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

//! Paged attention backward kernel launcher.

use crate::error::{Error, Result};
use crate::ops::cuda::kernels::{self, PAGED_ATTENTION_BWD_MODULE};
use cudarc::driver::PushKernelArg;
use cudarc::driver::safe::LaunchConfig;
use numr::dtype::DType;
use numr::runtime::Device;
use numr::runtime::cuda::{CudaClient, CudaRuntime};
use numr::tensor::Tensor;

use super::paged_attention::bwd_block_config;

/// Paged attention backward kernel launcher.
#[allow(clippy::too_many_arguments)]
pub(super) fn paged_attention_bwd_impl(
    client: &CudaClient,
    dout: &Tensor<CudaRuntime>,
    q: &Tensor<CudaRuntime>,
    k_blocks: &Tensor<CudaRuntime>,
    v_blocks: &Tensor<CudaRuntime>,
    output: &Tensor<CudaRuntime>,
    lse: &Tensor<CudaRuntime>,
    block_table: &Tensor<CudaRuntime>,
    num_heads: usize,
    num_kv_heads: usize,
    seq_len_q: usize,
    seq_len_k: usize,
    head_dim: usize,
    block_size: usize,
    causal: bool,
) -> Result<(
    Tensor<CudaRuntime>,
    Tensor<CudaRuntime>,
    Tensor<CudaRuntime>,
)> {
    let q_shape = q.shape();
    if q_shape.len() != 4 {
        return Err(Error::InvalidArgument {
            arg: "q",
            reason: format!("expected 4D [B, H, S, D], got {}D", q_shape.len()),
        });
    }
    let batch_size = q_shape[0];
    let dtype = q.dtype();

    let dtype_suffix = match dtype {
        DType::F32 => "fp32",
        DType::F16 => "fp16",
        DType::BF16 => "bf16",
        _ => {
            return Err(Error::InvalidArgument {
                arg: "dtype",
                reason: format!("unsupported dtype {:?} for paged attention backward", dtype),
            });
        }
    };

    let (block_m, block_n) = bwd_block_config(head_dim, dtype)?;
    let kernel_name = format!(
        "paged_flash_attention_bwd_{}_{}_small",
        head_dim, dtype_suffix
    );

    let device = q.device();
    let max_num_blocks = block_table.shape()[1];

    // dQ: contiguous [B, H, S_q, D]
    let dq =
        Tensor::<CudaRuntime>::empty(&[batch_size, num_heads, seq_len_q, head_dim], dtype, device);
    // dK, dV: same shape as k_blocks/v_blocks — zeroed for atomicAdd
    let dk_blocks = Tensor::<CudaRuntime>::zeros(k_blocks.shape(), dtype, device);
    let dv_blocks = Tensor::<CudaRuntime>::zeros(v_blocks.shape(), dtype, device);

    let dtype_size = dtype.size_in_bytes();
    let smem_size = (3 * block_m * head_dim + 2 * block_n * head_dim) * dtype_size;

    let device_index = device.id();
    let module =
        kernels::get_or_load_module(client.context(), device_index, PAGED_ATTENTION_BWD_MODULE)?;
    let func = kernels::get_kernel_function(&module, &kernel_name)?;

    let cfg = LaunchConfig {
        grid_dim: (
            (batch_size * num_heads) as u32,
            seq_len_q.div_ceil(block_m) as u32,
            1,
        ),
        block_dim: (block_m as u32, 1, 1),
        shared_mem_bytes: smem_size as u32,
    };

    let q_ptr = q.ptr();
    let kb_ptr = k_blocks.ptr();
    let vb_ptr = v_blocks.ptr();
    let o_ptr = output.ptr();
    let do_ptr = dout.ptr();
    let l_ptr = lse.ptr();
    let bt_ptr = block_table.ptr();
    let dq_ptr = dq.ptr();
    let dkb_ptr = dk_blocks.ptr();
    let dvb_ptr = dv_blocks.ptr();
    let scale = (head_dim as f32).sqrt().recip();
    let batch_i32 = batch_size as i32;
    let nh_i32 = num_heads as i32;
    let nkvh_i32 = num_kv_heads as i32;
    let sq_i32 = seq_len_q as i32;
    let sk_i32 = seq_len_k as i32;
    let mnb_i32 = max_num_blocks as i32;
    let bs_i32 = block_size as i32;
    let causal_i32 = if causal { 1i32 } else { 0i32 };

    unsafe {
        let mut builder = client.stream().launch_builder(&func);
        builder.arg(&q_ptr);
        builder.arg(&kb_ptr);
        builder.arg(&vb_ptr);
        builder.arg(&o_ptr);
        builder.arg(&do_ptr);
        builder.arg(&l_ptr);
        builder.arg(&bt_ptr);
        builder.arg(&dq_ptr);
        builder.arg(&dkb_ptr);
        builder.arg(&dvb_ptr);
        builder.arg(&batch_i32);
        builder.arg(&nh_i32);
        builder.arg(&nkvh_i32);
        builder.arg(&sq_i32);
        builder.arg(&sk_i32);
        builder.arg(&mnb_i32);
        builder.arg(&bs_i32);
        builder.arg(&scale);
        builder.arg(&causal_i32);
        builder.launch(cfg).map_err(|e| Error::KernelError {
            reason: format!("Paged attention bwd kernel launch failed: {:?}", e),
        })?;
    }

    Ok((dq, dk_blocks, dv_blocks))
}