boostr 0.1.0

ML framework built on numr - attention, quantization, model architectures
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

//! WGSL shader generators for quantized matmul operations
//!
//! Each thread computes one element of the output matrix by dequantizing
//! the weight row on-the-fly and computing the dot product with the activation row.

use super::common_helpers;

/// Generate WGSL shader for Q4_0 quantized matmul.
///
/// activation [M, K] × weight_q4_0 [N, K] → output [M, N]
/// Each thread handles one (m, n) output element.
pub fn generate_quant_matmul_q4_0_shader() -> String {
    format!(
        r#"// Quantized matmul: activation [M,K] x Q4_0 weight [N,K] -> output [M,N]
{helpers}

const BLOCK_SIZE: u32 = 32u;
const BLOCK_BYTES: u32 = 18u;

struct MatmulParams {{
    M: u32,
    K: u32,
    N: u32,
    _pad: u32,
}}

@group(0) @binding(0) var<storage, read_write> activation: array<f32>;
@group(0) @binding(1) var<storage, read_write> weight: array<u32>;
@group(0) @binding(2) var<storage, read_write> output: array<f32>;
@group(0) @binding(3) var<uniform> params: MatmulParams;

@compute @workgroup_size(16, 16)
fn quant_matmul_q4_0(@builtin(global_invocation_id) gid: vec3<u32>) {{
    let n = gid.x;
    let m = gid.y;

    if (m >= params.M || n >= params.N) {{
        return;
    }}

    let num_blocks_per_row = params.K / BLOCK_SIZE;
    let weight_row_base = n * num_blocks_per_row * BLOCK_BYTES;
    let act_row_base = m * params.K;

    var acc: f32 = 0.0;

    for (var b: u32 = 0u; b < num_blocks_per_row; b = b + 1u) {{
        let block_base = weight_row_base + b * BLOCK_BYTES;
        let delta = f16_to_f32((weight[block_base / 4u] >> ((block_base % 4u) * 8u)) & 0xFFFFu);
        let k_base = b * BLOCK_SIZE;

        for (var i: u32 = 0u; i < 16u; i = i + 1u) {{
            let u8_byte_idx = block_base + 2u + i;
            let byte_val = ((weight[u8_byte_idx / 4u] >> ((u8_byte_idx % 4u) * 8u)) & 0xFFu);
            let lo = byte_val & 0xFu;
            let hi = (byte_val >> 4u) & 0xFu;

            let w0 = (f32(lo) - 8.0) * delta;
            let w1 = (f32(hi) - 8.0) * delta;

            acc = acc + activation[act_row_base + k_base + i] * w0;
            acc = acc + activation[act_row_base + k_base + i + 16u] * w1;
        }}
    }}

    output[m * params.N + n] = acc;
}}
"#,
        helpers = common_helpers(),
    )
}

/// Generate WGSL shader for Q8_0 quantized matmul.
pub fn generate_quant_matmul_q8_0_shader() -> String {
    format!(
        r#"// Quantized matmul: activation [M,K] x Q8_0 weight [N,K] -> output [M,N]
{helpers}

const BLOCK_SIZE: u32 = 32u;
const BLOCK_BYTES: u32 = 34u;

struct MatmulParams {{
    M: u32,
    K: u32,
    N: u32,
    _pad: u32,
}}

@group(0) @binding(0) var<storage, read_write> activation: array<f32>;
@group(0) @binding(1) var<storage, read_write> weight: array<u32>;
@group(0) @binding(2) var<storage, read_write> output: array<f32>;
@group(0) @binding(3) var<uniform> params: MatmulParams;

@compute @workgroup_size(16, 16)
fn quant_matmul_q8_0(@builtin(global_invocation_id) gid: vec3<u32>) {{
    let n = gid.x;
    let m = gid.y;

    if (m >= params.M || n >= params.N) {{
        return;
    }}

    let num_blocks_per_row = params.K / BLOCK_SIZE;
    let weight_row_base = n * num_blocks_per_row * BLOCK_BYTES;
    let act_row_base = m * params.K;

    var acc: f32 = 0.0;

    for (var b: u32 = 0u; b < num_blocks_per_row; b = b + 1u) {{
        let block_base = weight_row_base + b * BLOCK_BYTES;
        let delta = f16_to_f32((weight[block_base / 4u] >> ((block_base % 4u) * 8u)) & 0xFFFFu);
        let k_base = b * BLOCK_SIZE;

        for (var i: u32 = 0u; i < 32u; i = i + 1u) {{
            let i8_byte_idx = block_base + 2u + i;
            let i8_u8 = ((weight[i8_byte_idx / 4u] >> ((i8_byte_idx % 4u) * 8u)) & 0xFFu);
            let qs = select(i32(i8_u8), i32(i8_u8) - 256, i8_u8 >= 128u);
            let w = f32(qs) * delta;
            acc = acc + activation[act_row_base + k_base + i] * w;
        }}
    }}

    output[m * params.N + n] = acc;
}}
"#,
        helpers = common_helpers(),
    )
}

/// Generate WGSL shader for Q4_K quantized matmul.
pub fn generate_quant_matmul_q4_k_shader() -> String {
    format!(
        r#"// Quantized matmul: activation [M,K] x Q4_K weight [N,K] -> output [M,N]
{helpers}

const BLOCK_SIZE: u32 = 256u;
const BLOCK_BYTES: u32 = 144u;

struct MatmulParams {{
    M: u32,
    K: u32,
    N: u32,
    _pad: u32,
}}

@group(0) @binding(0) var<storage, read_write> activation: array<f32>;
@group(0) @binding(1) var<storage, read_write> weight: array<u32>;
@group(0) @binding(2) var<storage, read_write> output: array<f32>;
@group(0) @binding(3) var<uniform> params: MatmulParams;

@compute @workgroup_size(16, 16)
fn quant_matmul_q4_k(@builtin(global_invocation_id) gid: vec3<u32>) {{
    let n = gid.x;
    let m = gid.y;

    if (m >= params.M || n >= params.N) {{
        return;
    }}

    let num_blocks_per_row = params.K / BLOCK_SIZE;
    let weight_row_base = n * num_blocks_per_row * BLOCK_BYTES;
    let act_row_base = m * params.K;

    var acc: f32 = 0.0;

    for (var blk: u32 = 0u; blk < num_blocks_per_row; blk = blk + 1u) {{
        let base = weight_row_base + blk * BLOCK_BYTES;
        let d = f16_to_f32((weight[base / 4u] >> ((base % 4u) * 8u)) & 0xFFFFu);
        let dmin_byte = base + 2u;
        let dmin = f16_to_f32((weight[dmin_byte / 4u] >> ((dmin_byte % 4u) * 8u)) & 0xFFFFu);
        let scales_base = base + 4u;
        let qs_base = base + 16u;
        let k_base = blk * BLOCK_SIZE;

        for (var sb: u32 = 0u; sb < 8u; sb = sb + 1u) {{
            var sc: f32;
            var mn: f32;

            if (sb < 4u) {{
                let sb_byte_idx = scales_base + sb;
                let scale_byte = ((weight[sb_byte_idx / 4u] >> ((sb_byte_idx % 4u) * 8u)) & 0xFFu);
                sc = d * f32(scale_byte & 63u);
                let min_byte_idx = scales_base + sb + 4u;
                let min_byte = ((weight[min_byte_idx / 4u] >> ((min_byte_idx % 4u) * 8u)) & 0xFFu);
                mn = dmin * f32(min_byte & 63u);
            }} else {{
                let sb_off = sb - 4u;
                let slo_byte_idx = scales_base + sb;
                let scale_byte_lo = ((weight[slo_byte_idx / 4u] >> ((slo_byte_idx % 4u) * 8u)) & 0xFFu);
                let shi_byte_idx = scales_base + sb + 4u;
                let scale_byte_hi = ((weight[shi_byte_idx / 4u] >> ((shi_byte_idx % 4u) * 8u)) & 0xFFu);
                sc = d * f32((scale_byte_lo & 63u) | ((scale_byte_hi & 0x03u) << 6u));
                mn = dmin * f32(((scale_byte_lo >> 6u) & 3u) | ((scale_byte_hi >> 2u) << 2u));
            }}

            let sub_qs = qs_base + sb * 16u;
            let sub_k = k_base + sb * 32u;

            for (var i: u32 = 0u; i < 16u; i = i + 1u) {{
                let qs_byte_idx = sub_qs + i;
                let byte_val = ((weight[qs_byte_idx / 4u] >> ((qs_byte_idx % 4u) * 8u)) & 0xFFu);
                let lo = byte_val & 0xFu;
                let hi = (byte_val >> 4u) & 0xFu;

                let w0 = f32(lo) * sc - mn;
                let w1 = f32(hi) * sc - mn;

                acc = acc + activation[act_row_base + sub_k + i] * w0;
                acc = acc + activation[act_row_base + sub_k + i + 16u] * w1;
            }}
        }}
    }}

    output[m * params.N + n] = acc;
}}
"#,
        helpers = common_helpers(),
    )
}

/// Generate WGSL shader for Q6_K quantized matmul.
pub fn generate_quant_matmul_q6_k_shader() -> String {
    format!(
        r#"// Quantized matmul: activation [M,K] x Q6_K weight [N,K] -> output [M,N]
{helpers}

const BLOCK_SIZE: u32 = 256u;
const BLOCK_BYTES: u32 = 210u;

struct MatmulParams {{
    M: u32,
    K: u32,
    N: u32,
    _pad: u32,
}}

@group(0) @binding(0) var<storage, read_write> activation: array<f32>;
@group(0) @binding(1) var<storage, read_write> weight: array<u32>;
@group(0) @binding(2) var<storage, read_write> output: array<f32>;
@group(0) @binding(3) var<uniform> params: MatmulParams;

@compute @workgroup_size(16, 16)
fn quant_matmul_q6_k(@builtin(global_invocation_id) gid: vec3<u32>) {{
    let n = gid.x;
    let m = gid.y;

    if (m >= params.M || n >= params.N) {{
        return;
    }}

    let num_blocks_per_row = params.K / BLOCK_SIZE;
    let weight_row_base = n * num_blocks_per_row * BLOCK_BYTES;
    let act_row_base = m * params.K;

    var acc: f32 = 0.0;

    for (var blk: u32 = 0u; blk < num_blocks_per_row; blk = blk + 1u) {{
        let base = weight_row_base + blk * BLOCK_BYTES;
        let ql_base = base;
        let qh_base = base + 128u;
        let sc_base = base + 192u;
        let d_offset = base + 208u;
        let d = f16_to_f32((weight[d_offset / 4u] >> ((d_offset % 4u) * 8u)) & 0xFFFFu);
        let k_base = blk * BLOCK_SIZE;

        for (var sb: u32 = 0u; sb < 16u; sb = sb + 1u) {{
            let sc_byte_idx = sc_base + sb;
            let sc_u8 = ((weight[sc_byte_idx / 4u] >> ((sc_byte_idx % 4u) * 8u)) & 0xFFu);
            let scale = select(i32(sc_u8), i32(sc_u8) - 256, sc_u8 >= 128u);

            for (var i: u32 = 0u; i < 16u; i = i + 1u) {{
                let elem_idx = sb * 16u + i;

                let ql_byte_idx = ql_base + elem_idx / 2u;
                let ql_byte = ((weight[ql_byte_idx / 4u] >> ((ql_byte_idx % 4u) * 8u)) & 0xFFu);
                var ql_val: u32;
                if (elem_idx % 2u == 0u) {{
                    ql_val = ql_byte & 0xFu;
                }} else {{
                    ql_val = (ql_byte >> 4u) & 0xFu;
                }}

                let qh_byte_idx = qh_base + elem_idx / 4u;
                let qh_byte = ((weight[qh_byte_idx / 4u] >> ((qh_byte_idx % 4u) * 8u)) & 0xFFu);
                let qh_shift = (elem_idx % 4u) * 2u;
                let qh_val = (qh_byte >> qh_shift) & 0x3u;

                let q = ql_val | (qh_val << 4u);
                let w = d * f32(scale) * (f32(q) - 32.0);

                acc = acc + activation[act_row_base + k_base + elem_idx] * w;
            }}
        }}
    }}

    output[m * params.N + n] = acc;
}}
"#,
        helpers = common_helpers(),
    )
}