wgml 0.2.0

Cross-platform GPU LLM inference.
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
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382

// Gpt-2 transformer, ported from ggml/examples/gpt-2/main-backend.cpp

use crate::gguf::Gguf;
use crate::ops::{
    BatchedMultiqueryAttention, BatchedMultiqueryAttentionParams, LayerNorm, UnaryOp,
};
use nalgebra::{DMatrix, DVector, Dyn, OMatrix, OVector, Vector4};

pub struct Transformer;

type VocabSize = Dyn;
type SeqLen = Dyn;
type NumHeads = Dyn;
type Dim = Dyn;
type Attn = Dyn; // 2304
type HiddenDim = Dyn; // 3072?

pub struct Gpt2Params {
    /// Size of the vocabulary.
    pub n_vocab: usize,
    /// Max sequence length.
    pub n_seq: usize,
    /// Token embedding length.
    pub n_embd: usize,
    /// Number of heads.
    pub n_head: usize,
    /// Number of layers.
    pub n_layer: usize,
    // Feed-forward length.
    pub ff_len: usize,
    pub attn_b: usize,
    pub ftype: usize,
}

impl Gpt2Params {
    pub fn from_gguf(gguf: &Gguf) -> Self {
        Self {
            n_vocab: gguf.metadata["tokenizer.ggml.tokens"].unwrap_array_len(),
            n_seq: gguf.metadata["gpt2.context_length"].unwrap_u32() as usize,
            n_embd: gguf.metadata["gpt2.embedding_length"].unwrap_u32() as usize,
            n_head: gguf.metadata["gpt2.attention.head_count"].unwrap_u32() as usize,
            n_layer: gguf.metadata["gpt2.block_count"].unwrap_u32() as usize,
            ftype: gguf.metadata["general.file_type"].unwrap_u32() as usize,
            ff_len: gguf.metadata["gpt2.feed_forward_length"].unwrap_u32() as usize,
            attn_b: gguf.tensors["blk.0.attn_qkv.bias"].dimensions()[0] as usize,
        }
    }
}

impl Default for Gpt2Params {
    fn default() -> Self {
        // Default params for GPT-2 117M
        Self {
            n_vocab: 50257,
            n_seq: 1024,
            n_embd: 768,
            n_head: 12,
            n_layer: 12,
            attn_b: 2304,
            ftype: 1,
            ff_len: 3072,
        }
    }
}

pub struct Gpt2Layer {
    // Normalization.
    pub(crate) ln_1_g: OVector<f32, Dim>,
    pub(crate) ln_1_b: OVector<f32, Dim>,
    pub(crate) ln_2_g: OVector<f32, Dim>,
    pub(crate) ln_2_b: OVector<f32, Dim>,

    // attention
    pub(crate) c_attn_attn_w: OMatrix<f32, Attn, Dim>,
    pub(crate) c_attn_attn_b: OVector<f32, Attn>,
    pub(crate) c_attn_proj_w: OMatrix<f32, Dim, Attn>,
    pub(crate) c_attn_proj_b: OVector<f32, Dim>,

    // KV cache
    pub(crate) key_cache: OMatrix<f32, Dim, SeqLen>,
    pub(crate) value_cache: OMatrix<f32, Dim, SeqLen>,

    // mlp
    pub(crate) c_mlp_fc_w: OMatrix<f32, HiddenDim, Dim>,
    pub(crate) c_mlp_fc_b: OVector<f32, HiddenDim>,
    pub(crate) c_mlp_proj_w: OMatrix<f32, Dim, HiddenDim>,
    pub(crate) c_mlp_proj_b: OVector<f32, Dim>,
}

pub struct Gpt2Model {
    // Normalization
    pub(crate) ln_f_g: OVector<f32, Dim>,
    pub(crate) ln_f_b: OVector<f32, Dim>,

    pub(crate) wte: OMatrix<f32, Dim, VocabSize>, // token embedding
    pub(crate) wpe: OMatrix<f32, Dim, SeqLen>,    // position embedding
    pub(crate) lm_head: OMatrix<f32, VocabSize, Dim>, // language model head

    pub(crate) layers: Vec<Gpt2Layer>,

    // scratch memory
    memory_q: OVector<f32, Dim>,
    memory_att: OMatrix<f32, SeqLen, NumHeads>,
    layer_input: DVector<f32>,
    curr_768: DVector<f32>,
    curr_768_b: DVector<f32>,
    curr_2304: DVector<f32>,
    curr_3072: DVector<f32>,
    curr_vocab: DVector<f32>,
}

impl Gpt2Model {
    pub fn from_gguf(gguf: &Gguf) -> (Self, Gpt2Params) {
        let params = Gpt2Params::from_gguf(gguf);
        let mut layers = vec![];

        for i_layer in 0..params.n_layer {
            let ln_1_g = format!("blk.{}.attn_norm.weight", i_layer);
            let ln_1_b = format!("blk.{}.attn_norm.bias", i_layer);
            let ln_2_g = format!("blk.{}.ffn_norm.weight", i_layer);
            let ln_2_b = format!("blk.{}.ffn_norm.bias", i_layer);
            let c_attn_attn_w = format!("blk.{}.attn_qkv.weight", i_layer);
            let c_attn_attn_b = format!("blk.{}.attn_qkv.bias", i_layer);
            let c_attn_proj_w = format!("blk.{}.attn_output.weight", i_layer);
            let c_attn_proj_b = format!("blk.{}.attn_output.bias", i_layer);

            let c_mlp_fc_w = format!("blk.{}.ffn_up.weight", i_layer);
            let c_mlp_fc_b = format!("blk.{}.ffn_up.bias", i_layer);
            let c_mlp_proj_w = format!("blk.{}.ffn_down.weight", i_layer);
            let c_mlp_proj_b = format!("blk.{}.ffn_down.bias", i_layer);

            let ln_1_g = gguf.tensors[&ln_1_g].data().as_f32().unwrap();
            let ln_1_b = gguf.tensors[&ln_1_b].data().as_f32().unwrap();
            let ln_2_g = gguf.tensors[&ln_2_g].data().as_f32().unwrap();
            let ln_2_b = gguf.tensors[&ln_2_b].data().as_f32().unwrap();
            let c_attn_attn_w = &gguf.tensors[&c_attn_attn_w].data().dequantize().unwrap();
            let c_attn_attn_b = gguf.tensors[&c_attn_attn_b].data().as_f32().unwrap();
            let c_attn_proj_w = &gguf.tensors[&c_attn_proj_w].data().dequantize().unwrap();
            let c_attn_proj_b = gguf.tensors[&c_attn_proj_b].data().as_f32().unwrap();
            let c_mlp_fc_w = &gguf.tensors[&c_mlp_fc_w].data().dequantize().unwrap();
            let c_mlp_fc_b = gguf.tensors[&c_mlp_fc_b].data().as_f32().unwrap();
            let c_mlp_proj_w = &gguf.tensors[&c_mlp_proj_w].data().dequantize().unwrap();
            let c_mlp_proj_b = gguf.tensors[&c_mlp_proj_b].data().as_f32().unwrap();

            let ln_1_g = DVector::from_row_slice(ln_1_g);
            let ln_1_b = DVector::from_row_slice(ln_1_b);
            let ln_2_g = DVector::from_row_slice(ln_2_g);
            let ln_2_b = DVector::from_row_slice(ln_2_b);

            let c_attn_attn_w =
                DMatrix::from_row_slice(params.attn_b, params.n_embd, c_attn_attn_w);
            let c_attn_attn_b = DVector::from_row_slice(c_attn_attn_b);
            let c_attn_proj_w =
                DMatrix::from_row_slice(params.n_embd, params.n_embd, c_attn_proj_w);
            let c_attn_proj_b = DVector::from_row_slice(c_attn_proj_b);
            let c_mlp_fc_w = DMatrix::from_row_slice(params.ff_len, params.n_embd, c_mlp_fc_w);
            let c_mlp_fc_b = DVector::from_row_slice(c_mlp_fc_b);
            let c_mlp_proj_w = DMatrix::from_row_slice(params.n_embd, params.ff_len, c_mlp_proj_w);
            let c_mlp_proj_b = DVector::from_row_slice(c_mlp_proj_b);

            let layer = Gpt2Layer {
                ln_1_g,
                ln_1_b,
                ln_2_g,
                ln_2_b,
                c_attn_attn_w,
                c_attn_attn_b,
                c_attn_proj_w,
                c_attn_proj_b,
                c_mlp_fc_w,
                c_mlp_fc_b,
                c_mlp_proj_w,
                c_mlp_proj_b,
                key_cache: DMatrix::zeros(params.n_embd, params.n_seq),
                value_cache: DMatrix::zeros(params.n_embd, params.n_seq),
            };
            layers.push(layer);
        }

        let ln_f_g = gguf.tensors["output_norm.weight"].data().as_f32().unwrap();
        let ln_f_b = gguf.tensors["output_norm.bias"].data().as_f32().unwrap();
        let wte = gguf.tensors["token_embd.weight"]
            .data()
            .dequantize()
            .unwrap();
        let wpe = &gguf.tensors["position_embd.weight"]
            .data()
            .dequantize()
            .unwrap();

        let ln_f_g = DVector::from_row_slice(ln_f_g);
        let ln_f_b = DVector::from_row_slice(ln_f_b);
        let wte = DMatrix::from_column_slice(params.n_embd, params.n_vocab, &wte);
        let wpe = DMatrix::from_column_slice(params.n_embd, params.n_seq, wpe);
        // NOTE: GPT2 shares the lm_head tensor with wte.
        let lm_head = wte.transpose();

        let model = Self {
            ln_f_b,
            ln_f_g,
            wte,
            wpe,
            layers,
            lm_head,
            memory_q: DVector::zeros(params.n_embd),
            memory_att: DMatrix::zeros(params.n_seq, params.n_head),
            layer_input: DVector::zeros(params.n_embd),
            curr_768: DVector::zeros(params.n_embd),
            curr_768_b: DVector::zeros(params.n_embd),
            curr_2304: DVector::zeros(params.attn_b),
            curr_3072: DVector::zeros(params.ff_len),
            curr_vocab: DVector::zeros(params.n_vocab),
        };

        (model, params)
    }

    pub fn logits_mut(&mut self) -> &mut DVector<f32> {
        &mut self.curr_vocab
    }
}

impl Transformer {
    pub fn forward(params: &Gpt2Params, model: &mut Gpt2Model, embd: usize, pos: usize) {
        // Positional encoding.
        model.layer_input.copy_from(&model.wte.column(embd));
        model.layer_input += &model.wpe.column(pos);

        for layer in model.layers.iter_mut() {
            // Layer norm.
            {
                // NOTE: in this implementation, we always have N = 1
                // [ 768, N]
                LayerNorm::run_cpu(&mut model.curr_768, &model.layer_input);

                // cur = ln_1_g*cur + ln_1_b
                // [ 768, N]
                model.curr_768.component_mul_assign(&layer.ln_1_g);
                model.curr_768 += &layer.ln_1_b;
            }

            // attn
            // [2304, 768] - model.layers[il].c_attn_attn_w
            // [2304,   1] - model.layers[il].c_attn_attn_b
            // [ 768,   N] - cur (in)
            // [2304,   N] - cur (out)
            //
            // cur = attn_w*cur + attn_b
            // [2304, N]
            {
                model
                    .curr_2304
                    .gemv(1.0, &layer.c_attn_attn_w, &model.curr_768, 0.0);
                model.curr_2304 += &layer.c_attn_attn_b;
            }

            // self-attention
            // TODO: refactor this so that both llama2 and gpt2 share the attn code with KV cache.
            // TODO: implement flash attention.
            {
                // [2304, 1]
                let mut k_cache = layer.key_cache.column_mut(pos);
                let mut v_cache = layer.value_cache.column_mut(pos);

                model
                    .memory_q
                    .copy_from(&model.curr_2304.rows(0, params.n_embd));
                k_cache.copy_from(&model.curr_2304.rows(params.n_embd, params.n_embd));
                v_cache.copy_from(&model.curr_2304.rows(2 * params.n_embd, params.n_embd));

                let head_size = params.n_embd / params.n_head;
                let attn_params = BatchedMultiqueryAttentionParams {
                    seq_len: params.n_seq as u32,
                    kv_dim: params.n_embd as u32,
                    kv_mul: 1,
                    n_heads: params.n_head as u32,
                    head_size: head_size as u32,
                    pos: pos as u32,
                };

                BatchedMultiqueryAttention::run_cpu(
                    &attn_params,
                    &model.memory_q,
                    &layer.key_cache,
                    &layer.value_cache,
                    &mut model.memory_att,
                    &mut model.curr_768,
                );
            }

            // projection
            // [ 768, 768] - model.layers[il].c_attn_proj_w
            // [ 768,   1] - model.layers[il].c_attn_proj_b
            // [ 768,   N] - cur (in)
            // [ 768,   N] - cur (out)
            //
            // cur = proj_w*cur + proj_b
            // [768, N]
            {
                model
                    .curr_768_b
                    .gemv(1.0, &layer.c_attn_proj_w, &model.curr_768, 0.0);
                model.curr_768_b += &layer.c_attn_proj_b;
            }

            // add the input
            model.curr_768_b += &model.layer_input;

            // prep input for next layer
            model.layer_input.copy_from(&model.curr_768_b);

            // feed-forward network
            {
                // norm
                {
                    LayerNorm::run_cpu(&mut model.curr_768, &model.curr_768_b);

                    // cur = ln_2_g*cur + ln_2_b
                    // [ 768, N]
                    model.curr_768.component_mul_assign(&layer.ln_2_g);
                    model.curr_768 += &layer.ln_2_b;
                }

                // fully connected
                // [3072, 768] - model.layers[il].c_mlp_fc_w
                // [3072,   1] - model.layers[il].c_mlp_fc_b
                // [ 768,   N] - cur (in)
                // [3072,   N] - cur (out)
                //
                // cur = fc_w*cur + fc_b
                // [3072, N]
                model
                    .curr_3072
                    .gemv(1.0, &layer.c_mlp_fc_w, &model.curr_768, 0.0);
                model.curr_3072 += &layer.c_mlp_fc_b;

                // GELU activation
                // [3072, N]
                model
                    .curr_3072
                    .apply(|x| *x = UnaryOp::Gelu.eval(*x, Vector4::zeros()));

                // projection
                // [ 768, 3072] - model.layers[il].c_mlp_proj_w
                // [ 768,    1] - model.layers[il].c_mlp_proj_b
                // [3072,    N] - cur (in)
                // [ 768,    N] - cur (out)
                //
                // cur = proj_w*cur + proj_b
                // [768, N]
                model
                    .curr_768
                    .gemv(1.0, &layer.c_mlp_proj_w, &model.curr_3072, 0.0);
                model.curr_768 += &layer.c_mlp_proj_b;
            }

            // finalize input for next layer
            model.layer_input += &model.curr_768;
        }

        // norm
        {
            // [ 768, N]
            LayerNorm::run_cpu(&mut model.curr_768, &model.layer_input);

            // inpL = ln_f_g*inpL + ln_f_b
            // [ 768, N]
            model.curr_768.component_mul_assign(&model.ln_f_g);
            model.curr_768 += &model.ln_f_b;
        }

        // inpL = WTE * inpL
        // [ 768, 50257] - model.lm_head
        // [ 768, N]     - inpL
        model
            .curr_vocab
            .gemv(1.0, &model.lm_head, &model.curr_768, 0.0);

        // // logits -> probs
        // SoftMax::run_cpu(&mut curr2); // NOTE: done by the sampler
    }
}