, tgt_len + seqlen_offset))?
345            .to_dtype(self.dtype)
346    }
347
348    pub fn forward(&mut self, input_ids: &Tensor, seqlen_offset: usize) -> Result<Tensor> {
349        let (b_size, seq_len) = input_ids.dims2()?;
350        let attention_mask = if seq_len <= 1 {
351            None
352        } else {
353            let mask = self.prepare_decoder_attention_mask(b_size, seq_len, seqlen_offset)?;
354            Some(mask)
355        };
356        let mut xs = self.embed_tokens.forward(input_ids)?;
357        for layer in self.layers.iter_mut() {
358            xs = layer.forward(&xs, attention_mask.as_ref(), seqlen_offset)?
359        }
360        xs.narrow(1, seq_len - 1, 1)?
361            .apply(&self.norm)?
362            .apply(&self.lm_head)
363    }
364}
candle_transformers/models/yi.rs

candle_transformers/models/
yi.rs
