candle_transformers/models/
quantized_stable_lm.rs

1//! Module for quantized StableLM implementation.
2//!
3//! StableLM is a series of open-source large language models
4//! optimized for performance and stability. This implementation
5//! provides quantization support for efficient model deployment.
6//!
7//! Key characteristics:
8//! - RMSNorm for layer normalization
9//! - Rotary positional embeddings (RoPE)
10//! - Support for 8-bit quantization
11//!
12//! References:
13//! - [StableLM](https://github.com/Stability-AI/StableLM)
14//!
15
16use crate::quantized_nn::{layer_norm, linear, linear_no_bias, Embedding, Linear};
17pub use crate::quantized_var_builder::VarBuilder;
18use candle::{DType, Device, Module, Result, Tensor, D};
19use candle_nn::{Activation, LayerNorm};
20use std::sync::Arc;
21
22pub use crate::models::stable_lm::Config;
23use crate::models::stable_lm::RotaryEmbedding;
24
25#[derive(Debug, Clone)]
26#[allow(clippy::upper_case_acronyms)]
27struct MLP {
28    gate_proj: Linear,
29    up_proj: Linear,
30    down_proj: Linear,
31    act_fn: Activation,
32    span: tracing::Span,
33}
34
35impl MLP {
36    fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
37        let hidden_sz = cfg.hidden_size;
38        let intermediate_sz = cfg.intermediate_size;
39        let gate_proj = linear_no_bias(hidden_sz, intermediate_sz, vb.pp("gate_proj"))?;
40        let up_proj = linear_no_bias(hidden_sz, intermediate_sz, vb.pp("up_proj"))?;
41        let down_proj = linear_no_bias(intermediate_sz, hidden_sz, vb.pp("down_proj"))?;
42        Ok(Self {
43            gate_proj,
44            up_proj,
45            down_proj,
46            act_fn: cfg.hidden_act,
47            span: tracing::span!(tracing::Level::TRACE, "mlp"),
48        })
49    }
50}
51
52impl Module for MLP {
53    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
54        let _enter = self.span.enter();
55        let lhs = xs.apply(&self.gate_proj)?.apply(&self.act_fn)?;
56        let rhs = xs.apply(&self.up_proj)?;
57        (lhs * rhs)?.apply(&self.down_proj)
58    }
59}
60
61#[derive(Debug, Clone)]
62struct Attention {
63    q_proj: Linear,
64    k_proj: Linear,
65    v_proj: Linear,
66    o_proj: Linear,
67    num_heads: usize,
68    num_kv_heads: usize,
69    num_kv_groups: usize,
70    head_dim: usize,
71    hidden_size: usize,
72    rotary_emb: Arc<RotaryEmbedding>,
73    kv_cache: Option<(Tensor, Tensor)>,
74    use_cache: bool,
75    rotary_ndims: usize,
76    span: tracing::Span,
77}
78
79impl Attention {
80    fn new(rotary_emb: Arc<RotaryEmbedding>, cfg: &Config, vb: VarBuilder) -> Result<Self> {
81        let hidden_sz = cfg.hidden_size;
82        let head_dim = cfg.head_dim();
83        let num_heads = cfg.num_attention_heads;
84        let num_kv_heads = cfg.num_key_value_heads;
85        let linear_layer = if cfg.use_qkv_bias {
86            linear
87        } else {
88            linear_no_bias
89        };
90        let q_proj = linear_layer(hidden_sz, num_heads * head_dim, vb.pp("q_proj"))?;
91        let k_proj = linear_layer(hidden_sz, num_kv_heads * head_dim, vb.pp("k_proj"))?;
92        let v_proj = linear_layer(hidden_sz, num_kv_heads * head_dim, vb.pp("v_proj"))?;
93        let o_proj = linear_no_bias(num_heads * head_dim, hidden_sz, vb.pp("o_proj"))?;
94        Ok(Self {
95            q_proj,
96            k_proj,
97            v_proj,
98            o_proj,
99            num_heads,
100            num_kv_heads,
101            num_kv_groups: cfg.num_kv_groups(),
102            head_dim,
103            hidden_size: hidden_sz,
104            rotary_emb,
105            kv_cache: None,
106            use_cache: cfg.use_cache,
107            rotary_ndims: cfg.rotary_ndims(),
108            span: tracing::span!(tracing::Level::TRACE, "attn"),
109        })
110    }
111
112    fn forward(
113        &mut self,
114        xs: &Tensor,
115        attention_mask: Option<&Tensor>,
116        seqlen_offset: usize,
117    ) -> Result<Tensor> {
118        let _enter = self.span.enter();
119        let (b_sz, q_len, _) = xs.dims3()?;
120
121        let query_states = self.q_proj.forward(xs)?;
122        let key_states = self.k_proj.forward(xs)?;
123        let value_states = self.v_proj.forward(xs)?;
124
125        let query_states = query_states
126            .reshape((b_sz, q_len, self.num_heads, self.head_dim))?
127            .transpose(1, 2)?;
128        let key_states = key_states
129            .reshape((b_sz, q_len, self.num_kv_heads, self.head_dim))?
130            .transpose(1, 2)?;
131        let value_states = value_states
132            .reshape((b_sz, q_len, self.num_kv_heads, self.head_dim))?
133            .transpose(1, 2)?;
134
135        let (rot_ndims, pass_ndims) = (self.rotary_ndims, self.head_dim - self.rotary_ndims);
136        let query_rot = query_states.narrow(D::Minus1, 0, rot_ndims)?;
137        let query_pass = query_states.narrow(D::Minus1, rot_ndims, pass_ndims)?;
138        let key_rot = key_states.narrow(D::Minus1, 0, rot_ndims)?;
139        let key_pass = key_states.narrow(D::Minus1, rot_ndims, pass_ndims)?;
140        let (query_rot, key_rot) =
141            self.rotary_emb
142                .apply_rotary_emb_qkv(&query_rot, &key_rot, seqlen_offset)?;
143        let query_states = Tensor::cat(&[query_rot, query_pass], D::Minus1)?.contiguous()?;
144        let key_states = Tensor::cat(&[key_rot, key_pass], D::Minus1)?.contiguous()?;
145
146        let (key_states, value_states) = match &self.kv_cache {
147            None => (key_states, value_states),
148            Some((prev_k, prev_v)) => {
149                let key_states = Tensor::cat(&[prev_k, &key_states], 2)?;
150                let value_states = Tensor::cat(&[prev_v, &value_states], 2)?;
151                (key_states, value_states)
152            }
153        };
154        if self.use_cache {
155            self.kv_cache = Some((key_states.clone(), value_states.clone()));
156        }
157
158        let key_states = crate::utils::repeat_kv(key_states, self.num_kv_groups)?.contiguous()?;
159        let value_states =
160            crate::utils::repeat_kv(value_states, self.num_kv_groups)?.contiguous()?;
161
162        let attn_output = {
163            let scale = 1f64 / f64::sqrt(self.head_dim as f64);
164            let attn_weights = (query_states.matmul(&key_states.transpose(2, 3)?)? * scale)?;
165
166            let attn_weights = match attention_mask {
167                None => attn_weights,
168                Some(mask) => attn_weights.broadcast_add(mask)?,
169            };
170            let attn_weights = candle_nn::ops::softmax_last_dim(&attn_weights)?;
171            attn_weights.matmul(&value_states)?
172        };
173        attn_output
174            .transpose(1, 2)?
175            .reshape((b_sz, q_len, self.hidden_size))?
176            .apply(&self.o_proj)
177    }
178}
179
180#[derive(Debug, Clone)]
181struct DecoderLayer {
182    self_attn: Attention,
183    mlp: MLP,
184    input_layernorm: LayerNorm,
185    post_attention_layernorm: LayerNorm,
186    span: tracing::Span,
187}
188
189impl DecoderLayer {
190    fn new(rotary_emb: Arc<RotaryEmbedding>, cfg: &Config, vb: VarBuilder) -> Result<Self> {
191        let self_attn = Attention::new(rotary_emb, cfg, vb.pp("self_attn"))?;
192        let mlp = MLP::new(cfg, vb.pp("mlp"))?;
193        let input_layernorm = layer_norm(
194            cfg.hidden_size,
195            cfg.layer_norm_eps,
196            vb.pp("input_layernorm"),
197        )?;
198        let post_attention_layernorm = layer_norm(
199            cfg.hidden_size,
200            cfg.layer_norm_eps,
201            vb.pp("post_attention_layernorm"),
202        )?;
203        Ok(Self {
204            self_attn,
205            mlp,
206            input_layernorm,
207            post_attention_layernorm,
208            span: tracing::span!(tracing::Level::TRACE, "layer"),
209        })
210    }
211
212    fn forward(
213        &mut self,
214        xs: &Tensor,
215        attention_mask: Option<&Tensor>,
216        seqlen_offset: usize,
217    ) -> Result<Tensor> {
218        let _enter = self.span.enter();
219        let residual = xs;
220        let xs = self.input_layernorm.forward(xs)?;
221        let xs = self.self_attn.forward(&xs, attention_mask, seqlen_offset)?;
222        let xs = (xs + residual)?;
223        let residual = &xs;
224        let xs = xs.apply(&self.post_attention_layernorm)?.apply(&self.mlp)?;
225        residual + xs
226    }
227}
228
229#[derive(Debug, Clone)]
230pub struct Model {
231    embed_tokens: Embedding,
232    layers: Vec<DecoderLayer>,
233    norm: LayerNorm,
234    lm_head: Linear,
235    device: Device,
236    span: tracing::Span,
237}
238
239impl Model {
240    pub fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
241        let vb_m = vb.pp("model");
242        let embed_tokens =
243            Embedding::new(cfg.vocab_size, cfg.hidden_size, vb_m.pp("embed_tokens"))?;
244        let rotary_emb = Arc::new(RotaryEmbedding::new(DType::F32, cfg, vb_m.device())?);
245        let mut layers = Vec::with_capacity(cfg.num_hidden_layers);
246        let vb_l = vb_m.pp("layers");
247        for layer_idx in 0..cfg.num_hidden_layers {
248            let layer = DecoderLayer::new(rotary_emb.clone(), cfg, vb_l.pp(layer_idx))?;
249            layers.push(layer)
250        }
251        let norm = layer_norm(cfg.hidden_size, cfg.layer_norm_eps, vb_m.pp("norm"))?;
252        let lm_head = linear_no_bias(cfg.hidden_size, cfg.vocab_size, vb.pp("lm_head"))?;
253        Ok(Self {
254            embed_tokens,
255            layers,
256            norm,
257            lm_head,
258            device: vb.device().clone(),
259            span: tracing::span!(tracing::Level::TRACE, "model"),
260        })
261    }
262
263    fn prepare_decoder_attention_mask(
264        &self,
265        b_size: usize,
266        tgt_len: usize,
267        seqlen_offset: usize,
268    ) -> Result<Tensor> {
269        // Sliding window mask?
270        let mask: Vec<_> = (0..tgt_len)
271            .flat_map(|i| (0..tgt_len).map(move |j| if i < j { f32::NEG_INFINITY } else { 0. }))
272            .collect();
273        let mask = Tensor::from_slice(&mask, (tgt_len, tgt_len), &self.device)?;
274        let mask = if seqlen_offset > 0 {
275            let mask0 = Tensor::zeros((tgt_len, seqlen_offset), DType::F32, &self.device)?;
276            Tensor::cat(&[&mask0, &mask], D::Minus1)?
277        } else {
278            mask
279        };
280        mask.expand((b_size, 1, tgt_len, tgt_len + seqlen_offset))?
281            .to_dtype(DType::F32)
282    }
283
284    pub fn forward(&mut self, input_ids: &Tensor, seqlen_offset: usize) -> Result<Tensor> {
285        let _enter = self.span.enter();
286        let (b_size, seq_len) = input_ids.dims2()?;
287        let attention_mask = if seq_len <= 1 {
288            None
289        } else {
290            let mask = self.prepare_decoder_attention_mask(b_size, seq_len, seqlen_offset)?;
291            Some(mask)
292        };
293        let mut xs = self.embed_tokens.forward(input_ids)?;
294        for layer in self.layers.iter_mut() {
295            xs = layer.forward(&xs, attention_mask.as_ref(), seqlen_offset)?
296        }
297        xs.narrow(1, seq_len - 1, 1)?
298            .apply(&self.norm)?
299            .apply(&self.lm_head)
300    }
301}
candle_transformers/models/quantized_stable_lm.rs

candle_transformers/models/
quantized_stable_lm.rs
