candle_transformers/models/
quantized_mistral.rs

1//! Mistral model implementation with quantization support.
2//!
3//! Mistral is a large language model optimized for efficiency.
4//! This implementation provides quantization for reduced memory and compute.
5//!
6//! Key characteristics:
7//! - Sliding window attention mechanism
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//! - [Mistral Paper](https://arxiv.org/abs/2310.06825)
15//! - [Model Card](https://huggingface.co/mistralai/Mistral-7B-v0.1)
16//!
17
18use crate::quantized_nn::{linear_no_bias, Embedding, Linear, RmsNorm};
19pub use crate::quantized_var_builder::VarBuilder;
20use candle::{DType, Device, Module, Result, Tensor, D};
21use candle_nn::Activation;
22use std::sync::Arc;
23
24pub use crate::models::mistral::Config;
25
26#[derive(Debug, Clone)]
27struct RotaryEmbedding {
28    sin: Tensor,
29    cos: Tensor,
30}
31
32impl RotaryEmbedding {
33    fn new(cfg: &Config, dev: &Device) -> Result<Self> {
34        let rope_theta = cfg.rope_theta as f32;
35        let dim = cfg.hidden_size / cfg.num_attention_heads;
36        let max_seq_len = cfg.max_position_embeddings;
37        let inv_freq: Vec<_> = (0..dim)
38            .step_by(2)
39            .map(|i| 1f32 / rope_theta.powf(i as f32 / dim as f32))
40            .collect();
41        let inv_freq_len = inv_freq.len();
42        let inv_freq = Tensor::from_vec(inv_freq, (1, inv_freq_len), dev)?;
43        let t = Tensor::arange(0u32, max_seq_len as u32, dev)?
44            .to_dtype(DType::F32)?
45            .reshape((max_seq_len, 1))?;
46        let freqs = t.matmul(&inv_freq)?;
47        Ok(Self {
48            sin: freqs.sin()?,
49            cos: freqs.cos()?,
50        })
51    }
52
53    fn apply_rotary_emb_qkv(
54        &self,
55        q: &Tensor,
56        k: &Tensor,
57        seqlen_offset: usize,
58    ) -> Result<(Tensor, Tensor)> {
59        let (_b_sz, _h, seq_len, _n_embd) = q.dims4()?;
60        let cos = self.cos.narrow(0, seqlen_offset, seq_len)?;
61        let sin = self.sin.narrow(0, seqlen_offset, seq_len)?;
62        let q_embed = candle_nn::rotary_emb::rope(q, &cos, &sin)?;
63        let k_embed = candle_nn::rotary_emb::rope(k, &cos, &sin)?;
64        Ok((q_embed, k_embed))
65    }
66}
67
68#[derive(Debug, Clone)]
69#[allow(clippy::upper_case_acronyms)]
70struct MLP {
71    gate_proj: Linear,
72    up_proj: Linear,
73    down_proj: Linear,
74    act_fn: Activation,
75}
76
77impl MLP {
78    fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
79        let hidden_sz = cfg.hidden_size;
80        let intermediate_sz = cfg.intermediate_size;
81        let gate_proj = linear_no_bias(hidden_sz, intermediate_sz, vb.pp("gate_proj"))?;
82        let up_proj = linear_no_bias(hidden_sz, intermediate_sz, vb.pp("up_proj"))?;
83        let down_proj = linear_no_bias(intermediate_sz, hidden_sz, vb.pp("down_proj"))?;
84        Ok(Self {
85            gate_proj,
86            up_proj,
87            down_proj,
88            act_fn: cfg.hidden_act,
89        })
90    }
91}
92
93impl Module for MLP {
94    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
95        let lhs = xs.apply(&self.gate_proj)?.apply(&self.act_fn)?;
96        let rhs = xs.apply(&self.up_proj)?;
97        (lhs * rhs)?.apply(&self.down_proj)
98    }
99}
100
101#[derive(Debug, Clone)]
102struct Attention {
103    q_proj: Linear,
104    k_proj: Linear,
105    v_proj: Linear,
106    o_proj: Linear,
107    num_heads: usize,
108    num_kv_heads: usize,
109    num_kv_groups: usize,
110    head_dim: usize,
111    hidden_size: usize,
112    rotary_emb: Arc<RotaryEmbedding>,
113    kv_cache: Option<(Tensor, Tensor)>,
114}
115
116impl Attention {
117    fn new(rotary_emb: Arc<RotaryEmbedding>, cfg: &Config, vb: VarBuilder) -> Result<Self> {
118        let hidden_sz = cfg.hidden_size;
119        let num_heads = cfg.num_attention_heads;
120        let num_kv_heads = cfg.num_key_value_heads;
121        let num_kv_groups = num_heads / num_kv_heads;
122        let head_dim = hidden_sz / num_heads;
123        let q_proj = linear_no_bias(hidden_sz, num_heads * head_dim, vb.pp("q_proj"))?;
124        let k_proj = linear_no_bias(hidden_sz, num_kv_heads * head_dim, vb.pp("k_proj"))?;
125        let v_proj = linear_no_bias(hidden_sz, num_kv_heads * head_dim, vb.pp("v_proj"))?;
126        let o_proj = linear_no_bias(num_heads * head_dim, hidden_sz, vb.pp("o_proj"))?;
127        Ok(Self {
128            q_proj,
129            k_proj,
130            v_proj,
131            o_proj,
132            num_heads,
133            num_kv_heads,
134            num_kv_groups,
135            head_dim,
136            hidden_size: hidden_sz,
137            rotary_emb,
138            kv_cache: None,
139        })
140    }
141
142    fn forward(
143        &mut self,
144        xs: &Tensor,
145        attention_mask: Option<&Tensor>,
146        seqlen_offset: usize,
147    ) -> Result<Tensor> {
148        let (b_sz, q_len, _) = xs.dims3()?;
149
150        let query_states = self.q_proj.forward(xs)?;
151        let key_states = self.k_proj.forward(xs)?;
152        let value_states = self.v_proj.forward(xs)?;
153
154        let query_states = query_states
155            .reshape((b_sz, q_len, self.num_heads, self.head_dim))?
156            .transpose(1, 2)?
157            .contiguous()?;
158        let key_states = key_states
159            .reshape((b_sz, q_len, self.num_kv_heads, self.head_dim))?
160            .transpose(1, 2)?
161            .contiguous()?;
162        let value_states = value_states
163            .reshape((b_sz, q_len, self.num_kv_heads, self.head_dim))?
164            .transpose(1, 2)?;
165
166        let (query_states, key_states) =
167            self.rotary_emb
168                .apply_rotary_emb_qkv(&query_states, &key_states, seqlen_offset)?;
169
170        let (key_states, value_states) = match &self.kv_cache {
171            None => (key_states, value_states),
172            Some((prev_k, prev_v)) => {
173                let key_states = Tensor::cat(&[prev_k, &key_states], 2)?;
174                let value_states = Tensor::cat(&[prev_v, &value_states], 2)?;
175                (key_states, value_states)
176            }
177        };
178        self.kv_cache = Some((key_states.clone(), value_states.clone()));
179
180        let key_states = crate::utils::repeat_kv(key_states, self.num_kv_groups)?;
181        let value_states = crate::utils::repeat_kv(value_states, self.num_kv_groups)?;
182
183        let attn_output = {
184            let scale = 1f64 / f64::sqrt(self.head_dim as f64);
185            let attn_weights = (query_states.matmul(&key_states.transpose(2, 3)?)? * scale)?;
186
187            let attn_weights = match attention_mask {
188                None => attn_weights,
189                Some(mask) => attn_weights.broadcast_add(mask)?,
190            };
191            let attn_weights = candle_nn::ops::softmax_last_dim(&attn_weights)?;
192            attn_weights.matmul(&value_states)?
193        };
194        attn_output
195            .transpose(1, 2)?
196            .reshape((b_sz, q_len, self.hidden_size))?
197            .apply(&self.o_proj)
198    }
199
200    fn clear_kv_cache(&mut self) {
201        self.kv_cache = None
202    }
203}
204
205#[derive(Debug, Clone)]
206struct DecoderLayer {
207    self_attn: Attention,
208    mlp: MLP,
209    input_layernorm: RmsNorm,
210    post_attention_layernorm: RmsNorm,
211}
212
213impl DecoderLayer {
214    fn new(rotary_emb: Arc<RotaryEmbedding>, cfg: &Config, vb: VarBuilder) -> Result<Self> {
215        let self_attn = Attention::new(rotary_emb, cfg, vb.pp("self_attn"))?;
216        let mlp = MLP::new(cfg, vb.pp("mlp"))?;
217        let input_layernorm =
218            RmsNorm::new(cfg.hidden_size, cfg.rms_norm_eps, vb.pp("input_layernorm"))?;
219        let post_attention_layernorm = RmsNorm::new(
220            cfg.hidden_size,
221            cfg.rms_norm_eps,
222            vb.pp("post_attention_layernorm"),
223        )?;
224        Ok(Self {
225            self_attn,
226            mlp,
227            input_layernorm,
228            post_attention_layernorm,
229        })
230    }
231
232    fn forward(
233        &mut self,
234        xs: &Tensor,
235        attention_mask: Option<&Tensor>,
236        seqlen_offset: usize,
237    ) -> Result<Tensor> {
238        let residual = xs;
239        let xs = self.input_layernorm.forward(xs)?;
240        let xs = self.self_attn.forward(&xs, attention_mask, seqlen_offset)?;
241        let xs = (xs + residual)?;
242        let residual = &xs;
243        let xs = xs.apply(&self.post_attention_layernorm)?.apply(&self.mlp)?;
244        residual + xs
245    }
246
247    fn clear_kv_cache(&mut self) {
248        self.self_attn.clear_kv_cache()
249    }
250}
251
252#[derive(Debug, Clone)]
253pub struct Model {
254    embed_tokens: Embedding,
255    layers: Vec<DecoderLayer>,
256    norm: RmsNorm,
257    lm_head: Linear,
258    sliding_window: Option<usize>,
259    device: Device,
260}
261
262impl Model {
263    pub fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
264        let vb_m = vb.pp("model");
265        let embed_tokens =
266            Embedding::new(cfg.vocab_size, cfg.hidden_size, vb_m.pp("embed_tokens"))?;
267        let rotary_emb = Arc::new(RotaryEmbedding::new(cfg, vb_m.device())?);
268        let mut layers = Vec::with_capacity(cfg.num_hidden_layers);
269        let vb_l = vb_m.pp("layers");
270        for layer_idx in 0..cfg.num_hidden_layers {
271            let layer = DecoderLayer::new(rotary_emb.clone(), cfg, vb_l.pp(layer_idx))?;
272            layers.push(layer)
273        }
274        let norm = RmsNorm::new(cfg.hidden_size, cfg.rms_norm_eps, vb_m.pp("norm"))?;
275        let lm_head = linear_no_bias(cfg.hidden_size, cfg.vocab_size, vb.pp("lm_head"))?;
276        Ok(Self {
277            embed_tokens,
278            layers,
279            norm,
280            lm_head,
281            sliding_window: cfg.sliding_window,
282            device: vb.device().clone(),
283        })
284    }
285
286    fn prepare_decoder_attention_mask(
287        &self,
288        tgt_len: usize,
289        seqlen_offset: usize,
290    ) -> Result<Tensor> {
291        let sliding_window = self.sliding_window.unwrap_or(tgt_len + 1);
292        let mask: Vec<_> = (0..tgt_len)
293            .flat_map(|i| {
294                (0..tgt_len).map(move |j| {
295                    if i < j || j + sliding_window < i {
296                        f32::NEG_INFINITY
297                    } else {
298                        0.
299                    }
300                })
301            })
302            .collect();
303        let mask = Tensor::from_slice(&mask, (tgt_len, tgt_len), &self.device)?;
304        let mask = if seqlen_offset > 0 {
305            let mask0 = Tensor::zeros((tgt_len, seqlen_offset), DType::F32, &self.device)?;
306            Tensor::cat(&[&mask0, &mask], D::Minus1)?
307        } else {
308            mask
309        };
310        mask.expand((1, 1, tgt_len, tgt_len + seqlen_offset))?
311            .to_dtype(DType::F32)
312    }
313
314    pub fn forward(&mut self, input_ids: &Tensor, seqlen_offset: usize) -> Result<Tensor> {
315        let (_b_size, seq_len) = input_ids.dims2()?;
316        let attention_mask = if seq_len <= 1 {
317            None
318        } else {
319            let mask = self.prepare_decoder_attention_mask(seq_len, seqlen_offset)?;
320            Some(mask)
321        };
322        let mut xs = self.embed_tokens.forward(input_ids)?;
323        for layer in self.layers.iter_mut() {
324            xs = layer.forward(&xs, attention_mask.as_ref(), seqlen_offset)?
325        }
326        xs.narrow(1, seq_len - 1, 1)?
327            .contiguous()?
328            .apply(&self.norm)?
329            .apply(&self.lm_head)
330    }
331
332    pub fn clear_kv_cache(&mut self) {
333        for layer in self.layers.iter_mut() {
334            layer.clear_kv_cache()
335        }
336    }
337}
candle_transformers/models/quantized_mistral.rs

candle_transformers/models/
quantized_mistral.rs
