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

//! WebGPU implementation of FusedFp8TrainingOps
//!
//! F32 only (WebGPU is 32-bit by design).
//! Uses a two-pass approach: pass 1 computes unscaled values + norm²,
//! pass 2 applies clipping if needed (after readback of norm²).

use crate::error::{Error, Result};
use crate::ops::impl_generic::training::dynamic_loss_scale_update_impl;
use crate::ops::traits::FusedFp8TrainingOps;
use numr::dtype::DType;
use numr::runtime::wgpu::{WgpuClient, WgpuRuntime, get_buffer};
use numr::tensor::Tensor;
use wgpu::BufferUsages;

const SHADER_SOURCE: &str = include_str!("../shaders/training/fused_grad_unscale_clip.wgsl");

#[repr(C)]
#[derive(Clone, Copy, bytemuck::Pod, bytemuck::Zeroable)]
struct UnscaleClipParams {
    inv_scale: f32,
    max_norm: f32,
    n: u32,
    _pad: u32,
}

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

impl FusedFp8TrainingOps<WgpuRuntime> for WgpuClient {
    fn fused_grad_unscale_clip(
        &self,
        grad: &Tensor<WgpuRuntime>,
        max_norm: f64,
        loss_scale: f64,
    ) -> Result<(Tensor<WgpuRuntime>, f64, bool)> {
        validate_f32(grad, "fused_grad_unscale_clip")?;
        let n: usize = grad.shape().iter().product();

        // Allocate a fresh output tensor — can't alias grad (read vs read_write conflict)
        let out = Tensor::<WgpuRuntime>::zeros(grad.shape(), DType::F32, grad.device());

        let out_buf = get_buffer(out.storage().ptr()).ok_or_else(|| Error::KernelError {
            reason: "out buffer not found".into(),
        })?;
        let grad_buf = get_buffer(grad.storage().ptr()).ok_or_else(|| Error::KernelError {
            reason: "grad buffer not found".into(),
        })?;

        // Create result buffer for atomics: [found_inf: u32, norm_sq: f32]
        let result_buf = self.wgpu_device().create_buffer(&wgpu::BufferDescriptor {
            label: Some("unscale_clip_result"),
            size: 8, // 2 x u32/f32
            usage: BufferUsages::STORAGE | BufferUsages::COPY_SRC | BufferUsages::COPY_DST,
            mapped_at_creation: false,
        });
        // Zero-initialize
        self.wgpu_queue().write_buffer(&result_buf, 0, &[0u8; 8]);

        let params = UnscaleClipParams {
            inv_scale: (1.0 / loss_scale) as f32,
            max_norm: max_norm as f32,
            n: n as u32,
            _pad: 0,
        };
        let params_buf = self.wgpu_device().create_buffer(&wgpu::BufferDescriptor {
            label: Some("unscale_clip_params"),
            size: std::mem::size_of::<UnscaleClipParams>() 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("fused_grad_unscale_clip_f32", SHADER_SOURCE);

        let layout = cache.get_or_create_layout(numr::runtime::wgpu::shaders::LayoutKey {
            num_storage_buffers: 3, // grad(read), out(rw), result(rw)
            num_uniform_buffers: 1,
            num_readonly_storage: 1,
        });
        let pipeline = cache.get_or_create_pipeline(
            "fused_grad_unscale_clip_f32",
            "fused_grad_unscale_clip_f32",
            &module,
            &layout,
        );

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

        let workgroups = (n as u32).div_ceil(256);
        let mut encoder =
            self.wgpu_device()
                .create_command_encoder(&wgpu::CommandEncoderDescriptor {
                    label: Some("unscale_clip"),
                });
        {
            let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
                label: Some("unscale_clip"),
                timestamp_writes: None,
            });
            pass.set_pipeline(&pipeline);
            pass.set_bind_group(0, Some(&bind_group), &[]);
            pass.dispatch_workgroups(workgroups, 1, 1);
        }

        // Readback buffer
        let readback_buf = self.wgpu_device().create_buffer(&wgpu::BufferDescriptor {
            label: Some("unscale_clip_readback"),
            size: 8,
            usage: BufferUsages::MAP_READ | BufferUsages::COPY_DST,
            mapped_at_creation: false,
        });
        encoder.copy_buffer_to_buffer(&result_buf, 0, &readback_buf, 0, 8);

        self.wgpu_queue().submit(std::iter::once(encoder.finish()));

        // Map and read results
        let slice = readback_buf.slice(..);
        let (tx, rx) = std::sync::mpsc::channel();
        slice.map_async(wgpu::MapMode::Read, move |result| {
            tx.send(result).ok();
        });
        let _ = self.wgpu_device().poll(wgpu::PollType::Wait {
            submission_index: None,
            timeout: Some(std::time::Duration::from_secs(60)),
        });
        rx.recv()
            .map_err(|e| Error::KernelError {
                reason: format!("readback channel error: {:?}", e),
            })?
            .map_err(|e| Error::KernelError {
                reason: format!("readback map error: {:?}", e),
            })?;

        let data = slice.get_mapped_range();
        let results: &[u32] = bytemuck::cast_slice(&data);
        let found_inf = results[0] != 0;
        let norm_sq = f32::from_bits(results[1]);
        drop(data);
        readback_buf.unmap();

        let norm = (norm_sq as f64).sqrt();

        // If clipping needed and no inf, apply clip_coef via a second dispatch
        if !found_inf && norm > max_norm && max_norm > 0.0 {
            let clip_coef = (max_norm / (norm + 1e-6)) as f32;
            // Simple scale: multiply all elements by clip_coef
            // We reuse the shader's clip pass entry point
            let clip_params = UnscaleClipParams {
                inv_scale: clip_coef,
                max_norm: 0.0, // unused in clip pass
                n: n as u32,
                _pad: 0,
            };
            self.wgpu_queue()
                .write_buffer(&params_buf, 0, bytemuck::bytes_of(&clip_params));

            // Zero result buf again
            self.wgpu_queue().write_buffer(&result_buf, 0, &[0u8; 8]);

            let clip_pipeline =
                cache.get_or_create_pipeline("clip_scale_f32", "clip_scale_f32", &module, &layout);
            // Binding 0 is read-only, can't alias with binding 1 (read_write out_buf)
            let dummy = Tensor::<WgpuRuntime>::zeros(&[1], DType::F32, grad.device());
            let dummy_buf =
                get_buffer(dummy.storage().ptr()).ok_or_else(|| Error::KernelError {
                    reason: "dummy buffer not found".into(),
                })?;
            let clip_bind_group =
                cache.create_bind_group(&layout, &[&dummy_buf, &out_buf, &result_buf, &params_buf]);

            let mut encoder2 =
                self.wgpu_device()
                    .create_command_encoder(&wgpu::CommandEncoderDescriptor {
                        label: Some("clip_scale"),
                    });
            {
                let mut pass = encoder2.begin_compute_pass(&wgpu::ComputePassDescriptor {
                    label: Some("clip_scale"),
                    timestamp_writes: None,
                });
                pass.set_pipeline(&clip_pipeline);
                pass.set_bind_group(0, Some(&clip_bind_group), &[]);
                pass.dispatch_workgroups(workgroups, 1, 1);
            }
            self.wgpu_queue().submit(std::iter::once(encoder2.finish()));
        }

        Ok((out, norm, found_inf))
    }

    fn dynamic_loss_scale_update(
        &self,
        found_inf: bool,
        loss_scale: f64,
        growth_tracker: i32,
        growth_interval: i32,
        backoff_factor: f64,
    ) -> Result<(f64, i32)> {
        dynamic_loss_scale_update_impl(
            found_inf,
            loss_scale,
            growth_tracker,
            growth_interval,
            backoff_factor,
        )
    }
}