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

//! Flash Attention v2 FP8 backward pass: dQ, dK, dV with per-tensor scales.

use crate::error::{Error, Result};
use crate::ops::cuda::kernels::{self, FLASH_V2_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::flash_utils::{AttentionParams, set_smem_attribute};

/// FP8 backward pass (E4M3/E5M2). Returns (dQ, dK, dV).
///
/// Two-step process with extra per-tensor scale args:
/// 1. FP8 preprocessing: D = rowsum(dO ⊙ O) with do_scale and o_scale.
/// 2. FP8 main backward kernel with full scale set.
#[allow(clippy::too_many_arguments)]
pub(super) fn flash_attention_bwd_fp8_impl(
    client: &CudaClient,
    dout: &Tensor<CudaRuntime>,
    q: &Tensor<CudaRuntime>,
    k: &Tensor<CudaRuntime>,
    v: &Tensor<CudaRuntime>,
    output: &Tensor<CudaRuntime>,
    lse: &Tensor<CudaRuntime>,
    p: &AttentionParams,
    causal: bool,
    q_scale: f32,
    k_scale: f32,
    v_scale: f32,
    do_scale: f32,
    o_scale: f32,
    dq_scale: f32,
    dk_scale: f32,
    dv_scale: f32,
) -> Result<(
    Tensor<CudaRuntime>,
    Tensor<CudaRuntime>,
    Tensor<CudaRuntime>,
)> {
    let dtype = q.dtype();

    let expected_o_shape = [p.batch_size, p.num_heads, p.seq_len_q, p.head_dim];
    if dout.shape() != expected_o_shape || output.shape() != expected_o_shape {
        return Err(Error::InvalidArgument {
            arg: "dout/output",
            reason: format!("expected shape {:?}", expected_o_shape),
        });
    }
    let expected_lse_shape = [p.batch_size, p.num_heads, p.seq_len_q];
    if lse.shape() != expected_lse_shape {
        return Err(Error::InvalidArgument {
            arg: "lse",
            reason: format!("expected shape {:?}", expected_lse_shape),
        });
    }

    let device = q.device();
    let device_index = device.id();

    // Allocate gradient tensors (dQ zeroed for atomicAdd)
    let dq = Tensor::<CudaRuntime>::zeros(
        &[p.batch_size, p.num_heads, p.seq_len_q, p.head_dim],
        dtype,
        device,
    );
    let dk = Tensor::<CudaRuntime>::empty(
        &[p.batch_size, p.num_kv_heads, p.seq_len_k, p.head_dim],
        dtype,
        device,
    );
    let dv = Tensor::<CudaRuntime>::empty(
        &[p.batch_size, p.num_kv_heads, p.seq_len_k, p.head_dim],
        dtype,
        device,
    );

    let d_buf = Tensor::<CudaRuntime>::empty(
        &[p.batch_size, p.num_heads, p.seq_len_q],
        DType::F32,
        device,
    );

    let module = kernels::get_or_load_module(client.context(), device_index, FLASH_V2_BWD_MODULE)?;

    // Step 1: FP8 Preprocessing — extra scale args
    {
        let preprocess_name = format!("flash_attention_preprocess_bwd_{}_fp8", p.head_dim);
        let func = kernels::get_kernel_function(&module, &preprocess_name)?;

        let block_size = 256u32;
        let grid_x = (p.batch_size * p.num_heads) as u32;
        let grid_y = (p.seq_len_q as u32).div_ceil(block_size);

        let cfg = LaunchConfig {
            grid_dim: (grid_x, grid_y, 1),
            block_dim: (block_size, 1, 1),
            shared_mem_bytes: 0,
        };

        let dout_ptr = dout.ptr();
        let out_ptr = output.ptr();
        let d_ptr = d_buf.ptr();
        let batch_i32 = p.batch_size as i32;
        let nh_i32 = p.num_heads as i32;
        let sq_i32 = p.seq_len_q as i32;

        unsafe {
            let mut builder = client.stream().launch_builder(&func);
            builder.arg(&dout_ptr);
            builder.arg(&out_ptr);
            builder.arg(&d_ptr);
            builder.arg(&batch_i32);
            builder.arg(&nh_i32);
            builder.arg(&sq_i32);
            builder.arg(&do_scale);
            builder.arg(&o_scale);
            builder.launch(cfg).map_err(|e| Error::KernelError {
                reason: format!("Flash Attention FP8 bwd preprocess failed: {:?}", e),
            })?;
        }
    }

    // Step 2: FP8 Main backward — extra scale args
    {
        let sm_suffix = if p.use_sm_kernel { "_sm" } else { "" };
        let bwd_name = format!("flash_attention_bwd_{}{}_fp8", p.head_dim, sm_suffix);
        let func = kernels::get_kernel_function(&module, &bwd_name)?;

        // FP8 is 1 byte per element
        let smem_size = 2 * p.block_n * p.head_dim + 2 * p.block_m * p.head_dim;
        set_smem_attribute(&func, smem_size)?;

        let grid_x = (p.batch_size * p.num_heads) as u32;
        let grid_y = p.seq_len_k.div_ceil(p.block_n) as u32;

        let cfg = LaunchConfig {
            grid_dim: (grid_x, grid_y, 1),
            block_dim: (p.block_n as u32, 1, 1),
            shared_mem_bytes: smem_size as u32,
        };

        let q_ptr = q.ptr();
        let k_ptr = k.ptr();
        let v_ptr = v.ptr();
        let o_ptr = output.ptr();
        let dout_ptr = dout.ptr();
        let lse_ptr = lse.ptr();
        let d_ptr = d_buf.ptr();
        let dq_ptr = dq.ptr();
        let dk_ptr = dk.ptr();
        let dv_ptr = dv.ptr();
        let scale = (p.head_dim as f32).sqrt().recip();
        let batch_i32 = p.batch_size as i32;
        let nh_i32 = p.num_heads as i32;
        let sq_i32 = p.seq_len_q as i32;
        let sk_i32 = p.seq_len_k 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(&k_ptr);
            builder.arg(&v_ptr);
            builder.arg(&o_ptr);
            builder.arg(&dout_ptr);
            builder.arg(&lse_ptr);
            builder.arg(&d_ptr);
            builder.arg(&dq_ptr);
            builder.arg(&dk_ptr);
            builder.arg(&dv_ptr);
            builder.arg(&batch_i32);
            builder.arg(&nh_i32);
            builder.arg(&sq_i32);
            builder.arg(&sk_i32);
            builder.arg(&scale);
            builder.arg(&causal_i32);
            builder.arg(&q_scale);
            builder.arg(&k_scale);
            builder.arg(&v_scale);
            builder.arg(&do_scale);
            builder.arg(&dq_scale);
            builder.arg(&dk_scale);
            builder.arg(&dv_scale);
            builder.launch(cfg).map_err(|e| Error::KernelError {
                reason: format!("Flash Attention FP8 bwd kernel launch failed: {:?}", e),
            })?;
        }
    }

    Ok((dq, dk, dv))
}