pllm 0.4.1

Portable LLM
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340

use crate::{
    errors::PllmError,
    tensor::{F32TensorExt, Tensor},
    util::FloatVec,
    Config, Weights,
};
use rayon::prelude::*;
use std::time::Instant;

#[derive(Clone)]
pub struct LayerCache {
    data: Vec<f32>,
    header_size: u32,
    kv_dim: u32,
    kv_mul: u32,
}

impl LayerCache {
    pub fn new(header_size: u32, seq_len: u32, kv_dim: u32, kv_mul: u32) -> Self {
        Self {
            data: vec![0.0; (kv_dim * seq_len) as usize],
            header_size,
            kv_dim,
            kv_mul,
        }
    }

    pub fn get(&self, position: u32, header_idx: u32) -> &[f32] {
        let start =
            (position * self.kv_dim + (header_idx / self.kv_mul) * self.header_size) as usize;
        &self.data[start..(start + self.header_size as usize)]
    }

    pub fn get_mut(&mut self, position: u32) -> &mut [f32] {
        self.data.get_mut_chunk(self.kv_dim, position)
    }
}

#[derive(Clone)]
pub struct Head {
    scores: Vec<f32>,
}

impl Head {
    pub fn new(seq_len: u32) -> Self {
        Self {
            scores: vec![0.0; seq_len as usize],
        }
    }

    pub fn calculate_activation(
        &mut self,
        xb: &mut [f32],
        q: &[f32],
        k: &LayerCache,
        v: &LayerCache,
        pos: u32,
        header_idx: u32,
        header_size: u32,
    ) {
        for t in 0..=pos {
            let keys = k.get(t, header_idx);
            let mut score: f32 = (0..header_size as usize)
                .into_iter()
                .map(|i| q[i] * keys[i])
                .sum();
            score = score / (header_size as f32).sqrt();
            // save the score to the attention buffer
            self.scores[t as usize] = score;
        }

        self.scores[..(pos as usize + 1)].soft_max();

        // weighted sum of the values, store back into xb
        for t in 0..=pos {
            let values = v.get(t, header_idx);
            for i in 0..header_size as usize {
                xb[i] += values[i] * self.scores[t as usize];
            }
        }
    }
}

#[derive(Clone)]
pub struct Layer {
    /// activations (dim,)
    xb: Vec<f32>,
    /// an additional buffer just for convenience (dim,)
    xb2: Vec<f32>,
    /// buffer for hidden dimension in the ffn (hidden_dim,)
    hb: Vec<f32>,
    /// buffer for hidden dimension in the ffn (hidden_dim,)
    hb2: Vec<f32>,
    /// query (dim,)
    pub q: Vec<f32>,
    /// key (dim,)
    k: LayerCache,
    /// value (dim,)
    v: LayerCache,

    heads: Vec<Head>,
    header_size: u32,
    kv_dim: u32,
    norm_rms_eps: f32,
    rope_dim: u32,
}
impl Layer {
    pub fn new(c: &Config) -> Self {
        let xb = vec![0_f32; c.dim as usize];
        let xb2 = xb.clone();

        let hb = vec![0_f32; c.hidden_dim as usize];
        let hb2 = hb.clone();

        let q = xb.clone();

        let k = LayerCache::new(c.header_size(), c.seq_len, c.kv_dim(), c.kv_mul());
        let v = LayerCache::new(c.header_size(), c.seq_len, c.kv_dim(), c.kv_mul());

        let heads = vec![Head::new(c.seq_len); c.n_heads as usize];

        Self {
            xb,
            xb2,
            hb,
            hb2,
            q,
            k,
            v,
            heads,
            header_size: c.header_size(),
            kv_dim: c.kv_dim(),
            norm_rms_eps: c.norm_rms_eps,
            rope_dim: c.rope_dim,
        }
    }
    pub fn forward(
        &mut self,
        x: &mut [f32],
        wo: &Tensor,
        w1: &Tensor,
        w2: &Tensor,
        w3: &Tensor,
        wq: &Tensor,
        wk: &Tensor,
        wv: &Tensor,
        rms_att_weight: &[f32],
        rms_ffn_weight: &[f32],
        pos: u32,
        xb_q: &mut Tensor,
        hb_q: &mut Tensor,
        is_gemma: bool,
    ) -> Result<(), PllmError> {
        let k = self.k.get_mut(pos);
        let v = self.v.get_mut(pos);

        // attention rmsnorm
        self.xb.rms_norm(x, rms_att_weight, self.norm_rms_eps);

        if xb_q.is_none() {
            let xq = self.xb.to_tensor();
            self.q.tensor_mul(&xq, wq);
            k.tensor_mul(&xq, wk);
            v.tensor_mul(&xq, wv);
        } else {
            xb_q.quantize(&self.xb);
            self.q.tensor_mul(xb_q, wq);
            k.tensor_mul(xb_q, wk);
            v.tensor_mul(xb_q, wv);
        }

        // apply RoPE rotation to the q and k vectors for each head
        if is_gemma {
            self.q
                .rope_rotate_neox(pos, self.header_size, self.rope_dim);
            k.rope_rotate_neox(pos, self.header_size, self.rope_dim);
        } else {
            self.q.rope_rotate(k, pos, self.header_size, self.kv_dim)?;
        }

        // multihead attention. iterate over all heads
        self.xb
            .par_chunks_mut(self.header_size as usize)
            .enumerate()
            .zip(self.heads.par_iter_mut())
            .for_each(|((h, xb_chunk), header)| {
                xb_chunk.iter_mut().for_each(|item| *item = 0.0);

                let q = self.q.get_chunk(self.header_size, h as u32);
                header.calculate_activation(
                    xb_chunk,
                    q,
                    &self.k,
                    &self.v,
                    pos,
                    h as u32,
                    self.header_size,
                );
            });

        // final matmul to get the output of the attention
        if xb_q.is_none() {
            self.xb2.tensor_mul(&self.xb.to_tensor(), wo);
        } else {
            xb_q.quantize(&self.xb);
            self.xb2.tensor_mul(xb_q, wo);
        }
        // residual connection back into x
        x.accum(self.xb2.as_slice());

        // ffn rmsnorm
        self.xb.rms_norm(x, rms_ffn_weight, self.norm_rms_eps);

        // Now for FFN in PyTorch we have: self.w2(F.silu(self.w1(x)) * self.w3(x))
        // first calculate self.w1(x) and self.w3(x)
        if xb_q.is_none() {
            self.hb.tensor_mul(&self.xb.to_tensor(), w1);
            self.hb2.tensor_mul(&self.xb.to_tensor(), w3);
        } else {
            xb_q.quantize(&self.xb);
            self.hb.tensor_mul(xb_q, w1);
            self.hb2.tensor_mul(xb_q, w3);
        }

        // F.silu; silu(x)=x*σ(x),where σ(x) is the logistic sigmoid
        // elementwise multiply with w3(x)
        if is_gemma {
            for i in 0..self.hb.len() {
                let item = self.hb[i];
                let tmp = 0.797_884_560_802_865_4 * item * (1.0 + 0.044715 * item * item);

                self.hb[i] = 0.5 * item * (1.0 + tmp.tanh()) * self.hb2[i];
            }
        } else {
            for i in 0..self.hb.len() {
                self.hb[i] = self.hb[i] * (1.0 / (1.0 + (-self.hb[i]).exp())) * self.hb2[i];
            }
        }

        // final matmul to get the output of the ffn
        if hb_q.is_none() {
            self.xb.tensor_mul(&self.hb.to_tensor(), w2);
        } else {
            hb_q.quantize(&self.hb);
            self.xb.tensor_mul(hb_q, w2);
        }

        x.accum(self.xb.as_slice());

        Ok(())
    }
}

pub struct Transformer {
    config: Config,
    layers: Vec<Layer>,
    /// activation at current time stamp (dim,)
    x: Vec<f32>,
    /// output logits
    logits: Vec<f32>,
}

impl Transformer {
    pub fn new(config: Config) -> Self {
        let x = vec![0_f32; config.dim as usize];
        let logits = vec![0_f32; config.vocab_size as usize];
        let layers = vec![Layer::new(&config); config.n_layers as usize];
        let layers = vec![Layer::new(&config); config.n_layers as usize];

        Self {
            config,
            layers,
            x,
            logits,
        }
    }

    pub fn run(&mut self, token: u32, pos: u32, w: &Weights) -> Result<&mut [f32], PllmError> {
        let c = &self.config;

        w.token_embedding_table
            .dequantize(&mut self.x, (c.dim * token) as usize);

        if self.config.is_gemma() {
            self.x.scale((c.dim as f32).sqrt());
        }

        let mut xb_q = w.make_quantize_tensor(c.dim as usize);
        let mut hb_q = w.make_quantize_tensor(c.hidden_dim as usize);

        for (lu, layer) in self.layers.iter_mut().enumerate() {
            let l = lu as u32;
            let before = Instant::now();
            layer.forward(
                &mut self.x,
                &w.wo[lu],
                &w.w1[lu],
                &w.w2[lu],
                &w.w3[lu],
                &w.wq[lu],
                &w.wk[lu],
                &w.wv[lu],
                w.rms_att_weight.get_chunk(c.dim, l),
                w.rms_ffn_weight.get_chunk(c.dim, l),
                pos,
                &mut xb_q,
                &mut hb_q,
                self.config.is_gemma(),
            )?;
            // if l == self.config.n_layers - 1 {
            //     println!("{}", l);
            // }
            // println!("Layer time: l={}, {:.2?}", l, before.elapsed());
        }

        // final rmsnorm
        let x_clone = self.x.clone();
        self.x.rms_norm(
            x_clone.as_slice(),
            w.rms_final_weight.as_ref(),
            self.config.norm_rms_eps,
        );
        // classifier into logits

        let output_weight = if w.output_weight.is_none() {
            &w.token_embedding_table
        } else {
            &w.output_weight
        };

        if xb_q.is_none() {
            self.logits.tensor_mul(&self.x.to_tensor(), output_weight);
        } else {
            xb_q.quantize(&self.x);
            self.logits.tensor_mul(&xb_q, output_weight);
        }

        Ok(&mut self.logits)
    }
}