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

//! WebGPU implementation of MlaOps
//!
//! Uses the SDPA WGSL shader to avoid 4D matmul limitation in WebGPU.

use crate::error::{Error, Result};
use crate::ops::traits::MlaOps;
use numr::autograd::Var;
use numr::dtype::DType;
use numr::runtime::wgpu::{WgpuClient, WgpuRuntime, get_buffer};
use numr::tensor::Tensor;
use wgpu::BufferUsages;

const SDPA_SHADER_SOURCE: &str = include_str!("../shaders/attention/sdpa.wgsl");

#[repr(C)]
#[derive(Clone, Copy, bytemuck::Pod, bytemuck::Zeroable)]
struct SdpaParams {
    batch_size: u32,
    num_heads: u32,
    seq_len_q: u32,
    seq_len_k: u32,
    head_dim_k: u32,
    head_dim_v: u32,
    scale: f32,
    causal: u32,
}

impl MlaOps<WgpuRuntime> for WgpuClient {
    fn scaled_dot_product_attention(
        &self,
        q: &Var<WgpuRuntime>,
        k: &Var<WgpuRuntime>,
        v: &Var<WgpuRuntime>,
        scale: f64,
        causal: bool,
    ) -> Result<Var<WgpuRuntime>> {
        let q_tensor = q.tensor();
        let k_tensor = k.tensor();
        let v_tensor = v.tensor();

        let q_shape = q_tensor.shape();
        let k_shape = k_tensor.shape();
        let v_shape = v_tensor.shape();

        if q_shape.len() != 4 || k_shape.len() != 4 || v_shape.len() != 4 {
            return Err(Error::InvalidArgument {
                arg: "q/k/v",
                reason: "expected 4D tensors [B, H, S, D]".into(),
            });
        }

        let batch_size = q_shape[0];
        let num_heads = q_shape[1];
        let seq_len_q = q_shape[2];
        let head_dim_k = q_shape[3];
        let seq_len_k = k_shape[2];
        let head_dim_v = v_shape[3];

        if q_tensor.dtype() != DType::F32 {
            return Err(Error::InvalidArgument {
                arg: "dtype",
                reason: format!("WebGPU SDPA requires F32, got {:?}", q_tensor.dtype()),
            });
        }

        // Create output tensor: [B, H, S_q, D_v]
        let output = Tensor::<WgpuRuntime>::zeros(
            &[batch_size, num_heads, seq_len_q, head_dim_v],
            DType::F32,
            q_tensor.device(),
        );

        let q_buf = get_buffer(q_tensor.storage().ptr()).ok_or_else(|| Error::KernelError {
            reason: "q buffer not found".into(),
        })?;
        let k_buf = get_buffer(k_tensor.storage().ptr()).ok_or_else(|| Error::KernelError {
            reason: "k buffer not found".into(),
        })?;
        let v_buf = get_buffer(v_tensor.storage().ptr()).ok_or_else(|| Error::KernelError {
            reason: "v 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 = SdpaParams {
            batch_size: batch_size as u32,
            num_heads: num_heads as u32,
            seq_len_q: seq_len_q as u32,
            seq_len_k: seq_len_k as u32,
            head_dim_k: head_dim_k as u32,
            head_dim_v: head_dim_v as u32,
            scale: scale as f32,
            causal: if causal { 1 } else { 0 },
        };

        let params_buf = self.wgpu_device().create_buffer(&wgpu::BufferDescriptor {
            label: Some("sdpa_params"),
            size: std::mem::size_of::<SdpaParams>() 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("sdpa_forward_f32", SDPA_SHADER_SOURCE);

        // Shader bindings: 0=q(read), 1=k(read), 2=v(read), 3=out(rw)
        let layout = cache.get_or_create_layout(numr::runtime::wgpu::shaders::LayoutKey {
            num_storage_buffers: 4,
            num_uniform_buffers: 1,
            num_readonly_storage: 3,
        });
        let pipeline =
            cache.get_or_create_pipeline("sdpa_forward_f32", "sdpa_forward_f32", &module, &layout);

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

        let total_queries = (batch_size * num_heads * seq_len_q) as u32;
        let workgroups = total_queries.div_ceil(256);

        let mut encoder =
            self.wgpu_device()
                .create_command_encoder(&wgpu::CommandEncoderDescriptor {
                    label: Some("sdpa"),
                });
        {
            let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
                label: Some("sdpa"),
                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(Var::new(output, false))
    }
}