candle_transformers/models/
qwen2.rs

1//! Qwen2 model implementation with quantization support.
2//!
3//! Qwen2 is a large language model from Alibaba optimized for efficiency.
4//! This implementation provides quantization for reduced memory and compute.
5//!
6//! Key characteristics:
7//! - Streaming decode support
8//! - Grouped query attention (GQA)
9//! - RMSNorm for layer normalization
10//! - Rotary positional embeddings (RoPE)
11//! - Support for 8-bit quantization
12//!
13//! References:
14//! - 🤗 [Qwen2 Model](https://huggingface.co/Qwen/Qwen2-7B)
15//!
16
17use crate::models::with_tracing::{linear, linear_no_bias, Linear, RmsNorm};
18use candle::{DType, Device, IndexOp, Module, Result, Tensor, D};
19use candle_nn::{Activation, VarBuilder};
20use std::sync::Arc;
21
22#[derive(Debug, Clone, PartialEq, serde::Deserialize)]
23pub struct Config {
24    pub vocab_size: usize,
25    pub hidden_size: usize,
26    pub intermediate_size: usize,
27    pub num_hidden_layers: usize,
28    pub num_attention_heads: usize,
29    pub num_key_value_heads: usize,
30    pub max_position_embeddings: usize,
31    pub sliding_window: usize,
32    pub max_window_layers: usize,
33    pub tie_word_embeddings: bool,
34    pub rope_theta: f64,
35    pub rms_norm_eps: f64,
36    pub use_sliding_window: bool,
37    pub hidden_act: Activation,
38}
39
40#[derive(Debug, Clone)]
41struct RotaryEmbedding {
42    sin: Tensor,
43    cos: Tensor,
44}
45
46impl RotaryEmbedding {
47    fn new(dtype: DType, cfg: &Config, dev: &Device) -> Result<Self> {
48        let dim = cfg.hidden_size / cfg.num_attention_heads;
49        let max_seq_len = cfg.max_position_embeddings;
50        let inv_freq: Vec<_> = (0..dim)
51            .step_by(2)
52            .map(|i| 1f32 / cfg.rope_theta.powf(i as f64 / dim as f64) as f32)
53            .collect();
54        let inv_freq_len = inv_freq.len();
55        let inv_freq = Tensor::from_vec(inv_freq, (1, inv_freq_len), dev)?.to_dtype(dtype)?;
56        let t = Tensor::arange(0u32, max_seq_len as u32, dev)?
57            .to_dtype(dtype)?
58            .reshape((max_seq_len, 1))?;
59        let freqs = t.matmul(&inv_freq)?;
60        Ok(Self {
61            sin: freqs.sin()?,
62            cos: freqs.cos()?,
63        })
64    }
65
66    fn apply_rotary_emb_qkv(
67        &self,
68        q: &Tensor,
69        k: &Tensor,
70        seqlen_offset: usize,
71    ) -> Result<(Tensor, Tensor)> {
72        let (_b_sz, _h, seq_len, _n_embd) = q.dims4()?;
73        let cos = self.cos.narrow(0, seqlen_offset, seq_len)?;
74        let sin = self.sin.narrow(0, seqlen_offset, seq_len)?;
75        let q_embed = candle_nn::rotary_emb::rope(&q.contiguous()?, &cos, &sin)?;
76        let k_embed = candle_nn::rotary_emb::rope(&k.contiguous()?, &cos, &sin)?;
77        Ok((q_embed, k_embed))
78    }
79}
80
81#[derive(Debug, Clone)]
82#[allow(clippy::upper_case_acronyms)]
83struct MLP {
84    gate_proj: Linear,
85    up_proj: Linear,
86    down_proj: Linear,
87    act_fn: Activation,
88}
89
90impl MLP {
91    fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
92        let hidden_sz = cfg.hidden_size;
93        let intermediate_sz = cfg.intermediate_size;
94        let gate_proj = linear_no_bias(hidden_sz, intermediate_sz, vb.pp("gate_proj"))?;
95        let up_proj = linear_no_bias(hidden_sz, intermediate_sz, vb.pp("up_proj"))?;
96        let down_proj = linear_no_bias(intermediate_sz, hidden_sz, vb.pp("down_proj"))?;
97        Ok(Self {
98            gate_proj,
99            up_proj,
100            down_proj,
101            act_fn: cfg.hidden_act,
102        })
103    }
104}
105
106impl Module for MLP {
107    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
108        let lhs = xs.apply(&self.gate_proj)?.apply(&self.act_fn)?;
109        let rhs = xs.apply(&self.up_proj)?;
110        (lhs * rhs)?.apply(&self.down_proj)
111    }
112}
113
114#[derive(Debug, Clone)]
115struct Attention {
116    q_proj: Linear,
117    k_proj: Linear,
118    v_proj: Linear,
119    o_proj: Linear,
120    num_heads: usize,
121    num_kv_heads: usize,
122    num_kv_groups: usize,
123    head_dim: usize,
124    hidden_size: usize,
125    rotary_emb: Arc<RotaryEmbedding>,
126    kv_cache: Option<(Tensor, Tensor)>,
127}
128
129impl Attention {
130    fn new(rotary_emb: Arc<RotaryEmbedding>, cfg: &Config, vb: VarBuilder) -> Result<Self> {
131        let hidden_sz = cfg.hidden_size;
132        let num_heads = cfg.num_attention_heads;
133        let num_kv_heads = cfg.num_key_value_heads;
134        let num_kv_groups = num_heads / num_kv_heads;
135        let head_dim = hidden_sz / num_heads;
136        let q_proj = linear(hidden_sz, num_heads * head_dim, vb.pp("q_proj"))?;
137        let k_proj = linear(hidden_sz, num_kv_heads * head_dim, vb.pp("k_proj"))?;
138        let v_proj = linear(hidden_sz, num_kv_heads * head_dim, vb.pp("v_proj"))?;
139        let o_proj = linear_no_bias(num_heads * head_dim, hidden_sz, vb.pp("o_proj"))?;
140        Ok(Self {
141            q_proj,
142            k_proj,
143            v_proj,
144            o_proj,
145            num_heads,
146            num_kv_heads,
147            num_kv_groups,
148            head_dim,
149            hidden_size: hidden_sz,
150            rotary_emb,
151            kv_cache: None,
152        })
153    }
154
155    fn forward(
156        &mut self,
157        xs: &Tensor,
158        attention_mask: Option<&Tensor>,
159        seqlen_offset: usize,
160    ) -> Result<Tensor> {
161        let (b_sz, q_len, _) = xs.dims3()?;
162
163        let query_states = self.q_proj.forward(xs)?;
164        let key_states = self.k_proj.forward(xs)?;
165        let value_states = self.v_proj.forward(xs)?;
166
167        let query_states = query_states
168            .reshape((b_sz, q_len, self.num_heads, self.head_dim))?
169            .transpose(1, 2)?;
170        let key_states = key_states
171            .reshape((b_sz, q_len, self.num_kv_heads, self.head_dim))?
172            .transpose(1, 2)?;
173        let value_states = value_states
174            .reshape((b_sz, q_len, self.num_kv_heads, self.head_dim))?
175            .transpose(1, 2)?;
176
177        let (query_states, key_states) =
178            self.rotary_emb
179                .apply_rotary_emb_qkv(&query_states, &key_states, seqlen_offset)?;
180
181        let (key_states, value_states) = match &self.kv_cache {
182            None => (key_states, value_states),
183            Some((prev_k, prev_v)) => {
184                let key_states = Tensor::cat(&[prev_k, &key_states], 2)?;
185                let value_states = Tensor::cat(&[prev_v, &value_states], 2)?;
186                (key_states, value_states)
187            }
188        };
189        self.kv_cache = Some((key_states.clone(), value_states.clone()));
190
191        let key_states = crate::utils::repeat_kv(key_states, self.num_kv_groups)?.contiguous()?;
192        let value_states =
193            crate::utils::repeat_kv(value_states, self.num_kv_groups)?.contiguous()?;
194
195        let attn_output = {
196            let scale = 1f64 / f64::sqrt(self.head_dim as f64);
197            let attn_weights = (query_states.matmul(&key_states.transpose(2, 3)?)? * scale)?;
198
199            let attn_weights = match attention_mask {
200                None => attn_weights,
201                Some(mask) => attn_weights.broadcast_add(mask)?,
202            };
203            let attn_weights = candle_nn::ops::softmax_last_dim(&attn_weights)?;
204            attn_weights.matmul(&value_states)?
205        };
206        attn_output
207            .transpose(1, 2)?
208            .reshape((b_sz, q_len, self.hidden_size))?
209            .apply(&self.o_proj)
210    }
211
212    fn clear_kv_cache(&mut self) {
213        self.kv_cache = None
214    }
215}
216
217#[derive(Debug, Clone)]
218struct DecoderLayer {
219    self_attn: Attention,
220    mlp: MLP,
221    input_layernorm: RmsNorm,
222    post_attention_layernorm: RmsNorm,
223}
224
225impl DecoderLayer {
226    fn new(rotary_emb: Arc<RotaryEmbedding>, cfg: &Config, vb: VarBuilder) -> Result<Self> {
227        let self_attn = Attention::new(rotary_emb, cfg, vb.pp("self_attn"))?;
228        let mlp = MLP::new(cfg, vb.pp("mlp"))?;
229        let input_layernorm =
230            RmsNorm::new(cfg.hidden_size, cfg.rms_norm_eps, vb.pp("input_layernorm"))?;
231        let post_attention_layernorm = RmsNorm::new(
232            cfg.hidden_size,
233            cfg.rms_norm_eps,
234            vb.pp("post_attention_layernorm"),
235        )?;
236        Ok(Self {
237            self_attn,
238            mlp,
239            input_layernorm,
240            post_attention_layernorm,
241        })
242    }
243
244    fn forward(
245        &mut self,
246        xs: &Tensor,
247        attention_mask: Option<&Tensor>,
248        seqlen_offset: usize,
249    ) -> Result<Tensor> {
250        let residual = xs;
251        let xs = self.input_layernorm.forward(xs)?;
252        let xs = self.self_attn.forward(&xs, attention_mask, seqlen_offset)?;
253        let xs = (xs + residual)?;
254        let residual = &xs;
255        let xs = xs.apply(&self.post_attention_layernorm)?.apply(&self.mlp)?;
256        residual + xs
257    }
258
259    fn clear_kv_cache(&mut self) {
260        self.self_attn.clear_kv_cache()
261    }
262}
263
264#[derive(Debug, Clone)]
265pub struct Model {
266    embed_tokens: candle_nn::Embedding,
267    layers: Vec<DecoderLayer>,
268    norm: RmsNorm,
269    sliding_window: usize,
270    device: Device,
271    dtype: DType,
272}
273
274impl Model {
275    pub fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
276        let vb_m = vb.pp("model");
277        let embed_tokens =
278            candle_nn::embedding(cfg.vocab_size, cfg.hidden_size, vb_m.pp("embed_tokens"))?;
279        let rotary_emb = Arc::new(RotaryEmbedding::new(vb.dtype(), cfg, vb_m.device())?);
280        let mut layers = Vec::with_capacity(cfg.num_hidden_layers);
281        let vb_l = vb_m.pp("layers");
282        for layer_idx in 0..cfg.num_hidden_layers {
283            let layer = DecoderLayer::new(rotary_emb.clone(), cfg, vb_l.pp(layer_idx))?;
284            layers.push(layer)
285        }
286        let norm = RmsNorm::new(cfg.hidden_size, cfg.rms_norm_eps, vb_m.pp("norm"))?;
287        Ok(Self {
288            embed_tokens,
289            layers,
290            norm,
291            sliding_window: cfg.sliding_window,
292            device: vb.device().clone(),
293            dtype: vb.dtype(),
294        })
295    }
296
297    fn prepare_causal_attention_mask(
298        &self,
299        b_size: usize,
300        tgt_len: usize,
301        seqlen_offset: usize,
302    ) -> Result<Tensor> {
303        // Sliding window mask?
304        let mask: Vec<_> = (0..tgt_len)
305            .flat_map(|i| {
306                (0..tgt_len).map(move |j| {
307                    if i < j || j + self.sliding_window < i {
308                        f32::NEG_INFINITY
309                    } else {
310                        0.
311                    }
312                })
313            })
314            .collect();
315        let mask = Tensor::from_slice(&mask, (tgt_len, tgt_len), &self.device)?;
316        let mask = if seqlen_offset > 0 {
317            let mask0 = Tensor::zeros((tgt_len, seqlen_offset), self.dtype, &self.device)?;
318            Tensor::cat(&[&mask0, &mask], D::Minus1)?
319        } else {
320            mask
321        };
322        mask.expand((b_size, 1, tgt_len, tgt_len + seqlen_offset))?
323            .to_dtype(self.dtype)
324    }
325
326    fn prepare_attention_mask(&self, attn_mask: &Tensor) -> Result<Tensor> {
327        let (b_sz, sql_len) = attn_mask.dims2()?;
328        let mut mask: Vec<Tensor> = vec![];
329        for b in 0..b_sz {
330            mask.push(attn_mask.i((b, ..))?.expand((1, 1, sql_len, sql_len))?);
331        }
332        let mask = Tensor::cat(&mask, 0)?;
333        let on_true = mask.zeros_like()?.to_dtype(self.dtype)?;
334        let on_false = Tensor::new(f32::NEG_INFINITY, &self.device)?
335            .broadcast_as(mask.shape())?
336            .to_dtype(self.dtype)?;
337        mask.where_cond(&on_true, &on_false)
338    }
339
340    pub fn forward(
341        &mut self,
342        input_ids: &Tensor,
343        seqlen_offset: usize,
344        attn_mask: Option<&Tensor>,
345    ) -> Result<Tensor> {
346        let (b_size, seq_len) = input_ids.dims2()?;
347        let attention_mask: Option<Tensor> = match attn_mask {
348            Some(mask) => Some(self.prepare_attention_mask(mask)?),
349            None => {
350                if seq_len <= 1 {
351                    None
352                } else {
353                    Some(self.prepare_causal_attention_mask(b_size, seq_len, seqlen_offset)?)
354                }
355            }
356        };
357        let mut xs = self.embed_tokens.forward(input_ids)?;
358        for layer in self.layers.iter_mut() {
359            xs = layer.forward(&xs, attention_mask.as_ref(), seqlen_offset)?
360        }
361        xs.apply(&self.norm)
362    }
363
364    pub fn clear_kv_cache(&mut self) {
365        for layer in self.layers.iter_mut() {
366            layer.clear_kv_cache()
367        }
368    }
369}
370
371#[derive(Debug, Clone)]
372pub struct ModelForCausalLM {
373    base_model: Model,
374    lm_head: Linear,
375}
376
377impl ModelForCausalLM {
378    pub fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
379        let base_model = Model::new(cfg, vb.clone())?;
380        let lm_head = if vb.contains_tensor("lm_head.weight") {
381            linear_no_bias(cfg.hidden_size, cfg.vocab_size, vb.pp("lm_head"))?
382        } else {
383            Linear::from_weights(base_model.embed_tokens.embeddings().clone(), None)
384        };
385        Ok(Self {
386            base_model,
387            lm_head,
388        })
389    }
390
391    pub fn forward(&mut self, input_ids: &Tensor, seqlen_offset: usize) -> Result<Tensor> {
392        let (_b_size, seq_len) = input_ids.dims2()?;
393        self.base_model
394            .forward(input_ids, seqlen_offset, None)?
395            .narrow(1, seq_len - 1, 1)?
396            .apply(&self.lm_head)
397    }
398
399    pub fn clear_kv_cache(&mut self) {
400        self.base_model.clear_kv_cache()
401    }
402}
candle_transformers/models/qwen2.rs

candle_transformers/models/
qwen2.rs
