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

//! WebGPU implementation of SpeculativeOps
//!
//! verify_speculative_tokens: delegates to impl_generic (uses numr philox_uniform,
//!   results are Vec<VerificationResult> — CPU-side — so serial loop on CPU is correct).
//! compute_acceptance_probs: WGSL shader (element-wise, no RNG, F32 only).
//! compute_expected_tokens: WGSL shader (one thread per batch element, F32 only).

use crate::error::{Error, Result};
use crate::ops::impl_generic::inference::speculative::verify_speculative_tokens_impl;
use crate::ops::traits::inference::speculative::{SpeculativeOps, VerificationResult};
use numr::dtype::DType;
use numr::runtime::wgpu::{WgpuClient, WgpuRuntime, get_buffer};
use numr::tensor::Tensor;
use wgpu::BufferUsages;

const SPECULATIVE_SHADER: &str = include_str!("../shaders/inference/speculative_verify.wgsl");

fn validate_f32(t: &Tensor<WgpuRuntime>, op: &str) -> Result<()> {
    if t.dtype() != DType::F32 {
        return Err(Error::InvalidArgument {
            arg: "dtype",
            reason: format!(
                "{}: WebGPU speculative requires F32, got {:?}",
                op,
                t.dtype()
            ),
        });
    }
    Ok(())
}

#[repr(C)]
#[derive(Clone, Copy, bytemuck::Pod, bytemuck::Zeroable)]
struct AcceptParams {
    total_elements: u32,
    _pad0: u32,
    _pad1: u32,
    _pad2: u32,
}

#[repr(C)]
#[derive(Clone, Copy, bytemuck::Pod, bytemuck::Zeroable)]
struct ExpectedParams {
    batch_size: u32,
    max_spec_tokens: u32,
    _pad0: u32,
    _pad1: u32,
}

impl SpeculativeOps<WgpuRuntime> for WgpuClient {
    fn verify_speculative_tokens(
        &self,
        draft_probs: &Tensor<WgpuRuntime>,
        target_probs: &Tensor<WgpuRuntime>,
        draft_tokens: &Tensor<WgpuRuntime>,
        seed: u64,
    ) -> Result<Vec<VerificationResult>> {
        // Delegate to impl_generic: generates uniform randoms via philox_uniform
        // (same algorithm as CPU/CUDA), then runs the serial accept/reject loop on CPU.
        // Results are Vec<VerificationResult> (CPU-side) regardless of backend.
        verify_speculative_tokens_impl(self, draft_probs, target_probs, draft_tokens, seed)
    }

    fn compute_acceptance_probs(
        &self,
        draft_probs: &Tensor<WgpuRuntime>,
        target_probs: &Tensor<WgpuRuntime>,
    ) -> Result<(Tensor<WgpuRuntime>, Tensor<WgpuRuntime>)> {
        validate_f32(draft_probs, "compute_acceptance_probs")?;
        validate_f32(target_probs, "compute_acceptance_probs")?;

        let dp_shape = draft_probs.shape();
        let tp_shape = target_probs.shape();

        if dp_shape != tp_shape {
            return Err(Error::InvalidArgument {
                arg: "target_probs",
                reason: format!("shape mismatch: {:?} vs {:?}", dp_shape, tp_shape),
            });
        }

        let total: usize = dp_shape.iter().product();
        let device = draft_probs.device();

        let acceptance = Tensor::<WgpuRuntime>::empty(dp_shape, DType::F32, device);
        let residual = Tensor::<WgpuRuntime>::empty(dp_shape, DType::F32, device);

        let dp_buf = get_buffer(draft_probs.storage().ptr()).ok_or_else(|| Error::KernelError {
            reason: "draft buffer not found".into(),
        })?;
        let tp_buf =
            get_buffer(target_probs.storage().ptr()).ok_or_else(|| Error::KernelError {
                reason: "target buffer not found".into(),
            })?;
        let acc_buf = get_buffer(acceptance.storage().ptr()).ok_or_else(|| Error::KernelError {
            reason: "acceptance buffer not found".into(),
        })?;
        let res_buf = get_buffer(residual.storage().ptr()).ok_or_else(|| Error::KernelError {
            reason: "residual buffer not found".into(),
        })?;

        let params = AcceptParams {
            total_elements: total as u32,
            _pad0: 0,
            _pad1: 0,
            _pad2: 0,
        };

        let params_buf = self.wgpu_device().create_buffer(&wgpu::BufferDescriptor {
            label: Some("accept_params"),
            size: std::mem::size_of::<AcceptParams>() as u64,
            usage: BufferUsages::UNIFORM | BufferUsages::COPY_DST,
            mapped_at_creation: false,
        });
        self.wgpu_queue()
            .write_buffer(&params_buf, 0, bytemuck::bytes_of(&params));

        let cache = self.pipeline_cache();
        let module = cache.get_or_create_module("speculative_accept", SPECULATIVE_SHADER);
        let layout = cache.get_or_create_layout(numr::runtime::wgpu::shaders::LayoutKey {
            num_storage_buffers: 4,
            num_uniform_buffers: 1,
            num_readonly_storage: 2,
        });
        let pipeline = cache.get_or_create_pipeline(
            "compute_acceptance_probs_f32",
            "compute_acceptance_probs_f32",
            &module,
            &layout,
        );

        let bind_group = cache.create_bind_group(
            &layout,
            &[&dp_buf, &tp_buf, &acc_buf, &res_buf, &params_buf],
        );

        let workgroups = (total as u32).div_ceil(256);

        let mut encoder =
            self.wgpu_device()
                .create_command_encoder(&wgpu::CommandEncoderDescriptor {
                    label: Some("acceptance_probs"),
                });
        {
            let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
                label: Some("acceptance_probs"),
                timestamp_writes: None,
            });
            pass.set_pipeline(&pipeline);
            pass.set_bind_group(0, Some(&bind_group), &[]);
            pass.dispatch_workgroups(workgroups, 1, 1);
        }
        self.wgpu_queue().submit(std::iter::once(encoder.finish()));

        Ok((acceptance, residual))
    }

    fn compute_expected_tokens(
        &self,
        acceptance_rates: &Tensor<WgpuRuntime>,
    ) -> Result<Tensor<WgpuRuntime>> {
        validate_f32(acceptance_rates, "compute_expected_tokens")?;

        let shape = acceptance_rates.shape();
        if shape.len() != 2 {
            return Err(Error::InvalidArgument {
                arg: "acceptance_rates",
                reason: format!("expected 2D [batch, K], got {}D", shape.len()),
            });
        }

        let batch_size = shape[0];
        let k = shape[1];
        let device = acceptance_rates.device();

        let output = Tensor::<WgpuRuntime>::empty(&[batch_size], DType::F32, device);

        let rates_buf =
            get_buffer(acceptance_rates.storage().ptr()).ok_or_else(|| Error::KernelError {
                reason: "rates buffer not found".into(),
            })?;
        let out_buf = get_buffer(output.storage().ptr()).ok_or_else(|| Error::KernelError {
            reason: "output buffer not found".into(),
        })?;

        let params = ExpectedParams {
            batch_size: batch_size as u32,
            max_spec_tokens: k as u32,
            _pad0: 0,
            _pad1: 0,
        };

        let params_buf = self.wgpu_device().create_buffer(&wgpu::BufferDescriptor {
            label: Some("expected_params"),
            size: std::mem::size_of::<ExpectedParams>() as u64,
            usage: BufferUsages::UNIFORM | BufferUsages::COPY_DST,
            mapped_at_creation: false,
        });
        self.wgpu_queue()
            .write_buffer(&params_buf, 0, bytemuck::bytes_of(&params));

        let cache = self.pipeline_cache();
        let module = cache.get_or_create_module("speculative_expected", SPECULATIVE_SHADER);
        let layout = cache.get_or_create_layout(numr::runtime::wgpu::shaders::LayoutKey {
            num_storage_buffers: 2,
            num_uniform_buffers: 1,
            num_readonly_storage: 1,
        });
        let pipeline = cache.get_or_create_pipeline(
            "compute_expected_tokens_f32",
            "compute_expected_tokens_f32",
            &module,
            &layout,
        );

        let bind_group = cache.create_bind_group(&layout, &[&rates_buf, &out_buf, &params_buf]);

        let workgroups = (batch_size as u32).div_ceil(256);

        let mut encoder =
            self.wgpu_device()
                .create_command_encoder(&wgpu::CommandEncoderDescriptor {
                    label: Some("expected_tokens"),
                });
        {
            let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
                label: Some("expected_tokens"),
                timestamp_writes: None,
            });
            pass.set_pipeline(&pipeline);
            pass.set_bind_group(0, Some(&bind_group), &[]);
            pass.dispatch_workgroups(workgroups, 1, 1);
        }
        self.wgpu_queue().submit(std::iter::once(encoder.finish()));

        Ok(output)
    }
}