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

use crate::models::gpt2::cpu::{Gpt2Model, Gpt2Params};
use crate::ops::{
    BatchedMultiqueryAttention, BatchedMultiqueryAttentionParams, LayerNorm, Unary, UnaryOp,
};
use naga_oil::compose::ComposerError;
use wgcore::kernel::KernelInvocationQueue;
use wgcore::tensor::{GpuMatrix, GpuScalar, GpuVector};
use wgcore::Shader;
use wgebra::linalg::{Gemv, OpAssign, OpAssignVariant};
use wgpu::{BufferUsages, Device};

pub struct Gpt2State {
    memory_q: GpuVector<f32>,
    memory_att: GpuMatrix<f32>,
    layer_input: GpuVector<f32>,
    curr_768: GpuVector<f32>,
    curr_768_b: GpuVector<f32>,
    curr_2304: GpuVector<f32>,
    curr_3072: GpuVector<f32>,
    curr_vocab: GpuVector<f32>,
    logits_readback: GpuVector<f32>,
    attn_params: GpuScalar<BatchedMultiqueryAttentionParams>,
}

impl Gpt2State {
    pub fn new(device: &Device, config: &Gpt2Params) -> Self {
        const STORAGE: BufferUsages = BufferUsages::STORAGE;
        const UNIFORM: BufferUsages = BufferUsages::UNIFORM;

        Self {
            memory_q: GpuVector::uninit(device, config.n_embd as u32, STORAGE),
            memory_att: GpuMatrix::uninit(
                device,
                config.n_seq as u32,
                config.n_head as u32,
                STORAGE,
            ),
            layer_input: GpuVector::uninit(device, config.n_embd as u32, STORAGE),
            curr_768: GpuVector::uninit(device, config.n_embd as u32, STORAGE),
            curr_768_b: GpuVector::uninit(device, config.n_embd as u32, STORAGE),
            curr_2304: GpuVector::uninit(device, config.attn_b as u32, STORAGE),
            curr_3072: GpuVector::uninit(device, config.ff_len as u32, STORAGE),
            curr_vocab: GpuVector::uninit(
                device,
                config.n_vocab as u32,
                STORAGE | BufferUsages::COPY_SRC,
            ),
            attn_params: GpuScalar::uninit(device, UNIFORM | BufferUsages::COPY_DST),
            logits_readback: GpuVector::uninit(
                device,
                config.n_vocab as u32,
                BufferUsages::COPY_DST | BufferUsages::MAP_READ,
            ),
        }
    }

    pub fn logits_readback(&self) -> &GpuVector<f32> {
        &self.logits_readback
    }

    pub fn logits(&self) -> &GpuVector<f32> {
        &self.curr_vocab
    }

    pub fn attn_params(&self) -> &GpuScalar<BatchedMultiqueryAttentionParams> {
        &self.attn_params
    }
}

pub struct Gpt2LayerWeights {
    // Normalization.
    ln_1_g: GpuVector<f32>,
    ln_1_b: GpuVector<f32>,
    ln_2_g: GpuVector<f32>,
    ln_2_b: GpuVector<f32>,

    // attention
    c_attn_attn_w: GpuMatrix<f32>,
    c_attn_attn_b: GpuVector<f32>,
    c_attn_proj_w: GpuMatrix<f32>,
    c_attn_proj_b: GpuVector<f32>,

    // KV cache
    key_cache: GpuMatrix<f32>,
    value_cache: GpuMatrix<f32>,

    // mlp
    c_mlp_fc_w: GpuMatrix<f32>,
    c_mlp_fc_b: GpuVector<f32>,
    c_mlp_proj_w: GpuMatrix<f32>,
    c_mlp_proj_b: GpuVector<f32>,
}

pub struct Gpt2Weights {
    // Normalization
    ln_f_g: GpuVector<f32>,
    ln_f_b: GpuVector<f32>,

    wte: GpuMatrix<f32>,     // token embedding
    wpe: GpuMatrix<f32>,     // position embedding
    lm_head: GpuMatrix<f32>, // language model head

    layers: Vec<Gpt2LayerWeights>,
}

impl Gpt2Weights {
    pub fn from_ram(device: &Device, w: &Gpt2Model) -> Self {
        const STORAGE: BufferUsages = BufferUsages::STORAGE;

        Self {
            ln_f_g: GpuVector::init(device, &w.ln_f_g, STORAGE),
            ln_f_b: GpuVector::init(device, &w.ln_f_b, STORAGE),
            wte: GpuMatrix::init(device, &w.wte, STORAGE),
            wpe: GpuMatrix::init(device, &w.wpe, STORAGE),
            lm_head: GpuMatrix::init(device, &w.lm_head, STORAGE),
            layers: w
                .layers
                .iter()
                .map(|l| Gpt2LayerWeights {
                    ln_1_g: GpuVector::init(device, &l.ln_1_g, STORAGE),
                    ln_1_b: GpuVector::init(device, &l.ln_1_b, STORAGE),
                    ln_2_g: GpuVector::init(device, &l.ln_2_g, STORAGE),
                    ln_2_b: GpuVector::init(device, &l.ln_2_b, STORAGE),
                    c_attn_attn_w: GpuMatrix::init(device, &l.c_attn_attn_w, STORAGE),
                    c_attn_attn_b: GpuVector::init(device, &l.c_attn_attn_b, STORAGE),
                    c_attn_proj_w: GpuMatrix::init(device, &l.c_attn_proj_w, STORAGE),
                    c_attn_proj_b: GpuVector::init(device, &l.c_attn_proj_b, STORAGE),
                    key_cache: GpuMatrix::init(device, &l.key_cache, STORAGE),
                    value_cache: GpuMatrix::init(device, &l.value_cache, STORAGE),
                    c_mlp_fc_w: GpuMatrix::init(device, &l.c_mlp_fc_w, STORAGE),
                    c_mlp_fc_b: GpuVector::init(device, &l.c_mlp_fc_b, STORAGE),
                    c_mlp_proj_w: GpuMatrix::init(device, &l.c_mlp_proj_w, STORAGE),
                    c_mlp_proj_b: GpuVector::init(device, &l.c_mlp_proj_b, STORAGE),
                })
                .collect(),
        }
    }
}

pub struct Gpt2 {
    layernorm: LayerNorm,
    gelu: Unary,
    matmul: Gemv,
    attn: BatchedMultiqueryAttention,
    copy_from: OpAssign,

    // TODO: merge the add/mul kernels with the layernorm kernel?
    mul_assign: OpAssign,
    add_assign: OpAssign,
}

impl Gpt2 {
    pub fn new(device: &Device) -> Result<Self, ComposerError> {
        Ok(Self {
            layernorm: LayerNorm::from_device(device)?,
            gelu: Unary::new(device, UnaryOp::Gelu)?,
            matmul: Gemv::from_device(device)?,
            attn: BatchedMultiqueryAttention::from_device(device)?,
            copy_from: OpAssign::new(device, OpAssignVariant::Copy)?,
            mul_assign: OpAssign::new(device, OpAssignVariant::Mul)?,
            add_assign: OpAssign::new(device, OpAssignVariant::Add)?,
        })
    }

    pub fn queue<'a>(
        &'a self,
        queue: &mut KernelInvocationQueue<'a>,
        state: &Gpt2State,
        weights: &Gpt2Weights,
        config: &Gpt2Params,
        embd: u32,
        pos: u32,
    ) {
        // Positional encoding.
        self.copy_from
            .queue(queue, &state.layer_input, weights.wte.column(embd));
        self.add_assign
            .queue(queue, &state.layer_input, weights.wpe.column(pos));

        for layer in &weights.layers {
            // Layer norm.
            {
                self.layernorm
                    .queue(queue, &state.curr_768, &state.layer_input);

                // cur = ln_1_g*cur + ln_1_b
                self.mul_assign.queue(queue, &state.curr_768, &layer.ln_1_g);
                self.add_assign.queue(queue, &state.curr_768, &layer.ln_1_b);
            }

            // attn
            {
                self.matmul.queue(
                    queue,
                    &state.curr_2304,
                    &layer.c_attn_attn_w,
                    &state.curr_768,
                );
                self.add_assign
                    .queue(queue, &state.curr_2304, &layer.c_attn_attn_b);
            }

            // self-attention
            {
                let k_cache = layer.key_cache.column(pos);
                let v_cache = layer.value_cache.column(pos);

                self.copy_from.queue(
                    queue,
                    &state.memory_q,
                    state.curr_2304.rows(0, config.n_embd as u32),
                );
                self.copy_from.queue(
                    queue,
                    k_cache,
                    state
                        .curr_2304
                        .rows(config.n_embd as u32, config.n_embd as u32),
                );
                self.copy_from.queue(
                    queue,
                    v_cache,
                    state
                        .curr_2304
                        .rows(2 * config.n_embd as u32, config.n_embd as u32),
                );

                // attention.
                self.attn.queue(
                    queue,
                    &state.attn_params,
                    &state.memory_q,
                    &layer.key_cache,
                    &layer.value_cache,
                    &state.memory_att,
                    &state.curr_768,
                );
            }

            // projection
            // cur = proj_w*cur + proj_b
            {
                self.matmul.queue(
                    queue,
                    &state.curr_768_b,
                    &layer.c_attn_proj_w,
                    &state.curr_768,
                );
                self.add_assign
                    .queue(queue, &state.curr_768_b, &layer.c_attn_proj_b);
            }

            // add the input
            self.add_assign
                .queue(queue, &state.curr_768_b, &state.layer_input);

            // prep input for next layer
            self.copy_from
                .queue(queue, &state.layer_input, &state.curr_768_b);

            // feed-forward network
            {
                // norm
                {
                    self.layernorm
                        .queue(queue, &state.curr_768, &state.curr_768_b);

                    // cur = ln_2_g*cur + ln_2_b
                    self.mul_assign.queue(queue, &state.curr_768, &layer.ln_2_g);
                    self.add_assign.queue(queue, &state.curr_768, &layer.ln_2_b);
                }

                // fully connected
                self.matmul
                    .queue(queue, &state.curr_3072, &layer.c_mlp_fc_w, &state.curr_768);
                self.add_assign
                    .queue(queue, &state.curr_3072, &layer.c_mlp_fc_b);

                // GELU activation
                self.gelu
                    .queue(queue, &state.curr_3072, &state.curr_3072, None);

                // projection
                self.matmul.queue(
                    queue,
                    &state.curr_768,
                    &layer.c_mlp_proj_w,
                    &state.curr_3072,
                );
                self.add_assign
                    .queue(queue, &state.curr_768, &layer.c_mlp_proj_b);
            }

            // finalize input for next layer
            self.add_assign
                .queue(queue, &state.layer_input, &state.curr_768);
        }

        // norm
        {
            self.layernorm
                .queue(queue, &state.curr_768, &state.layer_input);

            // inpL = ln_f_g*inpL + ln_f_b
            self.mul_assign
                .queue(queue, &state.curr_768, &weights.ln_f_g);
            self.add_assign
                .queue(queue, &state.curr_768, &weights.ln_f_b);
        }

        // inpL = WTE * inpL
        self.matmul
            .queue(queue, &state.curr_vocab, &weights.lm_head, &state.curr_768);
    }
}